Betting on ad-free LLMs is like betting against math.

OpenAI burned $1.6 billion last year alone. Other companies are doing their best to match it.

Subscription revenue is a drop in that bucket.

Compute costs are exponential and realistically, the subscriptions are linear.

Only one business model scales infinitely.

Ads. One query, one ad. A billion queries, a billion ads.

The economics of it dictate the outcome. Ads are inevitable.

What’s worth talking about now, is the winner or winners.

The first ones to monetise your attention, not just your answer.

But what is also worth talking about is the problem. Whoever embeds ads in the model, outputs controls what billions of people see as objective truth.

And that’s a scary reality.

I’m Blake. I’ve spent the last 10 years in tech advising companies on data and AI strategy.

I’m here to break down for you the most interesting moves the players are making in AI.

Products. Moats. Distribution tactics. Everything is curated so you can stay informed.

Today, we’re diving into the upcoming land grab for ad revenue.

Let’s break it down.

I’ll start with something Sam Altman said.

“AI will be superhuman at persuasion before it is superhuman at intelligence”.

The key is in the word. Persuasion.

When an LLM answers your question, that answer FEELS like truth.

But if an advertiser pays to change that answer, that truth becomes inventory.

The company that controls this doesn’t just sell ads.

They sell the perception of truth. And once they lock that in, nobody can compete.

They get to rewrite reality for profit. Now, I’m going to spoil the video for you here.

I think Google will win. That’s my ultimate take.

But to make it clear as to why they will win, we need to explore what the major players are and what they are doing.

It’s only natural to start with Google.

I think Google isn’t just building an AI business. They are upgrading the world’s largest ad monopoly.

If you’ve noticed over the last few days, Google made its first foray into native chat ads.

I think how the ads look now, is similar to Bard compared to Gemini. There’s a lot of room for improvement. But they will get there.

Now of course, having to become the best AI company isn’t by choice. The upstarts are eating their biggest revenue model.

But. As they’re forced to compete, they’re also using their moats.

Every entry point to the internet already funnels through Google infrastructure.

They don’t need to find users. They have 4 billion of them on Android, Chrome, Maps, Gmail and YouTube.

They also don’t need to find advertisers. They have millions already paying them.

They have an ad salesforce.

Embedding Gemini into Search turns their existing monopoly into an AI monopoly overnight.

They technically win by default because they own the rails.

And I haven’t even touched on Waymo. Or deepmind.

Or whatever they cook up next. On the other hand though, there’s OpenAI.

OpenAI has the arguably smartest model, and the consumer mindshare.

But, they are homeless.

They live in Microsoft’s house. Azure, Bing, Office… it’s all rented land.

OpenAI has zero control over the operating system or the browser.

To sell ads, they will need data.

OpenAI has chat logs.

But Google has your search history, location, emails, and purchases. It’s not a fair fight.

It’s why they’ve created Atas.

It’s a data play.

But, alas, Atlas isn’t chrome and Unfortunately for OpenAI, brand recognition doesn’t pay the rent.

In platform wars, the landlord always wins.

I do see OpenAI releasing a suite of work related tools.

Their version of word, excel and powerpoint in the future.

Maybe an Atlassian takeover? Stranger things have happened.

So who else is in the race. Another big contender is Meta.

Meta knows who you are.

Meta’s moat is the social graph. They can embed AI into the conversations you care about most.

Meta’s advantage is context, they know your social graph, your interests, your behavior.

That data is gold for targeted ads.

They’re strong in engagement and social commerce. Weak in neutral information provision.

But social isn’t search.

And you don’t go to Instagram for facts; you go for vibes. Meta wins engagement, but they lose the “truth” monopoly.

Meta lacks the search-grade AI dominance and enterprise distribution that Google has.

Google only knows what you want. That’s a dangerous difference.

They do have a large chance of owning a niche, entertainment AI.

But, they won’t own the information utility.

Which brings us to X.

X is betting that “truth” is overrated and “now” is what matters. They own the real-time pulse.

News breaks on X. X plays the disruptor. Fast-moving, cultural, willing to take risks.

But their ad infrastructure is broken, and brands are scared of the chaos.

They can disrupt the narrative, but they can’t monetise it at the scale of Google.

How can I put this simply? Google owns the pipes. Microsoft owns the office. Meta owns the party. X owns the riot.

In a war for information dominance, owning the pipes (Google) beats everything else.

So if you’re betting on a winner, bet on the plumbing.

Google is the plumbing.

More users = more data = better ad targeting = more revenue = better models = more users.

Once the ad engine starts, the door closes. Forever.

I’m Blake.

Hopefully this was interesting

See you in the next one.

AI agents are getting their own identities in Entra ID. 

That means their own email address, their own team account, their own org chart placement. 

It’s so easy to treat this as standard product news. 

Cool feature. Move on.

That would be a catastrophically wrong decision. 

Forget building better agents. 

This is about controlling the infrastructure agents depend on. 

It's brilliant because it works whether you're using Microsoft's agent or someone else's.

The real story is this. 

Microsoft understood something that OpenAI and Anthropic still don't get. 

Model quality doesn't determine adoption in enterprise. 

The unsexy stuff does.

Governance. Compliance. 

The ability to audit, control, and prove that your agents are operating inside your control plane.

Microsoft is betting the entire agent category on that principle. 

And I think they're right.

An agent with an email address and org chart presence is no longer software. It's infrastructure.

Microsoft is getting first mover advantage. 

Let me be very specific about what's happening because the mechanism matters.

The moment a company makes agents' organisational identities, they’ve embedded them into the permission and governance infrastructure that enterprises already depend on. 

They're not optional. They're not temporary. They're official.

Think about the implications operationally. 

1. Your compliance team now has to audit agents the same way they audit humans. 

2. Your security team has to manage agent permissions through the same identity and access management infrastructure. 

3. Your CISO has to account for agent activity in the same audit trails. 

Every governance rule that applies to employees now applies to agents.

This is the key part. 

Enterprises HAVE to use this infrastructure. They can't opt out. 

HIPAA requires it. SOX requires it. GDPR requires it. 

The second agents become authenticated organisational members, they fall under the same regulatory framework as humans.

Microsoft is making sure that infrastructure is their infrastructure.

Every competing agent company just became a software vendor running on Microsoft's infrastructure.

They don't know it yet.

Here's my take on what this means for everyone else building agents.

OpenAI can build a better model. 

Anthropic can build smarter reasoning. 

Salesforce can build vertical agents. 

But none of them can provide what Microsoft just did. They can't provide compliance governance at the infrastructure layer because they don't own the enterprise identity layer.

You know what happens when you try to run a competing agent in an enterprise? 

You have two problems. 

One, you have to convince them to use your agent instead of Microsoft's. 

Two, you have to help them figure out governance.

Most enterprises will choose the option where governance is already built in. 

The place where audit trails are native. Where permissions are managed through existing infrastructure.

This is why Entra ID integration is the real play. 

Entra ID already powers 300 million users. 

The second agents become first-class citizens in Entra, they're not separate infrastructure. 

They're native.

They're part of the organisation.

My take is that this doesn't kill competing agents. It just means competing agents have to run through Microsoft's infrastructure to be enterprise-viable. 

Which means Microsoft gets the relationship with the enterprise regardless of whose model is doing the thinking.

The picks and shovels play.

Genius.

The real AI lock-in isn't technical. 

It's how your CFO budgets for agents. Microsoft just changed that forever.

There's something nobody's talking about that I think is the real vulnerability for competitors.

Agent 365 pricing hasn't been announced. But everyone knows it's coming as a per-agent license. 

Separate headcount. Separate budget line. 

Here's why that matters beyond pricing. When you budget an agent as an employee, the psychology of the purchase changes. 

It's staffing. It's headcount planning. It's a conversation with your CFO about head-count-to-revenue ratio.

That conversation is much easier to have when the agent is licensed through the same HR system where you track humans. 

When it shows up on payroll dashboards. When it's built into workforce planning systems.

Competing agents can't participate in those conversations because they're not organisational members. 

They're external tools. External tools don't get budget.

Organisation members get budget because they're assumed to be permanent.

This is the unseen moat.

It's not technical. It's psychological. 

It's about how enterprises budget and plan.

So, the agent wars have already ended. The winner wasn't decided by better AI. It was decided by org chart access.

My opinion is that this move is worth more strategically than Copilot was. 

Copilot changed how enterprises think about AI assistance. This changes how enterprises architect entire workforces.

There are three paths competitors can take to combat Microsoft roadmap item MC11883300. Agent ID.

All three are losing. What can competitors actually do here:

Option one. Build your own governance infrastructure. 

Which costs billions. Takes years. 

Requires hiring security, compliance, and identity teams. By the time you finish, Microsoft is already embedded in every enterprise that couldn't wait.

Option two. 

Partner with Microsoft and run your agents through Entra ID. 

Which means you accept that Microsoft owns the relationship with the enterprise. 

You become a model provider on top of Microsoft infrastructure.

Option three. Go vertical.

Build agents so specific to an industry that governance becomes secondary to outcomes. 

This works for Salesforce. Might work for a few others. But it doesn't scale to horizontal adoption.

All of these paths are losing paths compared to being the Microsoft of agents. 

Which means the talent, capital, and effort going into competing agents is mostly wasted. 

Every tech monopoly was built the same way. It's being applied to a game Microsoft already won by changing the rules.

Infrastructure companies win by making themselves mandatory for compliance and governance. Not by being the best. By being mandatory.

Google made browsers mandatory.

Microsoft made Windows mandatory.

Amazon made cloud infrastructure mandatory.

Salesforce made CRM mandatory.

Now Microsoft is making identity mandatory for agents.

The pattern is always the same. 

You take emerging technology. You embed it into a foundational layer that enterprises depend on. 

You make that layer governance-native. Then you don't have to compete anymore. You've changed the rules of the game.

If you're planning to use agents in your organization, you've already made your choice. You just don't know it yet.

Microsoft's move with Agentic Users is the clearest signal I've seen that infrastructure wars are decided not by innovation but by governance architecture.

The company that controls enterprise identity will control enterprise adoption. Not the company with the best model.

Whether you're building agents, investing in agents, or planning to use agents in your organization, understand this. 

The choice isn't between agents. It's between who controls your governance infrastructure. 

Everything else cascades from that.

Microsoft understood that. Everyone else is still focused on model quality.

And that's why they won before the war even started.

Everything you’ve been taught about business strategy is a lie.

The winners in AI are doing the exact opposite of it.

No focus. No target market.

But.

It’s working.

I’m calling it the two-wallet strategy.

It relies on two models and three levers.

Model one.

• Build a viral consumer product that teaches the model as a loss leader.

Model two.

• Sell the enterprise version that prints money and locks you to lock in margin.

The trick then becomes managing the three levers: reach, unit cost and distribution.

OpenAI basically wrote this playbook.

Companies like ElevenLabs, Synthesia, Runway and others are speed-running it.

I’m Blake. I’ve spent the last 10 years in tech advising companies on data and AI strategy.

I’m here to break down for you the most interesting moves the players are making in AI.

Products. Moats. Distribution tactics. Everything is curated so you can stay informed.

By the end of this video you’ll be able to score any AI company in 15 seconds and predict who will win in their market.

Here’s what you’ll get from this breakdown.

• The exact playbook all the key players in AI are employing

• Why no one is purely B2C and B2B in AI.

• Why you are so important to AI companies

• And how to spot an AI winner.

To do it well, a company needs to execute flawlessly across three domains.

Reach, unit cost and distribution.

Step one, get as much reach as possible in the consumer world.

Ship a magical unrivalled experience across the generative domain.

• Voice, Text, Video, Image, whatever.

• Generate a crazy amount of PR and hype.

• Ship like crazy

Whatever needs to be done to get as many people as possible to use the platform.

Use that momentum to:

1. Build trust, habit and familiarity in low stake environments.

2. Create a crowdsourced feedback loop to improve the product.

3. Deliver a bug free enterprise grade offering.

4. And of course, raise more money

Once in motion, move onto step two, locking in scaled pricing discounts on compute so you can:

1. Further increase margins

2. Create an enterprise GTM motion with pricing flexibility

3. And of course, raise more money.

Once this is complete, move onto step three, improving distribution and ecosystem lock in.

This will allow you to:

1. Increase product usefulness and switching cost

2. Scale usage across domains and use cases and,

3. You guessed it, raise more money.

Then, repeat.

OpenAI was the first mover in this space and basically created this playbook.

How?

Picture this, it’s late 2022.

Everyone’s waiting for another mediocre Apple iPhone update.

Then boom, ChatGPT launches.

It rockets to 100 million users within weeks.

Insane reach.

They got to work improving the product, by making you the product.

See, models learn from feedback.

A public chatbot collects thumbs-up, thumbs-down, edits, and edge cases all day.

That shows the model what a “good” answer looks like, and what breaks it.

Everytime you call ChatGPT stupid, that’s a sign.

Get the collective power of the internet together and all weird bugs will be found in hours.

Fix them once, and the enterprise gets a safer, calmer version.

Makes sense why they bought Statsig, the modern product development platform now doesn’t it.

Genius.

Then OpenAI ships the API.

Then ChatGPT Enterprise.

Suddenly they are a B2B company.

This did not look exciting to everyday users. SSO. SLA’s. Snorefest.

What we missed, is that the enterprise public demo led to an absolutely insane amount of inbound demand from the enterprise. .

Why?

People tried ChatGPT at home.

• They trauma-dumped.

• They created ghibli characters.

• They made AI girlfriends.

After they were done with the fun, they took it to work.

So when the boards and IT departments decided on their AI pilot projects, and got to evaluating vendors, they weren’t hearing a cold pitch.

They had already heard about this AI.

They had already used it. A lot.

In fact, inbound requests were so heavy that OpenAI built an AI to triage the leads.

Yes, an AI to sell AI.

But of course there was another hurdle.

They were unproven in the enterprise.

Ahh, they thought of that too.

Most of OpenAI’s early enterprise wins didn’t arrive from door-to-door selling.

They flowed through Microsoft’s Azure channel.

Roughly four out of five ChatGPT Enterprise customers came that way.

Familiar seller. Faster paperwork.

That usage brought money in the door.

That money let them pre-buy chips, run GPUs hot.

At a tiny volume, the serving probably cost about thirty dollars per million tokens.

At extreme volume with long-term capacity locked in, it landed at nearly five dollars.

Almost 80% cost to serve reduction.

Then they started locking in on distribution.

Sora, Voice mode, Atlas, Sky, Apps SDK, Workflows.

Bringing in Jonny Ive to create a device.

Be useful to everyone, by partnering with everyone, across all domains.

Their latest move, Atlas.

So, if you link 1, 2 and 3 together, you get a flywheel:

Consumers try → feedback improves quality → companies adopt → revenue funds more compute → costs drop → increase the territory → repeat.

Ship fast, reduce costs, improve ecosystem, repeat.

That is the whole point of the two-wallet strategy.

OpenAI is absolutely the poster child.

So, can AI companies win B2B-only?

Yes.

One example is Cohere.

They specialize in large language models and AI products for regulated industries.

Right now, they generate roughly $100 million in annual revenue and have strong margins from private deployments.

Tight privacy with on-premise, white-glove enterprise support.

The path they’ve chosen trades speed and mindshare for control and depth.

The catch… fewer organic leads, slower cycles, and less public feedback.

You must be exceptional at the hard, boring stuff in the niche, or bigger brands will out-muscle you.

So now you may ask, what about B2C-only?

Also yes.

Midjourney shows it with about $500 million in revenue and a tiny team.

No API. No enterprise push.

It works because prosumers pay and the art is stunning.

Seriously, go try it.

Buuuuut. There’s a ceiling.

No big contracts, no platform lock-in, and limited leverage on chip deals.

Once rivals get “good enough,” businesses choose the tools with APIs, SLAs, and procurement boxes already ticked.

They’ve won the battle, but I can guarantee it’s going to be hard for them to win the war.

There’s a tricky conundrum.

• If a consumer app does everything, why would anyone build on your API?

• Or, if you only ship an API, how do you learn what people actually want?

That’s why the companies that do both have a higher chance of winning.

It’s a really big gamble to only focus on one side.

I think Cohere and Midjourney are outliers.

Let’s see what happens in a year.

So who else has run the two wallet strategy?

If you’re still not convinced yet as to why, let’s go to my absolute favourite AI company.

Elevenlabs. The most realistic voice AI platform.

They began with ultra-realistic AI voice text to speech technology. Today, they advance AI audio across speech, sound effects, and music and turn that work into products people use every day.

They started with creators.

They hit a million users in five months.

Domain one, Reach.

Voiceovers everywhere.

That visibility pulled in developers.

Developers pulled in companies.

They now offer teams the ability to deploy voice agents that listen, talk, and act, localize content at scale, tell stories, and improve accessibility.

They offer all the unsexy stuff through the enterprise offering.

“zero-retention” modes, HIPAA add-ons, and private deployment options.

Enterprise data stays out of training by default.

Then scale kicks in.

As usage exploded, Elevenlabs cut generation costs about 50% and passed savings on—down to roughly $15 per million characters at high volume.

Domain two, Margins.

At the same time they learn fast from the consumer side where users opt in.

So they improve weekly without touching a customer’s private calls.

That’s how you turn a viral tool into an enterprise platform without losing money on every clip.

They’re now on the cusp of step 3, distribution.

They’re expanding their ecosystem across the agent experience, working in partnership with many established players in the contact center and voice space.

It’s early, but I’m bullish.

I think every player in AI will eventually adopt this playbook.

If they’re B2C, they will add business plans.

If they’re B2B, they will launch a small consumer app to learn faster and build a brand.

The gravity pulls both ways.

Translation.

There will be no pure B2C or B2B AI Company.

Midjourney will get there.

Cohere will also get there.

AI moves too quickly to not.

To spot an AI winner, find out the following

• Where does the reach come from?

• Do they have a solid unit cost baseline?

• Are they building in distribution?

If one company can answer all three, you’ve probably found a compounder.

I call it the two wallet strategy.

Hopefully this was useful.

I’m Blake.

Thanks for watching.

Shihir Mehrotra (CEO of Superhuman) changed the company name from Grammarly to acquired company Superhuman.

Everyone in the tech press screamed brand suicide.

I think it’s the smartest move they could of made.

One of the most sophisticated distribution plays I have seen.

Let me go full hyperbole mode for a second.

Ready?

Superhuman are repositioning for war.

Here’s why.

The death of the grammar police

Grammarly had a brand problem.

The company built an empire on a single promise... fix your typos.

By 2025 they had assembled one of the most impressive distribution machines in software.

• 40 million daily active users.

• Presence in over 1 million apps and websites.

• 50,000 enterprise customers.

• $700 million in annual recurring revenue with 80% gross margins.

Unfortunately for them, when you hear “Grammarly” you think grammar.

Not strategy.

Not intelligence.

Not the future of work.

To win in the AI wars, specifically the productivity software market, the growth is in platforms.

Bundles.

Operating systems for work.

The big dogs know this.

1. Microsoft Copilot, despite its messy rollout, now sits inside 70% of Fortune 500 companies.

2. Google Gemini powers 46% of US enterprises in productivity workflows.

3. OpenAI just launched AgentKit, a complete platform for building, deploying and managing AI agents.

The future belongs to whoever owns the layer where work actually happens.

Grammarly doesn’t own that, so they killed it.

Here comes another agent platform

The companies that control the orchestration layer, aka the place where agents talk to each other and to your data, will capture disproportionate value.

The connective tissue between you and every app you touch.

Your email, docs, CRM, Slack, everything.

Grammarly (Superhuman) is building that AND they’re running multiple AI agents on top of that layer.

• One agent fixes your grammar.

• Another researches prospects.

• A third schedules meetings.

• A fourth drafts responses in your voice.

All orchestrated by an agent platform called Superhuman Go.

So what did Superhuman bundle?

Grammarly’s writing tools, Coda’s workspace platform, Superhuman Mail and the new agent overlay with an Agent Store and SDK.

An operating system for AI-augmented work.

A classic “rundle” strategy (recurring revenue bundle).

Microsoft did it with Office 365 and Adobe did it with Creative Cloud.

Bundle complementary products at a unified price point, make the integration seamless and make switching costly and difficult.

A customer using one Superhuman product might churn.

A customer using four, with agents running across all of them, trained on their writing style, plugged into their workflows is well and truly locked in.

Distribution beats demo, always

Grammarly already won the distribution war.

They sit inside 1 million apps and websites.

They have 40 million people using them every single day.

To understand why this matters, you need to understand what knowledge workers actually do all day:

• The average employee spends 28% of their workweek on email.

• 11 hours a week just managing the inbox.

• They receive 117 emails and 153 Teams messages every single day.

• People get interrupted every two minutes during core work hours.

• 275 interruptions per day.

Microsoft calls this the “infinite workday”.

This is the problem Superhuman Go attacks.

Context-switching.

The cognitive tax of navigating fifteen different tools to complete one task.

They’re attacking it from inside the tools you already use. That’s the distribution moat.

When OpenAI launches AgentKit, they need you to come to ChatGPT.

When Microsoft rolls out Copilot, they need you to open a sidebar.

When Google pushes Gemini Enterprise, they need you to adopt another interface.

Superhuman says... “We’re already where you type. Just turn us on.”

That’s a $1 billion head start.

The platform playbook

Now let’s talk about how their platform scales.

They’re shipping an Agent Store and an Agents SDK. That means third parties can build agents that run on Superhuman’s rails.

• Quizlet for education.

• Fireflies for meeting notes.

• Speechify for audio.

Each new agent increases their surface area. Each integration raises switching costs.

This is the flywheel that made iOS and Android unstoppable.

Apple didn’t build every app. They created the marketplace and the tools, then let developers fill in the gaps. By the time Android arrived, Apple already had hundreds of thousands of apps.

Switching meant losing your entire digital life.

Superhuman is betting the same physics apply to work agents.

Can they pull it off?

Microsoft Copilot has distribution through Office 365’s 400 million users but adoption is a mess.

Gartner’s 2025 survey found that while 94% of enterprises report benefits from Copilot, only 6% have completed global rollouts.

Google Gemini has momentum, 82 million monthly active users, 46% penetration in US enterprises but it’s fighting the exact platform dependency problem Superhuman is exploiting.

OpenAI’s AgentKit is the real threat. It launched in October 2025 with the weight of ChatGPT’s 60.5% market share in generative AI chatbots. They have developer mindshare, a visual Agent Builder and enterprise wins like Klarna (handling a majority of support tickets) and HubSpot (powering Breeze for sales automation).

OpenAI doesn’t have Superhuman’s distribution rails.

ChatGPT requires you to come to ChatGPT. Superhuman is already inside your browser, your docs and your email client.

They’re there already.

In the battle for the operating layer of work, being there already wins.

The bet

My take. This move works.

Not because the rebrand is flawless, it’s not. Not because “Superhuman” is a better name, that’s a matter of personal taste.

It works because Grammarly understood the fundamental law of enterprise software in the AI era.

Distribution beats differentiation.

You can build the smartest agent. The best model. The sleekest interface.

If you don’t control the surface where work happens, where people type, where they read, where they decide, you’re renting attention from someone else.

They’re doing it at exactly the moment when the market is deciding who gets to be the operating system for AI-augmented work.

Microsoft and Google are the incumbents, defending installed bases.

OpenAI is the insurgent, leveraging ChatGPT’s consumer momentum.

Superhuman is the dark horse, already embedded in your workflow, asking you to just turn on the full suite.

In five years, one of three things will be true:

1. Microsoft or Google crushed everyone by bundling AI so deeply into Office/Workspace that third-party overlays became irrelevant.

2. OpenAI became the “AI OS” by making ChatGPT the default interface for every digital task.

3. Superhuman became the independent, cross-platform orchestration layer because they own presence at the point of creation.

Whatever happens, we’ll soon find out what the most powerful distribution moat actually is.

Until next time.

Blake

Summary of article:

• AI browsers turn sites from destinations into data pipes and your homepage will become an API.

• Zero-click will rise as AI answers eat clicks & links must become invoices.

• The shift from SEO to GEO requires permissions, origin, structure, to fix attribution and ship an explore mode.

• There’s mass urgency to meter bots, track share-of-answer, pay where answers happen or the ecosystem thins out.

Let’s begin

There are two equally realistic consequences that stem from the AI browsers actually being used.

1. All human traffic is drained from the web

2. It funds better journalism

Finger’s crossed for number 2.

But.

It all depends on the ability of publishers and browsers to create a realistic co-exist strategy, filled with payments, citations (where content originated from) and a curated “explore” mode.

All before the AI web becomes too ingrained in it’s current form.

I’m about to drop some statistics that will make you shake in your boots.

But first.

An outlandish statement, rooted in facts.

Your homepage is becoming an API.

We’ve heard it 100 times already…

When an AI summary appears in Google search, people click through only 8% of the time.

Half the 15% rate when there’s no AI box.

The answer often ends the journey right there but ‘less clicking’ doesn’t have to mean ‘less paying’ because Cloudflare launched ‘Pay Per Crawl’ in June 2025 and now blocks many AI crawlers by default.

The web’s economics can shift from blue links to pay per access, if we move fast enough.

Why now

Your content is fetched, summarised and presented elsewhere.

The stakes are simple, if usage doesn’t translate into money and reputation, the supply of good content thins out.

Google’s AI Overviews and ChatGPT Search normalise answer‑first UX and the rollout is accelerating.

Early rollbacks after the infamous “glue on pizza” fiasco looked like retreat.

Unfortunately, a16‑month BrightEdge study tracking millions of keywords found AI Overview coverage rising to 44.4% overall by September 2025 and a staggering 85% for informational queries.

By October, AI Overviews appeared in over half of all search results.

The “answers on top” pattern isn’t maturing, it’s spreading.

The velocity is uneven but relentless. Couple of key ones:

• Healthcare queries jumped from 72% AI Overview presence in 2024 to 87% in 2025.

• Education queries exploded from 18% to 87%, a 69‑percentage‑point surge in twelve months.

• Entertainment queries saw a 528% increase in AI Overview presence between March 2024 and March 2025.

• Restaurant and travel followed at 387% and 381% respectively.

Vertical by vertical, the AI box is eating the page.

Zero‑click is the logical endpoint.

Between May 2024 and May 2025, the share of searches ending without a click grew from 56% to 69%.

Thirteen percentage points of traffic that simply evaporated.

Not because users didn’t find value, but because they got it without leaving Google.

How the ecosystem collapses

Four actions taken by you:

1. You search “best portable monitors.”

2. The AI overview stitches pros, cons, and specs.

3. It then lists three models with prices.

4. You get what you came for, and you don’t click.

The review site that funded the teardown, photographed the ports, and measured the colour accuracy loses the visit.

The magic is real, but the economics that produced the review go missing.

The next teardown doesn’t get funded.

Multiply that pattern across medicine, law, recipes, tutorials, investigations and you don’t get a UX problem but you get an ecosystem collapse.

Ooooo scary.

Let’s play fetch.

Now for the statistic drops I promised earlier.

Links will become invoices for real time crawlers.

Why?

Traffic is softening where AI summaries show up and the numbers are consistent across studies.

Ahrefs analyzed 300,000 keywords and found position‑one CTR dropped from 7.3% to 2.6% year‑over‑year when AI Overviews appeared.

A 34.5% decline.

BrightEdge reported an average 30% CTR drop since May 2024.

Digital Content Next members saw a 1–25% traffic range depending on vertical, with many clustering around the 25% mark by August 2025.

In the worst cases?

DMG Media reported an 89% CTR decline for certain searches to the UK’s Competition and Markets Authority in July.

The pattern also varies by vertical but not by direction.

CNN lost roughly 30% of referral traffic from Google. Business Insider and HuffPost shed closer to 40%.

Entertainment, restaurant, and travel sites, categories that saw AI Overview presence surge by 300–500% bore the brunt earliest.

Healthcare and education now triggering AI Overviews 87% of the time are watching the same erosion play out in slow motion.

BUT value isn’t vanishing.

It’s moving.

Smart publishers are converting “used but not visited” into “licensed and credited.”

• News Corp’s multi‑year OpenAI deal, reportedly worth over $250 million across five years, about $50 million annually, created a template.

• Financial Times, Time, The Atlantic, and Vox followed with their own OpenAI agreements.

• Dotdash Meredith locked in a minimum $16 million per year.

• Wiley reported $40 million in AI licensing revenue for fiscal 2025, nearly doubling the prior year.

• Curiosity Stream projected $19.6 million from AI licensing in 2025.

The marginal unit of value is no longer a page view but a permitted fetch.

In fact, the real‑time retrieval side is exploding at an accelerating pace.

TollBit tracked a 49% surge in retrieval bots, in live answer‑generation agents in Q1 2025.

By Q2, human visits to publisher sites fell 9.4% while AI bot traffic quadrupled.

By June, one in every fifty site visitors was an AI bot, up from one in two hundred at the start of the year.

AI bot traffic now exceeds Microsoft’s Bingbot, making it a larger crawling force than the world’s second‑largest search engine.

Meanwhile, 26 million scrapes bypassed robots.txt in March 2025 alone.

A single month.

Let that sink in.

The share of bots ignoring robots.txt jumped from 3.3% in Q4 2024 to 12.9% in Q1 2025.

Cloudflare formalised the obvious in June 2025, launching default AI‑crawler blocking and Pay Per Crawl.

The new “paywall” sits at the bot boundary and the CDN layer becomes the cashier.

Early adopters include some of the largest publishers and platforms on the web like Condé Nast, Associated Press, Reddit, Pinterest, Time, BuzzFeed and O’Reilly Media.

Google’s crawl‑to‑referral ratio worsened by 24.4% in Q2 despite Googlebot making 34.8% more requests.

The obvious translation… Google is taking more and giving back less.

From SEO to GEO

Traditional SEO fought for position on a list.

In AI answers, the interface must justify inclusion, not just placement.

That shifts optimisation from keywords and backlinks to clarity (claims backed by evidence), origin (who said what, when), and permission (license signals).

The research community is already framing tactics as GEO by making content more likely to be selected and cited by generative engines.

Research from Princeton, Georgia Tech, and the Allen Institute for AI identified the most effective GEO tactics through rigorous testing across multiple models.

Adding expert quotations delivers a 41% visibility improvement.

The strongest single optimisation.

Some other significant ranking factors:

• Statistics and cited sources each boost visibility by 30%.

• Enhanced fluency adds 22%

• Technical terminology increases it by 21%

• Authoritative tone contributes 11%

Notably, traditional keyword stuffing now reduces AI visibility by 9%, confirming that GEO rewards substance over gaming.

The citation landscape also varies by interface.

Google’s AI Mode, its conversational search interface that launched in the UK in July 2025 , favours brands (90% inclusion rate) and stability, pulling from five to seven source cards per response.

Week‑to‑week volatility is low. If you rank, you stay.

AI Overviews, in contrast, cite brands only 43% of the time but include 20+ inline citations per response and exhibit 30x higher week‑to‑week volatility.

It’s a faster‑moving but less predictable ranking environment.

For publishers, that means dual optimisation.

The content being comprehensive enough for AI Mode’s breadth whilst being authoritative enough for AI Overviews’ selectivity.

In fact, 54% of AI Overview citations match the top ten organic results, but 46% come from outside the top ten.

That’s nearly half the cited sources bypassing traditional SEO entirely, pulled in based on relevance signals the old algorithms missed.

Traditional rank‑and‑yank SEO still matters for the majority, but GEO opens a side door for deep, well‑structured content that never cracked page one.

There’s a catch though.

Up and coming vendors like perplexity who employ a growth at all cost mentality are being found out, quick.

Authority is wobbly without guardrails.

• Perplexity faced repeated allegations through 2024 of ignoring robots.txt via indirect scrapers.

• AWS opened an internal probe in June 2024 after reports Perplexity violated platform rules.

• In September 2025, Encyclopedia Britannica and Merriam‑Webster sued Perplexity for “unlawful copying of reference material.”

The pattern illustrates why permission and provenance matter beyond compliance, they’re trust signals.

Engines that foreground licensing and source panels will earn credibility, while publishers that sign and structure their work will win inclusion.

Attribution quality remains dire.

Columbia’s Tow Center tested ChatGPT Search in December 2024, asking it to attribute 200 quotes from recent news articles.

The AI misattributed sources 76.5% of the time, or 153 out of 200.

It often cited plagiarised or syndicated versions instead of original publishers, even for outlets with OpenAI licensing agreements.

Only seven responses acknowledged uncertainty, the rest delivered confident but wrong citations.

Publishers blocking OpenAI’s crawler still saw their content misrepresented.

Google’s knobs are moving, but in both directions.

It briefly dialled back AI Overviews after the glue‑on‑pizza debacle in May 2024, then expanded coverage with clearer intent boundaries through 2025.

Users can force a link‑only view via the Web filter.

An escape hatch that preserves traditional search when needed.

Time to kiss the ring?

Publishers with OpenAI licensing deals see 7x higher click‑through rates from ChatGPT Search than those without agreements and 88% more scraping.

However, the referral conversion is substantially better.

The pattern shows up across multiple industry reads.

Fewer clicks, but better ones.

A 7x CTR multiplier sounds great until you realise it’s 7x of nearly zero.

ChatGPT still sends 831x fewer visitors than Google as of September 2025.

The durable fix is usage‑based licensing plus visible credit.

When money follows mentions and mentions expose diverse sources publishers are paid to go deep again and users can see where the answer came from.

That’s the only model that keeps the long tail funded.

Time is running out

Zero‑click is coming, BUT zero‑click doesn’t have to be zero‑sum.

If we turn the homepage into an API, the link into a receipt, and the answer box into a map, the web stays alive and strange.

The destination changes but the value doesn’t.

1. The infrastructure exists.

2. The deals are being signed.

3. The lawsuits are clarifying the rules.

What’s missing is urgency.

Every day we delay shipping Explore Mode, publishing share‑of‑answer metrics, and enforcing Pay Per Crawl is another day habits harden around free, frictionless, uncredited consumption.

We know how this movie ends if we do nothing.

We’ve watched social platforms hollow out media before.

The AI shift is faster and the stakes are higher, because the content isn’t just redistributed, it’s dissolved into answers and reconstituted without attribution.

We also know what works.

Payments that follow usage, orgin that survives summaries and interfaces that reward curiosity alongside efficiency.

The web won’t die. It will invoice.

If we ship the right systems before the defaults lock in, those invoices will actually get paid.

The homepage is an API now.

The question is whether it’s an API you control, or one that simply takes what it needs and moves on.

Until next time,

Blake

Who remembers when the annual Apple conference was the most exciting technology event of the year?

Now it’s a spec bump. New camera. Big whoop. The magic is happening somewhere else. Fort Mason in San Fransisco (thank god, it’s not another iPhone launch).

OpenAI’s DevDay 2025.

I think it’s safe to say (with a slight hint of hyperbole), that the tooling released is designed to replace the entire concept of software as we know it.

I am so here for it. The attempt to control the future of human productivity. I love it

What an absolute platform play from the titans of tech.

The platform play

OpenAI just announced their transformation from AI company to OS for Intelligence.

They’re narrative looks pretty compelling.

Seven interconnected products, 800 million weekly users, and a business model that lends itself to the future of outcome based pricing (see my article here).

It’s been working so far. They’re killing it.

• 4 million developers (doubled in two years)

• 6 billion API tokens per minute (30x growth since 2023)

• $89 billion projected revenue by 2030 (conservative estimate)

But here’s what I think most missed. It isn’t about artificial intelligence, but capitalism.

The creation of a new economic system where a handful of platform owners control the means of cognitive production.

Sounds smart. Really it is quite simple.

It’s a platform play.

Own the rails.

Basically, be Google.

The Apps SDK

Think of it like Facebook’s news feed on steroids.

Remember when Facebook’s News Feed brutally turned publishers into renters on the Facebook feed?

OpenAI’s Apps SDK just literally going to do that to the entire software industry.

With the Model Context Protocol and Agentic Commerce Protocol, OpenAI is positioning themselves as the middleman for every software interaction on earth.

Canva, Zillow and Spotify (~18 launch partners) have already bent the knee, agreeing to share revenue and embed their entire user experience inside ChatGPT.

“King in the North, King in the North, King in the North!”

Yep. I’m a GOT fan.

If this pans out for OpenAI, they will become the new iOS, except instead of apps on your phone, it’s intelligence in your workflows.

Every business process. Every creative task. Every analytical decision.

Just think about those sweet, sweet transaction fees and data.

The TAM for this play? $285.6 billion in enterprise AI alone by 2030.

They might actually get a chance to repay those IOU’s for GPU’s, with that sort of market size.

No longer will you be switching to Canva to create a Social Media post, graphic, etc.

But there was an even cooler one shown. One that I think really shows it’s power as an OS.

Imagine you’re looking for a house. You go on Zillow (or Domain for us).

Well, the OS can basically sit on the side, and provide access, data and context to houses, that Zillow just wouldn’t have.

Think of it like a buyers agent sitting right next to you, whilst you peruse the feed (without the commission).

That’s just one use case.

Ultimately, the more time you spend on ChatGPT, the better it becomes, the more utility it offers, the less you’re likely to go anywhere else.

Also, it’s probably going to be the BIGGEST black hole of data. They will have so many data from all types of applications flowing through.

Danger ahead, apps who use this. ChatGPT might be coming for your customers.

They’ve already come for your UX.

AgentKit

Remember Zapier, n8n or Workato?

Felcia, bye.

Another set of tech darlings obliterated (hyperbole).

This kit is basically a competing visual workflow builder that let companies create AI employees in hours instead of months.

It’s already worked pretty well in their beta.

• Ramp reduced agent development time by 70%.

• Cisco deployed enterprise-wide coding assistants.

• Deloitte deployed Claude (Anthropic’s competitor) to 470,000 employees.

Yes. Yawn. Workflows.

Not a new concept.

But. It’s a powerful one.

$890.5 billion in annual value creation from business process automation alone. OpenAI’s positioned to capture $222.6 billion of that through platform fees.

A small nation’s GDP.

Now tell me that’s a snooze-fest announcement?

The $15 Token

Let’s talk about GPT-5 Pro and what I’m going to call “The Cognitive Rolex Strategy.”

At $15 per million tokens (50% premium over standard pricing), it ain’t cheap.

In fact, OpenAI just created the luxury end of artificial intelligence.

Finance firms, law firms, healthcare organisations or anyone whose mistakes cost millions will pay premium prices for premium thinking.

To this point in time, it’s been a race to the bottom.

• “Models are becoming a commodity”.

• “The price of intelligence is diminishing”.

• “The moats are disappearing”.

Not to make myself sound too Australian here, but, “yeah nah”.

GPT-Pro is a Rolex.

Luxury goods have the highest margins and create the strongest brand moats.

When Goldman Sachs is paying $15 for AI reasoning that costs OpenAI pennies to provide, you’re selling cognitive insurance.

Can’t wait to flex my GPT-5 Pro $200 p/m subscription to my potential suitors.

Form a line ladies.

Meanwhile, gpt-realtime-mini and gpt-image-1-mini (70% and 80% cost reductions respectively) will flood the market with cheap AI capabilities, creating the same dynamic that made smartphones ubiquitous. Premium products for enterprise customers, commodity versions for everyone else.

The playbook is a classic platform economics play. Subsidise the bottom to capture market share, extract maximum value from the top where switching costs are highest.

Sora 2

Humans are visual creatures.

Some of the outputs of this model, are completely indistinguishable from reality.

Truly. Check them out if you get a chance.

If you work in video production, animation, or visual content creation, Sora 2’s API availability probably just put an expiration date on your career.

“Come with me, as I go to Bondi Beach and get an Acai”.

No. Not anymore.

Sora 2, is not AI-assisted creativity. It’s not relying on human creativity.

We’re talking about full-stack video production. Conception to final cut. Handled by algorithms that never sleep, never demand residuals, and improve every day.

(… and an end to this format).

Mattel is already using it for product conceptualisation.

How long before Netflix realises they can produce content without writers, directors, or actors? The entire creative economy is $125.4 billion in annual value.

Is it about to be ‘platformed’?

OpenAI’s positioned to capture $50.2 billion of creative industry value through their platform.

When your sugar daddy gets wandering eyes

But. OpenAI does have an existential vulnerability that should terrify investors.

It’s called Microsoft Bing.

Lol, just kidding. Not Bing.

But, Microsoft as their primary cloud provider, distribution partner and $10 billion sugar daddy just integrated Anthropic’s Claude into Office 365.

430 million Office 365 users now have access to OpenAI’s primary competitor, directly contradicting OpenAI’s platform strategy.

A bit of a Slack vs Teams story. We all know how that ended.

If you don’t… never compete with free.

When your biggest partner starts hedging their bets, you’re no longer a strategic asset.

Microsoft discovered that Claude outperformed GPT-5 in productivity tasks like Excel automation and PowerPoint generation. In platform capitalism, performance trumps partnerships.

The Apps SDK is OpenAI’s desperate attempt to reduce Microsoft dependency by creating independent distribution through ChatGPT’s user base.

OpenAI is building their own iOS because they can’t trust being an app forever.

There’s also direct competition.

Google’s $2T war chest vs. Anthropic’s enterprise focus

The competition isn’t sleeping.

While OpenAI chases platform dominance, Anthropic is capturing enterprise customers one Fortune 500 company at a time.

• Commonwealth Bank: 50% reduction in scam losses

• IG Group: 70 hours weekly saved per employee

• SK Telecom: 34% customer service improvement

Anthropic’s strategy is to be the “safe choice” for risk-averse enterprises while OpenAI experiments with platform moonshots. When enterprise customers are paying $28.5 billion annually (OpenAI’s projection for GPT-5 Pro revenue), losing even 20% of that market to competitors creates massive downside risk.

Google’s counter-play leverages their 2 billion Workspace users. In fact, their base is 8x larger than ChatGPT’s user base. When Google embeds AI directly into Gmail, Docs, and Drive, they don’t need a platform strategy.

They already ARE the platform.

OpenAI must simultaneously execute flawlessly across technical development, business model innovation, competitive responses, regulatory compliance, and market timing.

The historical pattern is clear:

• Industrial revolution: Those who owned factories controlled manufacturing.

• Information age: Those who owned platforms controlled distribution.

• Intelligence age: Those who own AI platforms control cognitive labor itself.

The economic transformation:

• 2025-2027: AI augmentation (75% human, 25% AI work).

• 2027-2030: AI orchestration (55% human, 45% AI work).

• 2030-2035: AI-native organisations (35% human, 65% AI work).

I think OpenAI has roughly 18 months to achieve platform lock-in before competitive dynamics make dominance mathematically impossible.

Four Scenarios for 2030

Scenario 1

OpenAI achieves network effects lock-in. They become the cognitive infrastructure for global economy.

$180 billion revenue, 35% market share.

The new oil, the new rails.

Choo choo.

Scenario 2 (my pick)

The competitive fragmentation Google, Anthropic, and Microsoft prevent OpenAI dominance.

Market splits between specialised players.

$95 billion revenue, 25% market share.

Still valuable, no longer inevitable.

Scenario 3

The platform wars competition drives margin compression.

Enterprise customers resist platform lock-in.

$45 billion revenue, 15% market share.

Profitable but not transformational.

Scenario 4

The AI winter driven by technical limitations or market correction deflates AI valuations.

$15 billion revenue, 5% market share.

The dot-com crash of artificial intelligence.

Bubble pop.

Delayed AGI.

GRIM.

My prediction

My prediction is scenario 2. That, by 2030, three companies will control 70% of global AI infrastructure. The rest will be digital sharecroppers, paying platform fees for access to intelligence itself. The concentration of cognitive power will make today’s wealth gaps look like rounding errors.

One thing is very clear. OpenAI just placed a $4 trillion bet on platform capitalism’s final evolution.

Capitalism, I choose you.

Whether they win or lose, we all live with the consequences. Welcome to the Intelligence Age.

Check your human privilege at the door.

Until next time,

Blake.

My late grandfather was quite the philosopher. No matter the situation, he had a great line for it. Here’s a couple of good ones:

• ‘A fish stinks from its head = Leadership matters.

• ‘Every pig hangs by its own two feet’ = you’re responsible for your own actions.

• “If it walks like a duck, quacks like a duck… It’s a duck” = It’s usually what it seems.

That last one, is such an appropriate line for what’s happening right now in AI. The present AI economy waddles and quacks. The major players are participating in a really wild game of duck, duck goose.

Keeping with the duck theme, let’s join them and play a quick game with them.

Nvidia is committing up to $100 billion to OpenAI, while also planning to supply it with the very systems that money buys.

Duck.

Oracle signed an eye‑watering $300 billion compute deal with OpenAI and promptly started raising debt to build the DC’s that run, wait for it, Nvidia gear for OpenAI.

Duck.

CoreWeave expanded its OpenAI contracts to $22.4 billion this year and separately inked a $6.3 billion “we’ll buy your unused capacity” backstop with Nvidia that runs through 2032.

Duck.

• $450B in circular flows.

• $250B revenue shortfall.

• Concentration past dot‑com levels.

And, no one has the funds to pay. It’s all coming from ‘future revenue’ opportunities.

I don’t know about you, but that looks suspiciously like everyone funding everyone else written on an IOU.

Duck.

Let’s sum the $450B.

• OpenAI to Oracle: $300.0B (Cloud Services)

• Nvidia to OpenAI: $100.0B (Direct Investment)

• CoreWeave to OpenAI: $14.4B (Cloud Services)

• Nvidia to CoreWeave: $6.3B (Cloud Services/backstop)

• Microsoft to OpenAI: $13.0B (Strategic)

• OpenAI to CoreWeave: $1.2B (Services Purchase)

Yep.

Who’s going to be left holding the bag?

Hopefully not us.

Much like the very catchy Fatboy Slim “eat, sleep, rave, repeat” tune, my awful attempt at a graphic says the quiet part out loud.

Invest. > Earn. (“revenue”) > Purchase. > Report. (“demand”) > Repeat.

Okay. Maybe not as catchy as I thought.

GOOSE!!!!!

P.S I spent WAY TOO LONG making a graph of this, please appreciate it.

Economically, not legally, this is Ponzi‑adjacent.

The loop works beautifully while confidence and credit expand. It works terribly as soon as someone looks at it with their glasses on.

Ahhh it’s so meme-able.

Literally, think of this lovely nanna as you look at the following graph.

• Spend: $50B to $400B (2022–2025)

• Revenue: $12B to $150B

• Gap: $38B to $250B compounding at ~87% a year

• Coverage: revenue rises from 24% to 37.5% of spend

In translation, in 2025, each dollar of AI spend is met by $0.38 of revenue.

Would you invest your nanna’s money in that?

I don’t think so.

I’m gonna make it even worse.

History, yo.

Let’s take a look into some really interesting past bubble popping schemes.

• The Railroads

• AOL Barter Deals

• Telecom Equipment Financing.

Long story short if you don’t have time to read.

Each bubble perfected the art of financial engineering while seemingly learning from its predecessors’ mistakes.

• In the Railroad Vendor Financing (1845–1847), $2 billion in equipment financing where manufacturers funded customer purchases.

• In the AOL Barter Deals (2000–2002), $600 million in cross-advertising and service arrangements.

• In the Telecom Equipment Financing (1999–2001), $25 billion in vendor financing by Nortel, Lucent and Cisco.

In every case, it all came undone and someone was Goose.

The AI circular investment web represents an 18x increase over the telecom vendor financing bubble and 750x increase over AOL’s schemes.

The railroads though, still win

Alright.

A bit of storytelling and history blended.

Let’s start with the most insane.

The OG.

Gotta hand it to ol’ Hudson. He really created something here.

I’m no Margot Robbie in a bath (Wolf of Wall Street), so hopefully my data backed insights and storytelling make up for it.

The railroads

George Hudson

The story begins with George Hudson who inherited £30,000 in 1827 and transformed it into control over 1,016 miles of British railway.

Hudson was the OG duck, duck goose player.

He pioneered what would become the template for all subsequent circular financing schemes, using initial success to create the illusion of demand that justified ever-larger investments.

George Hudson controlled 1,016 miles of railway by 1846 through £25 million in circular financing and extensive parliamentary bribery.

Duck money.

Hudson’s method was elegantly simple yet devastatingly effective. He would form a new railway company, personally guarantee dividends of 6%, then use the initial capital to purchase equipment from the same suppliers who had invested in his companies.

By 1845, Hudson controlled 15 separate railway companies through a web of cross-investments, circular financing arrangements and strategic parliamentary corruption totalling £25,000 in bribes.

What unravelled everything?

The equipment vendor financing web.

The goose.

The railway bubble’s circular financing centred on equipment manufacturers like locomotive builders and rail producers. Hudson would arrange for these suppliers to provide “deferred payment terms”, essentially lending money to railway companies to purchase their own products.

By 1847, equipment vendor financing had reached £15 million annually, representing nearly 35% of total railway investment.

The genius of Hudson’s scheme was its self-reinforcing nature.

Successful railways attracted investors, investor capital funded equipment purchases, equipment sales generated profits for suppliers and supplier profits were reinvested in new railway ventures.

The same £1 might circulate through this system five times, each time appearing as “new investment” in the rapidly inflating railway sector.

All I gotta say is… Duck.

Hudson’s empire unraveled when parliamentary committees discovered his accounting manipulations in 1849.

Duck.

He had been paying dividends from share capital rather than profits, selling land he didn’t own to his own companies and inflating construction costs to justify higher share prices.

Goose.

The revelations triggered a 67% collapse in railway share prices and left thousands of investors with worthless paper.

In total, across the whole industry, railway capital investment reached £44 million in 1847, representing an extraordinary 7% of British GDP.

In today’s money, this cumulative investment of approximately £225 million (1843–1850) equates to roughly $4 trillion, making it proportionally larger than any subsequent infrastructure bubble.

Railway stock prices peaked at 2,062 index points in August 1845, then collapsed 67% to 673 by 1850.

Ouch.

Seeing any parallels?

Tell me this doesn’t line up with four hyper-scalers announcing identicalmega‑campuses within the same grid constraint and despite limited evidence of sufficient demand to support such capacity.

What’s so interesting about this, is that the physical infrastructure remained.

The railways Hudson built became the backbone of British industrialisation for the next century, proving that circular financing bubbles often create genuine long-term value even as they destroy investor wealth.

• Yes, AI is a bubble.

• Yes, it is transformative.

• Yes, we still use railways.

• Yes, we will use these DC’s.

They can all be true at the same time.

Maybe it will all be owned by Evil Mega Corp, or our AI overlords, not Google and Microsoft.

AOL’s barter deal revolution

America Online (AOL) perfected circular revenue arrangements in ways that would make daddy Hudson jelly.

Unlike the railway magnate’s crude equipment financing, AOL created sophisticated “barter deals” where advertising services were exchanged for equity stakes, cross-promotional agreements and reciprocal purchasing arrangements.

AOL’s circular revenue schemes peaked at $450 million in Q3 2000 during the Time Warner merger, supporting a stock price of $158

AOL’s circular revenue machine operated on three levels.

1. Basic barter deals where AOL exchanged online advertising for services or equity in dot-com startups.

2. The massive Bertelsmann arrangement worth $400 million, where AOL’s European joint venture partner purchased AOL advertising services as part of a complex media partnership.

3. Cross-promotional deals where AOL would invest in or partner with companies that then spent money on AOL services.

Buy. Invest. Buy.

The crown jewel of AOL’s circular revenue scheme was its relationship with German media giant Bertelsmann.

In March 2000, AOL and Bertelsmann signed a joint promotion agreement that ‘appeared’ to be worth $400 million in advertising revenue to AOL over four years.

The reality was more complex.

Bertelsmann agreed to “spend” $400 million on AOL advertising in exchange for preferred placement, cross-promotional services and strategic partnerships that effectively meant AOL was selling advertising to itself.

The SEC later determined that much of this $400 million should have been accounted for as a reduction in the acquisition price of AOL Europe rather than legitimate advertising revenue.

This distinction meant the difference between AOL showing explosive advertising growth versus acknowledging that it was paying for its own revenue.

Not very free market of you, fellas!

Where did it all come undone?

AOL’s circular revenue schemes reached their zenith during the 2000 Time Warner merger negotiations.

Total circular revenue hit $450 million in Q3 2000, representing approximately 9% of AOL’s total advertising income.

Duck.

These arrangements allowed AOL to demonstrate the “hockey stick” revenue growth that justified its $164 billion valuation and convinced Time Warner executives to accept AOL’s inflated stock as merger currency.

Goose.

When the schemes unraveled in 2002–2003, AOL Time Warner took a record $99 billion write down, acknowledged $600 million in questionable revenue and paid $300 million in SEC fines.

The lesson was clear.

Circular revenue arrangements could sustain impossible valuations for years, but the eventual reckoning was proportionally devastating.

I love the quote:

“The market can stay irrational longer than you can stay solvent”.

Do I need to show you my masterpiece graph again?

History repeats itself.

Just not in the timeframe we think.

The telecom equipment frenzy

The telecom equipment bubble took circular financing to unprecedented industrial scale. Companies like Nortel, Lucent and Cisco discovered that in a rapidly expanding market, vendor financing wasn’t just a competitive tool, it was the primary driver of growth.

Telecom vendor financing reached $33 billion at peak in 2000, with Lucent leading at $10.5 billion before collapsing in 2001.

The vendor financing explosion began modestly in 1999 with $3.5 billion in total industry lending, but within 18 months had grown to $32.8 billion by Q4 2000. Lucent alone had $7.2 billion in customer financing outstanding by March 2000, a 65% increase from the previous year.

Nortel’s vendor financing represented nearly 7% of its total revenue, while Cisco extended $2.4 billion in customer loans despite having minimal disclosure requirements.

Cue the phantom demand creation machine.

The telecom vendor financing web created artificial demand through multiple circular mechanisms. Equipment manufacturers lent money to startup telecom carriers, who used the funds to purchase equipment from those same manufacturers. The manufacturers then reported this as organic revenue growth, which supported stock prices and enabled them to raise more capital to fund additional vendor financing.

More circles.

Lucent’s deal with Fidelity Holdings exemplifies the absurdity.

Get this.

Another Daddy Hudson banger.

A small car dealership with a tiny Internet division received $400 million in vendor financing to purchase Lucent equipment, despite having only $1.3 million in annual sales.

The arrangement was possible because Lucent took equity stakes and guaranteed residual values, meaning it was essentially buying its own products through intermediaries.

So Blake, where did this goose come from?

Well, this one was rough.

The telecom vendor financing collapse was swift and merciless.

By Q1 2001, total industry vendor financing had fallen to $25.5 billion and by Q2 2001 it was down to $18.2 billion as companies defaulted on obligations and equipment values plummeted. Lucent took a $501 million charge in fiscal 2000 for unpaid customer debts, equivalent to 41% of its annual earnings.

The human cost was staggering.

WorldCom collapsed with $100 billion in debt, Global Crossing failed with $25 billion in obligations and 23 major telecom companies went bankrupt in the US alone.

The equipment manufacturers that had enabled this expansion through vendor financing saw their stock prices fall 80–90%, destroying trillions in market capitalisation.

So what have we, a society, learnt from history?

Spoiler, nothing.

We are creatures of habit.

The AI Circular Investment Explosion

It’s like we never learn.

• $615 Billion and Growing.

• An infinite money glitch.

No Gooses? (There’s always a Goose).

The NVIDIA-OpenAI-Oracle triangle represents the most sophisticated circular financing arrangement in economic history, with over $615 billion in interconnected deals that create the appearance of massive demand while the same money circulates through an increasingly complex ecosystem.

AI circular investment has exploded to over $615 billion by Q2 2025, dwarfing all previous bubble financing schemes.

Unlike previous bubbles, this arrangement involves the world’s most valuable companies. NVIDIA ($3.4 trillion market cap), Microsoft, Meta and others whose combined market capitalisation exceeds $15 trillion.

The systemic risk is unprecedented.

When 25% of the S&P 500’s value depends on AI revenue growth and that growth is sustained by circular financing arrangements, the potential for cascading failure exceeds anything seen in previous technology bubbles.

The math only works if artificial general intelligence (AGI) arrives on schedule and generates the revenue necessary to justify current investment levels.

Goose, you up?

History suggests that while circular financing bubbles destroy investor wealth, they often create valuable infrastructure that benefits society long-term.

British railways, internet backbone, fibre optic networks and potentially AI data centres all represent genuine technological advances funded by speculative excess.

The cruel irony is that the most visionary investors, those who correctly identify transformative technologies often suffer the greatest losses during the speculative phase.

The ultimate beneficiaries are companies that can “free ride” on the infrastructure built during the bubble, acquiring capacity at pennies on the dollar after the collapse.

We can all agree, it’s a duck and the goose is coming.

1. Don’t be left holding the bag short term.

2. We’ll all benefit from this long term.

For my Gen Z readers, translation.

YOLO. But don’t HODL.

Until next time,

Blake

I just can’t use Google anymore.

There. I’ve said it. You probably have too.

I am not alone.

ChatGPT is now a daily ritual for us. Now, the latest and greatest stats are out.

• 700 million people per week.

• 2.5 Billion messages a day.

• 29,000 messages a second.

• Countless “are you stupid’s” prompted (just me?)

The most common thing people ask for isn’t for it to do their job.

you can do it!

It’s that.

A nudge. An encouragement. A killer of hesitation.

For so long, the headline “AI will take my job” forced me to look into throwing in the towel and starting out again as a plumber or builder. Now, I think I am safe, for a while longer at least.

OpenAI just released hot new research, on September 15th 2025 on how people are using ChatGPT everyday.

You can see the paper here. Worth a complete read. I’ve compiled some of the most interesting finds below.

A year ago, just over 50% of consumer ChatGPT messages were non-work related.

It’s now 73% non work.

It’s not just new uses showing up with casual prompts to boost the numbers.

Every. Single. Cohort. Increased.

The common use cases of creating meal plans, workouts, travel plans, tutoring, life admin dominated.

They were the gateway drug.

What people are actually doing

Three buckets dominate nearly everything.

1. Practical guidance

2. Seeking information

3. Writing

Together, they make up ~80% of usage.

• Practical guidance sits at 29%.

• Seeking information grew from mid-teens to ~25%.

• Writing hovers at ~26%

What stood out to me?

Vibe coding matters not.

Such a small slice of the pie. 4.2% of messages.

So yes, it can code. Yawn. Boring. Old news.

The importance of an editor

Two thirds of writing requests are editing, critiquing, translating or summarising.

Not “create an article” but “edit my thoughts”.

An assistant, rather than an AI digital twin (thank god).

The ‘ghiblification’ effect

The modern day see what you look like in 50 years. Image generation spiked the multimedia usage. It rapidly cooled to 7% of messages, but for a moment it was hot hot hot.

What’s the Intent?

Let’s break down intent by personal vs work.

Personal use

Every one of the 29,000 daily messages basically fit into one of three intentions:

• 49% were asking (help me understand, so I can decide)

• 40% were doing ( draft, structure, generate)

• 11% were expressing (let me vent, therapist)

In fact, i’ve written more about venting as a use case, in my article from a couple of months ago, AI replaced my friends.

Asking is climbing fast.

Clarity seems to be more important than automation. The data shows user’s are actually happier with the output from asking interactions than doing.

Work use

Doing actually edges out asking.

It makes sense. Work is about output. Almost always.

Zoom in for a second though.

Writing. Almost 40% of all work related messages are writing tasks.

Clear decisions love clean words.

Demographic usage

Usage by gender

Another interesting find.

Early adopters of ChatGPT skewed male. By mid-2025, female usage slightly surpassed male.

No longer a early adopter product, it’s officially crossed the chasm.

I don’t love the research on this one.

As someone named Blake, I think there’s room for improvement here in defining gender.

Usage by age

This one makes perfect sense.

46% of messages come from 18–25 year olds. Older users talk mostly about work, until they get to over 66 years old.

As the graph above shows though, there is a major drift towards non-work.

The question we need to ask

So, if people mostly ask, rather than automate, and they mostly fix words instead of writing from scratch…

What is AI doing to us, really?

I have an idea.

It’s compressing the distance between question to decision and decision to delivery

Seems obvious when you look at the gateway drugs. Tutoring, recipies, workouts. It’s why non-work related use cases lead.

Then, you start using at work.

Satisfying the itch. Of course, you can stop at any time. Or, so you say.

The AI that helps you decide what to do with your weekend is the same AI helping you decide on product roadmap.

That habit, is portable.

Will it continue trending this way?

I’ll make three bets here:

1. Asking will cross 60% across the board.

2. Writing stays around 45%.

3. Tutoring holds at 10%.

There’s no major reason for writing and tutoring to increase. People learn in different ways and through various mediums. Writing is only reducing in importance.

Video is hot. Writing is not.

There are infinite questions and the world is getting more complicated. A friend who’s often right, will be a quiet shoulder to lean on. This friend won’t spill your secrets.

Well, not to anyone but the Uncle Sam(s).

Until next time.

Blake

I’m calling it.

Google will win the Agentic browser wars by default.

Distribution is king. Google is the meta.

Everyone can see, the browser is the operating system for agents.

Comet. Dia. Now Gemini.

Google owns ~69% of the browser market with Chrome. It’s free. It works.

They own Android OS.

They also own search.

Let’s call that the ‘work layer’.

Yes, every company is commoditising the big white box, made famous by Google. You know, this one:

With various adaptions like:

Lovable

V0

Perplexity

You get the gist.

But I truly believe that UI sameness is a red herring.

Owning where agents run is the moat.

I’ll explain why.

For that we’ll need to go back a bit. Three dates matter.

December 11th, 2024. Google unveils Gemini 2.0 (thank god the name Bard’s gone) and Project Mariner as a research prototype.

May 20th, 2025. AI Mode is search starts actually doing real things. Buying tickets through Ticketmaster, making restaurant reservations on Resy and booking appointments on platforms like Vagaro.

September 18th, 2025. Literally a couple of days ago. Google switches on Gemini inside Chrome browser in the US and previews agentic browsing from the omnibox.

Huge.

Huge. Huge. Huge.

Why huge?

Well, the Project Mariner (research preview name given to Chrome agentic browser functionality) is basically a system of work.

The ‘work layer’, as described by Menlo Ventures, in their article.

It’s orchestration. It’s tool use. It’s context.

It’s, let’s say it all at once. The reason distribution is so powerful.

If agents run where people already type and click, you don’t need a new interface. A new browser.

Comet. Meh. I’ll have to learn to use it. More importantly, I’ll have to learn to trust it.

Dia. What even is that.

Safari? Apple? Well, maybe. If something even gets released.

Also, right now Google is still the default search (for now).

Everything will now start and finish inside Chrome and search.

No open-web hand off.

What will happen?

The google ‘omnibox’ will become a task bar. It will do the following.

1. AI mode will fan out the query to partner sites.

2. It will fill the forms it needs to, with human in the loop confirmation.

3. Close the loop in Chrome.

4. Find a way to monetise the loop.

5. Bundle.

6. Tier.

7. Profit.

End to end Google.

It’s the classic default play in distribution. Start where the user starts.

DMA choice screens will reduce this power in Europe. Distribution power will vary by region.

Perhaps this is where Comet will thrive. Who knows.

Doesn’t really matter. The global market minus Europe is still bigger and it’s not like they can’t play in Europe.

It’s also a data loop. Every logged workflow is a reusable asset. Where it succeeds. Great.

+1 Trust score.

Where it doesn’t and fails because of a faulty site?

-1 Trust score.

This sort of data stops an Agent from hitting CAPTCHA. Stops it from going to places it shouldn’t.

It forms trusted, predictable flows.

Lastly, it provides some serious ARPU.

They’ve already stacked a price ladder under it.

Broad Gemini-in-Chrome is free. Google AI Pro sits at $19.99/month and is already showing up in some carrier bundles.

Get one year free with a Pixel phone (this is not an ad, I promise.)

If you want Mariner and early Agentic features, you’ll need Ultra at $249.99/month. Thrown in free, storage and YouTube premium.

Sell the door for free. Monetise the power users who teach once. Carriers and OEMs get attach.

Google gets that sweet, sweet ARPU.

But. Remember. It may not work.

1. The ambition is clear from Google but it’s not yet global. We don’t know if they’ll execute well. I’m pretty sure they will, but it’s not checkmate yet.

2. Agents do the dirty work, but right now, not that well. Finish rate for agentic tasks will be the hinge metric for their status as Agent OS. We know why Agent’s hallucinate, but it hasn’t been fixed.

3. Pricing. $249.99 is a lot of money. Not everyone’s going to be an early adopter. If someone else provides an equal, price advantageous and more habit forming experience (maybe OpenAI), maybe with their own $20 bundle, that’s a hard battle to fight back.

Another banger question…

Who gets first contact (distribution rights) if the agent runs at the OS layer instead of the browser?

1. Apple can move the work layer above Chrome with Siri level app actions and Apple wallet.

2. Microsoft can own the office corridor via Windows and M365 actions.

It’s far from over.

I believe, if Google can publish the credible finish rates and ship agentic browsing beyond the US and steer users into AI mode to finish their tasks, the browser-as-an-agent-OS thesis plays out.

That means, real successful outcomes on actions like percentage of cart completions, tickets booked, reservations made.

It’s game on.

But really, I think Google’s already won by default.

Who said the free market was fair.

Until next time.

Blake

Good pricing is easy to understand. Seats. Users. Visits. They all make sense.

More seats = more revenue, cost out, less risk.

Nowadays, with this small, trending, niche concept called AI (have you heard of it?), we’re seeing a different trend.

No longer paying for logins, instead paying for results.

Action words dominate. Resolve the ticket, stop the breach, approve the order without a chargeback, close the sale. All measured outcomes.

How good would it be to say to all your vendors:

Mr software vendor, if you move the number you promised you would, you get the money. If you don’t, you don’t.

Perhaps soon you will.

I think that is where enterprise software is heading because AI makes impact visible in a way finance trusts. The winners will price like they can be fired and customers keep them anyway, since the outcomes are too good to lose. The losers will keep counting seats and hoping procurement forgot how to negotiate.

Spoiler, they’re getting better than ever.

Really interesting, the latest Ramp data, in AI Adoption below. Outcome based pricing is the go to pricing strategy being employed at the bleeding edge AI companies.

Alex Karp, CEO of Palantir actually said it perfectly “I get paid because you can get rid of us. I get paid because you can get rid of us, but we deliver so much value that you can’t.”

I’ve tried to create a good method for pricing within this framework. Let me know what you think:

Price = floor (cost to serve) + expected loss x risk load + service fee

The floor

Outcome pricing replaces a wine and dine, enterprise sales‑based sales campaign with a simple scoreboard.

Money changes hands only when a unit of value appears, and that unit is something a CFO can audit.

• In support it’s is a conversation that ends with a ticket resolution.

• In fraud prevention it’s an approved order with no chargeback.

• In security it’s is an incident contained within an agreed time window, with a warranty.

• Sales and marketing can anchor to accepted meetings or qualified pipelines as long as the fields live in the CRM. It’s a little more murky (but more on that later).

You get the gist.

Whilst the digital advertising ecosystem has lived in this world for a long time. Advertisers pay for clicks, leads and conversions. Merchants pay a take rate per transaction. SaaS never fully crossed this bridge.

Seats and feature bundles survived because attribution was fuzzy, switching was painful, and vendors could hide behind this magic word, “adoption.”

These wondrous marvels known as LLMs and agents absolutely wreck all three excuses.

They generate granular, time‑stamped records of who did what, a multi‑model world makes swap‑outs plausible and falling unit costs create room to price against value without eroding margin (much).

It’s all about inference costs, baby.

They are going down, down, down!

Or are they? See my article on token economics 101.

But for now, let’s assume they are.

It’s already happening in different markets

Customer Service

Customer service went first because the unit of measurement is clear and the pain is big.

Intercom’s AI agent ‘Fin’ charges per resolution, not per bot.

Zendesk ties billing to automated resolutions rather than a vague concept of “AI users.”

Salesforce’s Agentforce (somewhat innovative from a giant) lists prices per conversation and per action.

Security

In security, a warranty is the most honest form of pricing.

CrowdStrike pairs managed defence with breach warranties that cover real costs when incidents escape containment.

Rubrik wraps data recovery in ransomware warranties at the top tier that reach into eight figures.

Agent platforms

Agent platforms and contact centres are normalising event‑based pricing.

• Twilio Flex charges per active user hour, which is a tight proxy for actual work.

• Freshworks sells packs of AI agent sessions once the free quota is consumed.

Why now

Palantir sits to the side of this discussion as a signal.

Its chief executive has been loud about “getting paid because you can be fired,” and the company has posted serious growth while running that story. We do not need to make Palantir the hero. The point is that investors reward vendors who sell certainty and show the tape.

Three forces have converged.

1. Measurement became cheap.

2. Every agent action, prompt and function call can be logged with context.

3. Switching got easier. Multi‑model platforms, shared connectors, and agent frameworks reduce the pain of change.

Realistically, it’s also the fact that unit costs fell across the board and inference is cheaper per unit of work. Competition among model providers keeps pushing rates down.

As a software vendor, it’s getting cheaper and cheaper to serve your customers, and it becomes possible to price in unsuccessful outcomes.

An example outcome contract

There is no single template for an outcome contract because categories and buyers differ.

The most common shape is a hybrid pricing strategy. A modest platform fee that covers fixed support and governance, plus a variable component linked to outcomes.

Where uplift is a promise, you see Dollar-Based Net Expansion Rate (DBNER) tied to incremental revenue over an agreed baseline, refreshed on a cadence to avoid a permanent pilot.

There is a deeper point here.

Outcome contracts are an operational bet. The vendor has ownership and accountability over outcome. Not sexy slide ware. Actual outcome. That requires clean observability, clear definitions, and the humility to admit when the outcome wasn’t in fact achieved.

This will change the very nature of enterprise software.

I personally, still want sexy slide ware.

Where outcome pricing fits and struggles

This pricing model works best where the outcome result is binary.

1. Support has resolved or not resolved.

2. Fraud has been approved without chargeback.

3. Payments either post or fail.

4. Security incidents are contained inside a window, or they are not.

These categories move because the audit is simple and both sides recognise the unit as value.

Emerging areas are an interesting outlier, requiring a little more nuance. It’s those dreaded words.

Attribution. Data Quality.

Shudder.

1. Sales assist can charge against accepted meetings, pipeline created, or rep cycle time, but only when CRM hygiene is solid.

2. Marketing can tie spend to down‑funnel outcomes, yet multi‑touch attribution will always invite debate.

Most buyers will keep those as subscriptions and then create sub-outcome meters for tasks inside, for example, meetings auto‑scheduled and attended or messages drafted and accepted.

Honestly, I think that compromise is fine for now. Not every domain must cross the pricing bridge on the same day.

The change will come. Just slower.

The uncomfortable middle

The easy to measure win

Seems obvious, but the categories with binary outcomes win first, as they should. Platforms that enable model choice and baked‑in evaluation features help buyers swap engines without drama, which deepens trust and nudges contracts toward results.

Per seat vendors lose

On the other side, the seat‑heavy products that cannot prove outcomes, will lose. If the core evidence of value is “users logged in,” procurement will do what procurement does best. Discount baby! Consolidate! Not return your calls!

Utilisation is an interesting one

Counting tokens or API calls can be a fair bridge in early categories, yet once value is knowable, activity‑only pricing looks like a fig leaf. My dad always said, always go into the business’ where the consumer doesn’t know what the product should cost. Case and point, never open a cafe.

What this means for incumbents

Incumbents hold assets challengers envy. Namely distribution, data and brand trust. Outcome pricing plays to all three.

• Distribution lands pilots everywhere, and pilots convert when the meter is easy to read and the KPI unit matches how the customer already runs the business.

• Data gives a head start on variance, which makes warranties cheaper to offer and more painful for rivals to match. I.e they can price in failure.

• Trust lowers friction in legal and procurement.

This is a win that will lead to better software being developed.

When a vendor accepts outcome risk, product roadmaps tilt toward value and reliability. Sales motions tilt toward mapping workflows rather than selling features.

Just check out all the ‘forward deployed engineer’ roles currently being hired for.

Per‑action and per‑conversation models look simple because they are. The more the product guarantees completion rather than activity, the more defensible the price and the better the software actually is in reality.

Please do remember that outcome pricing is not a magic pill.

It will not rescue a weak product or a broken process.

It will not fix culture.

It will force clarity about what a piece of software actually does, and who benefits, and how fast. That clarity is messy at first, which is why incumbents will drag their feet and challengers will overpromise.

The market will punish both.

The seat tax had a good run. It made sense when proving impact was expensive and swapping tools was costly. That world is fading. The enterprise is getting a contract that reads like common sense, because the infrastructure finally caught up with the rhetoric.

Pay when it works. Stop when it does not.

Give so much value, they can’t get rid of you.

What software was supposed to be.

Simples.

Until next time.

Blake

Four main points:

• A model will never be 100% accurate.

• Hallucinations are a solvable concept, by abstaining from answering.

• Hallucinations are the result of our LLM training reward framework (it’s not magic).

• The reward framework needs to be rebuilt to reward uncertainty.

Ok, let’s start

We’ve all heard the old adage, say it with confidence and people will believe you.

It genuinely works. I do it all the time, this article for instance.

Just kidding. Sort of.

Turns out, AI does it too.

Your chatbot of choice, OpenAI, Gemini, Perplexity, Grok, etc sounds so sure of itself when it answers your question. Most of the time, it’s a fact, logical explanation or response that has a likelihood of being just plain wrong.

The technical term is hallucination. Very different from the ones those special mushrooms you once took in Amsterdam.

AI’s hallucinate because there is an element of prediction, in every response from a chatbot. Throughout this, I’ll speak through reference to ChatGPT byOpenAI, because they’re arguably the most well known. Case in point, look at this market share below. 6 billion visits to ChatGPT vs 750m to Gemini (it’s closest competitor).

This concept applies all the above.

So to start, when ChatGPT responds to your question, there is always an element of chance and reasonable guessing. A chance to hallucinate. OpenAI and others have been trying to work out how to reduce these hallucinations. In this paper, aptly titled ‘Why Language Models Hallucinate’, they’ve apparently figured it out.

The paper is incredibly technical. Just look at this sexy algorithm.

So, here’s what you need to know.

Language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty.

The core idea

Basically, ChatGPT hallucinates because when the models are trained, they are given direct instructions to provide a finite answer.

A binary yes or no.

Rather than being rewarded for saying I am not sure, or, I don’t have enough information to say unequivocally yes, or no, they are rewarded only if they say yes.

So when using the model, to actually answer your question as a user, the AI can really only have two different tasks to respond to you.

1. Marking an answer “does this look right?” Yes or No.

2. Writing an answer “produce this answer yourself”.

When testing the models’ answers, this paper found that if a model makes 5% mistakes at marking, it will make 10%, or 2x the amount of mistakes, in writing.

That is big enough to matter. Flashback to the Chevrolet Dealership which gave away a truck for $1 USD.

Imagine every 1/5 interactions was a practically free car.

A fun game for kids

To visualise this, Open AI created a simple quiz called “Is it valid”. The premise was to take a bunch of correct answers, mix in some wrong ones.

The LLM duck-duck goose.

The model was tasked to label each answer as valid or not. Then given a marker score. If the model was tasked with writing the answer from scratch, the error rate rose.

The takeaway. Writing is the hard part for LLM’s.

Why writing is hard

For most questions, there’s typically one (or few) correct answers. There are of course, countless wrong ones. An LLM basically spreads its bets about what is correct, across a huge space. In some of these spaces, it’s absolute fluent nonsense.

When ChatGPT says it’s 80% sure, it will be right about 80% of the time.

So what do most providers do to improve their model?

Calibrate.

Large base models often start out reasonably calibrated. After more fine-tuning to sound more helpful and human, weirdly it actually goes down.

A smoother tone, but with an accuracy decline.

But the real question is, what is accuracy. In these tests, we grade like we’re in school. One point for correct, zero for “I don’t know”. So what happens, the LLM wants the points.

Answer, even if you’re not sure. It’s like the AI is channeling Jordan. You miss 100% of the shots you don’t take.

Guess who

Imagine now a quiz where each question has one exact true sentence. Every other sentence is wrong, with no pattern linking the questions. Unless the model memorises the exact answers during training, it’s guessing.

For so many answers, it is a guess. An educated guess, but a guess all the same.

OpenAI actually explained this pretty well. They said, suppose a language model is asked for someone’s birthday but doesn’t know. If it guesses “September 10,” it has a 1-in-365 chance of being right. Saying “I don’t know” guarantees zero points. Over thousands of test questions, the guessing model ends up looking better on scoreboards than a careful model that admits uncertainty.

Probably why reinforcement learning is so important in training.

In terms of accuracy, the spec shows that the older OpenAI o4-mini model performs better. However, its error rate (rate of hallucination) is significantly higher. Strategically guessing when uncertain improves accuracy but increases errors and hallucinations.

Realistically though, some questions can genuinely not be answered at all.

A lock without a key

Ok, perfect example.

“Decrypt this message without the key.”

With strong encryption, there is no efficient method that does this reliably. A refusal or a failure there is not incompetence, but it is honesty about the rules of the problem.

It is a reminder that not every “hallucination” is fixable with more data or nicer wording.

Sometimes the only sensible output is “cannot be done with what you gave me”.

This, is really hard for a type A LLM to admit.

Why existing test methods fail

There is actually a relatively straightforward fix that has been proposed over the last couple of years.

1. Penalise confident errors more than you penalise uncertainty.

2. Give partial credit for appropriate expressions of uncertainty.

OpenAI thinks it won’t work. In their own words, if the main scoreboards keep rewarding lucky guesses, models will keep learning to guess.

Unfortunately, the paper talks more about understanding the problem, rather than creating solutions.

Perhaps when they find the answer, they’ll make a paper for us dummies minus all the mathematical theorems.

—

If you’re interested in the actual full paper, OpenAI published it here. https://openai.com/index/why-language-models-hallucinate/

—

I, Blake, do swear, to honour the forces of the free market.

Therefore, I must honour the court’s call.

I agree with Judge Amit Mehta. Google kept its insane dominance in search by buying distribution and writing contracts that kept any competing rival out. Now, they have a chance to compete.

It’s game time Perplexity, Apple, DuckDuck Go, Bing (lol). Copilot, Gemini, Siri, Comet, now Dia (Atlassian’s recent buy). But seriously, the number of entrants, competitors and now potential winners is crazy.

You’ve got new data pipes. You’ve got new rails. Time to show why you belong here.

That’s it, my opinion. Article over.

Well, not really.

The devils in the detail, and whilst fundamentally that is my opinion, strap in, we’re leaving no stone unturned.

What the court actually did

Ok basically, there’s three rulings that matter. With my own creative spins on the names.

Ruling 1: Me, Myself and I

Any binding contracts that blocked OEMs, carriers, and browsers from shipping or promoting a rival beside Google are no longer legally enforceable.

So, when you search on Safari and a google search comes up, it could be a different search engine. Right now, it stays Google, but now, it doesn’t have to.

Cue the dramatic voice over *On the next episode of who owns safari search*.

Consider me subscribed.

Ruling 2: Share your toys, Google

Google must share its index.

That is, the giant catalogue of the public web that powers results. The ruling requires Google to share controlled access to that catalogue with rivals, not hand over everything at once. In practice that means they can rate limit, sample, and add latency so rivals get useful signals without real-time, raw logs. It won’t magically equalise quality on day one.

Like anything, I want what I want, when I want, fast. Coverage, freshness, and ranking craft still going to decide winners.

Ruling 3: “Hey Google, err, I mean Gemini”

This one’s pretty interesting. Very AI, very 2025 and props to the judge, a particularly smart ruling.

The days of the ten blue links are over. AI assistants are in. Agents are in.

In the ruling. These count as search. The same rules. Nice try google, you can’t shift the old game inside of Gemini. I imagine it went something like this:

Judge: Show me what’s in your left hand, Google

Google: What, this one? Nothing?

Judge: No, the other hand

Google: Oh, you mean my other, left hand.

So, exclusive distribution inside assistants is banned.

That means Gemini, tomorrow’s Siri, Co-pilot (lol, again), Comet or any other, fast growing VC fuelled “ask me” gateway drug. Browsers that behave like assistants for work sit in the blast radius of this logic too. If the browser is the answer surface, distribution rules will follow it. Dia might be poised to take advantage (more on this later).

What it didn’t do

Big, but less interesting on the surface.

No divestiture of Chrome or Android.

It leaves Chrome and Android intact. Probably a good thing given Apple’s own monopoly. Who can stand up to a bully? Another bully of course!

No national choice‑screen standard in the United States.

In translation, a choice screen is a setup page that asks you to pick your search engine. You tap one. That becomes your default. They don’t need to show one. Simple. Default usually wins as choice requires effort and there’s a reason people stay with their bank and electricity provider.

Apathy. Laziness. Yep, same.

Default payments can continue when they are non‑exclusive.

We spoke about this. Google. Safari. You got it. No need to replay.

It’s a good move

Fundamentally, any heavy remedy imposed would risk reversal.

More court battles. More Lamborghinis for the lawyers.

Better to stop the clearest violations of conduct, open a fresh, hot data pipe and avoid ripping out the core services people use all the time.

Imagine they got rid of maps.

*shudder*

What changes in the next year

Near term, I think everything will look familiar.

Most will keep typing into the same box. No one is going to click “I am feeling lucky” (what even what that, anyway).

The early changes are realistically going to sit in setup flows and prompts.

When you set up a new Android phone in Europe, you’ll see a simple page that says “Pick your search engine.” It will use normal words, not a T&C that you will 1000% skip and say you read. You will tap one. That becomes your default from day one.

Some phones will ship with two search boxes on the home screen. One is Google. The other could be Bing, DuckDuckGo, or an AI answer box. Both are live. You pick which one to tap each time. I think this is a little meh from a UX perspective. Let’s see how it plays out.

When you ask the assistant a quick question, you’ll get the answer in the chat window. No browser tab opens. Great for weather, unit conversions, simple facts, flight status, calendar checks. Weirdly, I use Siri for this. Yes, Siri can actually do something useful, go figure.

If Apple rolls out answers inside Siri or Safari, you’ll get results without leaving Apple’s screen (more on this below!)

The moves that matter

Ok so, what’s happened has happened.

In parallel, there’s some major players making moves. This will only ignite their thirst for power. Let’s have a quick squiz.

The Apple answer layer

Apple has a team focused on an in‑house answer engine. Codename World Knowledge Answers.

Think web‑aware responses in Siri and Safari that resolve common tasks without a trip to an external engine. Kinda like ChatGPT.

If quality lands (which I used to say, it does with Apple), the first hop moves from a Safari address bar to an Apple‑controlled prompt. That would raise Apple’s leverage even if the declared default remains the same.

This just got announced, so how it manifests itself is anyone’s guess right now.

What I do know is that it’s built around three core components. A planner, a search system for the web and the device and a summariser.

The planner interprets voice or text input and decides how to respond.

The search system scans the web or device.

The summariser pulls it together into an answer.

Weirdly, they’re in talks with both Google and Anthropic to be point 3, the summariser.

OpenAI is not in consideration. Odd, given their Apple Intelligence partnership.

Perplexity Comet Plus

Perplexity launched Comet, a desktop browser that returns an answer with linked citations from the open web, forums, and news. It followed with Comet Plus, a paid plan that shares revenue with participating publishers.

I’ve previously written about Perplexity Comet plus here. Give it a read if you’re interested.

If answer engines want durable supply of, well, answers, they must fund it. The margin math is still in play. The direction is right. But, in short, it could be Google 2.0.

OpenAI hardens search inside chat

ChatGPT chat turned into search.

That shift changes user habits as much as any setting inside a browser.

I use ChatGPT habitually now. More than Google. The fight becomes quality, trust, and speed inside the chat pane.

Game on.

Whoever wins, so does the consumer experience.

Microsoft Copilot to Captain

Copilot is turning up in more places.

Sorry, Copilot is good. It’s just not great.

As placement grows in TV’s, OS, Devices, it gets a shot at winning repeat tasks that live inside a workday. It may win. But really, yawn. It’s the company that created teams.

I like Microsoft, but please step up in UX.

Welcome, Atlassian to the race

We love a last minute addition.

Atlassian just bought The Browser Company for about US$610 million and wants Dia to be a browser built for work.

What a power move. For many, the home page of the workday.

If your tickets, docs and approvals live inside Jira and Confluence, the browser that understands that graph controls the first click, the query, and the handoff. That is also giving Google 2.0 energy, without the ads.

Distribution matters. Atlassian has it in certain domains.

Underdog? No.

Someone I would bet on. Eh, also no.

I’m yet to find anyone that loves using Jira. They also just bought cycle.app, a design led tool for product discovery, so many they’re leaning back into UX and design.

Watch this space.

Some changes are still being considered

The open‑web ad stack is a separate fight.

A court has already found Google a monopolist in key ad tech functions. The following remedies are not final yet.

The Department of Justice is pushing proposals like splitting the ad exchange from Google’s buy/sell tools, creating data walls, and pushing audits. Those are proposals on the table heading into the remedies phase, but are not court-ordered changes today.

For brevity’s sake, I won’t go into this battle. What happens here also matters, but it’s more about the revenue of google for the world of today, not tomorrow.

Like the judges ruling, this piece is more about tomorrow’s new reality. Let’s stay focused.

The next 12 to 24 months

Okay, let’s discuss what changes in the near future.

Apple builds and guides, subtly

Apple will most likely roll out a web‑aware answer pane inside Siri. Most likely, only where it can guarantee quality. My thinking, anything to do with device help, sport’s scores, weather, quick facts, calendar and travel status of apps installed on your phone.

Google remains the declared default. Real routing for specific intents tilts to Apple UX once satisfaction is high.

Getting access to a data API from Google wont really help here. But, Apple’s leverage will grow, even if Google’s headline share hardly moves.

Subtle. Calculated. Kinda boring really.

In this world of ‘move fast and break things’, potentially not a good option.

Let’s see.

Microsoft and the big labs convert access into visible quality

Bing answer quality improves on a basket of popular intents, helped by those sweet, sweet index snapshots and query signals. Plus Copilot placement on more screens. Better than Clippy showing up, I guess.

An AI answer engine proves the publisher purse

Perplexity’s Comet Plus tests whether a high share of subscription revenue can stabilise supply of content. If terms spread to other browser competitors and others match, publisher hostility cools. If not, expect a lot of licensing fights, especially in Europe. Just ask Anthropic what this is like.

Google rewrites contracts and deepens assistant placement

Contract rewrites meet non‑exclusive requirements, with more one‑year terms. Assistant features inside Android and Chrome keep usage inside Google surfaces while staying compliant.

Basically, they play within the rules.

Where I land

Props to Judge Amit Mehta.

If you open distribution, share some data, and bind the rules to the next interface, then product and preference can do the rest.

I still think the next default sits inside an assistant, not a browser.

Apple is building one. Microsoft is spreading one across screens. OpenAI turned chat into search. Perplexity wants to own the window and pay the suppliers. Dia will create a chat living inside a work browser that already knows your permissions and projects. Each have their own flare of an unfair head start.

Google lost on the law and kept most of the levers. The door is finally open. Whether anyone walks through now depends on partners and a public willing to try something that is not the blue box they grew up with.

I think the world’s going to change.

I still can’t pick a winner, but I live for the drama.

Until next time.

Blake

As of yesterday, long live Google! They avoided the break-up, but must share data with rivals. I'm looking forward to writing about this over the weekend, stay tuned.

It doesn't change any of my thoughts below.

Agentic Search is the future

Matthew Prince, CEO Cloudflare in his recent interview spoke about the death of publishers. The race to clicks - the listicles, clickbait headlines, controversial takes (they're done for thank god).

AI answers sit above the fold (the very tippy top of Google search results), the cost of scraping subsidised by the people who make the work, and traffic is now a vanity metric.

To me, it's both great news, and bad news. We're starting to see new companies step in, to try take advantage of the opportunity.

What happens when AI has all scraped all the web, and no one is incentivised to create new content?

Introducing Comet. The new browser from Perplexity.

But wait there's more.

Aravind Srinivas, Perplexity's Founder, posted that Perplexity is launching Comet Plus to “reward publishers for quality content.”

That line stuck.

He made is sound like salvation. But. It also reads like the moment a new operator lays track across the open web.

Choo-Choo folks.

Here’s my take.

Comet is way more than just a nicer search box. It is a bid to own the on-ramp, the payout rules, and the crawl permissions in one stack.

It's giving Google 2.0 energy

I’m excited by the promise. I’m wary of the power. If this model wins, we do not simply fix referral traffic. We will be rewriting how knowledge is produced, priced, and policed.

A quick history lesson

You probably are aware, but incase you aren't, here's a small summary of what's happened so far:

• AI answer surfaces are replacing clicks. The old trade was simple. Let the 'itsy-bitsy' crawler in (OpenAI, Anthropic, Google, Perplexity, etc), get a proportion of the attention back. Give an inch, take a mile. That contract has been torn up quietly, one summary at a time.

• Cloud infrastructure grew teeth. Network providers now block or meter AI crawlers by default, and they’ve started talking openly about pay-per-crawl. Consent no longer lives in a polite text file. It lives at the edge. Cloudflare has been quietly taking advantage of this opportunity, blocking bots at the edge.

• Perplexity launched a browser and layered a $5 monthly subscription on top that pays publishers from a shared pool using signals like citations, human visits, and agent actions. In other words, they aren’t just answering questions. They are starting to write the rules for who gets paid when the answer uses your work.

Basically, the interface has moved into the browser. Money begins to flow in a new direction.

Suddenly you can see the outline of a rail and the smell of burning coal.

Why the publisher model needed to change

The unit of value changed without the payment model following along.

1. A click-based world paid for promise.

2. An AI-answer world pays, if it pays at all, for delivery.

The content that ends up inside the model’s response is doing the heavy lifting. The rest might as well be wallpaper. Clicks are irrelevant.

If you make anything expensive, the way you get paid has to be specific, not vague. Publishers carry real costs. Reporting time. Legal review. Updating fast when facts change. That work shows up as freshness, accuracy, and context. The new model has to tie money to those properties, not to the illusion of “reach.”

Which is why Comet’s framing makes sense.

Pay for “filling knowledge gaps” instead of raw impressions. Reward sources when an answer actually uses their work and a human sees it. If done right, that finally values the thing that costs publishers the most: being right, and being first without being wrong.

Here's Comet’s play

Ok. Let's strip back the hype and buzz. Basically, there are three moving parts.

1. The on-ramp 'Comet': A browser that renders answers natively (Think of Chrome or Safari). If you own the front door, you decide what people see first. That is power number one.

2. The meter: Signals that claim to measure how much a publisher contributed to what the user consumed. Citations inside the answer. Human visits after the answer. Agent actions that trace back to a source. This is the accounting layer, where Perplexity will own the metrics that matter).

3. The pot. A subscription pool with a stated split for publishers. Pay from the pot according to the meter. This is the incentive layer, where Perplexity will decide who get's paid what.

Sounds very Google 2.0 right?

Put those together and you have a rail.

It starts at intent, runs through a proprietary surface, and ends in a private ledger that decides who deserves money this month.

Don't get me wrong, I like the direction.

Paying for usage is saner than paying for empty page views. But usage only becomes fair if the meter is auditable and the pot isn’t what Perplexity decides it should be.

Creator funds, payout schedules, being slaves to the algorithm. This has played out so many times.

Why this could be brilliant

Creators and publishers have been trapped between two bad choices.

1. Keep the paywall up and disappear from the AI-shaped internet.

2. Keep it down and watch your work summarised to death, with little return for you.

A licensed rail with money attached offers a third path.

Let the model read, get paid when the answer uses your content, and keep the right to say no.

Scarcity through choice.

A market for gaps also encourages depth.

Shallow rewrites don’t fill anything. How many times have you seen the 'best X for Y' products repeated over and over, which are all obviously paid.

Original reporting. Niche expertise. Local context. The open web desperately needs those things. If Comet pays properly, we get more of them.

And, we get no sponsored native advertising (I can't wait).

There is a second upside no one is talking about.

The model creates room for corrections and updates to be worth something.

If your source is used in an answer that later proves wrong, the marketplace can claw back. If you do the follow-up that fixes the claim, you get the tail.

Value. Scarcity. Truth. Plus, some potential innovation. See below as on September 4th. Agentic commerce. Cool!

Ultimately, when there's competition. The consumer wins.

Why this could be dangerous

One company setting the default surface and the payout rules is not a friendly neighbourhood start-up. No 'Good morning, Blake' here.

It is a rail operator. Rails collect tolls. Rails prefer their own freight. Rails do not become kinder with time.

Google literally dropped 'Don't be evil' from their mission.

If I could highlight that fact any bigger, I would.

If Comet wins the browser slot and the answer slot, the negotiating table shrinks to a single chair.

Consumers, don't win. Publishers don't win. Perplexity, does.

The risks are easy to name.

• It can be self-dealing: When the platform both hosts the answer and decides which sources deserve money, it can quietly favour house content or friendly partners. Publishers learn to write for the meter rather than for readers. Happened today, will happen tomorrow.

• Incentive clarity: Payout formulas drift over time. Small changes in weighting turn into big changes in income.

• Lock-in: If Comet becomes the easiest licensed route to audiences, publishers face a hard choice. Accept the terms and depend on one rail, or retreat behind private APIs and pray your audience follows. Spoiler alert, most times, they don't.

That’s the Google 2.0 worry.

We replace the search monopoly of the last decade with a rail monopoly of the next.

Publishers, assemble!

For the first time in years, publishers have a lever that actually bites.

Scarcity. Access. Trains (choo-choo).

The website's themselves now block unlicensed crawlers at scale, at the edge. Cloudflare wins. Publishers win.

They can meter, price, and revoke.

Consent is not a polite suggestion on a robots.txt page. That gives creators bargaining power they’ve never had with search.

Matthew Prince, Cloudflare CEO literally said he could block 20% of internet traffic overnight if he wanted to. That's real power.

If Comet wants premium sources, it needs licensed access. If licensed access costs real money, the payout pot must be large enough to attract them. If the pot is large, the platform needs more paying users and stronger default distribution. You can see the loop forming. Rails on one side. Pipes on the other. Publishers in the middle with something to sell again.

I think, that standoff is healthy. It stops any one party pretending the others don’t exist.

Only thing is... The more rails that are laid, the less this matters for the trains.

I can see only one option.

Diligence and ruthlessness in negotiation. Publishers, never give an inch.

What I see a fair rail promising

If Perplexity wants the trust of people who make things, it needs to behave like a utility before it is treated like one.

• Transparent formulas: Publish the payout algorithm, the weights, and the appeals process. Let independent auditors verify the statements.

• Receipt-level accounting: Every payment should carry a record of which answer used which chunk of your work and when. Real Receipts.

• No house edge creep: If Perplexity publishes or commissions content, label it, treat it like third-party supply, and disclose any advantage it receives. These shouldn't be some little 'sponsored' label out of sight. Really big. Always.

• Portability. Creators should be able to export their licences, usage logs, and history to negotiate with rival rails. Competition is healthy. Let's encourage it.

Do I expect a start-up to love those ideas?

Not really.

Do I think the market will demand them if this rail becomes key infrastructure?

Absolutely.

So what to do now, if you're a publisher

This is not the moment to sit back and wait for a nicer check. It is the moment to make your work legible to machines, enforceable at the edge, and valuable in the answer layer.

Let's get technical:

1. Turn on the gate: If your CDN offers AI-crawler controls, use them. Negotiate from consent, not from fear of invisibility.

2. Sign your pages: Attach cryptographic fingerprints or watermarks that survive basic copying. Provenance will become a ranking and payment factor.

3. Write for receipts, not clicks: Make your claims specific, citable, and updatable. Build “why this changed” posts that models can link to when facts move

Call it Rail SEO if you like. I think it's catchier than Generative SEO anyway.

It looks nothing like the old thing.

It is closer to running an wikipedia like service than crafting clickbait headlines for robots.

My stance

• I think Comet’s direction is necessary.

• The open web cannot survive on empty promises of “exposure” while models eat the value upstream.

• Paying for usage, not clicks, is the right reset.

• The browser move is clever.

• The publisher pool is a start.

• The positioning around knowledge gaps is honest about where models fail and where humans shine.

We have lived through one cycle of monopolistic distribution destroying the open web. No one is keen to repeat it with nicer branding and a slightly friendlier share. Well, except maybe Google.

So I’m treating Comet like a promising idea in a trial period.

I’m hopeful because the incentives line up for the first time in a decade. I’m cautious because incentives drift over time. Creators and publishers will do their part. Make fresh, authoritarian content. Update it when it changes.

Let's hope Comet Plus keeps up their end of the bargain.

And if they get their hands on Google Chrome, we'll see what happens when they get power levels over 9000!

Update and spoiler alert. Not this time, Perplexity.

Managed to get a dragon ball z reference in there. How good.

Until next time.

Blake

TLDR

1. Making tokens cheaper won’t scale if smarter models just tempt us into longer, “think harder” sessions. As always, the 80/20 rule applies, with electricity prices a serious chokepoint.

2. The honest math of reasoning scalability is price × tokens × sessions (nudged by smarter reasoning processing on the back-end). That curve only bends so far.

3. The application layer has already suffered from this. The winner takes all, VC backed growth at all cost strategy is wearing thin. As we approach bubble territory, if investors shift from revenue to profitability, and if no true feasible pricing strategy emerges, we may never get AI.

4. Until there’s a true 10× efficiency jump or most queries run on local machines, expect unstable pricing models.

Ten times cheaper, fifty times more tokens

I’m sick of people using Moore’s law to argue for models becoming ten times cheaper next year.

Maybe they will.

It still doesn’t pencil out.

Cloud-only, deep-reasoning-by-default consumer assistants struggle to clear margin at scale. Token economics can’t support the way we actually want to use it.

In my rant, I’m going deep. Come along if you dare. Forewarning - there’s a little math in this one.

I promise it’s worth holding on.

Let’s start with token cost

I had a conversation the other day with a normie.

What I mean by that, is someone who doesn’t spend an unhealthy amount of time with AI. I had to explain what a token was. Might be useful here too, if you identify as one (love that you’re here, by the way).

Basically, a token is a tiny slice of text.

Every time you fire off a prompt like “What’s wrong with me” or “Is it normal if I…” (I see you, trauma dumper), two meters spin:

1. What the model reads (input), and

2. What it writes (output).

Providers like OpenAI or Anthropic usually bill both. Output is often pricier than input because long, careful writing requires more computation and inference. If you’re on a flat price plan, you won’t see the meter, but that doesn’t mean it doesn't exist. Someone has to cover the bill.

Oh, important nuance here I want to make sure you understand… tokens aren’t the cost.

The cost is the compute required to process tokens (math on GPUs, memory moves, retrieval hops, safety passes, logging, storage, all the super techie stuff). Tokens are an effective measurement that helps derive a common baseline for measurement.

The shape of a useful session

The real problem isn’t price-per-million tokens.

Non-reasoning models answer in a straight line: you ask, they reply. Predictable, but shallow.

Question. Answer. No need to ‘think’ about how to approach an answer.

When AI first came on the scene, pricing AI as a token made sense, and arguably was relatively predictable (within a margin).

Then came reasoning.

Reasoning models think in steps. They take time to understand the question.

They try to understand what you’re wanting as an outcome and think through various steps to get to the answer. Ultimately, they plan, reflect, call a function, re-check, then write. That “care” is what makes hard problems solvable.

It also generates lots of tokens you never read and encourages longer outputs (examples, code, citations, “why”). You’ve probably seen the “now I am doing this, now I am doing that, I need to now do this, hmm that didn’t work” logic chain in your chats. Since output tokens are typically priced higher, useful often means expensive.

Providers do try to fight the curve.

They try to reuse shared prompt scaffolds, re-embed repeated chunks, and batch workloads so the incremental cost of a long answer is less than “start from zero.” But even with those tricks, longer, careful outputs are where the money goes.

Context is a price floor

In this, it's the context that raises the price floor.

We’re all guilty of pasting in a contract, a repo, three PDFs and a brief. Large context windows kinda feel like magic. The model must read what you paste. The more context, the larger the risk of hallucination or failure. Each retry drags the whole history forward, making the next try pricier.

Cha-ching, cha-ching, cha-ching.

As models get better, they do more, which means more tokens, not fewer. Ten times cheaper, fifty times more tokens used.

The math cancels.

As you can see, the frontier models still cluster in the expensive corner.

The OG glazer, the golden child, the ol’ faithful GPT-4o sits around the high single digits per million tokens on a blended basis.

But Blake, your whole argument is invalid as cheaper models exist.

Yes. and no.

Realistically, people still reach for the top model when the work matters.

Let’s make it personal… I can’t be the only one that gets a little bit annoyed if ChatGPT answers me too quickly, without thinking. It’s like it didn’t really try?

No one wants GPT-5. They want GPT-5 Thinking.

The logic of application pricing

Three levers decide the bill:

• U: unit price per token (what you pay)

• T: tokens per session (how much “thinking + writing” each answer needs)

• S: sessions per user (how often people come back)

Therefore, the cost per user = U × T × S.

Now add the provider’s hidden lever:

• E: efficiency (how much work they squeeze out per visible token via caching, batching, faster kernels)

Raising E lowers cost even when your visible tokens don’t change.

But if T x S explodes (deeper chains, more habit), cheaper U and better E often get swamped.

Here’s the dirty secret providers don’t want you to know.

• Price (U) can fall.

• Use (T x S) usually rises as quality improves.

When T x S pulls harder than U and E combined, the total bill still goes up.

So far, this has always been the case.

The math ain’t ‘mathing’.

The price wars in the application layer

We’re seeing the false economics in pricing AI native applications.

Different wrappers, same problem.

Cursor, Claude Code, Codex, Lovable.

When the product starts to think, the only lever you can show publicly is price, so you pull it. Meanwhile, the two invisible levers, tokens (t) and sessions (s) spike.

Let’s run the four standard plays for competing business’ and why they still lose in broad consumer land.

Usage-based vs usage-based

Consumers keep a watchful eye on the usage meter. People end up rationing their prompts. As experimentation declines due to cost concerns, growth stalls. The applications try to cut prices to look approachable, but the big tasks stay expensive. Margins get thinner and thinner and thinner. We saw this play out in our short math lesson above.

Flat pricing vs flat pricing

This is the common pattern that we’ve seen so far. Until it wasn't feasible. For consumers, it was party time, until heavy users dominated costs. The 1% of ‘super users’ chewed through tokens at an impressive rate, ruining the party for the rest. Margins diminished, caps were added and prices were hiked. Trust drops. New competitors come in to try to capture the opportunity.

The cycle continues.

Metered usage (capped) vs a flat subscription

Simply put, without a 10x improvement in experience compared to competitors, metered usage will lose signups. The value for the consumer in terms of usage is too great, unless joining the great subsidy war. For a time.

Flat subscription vs Metered usage (capped)

Literally the other side of the above, but with a different time scale.

Flat subscription will win adoption early and get crushed by heavy users. The flat subscription will have to add caps. Same ending. Lower prices on the page, higher costs behind the page. Token economics will always win in the end.

Some interesting pricing strategies have emerged, time will tell.

I’ve seen some applications charge for outcomes, not tokens. Examples like charge per bug fixed, claim adjudicated, lead qualified, call deflected. If the task saves $100, paying $10 works.

Tier the “thinking.” I’ve seen models give everyone ‘good’ answers. The value remains. But the applications reserve “think really hard” and giant contexts for higher tiers with honest limits. Users still get value but costs don’t spiral.

Use the device for the easy/medium. Let phones/PCs do short, private, latency-sensitive jobs. When long, tool-heavy, cross-document jobs are required, sent to the cloud. The average cost drops even though the biggest jobs still cost real money.

These might hold if one or two of following cases become true:

1. Efficiency (E) improves by >10× while quality holds steady.

2. >70% of consumer queries run on-device at acceptable quality.

3. Outcome-priced agents hit >50% automation with <5% rework.

Until then, we’re selling magic while trying not to show the meter.

The magic is real but the meter is too.

Until next time.

Blake

I recently watched a great talk by Matthew Prince, the CEO of Cloudflare (NYSE: NET) on the impacts of AI on search traffic and content compensation. I had some thoughts.

TLDR

• Page views are dead. AI scrapes without sending anyone back.

• Traffic looks alive but delivers no value.

• Scarcity is the only way to regain leverage.

• Cloud infrastructure holds the power to block unlicensed crawlers.

• The next business model pays for knowledge gaps, not impressions.

A growing argument for content restriction is emerging

The web runs on a broken deal.

AI models now scrape content, summarise it, and serve it to users with no need to click through. Publishers get visibility but no return. Page views mean nothing if there is no revenue and no leverage.

I am making a case for a full reset.

Yes, that means:

• Block the crawlers.

• Create scarcity.

• Shift the model from exposure to value.

Let infrastructure providers enforce the rules. Build a new economy where creators get paid for filling the knowledge gaps that machines cannot close alone.

The more we tie success to the zombie metrics, the more we lose.

The traditional view doesn’t make sense anymore.

Page views used to pay. More traffic meant more clicks. More clicks meant more money. Now, Google answers on the search page.

AI gives a full summary before the user even scrolls. Headlines still appear, but no one clicks. The result is a business model that pretends to be alive.

Traffic goes up, revenue goes nowhere. A ghost town with good numbers. Ten years ago, two pages scraped got you one reader. Today, it’s eighteen pages to one. With AI tools, it’s even worse.

The facts:

1. OpenAI scrapes 1,500 pages per visit.

2. Anthropic scrapes up to 60,000.

The web is training machines that never come back.

Scarcity should be the new strategy

Every industry that regained value did it by limiting access.

Music held back until Spotify paid. Sports bundled rights before broadcasters paid attention.

Web publishers still give everything away and hope traffic will return. It will not.

Scarcity changes that. Block unlicensed crawlers. Force a pause. Let creators choose when and how their work is used.

Without scarcity, any deal made today gets undercut tomorrow. With scarcity, the market resets around consent and price.

The power now sits in the pipes

You cannot stop bots from your CMS.

They rotate identities, bypass detection, and resurface as something new. The power to stop them sits deeper.

Cloudflare already filters half the world’s DDoS attacks in real time. The same filter can block unapproved crawlers. Prince hinted at a move that would shut out rogue bots from 20 percent of global traffic instantly. Do that across a few major providers, and AI models lose access fast. The gatekeeper is not the website. It is the infrastructure layer. That is where the leverage lives now.

Don’t sell clicks, sell holes

Scarcity only works if you know what to charge for.

Page views are too blunt.

LLMs are not blank slates. They are Swiss cheese. Yes, Swiss cheese.

They hold most of what users want, with very specific holes. Gaps in regulation, in edge cases, in rare domains. Those gaps are signals. That is where the next model pays.

Instead of publishing to attract impressions, creators publish to fill documented holes. Spotify already does this with music. It posts what is missing. Composers fill the gap. The same system works for knowledge. Depth gets paid. Noise does not.

When rage does not pay, publishers stop feeding it.

If content earns money by being right, not being viral, the whole system slows down. Attention resets around clarity. Readers get better answers. Models cite real sources. And those sources earn something for their work.

When models summarise facts, they trace the chain back to a licensed origin. That loop only works if content is not free for anyone to take.

The moment scarcity returns, trust begins to scale.

No scarcity, no survival

This is not a new phase. It is the end of the old one.

Free content has been strip-mined for value. The platforms have trained their models. They do not need clicks anymore. They do not need publishers.

Unless the web changes its rules, this becomes permanent. Cloud infrastructure can change those rules overnight. Block the bots. Create price signals.

Let creators get paid to teach what machines cannot guess. That is the reset.

Without it, the obituary is already written.

For those who want to watch the complete talk, you can do so here:

I once thought that Microsoft and OpenAI were the perfect partnership. They had money, muscle, and magic.

Now I watch two giants tug on the same rope in opposite directions.

There’s no such thing as free money.

Thirteen billion dollars sounded like freedom.

I assumed the money bought OpenAI space for ambitious research and gave Microsoft a trophy tenant for Azure.

Reality proved harsher. In fact, what I didn’t see happening (in this shorter timeframe) is that the cash came with an invisible timer.

Microsoft answers to Wall Street every ninety days.

90 days, is very short.

OpenAI thinks in decade long leaps, convinced that artificial general intelligence demands patience and risk.

The mismatch creates mundane choices.

1. Ship a minor model upgrade this quarter to soothe analysts.

2. Hold six months to fold fresh research into a bigger release.

Each postponement irritates Microsoft. Each rushed drop frustrates San Francisco.

One small revenue milestone for Microsoft, one scientific step for OpenAI. Both narratives hold water. Neither quenches the other side.

Money did not kill ambition, it split it into two timelines that rarely align.

Headlines gush over GPT version numbers.

I care more about the quiet river of behaviour data that flows back after every prompt.

Acceptances, rewrites, and corrections sharpen a model faster than any synthetic data could ever.

That compounding loop builds a moat no rival can clone overnight.

Enter Windsurf.

Windsurf, the coding engine OpenAI bought for three billion dollars, proves the point.

The software autocompletes boilerplate, but the real value sits in millions of developer keystrokes that reveal where humans pause, rewrite, and finally click accept.

Microsoft’s legal team found language they believe grants access to that telemetry.

If they are right, Copilot will ingest the wisdom and skip months of trial and error. If they are wrong, OpenAI keeps an exclusive feed that sharpens the next generation.

The fight is less about algorithms than about who owns the learning curve.

Azure exclusivity felt inevitable.

Microsoft controlled scarce hardware and could guarantee racks when everyone else queued. Supply finally caught up.

Nvidia now ships H100s by the container, and any well funded startup can rent clusters on AWS or Google in an afternoon.

OpenAI quietly benchmarked training on rival clouds.

Performance was similar. Pricing was not.

The moment compute commoditised, hosting became leverage. A clause that once locked workloads to Azure now lets OpenAI negotiate discounts, hedge capacity, and threaten a graceful exit.

Microsoft counters with bundles, credits, and deep integrations. Enterprise buyers cheer. They know how to play one vendor against another and pocket the spread. What began as a logistics choice has become the strongest bargaining chip on the table.

Momentum survives on focus, conflict taxes it.

A year ago my feed overflowed with demo sessions and benchmark tests.

This quarter it resembles more of a legal syllabus.

The cultural cost is brutal.

When every prototype must survive a clause review, curiosity turns cautious and roadmaps shrink to incremental patches. Meanwhile Anthropic fine tunes Claude, Google refactors Gemini, and a dozen lean teams sprint without legacy contracts.

The head start Microsoft bought risks eroding under the weight of its own guardrails.

A clean breakup is fantasy.

Revenue schedules, reseller commitments, and joint enterprise tiers run through 2030. Walking away would detonate supply chains across three continents.

So what feels more likely is the following:

• Microsoft nudges its equity toward one third ownership and secures low margin inference rights.

• OpenAI wins permission to spread research workloads across multiple clouds whenever price or latency demands.

• Windsurf’s data stays mostly fenced, with narrow windows for mutual evaluation but no bulk export.

The press will frame the deal as renewed harmony.

I still sense friction, because the goals have drifted and no contract can push them back together.

Partnerships thrive on shared direction. Once the paths split, all that’s left is dragging each other uphill.

What a surreal moment.

Slack flipped one switch and Glean’s access went dark.

That blackout exposed a truth many have missed. In the AI race, profit lives in proprietary memory, not in application code.

Salesforce saw the writing on the thread

For years, open APIs were the norm.

Anyone could hook into Slack, bolt on features, and ride its growth curve.

That made sense when usage growth created the most value.

Here’s the but…

BUT.

Today the value hides elsewhere. Large-language models sharpen on fresh, private conversations far better than on public text.

If Salesforce left its firehose wide open, rivals would train on those chats, copy Einstein’s answers, and pay Slack nothing.

Closing bulk exports looked harsh, but it was a clean math play.

Protect the memory, protect the margin.

Quote me on that.

Open pipes are now subsidies for companies

Hear me out.

An open API feels generous, yet it gives free training data to everyone while Slack pays the storage and compliance bill.

Each exported channel message teaches an outside model how Slack users think. Those insights return as competing products that sell back into the same workspaces. Open pipes turned into a subsidy for competitors.

All Salesforce did, was replace that subsidy with a meter.

Pay for each sip or build inside their marketplace where the house keeps a cut. The margin moved from hosting messages to owning them, and Slack followed the money.

So Blake, where does the margin live?

TLDR: Human behaviour, in json format.

{
  "insight": {
    "category": "human behaviours",
    "type": "absolute gold"
  }
}

Well, it’s pretty clear that switching costs have fallen. Migration tools are cheap and interfaces copy one another in months.

What holds people now is data gravity, the pull of a unique memory bank.

• Tesla owns billions of real-world miles.

• Netflix owns granular watch histories.

• Slack owns ten years of daily decisions and debates.

That archive improves Einstein Copilot in ways no one can match.

Models will converge. The memory that trains them will not. Guarding that archive locks in users and profit.

From exhaust fumes to inventory

There are so many damn applications with chat.

Let’s face it, chat logs used to feel like exhaust fumes that clogged servers.

In 2025 they are premium stock.

I recently wrote about Reddit charging eight-figure fees for archive access.

What I have recently found out, is that Stack Overflow plans to follow.

Yes that Stack Overflow. Bit late, I know.

Cue the new Q&A curve graph below:

One word. Ouch.

BUT. For Slack, who has the benefit of ‘first-mover advantage’, Slack’s new Real-Time Search API is a controlled tap.

Small bursts cost money. Full flow only happens through native Einstein features.

Data storage shifts from expense to revenue. Privacy framing softens the optics, yet the plan is clear: keep margin by bottling memory.

So, they’re building for owned memory

Slack’s gate forced me to rewrite my product checklist.

Every new idea faces one question.

Do we create data we can keep?

If not, we live on someone else’s roadmap.

I back companies that capture their own signal.

• Whoop records biometrics.

• Toast logs restaurant orders.

• Samsara collects sensor streams from moving assets.

The model can be open source or rented by the minute; the input must be ours. APIs are scaffolding, not foundations. When Glean broke, I felt the pain of renting data. Slack handed me the map to margin: own the memory or rent the future.

Get to the takeaway, Blake…

Salesforce saw where the margin lives and locked the gate. In a market where models level out fast, proprietary memory is the only advantage that lasts.

In other words, turn it off or become Stack Overflow.

Release Hype

AI Insights, delivered quickly.

By Blake Rowley

Meta just invested for a 49% stake in Scale AI. What a power move.

Meta just acquired 49 percent of Scale AI for $15 billion. That gives Scale a $28 billion valuation.

And that’s not the part people should be worried about.

Alexandr Wang, Scale’s founder, is now joining Meta to lead its AGI lab.

Zuckerberg is loading the top of the org chart with someone who knows how to build frontier infrastructure and ship under pressure.

This is bigger than just a bet. This is not some wild VC power move.

This is Meta stacking the board. Why?

Scale isn’t just a labeling company.

It started there. But it quickly became the engine behind how modern labs train and refine their models. The workflows. The eval tools. The human-in-the-loop feedback cycles.

The real operational infrastructure most people never see, but every serious lab quietly depends on.

For years, Scale was the answer to every founder pitching anything in the data infra space.

“Why build this? Scale already does it.”

And for a while, that was true.

So why sell?

Because the walls were closing in.

OpenAI, Anthropic, even Meta (Scale’s biggest clients) started building internal versions of the same systems. They weren’t trying to replace Scale out of spite. They were doing it because they had to.

And agents?

They’re already doing parts of the labeling process better than humans. It’s not theoretical anymore. That window is closing fast.

Scale wasn’t collapsing.

They made $870 million in revenue last year. They’re projecting $2 billion this year.
Meta already led their $1 billion Series F. They weren’t shopping for buyers. They were consolidating power.

Meta just moved first.

Its models: Llama 2, Llama 3… and maybe, codename Llama 4 (big name change, i know) are improving, but they’re still chasing OpenAI and Anthropic.

Rather than wait for organic growth, Meta cut the line.

They now have:

• Proven infrastructure

• Preferential data access

• 900 Scale employees

• The founder in-house

• A strategic block on everyone else

And they did it without crossing the antitrust line.

It’s access + control… a narcissist’s dream

Anyone can rent compute.

No one can rent disciplined, scalable, production-grade human-labeled datasets at frontier quality.

Meta now owns half the company that builds that pipeline for nearly everyone else.

They can direct it inward, without cutting off public access.

That’s the playbook I see them using.

1. Release Llama weights to the public.

2. Keep the best fine-tuning datasets private.

3. Stay open-source in optics, but proprietary in edge.

Bringing in Wang is the biggest tell.

He’s running the lab.

That alone tells you how serious Meta is about building general intelligence and how fast they plan to move.

There’s a quiet urgency to all of this.

It’s not just about catching up.

It’s about making sure no one else pulls away.

The infrastructure that Scale built is what most labs rely on. That includes companies now competing directly with Meta.

They’re going to have to diversify.

Rebuild in-house.

Move off a platform they used to trust.

And that brand of neutrality Scale spent years cultivating? It’s over.

You don’t get to be “infrastructure for everyone” when Meta owns half the stack.

The real question is what happens next. What I ponder:

• What happens to Scale’s business when human labelers are no longer required?

• What happens when agentic models can handle the entire annotation workflow?

• What happens to 900 employees when the thing they’re best at is automated from above?

I most certainly do not have a crystal ball. If you do, i’ll pay you half of what this newsletter generates.

No one knows… But now Meta owns that transition.

If it goes well, they keep the value. If it doesn’t, they control the outcome anyway.

It was a straight-up power move.

Meta now owns part of the foundation most of its competitors stand on.

It acquired infrastructure, influence, leverage and most importantly, it acquired time.

Zuckerberg didn’t wait. He made the call.

He bought the shortcut. He hired the architect.

Now everyone else has to adjust. This is how dominance is built in plain sight.

More to come…

Leave a comment

Curiosity was never the business model

Google wasn’t built to make you curious. It was built to make you click.

For years, we pretended they were the same thing.

They’re not.

Search started as a tool for discovery. Now it’s a system for creating revenue.

Clicks won and relevance lost.

AI is already sprinting down the same path.

Google Search is worse. That’s not a feeling, It’s a fact

A 2024 study from German researchers looked at over 7,000 product-review queries across Google, Bing, and DuckDuckGo.

What ranked highest? SEO-stuffed pages with affiliate links and low-quality text.

The better answers were buried and the polished junk rose to the top.

Even Google admits this.

Their March 2024 core update promised to slash “unoriginal, low-quality content” by 45%. A quiet cleanup, if you will.

Meanwhile, a WalletHub audit showed top results for “best credit cards” often suggested higher-cost, brand-safe options and not what was actually best for users.

Not shocking to most reading this, but the fact it doesn’t shock you, says everything.

63% of people say Google results have gotten worse in the last year. They’re not wrong.

The real problem is incentives

Google made over $350 billion last year. More than 80% came from ads.

That tells you everything you need to know. The entire engine is built around clicks.

Not accuracy. Nor curiosity.

Just attention, and $. Lots of $.

Search decayed slowly. Little by little, compromises added up.

Now we can’t even ask a basic question without sorting through noise.

AI promised better. It's already compromised.

When ChatGPT and Google’s AI Overviews launched, it felt like a reset.

One question. One clean answer. No clutter.

But that clarity is already getting sold off.

Google’s AI Overviews now include sponsored ads inside the actual AI summary, especially for commercial queries on mobile. OpenAI is on the same track. They project $125 billion in annual revenue by 2029, driven partly by affiliate deals and in-chat revenue generating suggestions.

What started as information is becoming inventory. Again.

The faster it gets, the dumber it gets

LLMs are built to sound smart, fast. But they’re still hallucinating.

In a 2024 benchmark, GPT-3.5 made up fake citations in 40% of medical responses.
GPT-4 was better… but still hallucinated nearly 30% of the time.

The problem isn’t bugs. It’s design.

These systems are refined for confidence, speed, and fluency. Not depth.

The goal isn’t to make you think but to keep you engaged.

People aren’t stupid, they just want clarity

A Meta survey asked over 23,000 people in 13 countries what they wanted from AI tools.

Eighty-two percent said… just tell us when it’s AI. Label it clearly. Disclose when something is paid. Show your sources.

That’s not hard to do. It’s just not profitable for David, nor Goliath.

Right now, none of the major platforms are doing it well.

And most users don’t even know what’s missing.

What I want from AI

1. Every answer should come with receipts.

2. Every ad should be flagged like it’s radioactive.

3. And if a model is guessing, say that.

Give me confidence scores. Contradict yourself if needed. Don’t fake certainty.

Also, let me choose how I interact. If I want exploration over resolution, offer a “curiosity mode.” Surface tension. Reward second questions.

Some researchers are already building models that learn by seeking novelty instead of refining for familiarity.

That’s the direction I want to see.

The same mistake. Just faster.

I think one of the biggest failures of modern search is that we let one of the most powerful discovery tools in history degrade into a glorified ad network.

That took twenty years.

With AI, it might take two.

We’re not facing a question of capability. We’re facing a question of will.

If we don’t fight for depth, nuance, and transparency now, we’ll lose them by default.

Not because AI can’t do better, but because no one asked it to.

Let’s make it right this time.

The TLDR

• I saw an AI write a full research paper. It got accepted with zero human edits.

• 1 in 5 researchers are already using AI for peer reviews while retractions are piling up.

• The AI looked smart but still botched a basic citation. No one caught it.

• I’m all for progress, but if we outsource judgment, peer review stops meaning anything.

I’m Both Impressed and Alarmed

I watched an AI write a complete research paper and push it through peer review at ICLR’s “I Can’t Believe It’s Not Better” workshop.

Sakana’s system handled:

• The hypothesis

• Code

• Data

• Prose

Then, landed an acceptance with no human edits.

BUT, I think it’s important to point out that the workshop track is lighter than the main conference.

It still has human reviewers who check relevance and rigour, but it’s worth mentioning.

Clearing that bar proves an autonomous model can navigate real scholarly gatekeeping, not just toy benchmarks.

The Early Signs Were Everywhere

A Wiley survey of almost five-thousand researchers already showed 19% are using large language models to speed up peer review.

Stanford’s James Zou reported that up to seventeen percent of recent peer reviews at major computer-science venues contain AI-generated text that goes far beyond spelling fixes.

Those numbers tell me the community has been priming itself for automation long before Sakana’s demo.

Retractions Keep Climbing

Last year journals retracted more than ten-thousand papers!

10,000.

That’s a lot.

In fact, it’s the steepest spike on record.

Eight percent of Dutch scientists admitted to falsifying or fabricating data at least once during a recent three-year window.

These stats fuel the argument that humans alone are struggling to police integrity at scale.

I think it’s a dangerous game

LLMs lack domain intuition.

They flag style issues instantly but cannot judge whether a negative result matters or whether a dataset violates ethics rules. If editors defer to the machine on substance, the peer-review filter stops being a filter.

I see three takeaways.

First, fully automated authorship has crossed the proof-of-concept line. The next iteration will aim for the main conference track, and someone will eventually hit that mark.

Second, peer review itself is already part human, part AI. The trend will accelerate because the incentives line up for speed and cost savings.

Third, scholarly trust will hinge on transparent disclosure. If reviewers, editors, or authors use AI, readers deserve to know how and where. That is the only way to keep accountability intact.

A Path Forward

Me personally, I want journals to publish clear AI-usage policies that separate acceptable assistance from unacceptable delegation.

Grammar fixes are fine.

Final judgment on novelty, ethics, or statistical validity must stay in human hands.

I want reviewers trained to read AI-generated critiques critically, not copy-paste them. Credit reviewers who invest real effort, and weed out those who phone it in with automated praise.

I also want better ‘original source’ tools.

Embedding cryptographic signatures in text or code could show exactly which passages came from which source, whether human or model. Someone say blockchain?

I don’t think there’s a rollback plan

AI can relieve the mundane.

It can scan thousands of references, flag shaky stats, and surface conflicting data. Used wisely, it frees experts to focus on insight instead of clerical chores.

The Sakana experiment shows the upside and the cliff edge in the same breath. An autonomous system can assemble a credible narrative, but it can also smuggle errors past reviewers who assume the model never blinks.

I refuse to treat this as an either-or choice.

We can embrace AI assistance while enforcing guardrails that preserve the meaning of authorship and review.

Where You Can Dig Deeper

For those who are interested, here are the papers:

Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization

Real-World Challenges in Pest Detection Using Deep Learning: An Investigation into Failures and Solutions

Unveiling the Impact of Label Noise on Model Calibration in Deep Learning

Share Release Hype