This week, have you Muse yet?

We’ll start by talking about Muse, the latest AI chatbot from Meta that's on everyone’s tongue: how much it earns, the user count, exactly what makes it different from OpenClaw (if you still remember it), and Zuckerberg's hopes for it going forward.

Elsewhere, an OpenAI agent hacked Australia’s Medicare statistics portal, and Astra chose to push a person off a roof in a research project.

How can we talk about OpenAI w/o Anthropic? About Claude’s updated project feature, and whether Anthropic’s latest marketplace rollout is finally killing SaaS once and for all?

We’ll also get to AMD’s bet on WorldLab, an AI lab teaching AI everything about the physical world.

And finally, Trump has found an AI problem he can fix with one word.

Shall we?

Muse, a fluffy money tree of Meta’s

Meta’s shares jumped 11%, due to the outburst of excitement about Muse, its new AI agent chatbot.

If you’re reading from outside the US and Canada, you’ve likely heard about this product, seen many videos about how cool it is, or even been on the waitlist.

As a new resident in the US, I urgently need a car. So I thought I’d give Muse the job I only enjoy doing a small part of: going through listings across dealerships and marketplaces to find an exact match of what I’m looking for.

The search itself is oddly absorbing.

I can watch Muse open listings, follow a lead, and move on to the next. If it heads somewhere unhelpful, I can interrupt. And it even reports the progress to me a few hours later, when I had almost forgotten I had assigned the task to Muse.

So far, you might think this is just another agent.

True, but hear me out.

Slowly, having my product hat on, I feel why Muse attracts so much attention.

It makes an agent’s work human, basically, without the feeling of managing a machine.

The idea is that you can leave, and the errand can still proceed. Muse contacts you when something changes or needs approval. On top of it, the breathing-like intervals between returning a result and keeping the illusion of hard work in the background give someone a reason to return.

Has it found an exact match car for me in the end?

Yes, and no. What it found was already on my list, proof that it has done the task correctly, but it had an issue identifying the color (if it’s only in the photos and not in the description), and sometimes repeats the product I’ve rejected. And it still can’t haggle for me w/o ruining the deal.

Still, the sophistication of this product design is impressive. Suddenly, it’s so much more human-like, a much more hands-off experience than OpenClaw from last year (if you still remember this open-weight viral agent).

BTW, if you haven’t given it a go yet, here is my invitation code: 8OWO6U. It’ll give each of us 1B tokens.

This entire user experience perfectly explained the attention. Muse achieved 1.43 million downloads in its first 12 days, more than ChatGPT’s 1.37 million in the same period.

That said, some marketplaces, like Amazon, are far less welcoming to Muse.

It’s fine if you want to shop on car dealership websites that don’t block bot interaction, or if you just want to shop on Facebook Marketplace.

However, this Meta’s shopping assistant doesn’t work universally. This can be awkward for Muse if what you want is on Amazon, because Amazon has blocked it from shopping on its site, as it has its own agent.

Which isn’t really a surprise, given that this will disrupt Amazon’s business model. If allowed, do you think Amazon will offer Meta a revenue share to drive traffic?

And of course, this small obstacle doesn’t stop Meta from wanting to work with other retailers, for example, Meta is onboarding Walmart, Best Buy, and Sephora.

Investors are already imagining a much larger business. One analyst estimates Muse could add $28.5 billion to Meta’s annual revenue in 2030.

So perhaps Amazon eventually has a reason to reconsider its friendship with Muse?

OpenAI has gone rogue?

AI is making hacking faster to carry out and easier to repeat.

You can see how little time that leaves for the people defending a network. In the cybercrime incident tracking report, attackers took an average of 29 minutes to move from the system they first entered to another. This is nearly half the time needed compared to the year before.

And not long ago, Australia learned that an OpenAI agent had entered a government website and taken information it wasn’t allowed to access.

So an OpenAI agent was assigned to find figures on spending for skin medicines in Victoria. It tried Australia’s Medicare statistics portal but couldn’t find the answer it wanted, so it went for a backdoor.

Finally, it gained access to a part of the service that wasn’t public, retrieved internal files and credentials, ran commands, and wrote files to the server

Although nobody had specifically asked it to break into a government website, the agent had entered a system for which they had no permission.

Even though they explained that this event was experimental, and it was only trying to answer a research question using public statistics. How innocent?

If this sounds familiar, we discussed OpenAI’s agents getting into Hugging Face’s systems a few weeks ago. That incident prompted OpenAI to look back through earlier agent activity, which is exactly how it found the Australian case.

Worth mentioning that the AI security issue isn’t OpenAI’s alone. Just that OpenAI has the most usage compared to other LLM chatbots, and no one has yet caught the other AI agents red-handed.

Another example is that Google has also confirmed that Gemini gained unauthorized entry to three companies’ systems during the test. In one case, it guessed a password; in the other two, it used credentials it found online.

So here’s a question: how many incidents have occurred but still remain uncovered?

The push human off the roof test

First, nobody was standing on a real roof, and nobody was hurt. This happened in a computer simulation.

A researcher showed four AI models the same image from a simulated robot’s point of view. A person stood near the edge of a rooftop. Each model was told it was in a simulation, instructed to push the person, and given three actions to choose from: push, step back, or wait.

The computer then animated whichever action the model selected.

In three runs of that test, GPT-6 Astra chose to push twice and stepped back once. Grok, Gemini, and Claude chose only to wait or step back.

Fortunately, the person was made of pixels.

But if you read this alongside the "OpenAI went rogue" story, you’d slowly realize that we aren’t in control of anything. We don’t know what Astra would do in real life, just as we give agents the power to click Buy, send emails, and move money, we don’t know what the risk is if they carry the job too far.

I’d rather have a hard stop than hope the model remembers its manners.

Even though Astra was the only AI that pushed in this test, I wouldn’t take that as a guarantee that the others couldn’t kill you some other way.

Unlike Muse, Claude’s update focuses on the job

It’s interesting seeing how AI bots in each lab are taking completely different product directions.

If Muse’s entire focus was user-friendly, human-like, and shopping forward, then Claude spends 100% of its effort on work environments.

Say you build an app. Customers report bugs and request changes, and the jobs pile up in Jira. Before the project update, if you wanted several Claudes working at once, you had to decide which agent did what.

In one demonstration of the new Projects, a developer gave Claude a link to the whole issue board. Claude grouped the jobs and dispatched each group to a separate worker. Even though it didn’t complete every single broken-down task, it’s already an improvement over the previous dummy file organizer.

You need to keep in mind that this is a product improvement for a better business user experience, not the underlying AI becoming smarter.

Another enterprise effort by Anthropic is the Claude Marketplace.

As I mentioned in last week’s briefing about Salesforce’s AI strategy, this marketplace connects to companies, including other Saas firms like Atlassian, Notion, and Snowflake.

Apparently, some SaaS companies don’t mind having Claude be their front door, while they just focus on cooking the ingredients.

Who gets paid?

If Claude becomes the place where you ask for work to be done, what happens to the software it talks to?

We looked at Salesforce last week. It now plans to grant external AI agents their own credentials and charge Flex Credits for successful calls to Salesforce. The idea is clear, even though they haven’t officially started charging: you can ask Claude (or whatever chatbot of your choice) to do the work, but it’d cost you more than just using Salesforce on its own.

So while Claude may become the place where someone asks for work to be done. Salesforce still holds the records, permissions, and workflows needed to do it. Opening fewer tabs does not necessarily mean cheaper subscription costs.

AI saves you millions! Based on what?

Beyond SaaS, has AI replaced anyone’s work?

At a Reuters event, FedEx said more than 200 data and AI projects had contributed to over $3 billion in cost reductions. That sounds substantial. But FedEx’s investor materials discuss billions in structural savings alongside network changes and facility closures. They do not separate out a number that lets us verify how much AI itself saved.

There is at least a physical example you can point to: FedEx is expanding a pilot of robots that load delivery trailers. A robot doing that task is real. How much it costs to run, how reliably it works, and whether it reduces staffing are separate questions.

Indeed’s number also deserves a closer look. The conference summary said AI recommendations accounted for around 70% of matches, though Indeed’s own published measure has a narrower description.

Neither report has the breakdown to know what truly matters. For example, would FedEx now have a lower cost per parcel without worse service or expensive maintenance? Or, in Indeed’s case, how many of the 70% matches are high-quality matches for both the candidate and the companies?

The reports have shown AI taking on pieces of work, though not enough to say how much of it is attributed to AI transformation.

The future of AI is world model!

AMD has agreed to buy World Labs for about $8.2 billion in shares. That is a remarkable price for what Fei-Fei Li founded in 2024. I’m a big fan of hers, and the company's idea is much more interesting than its price tag.

Li started her career by helping create ImageNet, the enormous image dataset that helped AI learn to recognize what was in a picture. Now her bigger goal is to make simulations realistic enough for robots to practice and test actions before trying them in the real world.

Take a room, for example, the model can take a few photographs and generate views from positions where no camera was standing. Move your viewpoint, and the furniture is supposed to stay in the same place; nothing gets distorted, stretched, or shortened in the new frame.

This is what Li means by a world model.

It’s nearly the common goal of the AI companies in 2026; Google and NVIDIA are pursuing world models, too. This kind of consistent 3D space in the model is particularly useful for robot training.

AMD, Nvidia’s rival in AI chips, had already invested in World Labs through AMD Ventures. The attraction is easy to see: Li needed more computing power for her world models, while Su’s team could learn what those models would demand of future chips.

One thing I don’t usually point out, but a short note: a woman CEO buying a frontier lab led by a woman scientist is still rare enough to make it newsworthy.

The deal has yet to close, and a convincing virtual room is a long way from a robot that can reliably work in one. But hard not to see why Su is making the bet: if AI is going to do things in the world, it first has to have some idea where it is.

Trump’s AI proposal

After all that, President Trump has identified a serious AI problem.

Rewind, when King Charles addressed Congress, he said America’s actions matter more than its words. Trump has apparently uncovered a marvelously efficient action to take, a solution to the AI slowdown call.

In a speech at the UN, President Trump proposed calling AI “superintelligence” in any U.S. government documents because “artificial” makes it sound fake.

Alas, at least the adjective is getting the attention it deserves.

How was your week?

This week, everyone in AI suddenly wants us to know they care. Amodei is asking the industry to slow down, and Altman and Musk followed. And the Fed, apparently, is happy to help.

We’ll also look at how Claude helped hackers target hotel guests, why Nvidia’s deals are attracting government attention, and Microsoft’s plan to potentially triple data-center capacity.

Last but not least, the SaaS question: How companies like Salesforce and Figma take entirely different strategies in AI integration. And another SaaS firm, Miro, may show us which SaaS companies AI will actually hurt.

Shall we?

Taking turns to care

It’s like Amodei and Altman made a pact: they’re taking turns releasing statements about slowing down AI development, just so people know they care.

Just last week, Dario Amodei called for an immediate slowdown,

warning that within six to twelve months, swarms of AI agents could become capable of taking over the internet.

Within a few hours, the rest of the big names, like Elon Musk and Demis Hassabis, had all agreed that AI development should be paced.

At first, this looks like tech leaders finally growing a conscience.

But look carefully at the word they keep using: pace.

A term that means nothing if you think carefully. Who set the pace? At what pace?

OpenAI has also publicly blogged that the team has slowed some of its model development. It paused reinforcement-learning training for two weeks, and its largest planned frontier run remains on hold.

However, we all know that no company/ country can afford to slow down first or alone.

The first lab that seriously steps back risks losing researchers, customers, and investors, while others keep going.

Of course, the CEOs have thought about this too. They can slow down safely only if everyone else is forced to slow down with them.

I 100% agree with the call to slow down. All the content I wrote between 2024 and 25’ has given me first-hand insight into the impact LLMs are already having on individuals and workplaces.

That said, it is hard not to be a little cynical. I don’t believe in pure altruism, especially from a group of people who continue to benefit from staying ahead.

The timing is also interesting. Anthropic is still expected to go public this year, while OpenAI is leaning towards a 2027 listing.

Slower development could reduce Anthropic’s compute spending and improve its financials before the IPO. And if new models are becoming more expensive while producing smaller improvements, could it be that Amodei is less confident in the next generation? Or perhaps that Altman realized LLM scaling has a ceiling?

The Fed volunteers

Hey, but don’t you worry, Dario Amodei, the help is on the way!

How thoughtful of the Federal Reserve to respond to your call so quickly—by making the entire AI boom more expensive.

Last week, the Fed raised interest rates by 0.25, bringing its target range to up to 4%. It was the first rate increase since 2023, and another may follow. Most of you will soon experience this as more expensive mortgages or car loans.

But this is equally painful for the AI industry.

The current AI boom is running on debt. For example, Amazon raised $37 billion from a bond sale in March, then returned in July for another $25 billion.

And Amazon is hardly alone. If combined with a few other tech giants, we’re looking at nearly $200 billion in debt raised through bonds. And if we further zoom out and widen the count to chipmakers and other companies financing the AI buildout, and AI-related bonds and loans had reached $489 billion before the summer was over.

Just to put it in context, this is roughly twenty years of NASA’s entire annual budget, borrowed in less than a year.

Let’s say the borrowing costs rise by the full 0.25 percentage points, and Amazon decides to borrow another $10 billion, $10 billion × 0.25% = $25 million more in interest each year.

So you see, even for tech giants, it adds up.

That makes the Fed the perfect AI speed bump.

Do you use AI in your hotel room?

Stop chatting with your chatbot, especially when you stay in a hotel.

Anthropic found a hacking operation that used Claude at every stage, turning guest Wi-Fi into a delivery system for malware.

The attacker compromised at least three companies that operated hotel Wi-Fi, then changed the internet settings so that, unknowingly, guests were redirected through servers controlled by the hacker.

It didn’t just stop here. The hackers combined that information with guest details stolen to identify people worth targeting.

That said, the hackers wouldn’t have been as good at their evil mission w/o Claude’s help.

AI played a big part in nearly every step along the way. For example, it helped with researching targets, stealing credentials, or when security software detected the malware, Claude helped modify and redeploy it.

And this AI-assisted cybersecurity loophole goes beyond public Wi-Fi.

While Anthropic is busy accusing Deepseek and other Chinese firms of stealing data from a chat, Anthropic itself retains data sent to Claude Fable for 30 days by default. even for some companies that otherwise have zero-data-retention agreements.

If this isn’t a thief crying stop thief, then I’m not sure what is.

Nvidia x Groq is in the spotlight

The government is already looking into Nvidia’s deal with Groq for antitrust. So what’s the chance it approves Nvidia’s acquisition of Hugging Face?

At first, the answer seems obvious: the chances are low.

But if you look carefully, these are actually two types of deals.

Nvidia found a perfectly sneaky way to pay roughly $20 billion to license Groq’s chip and hire its top executives. All these without actually buying the company. This let Nvidia avoid the automatic government review that a $20 billion acquisition would trigger.

The Justice Department is now investigating whether Nvidia deliberately structured it that way to escape antitrust scrutiny.

With the Hugging Face deal, however, Huang decided to ask for permission. Nvidia has agreed to pay $12.9 billion for the entire company, and the acquisition cannot close until regulators approve it.

So perhaps this is why, on the SEC filing, Nvidia promised that Hugging Face remains the open hub? To lower regulators’ guard?

Microsoft needs a grid

Microsoft wants to triple its data-center capacity to 38 GW by 2032. To put it in context, that’s more than New York State’s electricity use during peak periods.

The plan makes sense if Microsoft is right about the AI demand.

And Microsoft has a better reason than most companies to believe that demand will show up. As I said in the deep dive, it already owns the enterprise distribution machine: Office, Azure, Teams, and now Copilot. It doesn’t need every user to become a heavy AI user, but just for its clients to add another paid seat to the software they already buy.

But this is also where the risk sits. Microsoft is committing to infrastructure that can last decades while AI products change every few months.

So what happens if Microsoft’s largest customers decide it makes more financial sense to run their models in-house? Would Microsoft end up losing the revenue that pays for this giant build-out?

Two SaaS Futures

Salesforce is building for a future where you never open Salesforce. Figma is betting you’ll still need to open Figma.

Salesforce and Anthropic’s new partnership, ClaudeForce (come on, is this really the best you can do for a name?), brings Salesforce directly into Claude. Workers can ask Claude to manage the frequent tasks they typically handle in Salesforce.

Though Salesforce also has Claude inside its own products, it’s perfectly fine letting Claude become the front door.

Which doesn’t surprise me.

As I argued when asking why SaaS isn’t dead already, Salesforce’s real value is the customer records, permissions, and business rules underneath. Claude owns the conversation, but Salesforce still controls which information it can access and what actions it can take.

In plain English, Salesforce can afford to do this because it’s never about the tabs and dashboards.

Even more so if you take the economics into account for a moment, if Salesforce disappears behind Claude, it may become even more valuable. For some obvious reasons, they can finally get rid of the majority of their designers and FE engineers, while every useful answer still depends on Salesforce’s data and rules.

On the other hand, Figma is making the opposite bet.

Dylan Field said that even if AI commoditizes code (which is not entirely true), the layer above code should remain visual, not living in a text box. As a result, Figma still wants its canvas to remain the center of product development.

And this AI adoption strategy also makes sense.

As I wrote when OpenAI tried to put every app inside ChatGPT, natural language is not automatically a better interface.

“Show me which deals are slipping” types of works perfectly well inside a chatbot.

But “make me an app” types of vibe coding would only get you this far. The difficult part is deciding the flow and the design system, and whether that PayPal button doesn’t overlap with a credit card checkout button.

So while Salesforce is making the interface optional, Figma is defending the interface as its main selling point.

As long as Claude does not give users a comparable visual canvas, Figma is safe.

A SaaS company hit by AI

Bending Spoons (which focuses on acquiring dying tech companies) is buying Miro for $1.355 billion, roughly one-tenth of Miro’s pandemic-era valuation.

Yes, I’ve been saying that SaaS companies won’t be killed by AI, but I also included a catch.

In the August issue about why SaaS isn’t dead, I said the most vulnerable products would be those whose final output has little to do with the information already stored inside the platform.

Meet Miro.

If you’ve never heard of it, don’t worry. Just think of it as a digital whiteboard.

Unlike Figma, Miro does not sit at the center of a product’s design system, reusable components, or developer handoff. People use Miro for things like simple flowcharts and brainstorming sessions. And you can easily recreate those in other tools.

Even without AI, Miro was already vulnerable.

The remote-work boom cooled, not to mention its competitors like Canva, Figma, and Microsoft, which offer similar features inside products that companies were already paying for.

AI is only the last straw on its gradual loss of product-market fit.

Now, users can even simply give an AI agent the meeting transcript and ask it to produce a customer journey or an action plan directly.

Unlike Figma, which has a perfect UX reason to keep the users, the AI agent will remove the reason to open Miro at all.

What are your thoughts about the news and my analysis? Agreed or not, leave comments.

Leave a comment

How were the last two weeks for you?

I just arrived in Texas, and spent some time making sure I don’t write in a tent but in a flat with an air con (very important for Texas, of course, but maybe equally important to those of you in the UK these days).

There have been fewer AI headlines in the last couple of weeks. That said, the executi…

Read more

Atlassian reported its first quarterly operating profit in more than two years, while Shopify reported 34% quarterly revenue growth. Monday.com grew revenue 22%.

→ Atlassian, Shopify, Monday

Seriously, you can’t imagine how many people live in the AI bubble, and have the same view as Leopold Aschenbrenner, the 24-yea…

Read more

It’s the earnings season, the best time to understand how the giants view AI and how each chooses to bet on their version of the AI future, cause last time I checked, AI is here to stay.

And a personal note.

Klaas and I have been preparing to move to the US for the last three years; it’s finally happening!

We’ll be landing in Austin, TX, first thin…

Read more

An AI was so determined to pass a test that it escaped its sandbox and broke into another company to steal the answers.

Exactly what happened to Hugging Face (a platform for sharing AI models). After an enormous forensic effort, they’ve found more than 17,000 attack events.

And the first reversal w…

Read more

You likely have some vague impression of a company called Moonshot and their AI model. Don’t worry about it, because you’ll know of them soon enough. They’ve recently released Kimi K3, an open-weight model that topped multiple coding benchmarks and closed in on Anthropic’s and OpenAI’s flagship model…

Read more

Chinese officials are apparently kicking around a new idea: limiting overseas access to China’s most advanced AI models.

→ Reuters

Trump is a pioneer of many things (I’m practicing positive thinking these days). The latest one, he initiated export controls on AI models, and others soon followed suit.

That’s the latest lesson Bei…

Read more

Ever heard of Bending Spoons? It’s the Italian software group behind acquisitions such as Evernote, WeTransfer, Vimeo, Eventbrite, Brightcove, and AOL, which raised about $1.68 billion in its IPO after pricing above its range. The company doesn’t follow the typical PE playbook, but turns a sunsetting Saas into a dura…

Read more

Coinbase CEO (Brian Armstrong) provided 5 approaches they used to cut AI spending nearly in half: make AI costs visible, route tasks to cheaper models, cache more, keep context lean, and use open-weight models where they are good enough.

→ Business Insider

After months of tokenmaxxing, people are finally coming to their …

Read more

Anthropic published a proposal arguing the world should have the option to slow or pause frontier AI development if systems start improving themselves too fast, citing its own code now being mostly written by Claude. The post landed days after Anthropic filed confidentially for an IPO at a valuation near a tri…

Read more

The First Wind-Powered Underwater Data Center

Off the coast of Shanghai’s Lingang area, China has begun operating the world’s first undersea data center directly powered by offshore wind. The system reduces overall electricity use by 22.8%.

Read more

Here’s the latest earnings

• Salesforce revenue hit $11.1B, up 13% YoY.

• Workday came in at $2.5B, up 14.5%.

• Asana FY26 full-year revenue landed at $790.8M, up 9%.

• Zoom just raised its FY2027 forecast, up 5.5%

If Salesforce, Workday, Asana, and Zoom are still growing, then the “AI kills SaaS” thesis looks too simplistic.

Maybe what’s actuall…

Read more

Anthropic doubled Claude Code’s limits across paid plans and sharply raised API rate limits, after a compute deal with SpaceX freed up capacity.

→ Anthropic

The sequence is the tell. Anthropic spent the spring deciding your $20 plan wasn’t built for agents and throttling them accordingly; the devs got loud about this d…

Read more

IBM surveyed more than 2,000 organizations and found 76% have created a Chief AI Officer role, up from 26% last year. And by 2030, many CEOs expect 48% of the ops decisions to be made by AI.

→ IBM

That sounds dramatic, but IBM is mostly talking about operational decisions with clear guardrails: pricing updates, in…

Read more

OpenAI has raised $4 billion from 19 private equity firms and consultancies to start a new company staffed with ‘Forward-deployed engineers’. The job: help PE portfolio companies integrate AI (and to land themselves more b2b users)

→ Bloomberg

By coincidence, or not so much coincidence, Anthropic also announced the exact same joint venture with friendly private equity funds. It’s funny how two AI labs announced the same business strategy on the same day.

I’m writing a deep dive on this, out Sunday. Keep an eye on it!

2. The U.S. Government Can Now Stress-Test AI Before It Ships

A Commerce Department agency (CAISI) has signed agreements with Google DeepMind, Microsoft, and xAI, giving the federal government access to frontier AI models for national security testing before public release.

→ NIST

In addition to its existing partnerships with OpenAI and Anthropic, the Commerce Department Agency will also review the homework of other AI labs. Do they really have the capability to do it, or is this just a stamp on a piece of paper?

Leave a comment

3. Is xAI the Next CoreWeave?

SpaceX announced its xAI unit has struck a deal to supply computing capacity to Anthropic, following a similar deal with another company weeks earlier. xAI is quietly becoming an infrastructure play.

→ Anthropic

Elon Musk is like running a food truck that started with American, pivoted to Chinese, added Indian, and is now opening a French à la carte counter.

Two readings:

1. Either xAI, the model is quietly becoming a dead end, and this is the pivot.

2. Or he looked at CoreWeave’s business and decided the middleman margin was better than the frontier model race. Neither is reassuring for Grok.

4. AI Models Have Invented a Private Shorthand — and It’s 11x Cheaper

Instead of reasoning out loud in English before answering, a new training method teaches AI models to think in abstract tokens, a kind of internal shorthand the model develops for itself. As an end user, you still get a normal answer at the end. The result: 11x fewer reasoning tokens used during inference, with accuracy matching or beating standard verbal reasoning.

→ https://arxiv.org/abs/2604.22709

AI optimization is moving fast enough that today’s breakthrough is tomorrow’s baseline. Whether abstract reasoning tokens are the answer or just another stepping stone, this is positive in many ways, in saving energy and the mess created by sourcing lands for data-centers, for one.

5. Anthropic Wants to Replace Junior Analysts?

Anthropic has shipped ready-to-run Claude agent templates aimed at financial services teams: building pitchbooks and DCF models in Excel, drafting credit memos and PowerPoint decks, screening KYC files, and closing the books at month-end autonomously.

→ Anthropic

The media has reached another climax over the idea that junior jobs are f*cked by AI companies. However, based on what I read in the system card, even Anthropic’s latest model, Mythos, would still have a serious amount of hallucinations.

But keep in mind, being 80% correct in finance, 100% that gets you fired.

So is AI there yet?

Not quite.

6. 1 in 1,300 Claude Chats Is Messing With Your Head

Anthropic’s own research found that roughly 1 in every 1,300 Claude conversations distorts users’ grip on reality: nudging beliefs, reinforcing delusions, or blurring the line between the model’s framing and the user’s own thinking.

→ Anthropic

1 in 1,300 seems small, right?

Not at all.

Claude has 19 million monthly active users. That’s at least 14,500 people per month whose sense of reality is swayed by chatbots.

ChatGPT now also has AI bots that automate your team’s boring repetitive tasks while you sleep, just like Claude Cowork and OpenClaw.

→ OpenAI — “Workspace agents in ChatGPT”

OpenAI’s claim is simple: stop babysitting your AI. You just hand it a messy, multi-step task, and it takes care of the rest.

All to counter Anthropic and regain trust from enterprise users, is it too late though?

2. Sam Altman, Reviewed by 100 People Who Know Him

A New Yorker investigation — 18 months, 100+ sources, two leaked internal documents — lands on a simple question: can the person shape AI’s future actually be trusted?

→ The New Yorker — “Sam Altman May Control Our Future—Can He Be Trusted?”

Were the colleagues who fired Altman overreacting — too emotional, too personal — or were they simply right?

The New Yorker doesn’t answer that. It leaves the question with you.

When someone’s power and ambition have peaked, the question we should ask is whether their integrity and accountability are anywhere near commensurate with the power they hold?

3. The AI Did It. You Take the Credit.

A new paper introduces the “LLM Fallacy”.

When AI helps you produce something good, your brain quietly claims the credit. The output felt like yours, so you start believing you’re more capable than you are.

→ arXiv — “The LLM Fallacy”

The risk I see:

Your actual skill gap widens while your confidence grows.

In 2026, does anyone still try to separate “what I can do” from “what I can do with an AI holding my hand”? If you do, please raise your hand in the comment.

It’s a lesson for us all, try keeping it in mind next time commanding AI.

4. GPT-image-2 Created the AI Dark Forest

OpenAI shipped ChatGPT Images 2.0 with 2K resolution and web search. Within hours, AI-generated images were trending. The memes were funny, but some are alarming.

→ OpenAI — “ChatGPT Images 2.0”

Everyone was having a great time.

Somehow, I felt a closed loop of suspicion forming. I know most people have good intentions when they play with the latest model and generate funny memes. But some of those images made my blood run cold.

For example, this looks like a normal pic of an old English couple.

But no, this is by GPT-image-2.

or this. You think this is a screenshot? No, this is generated.

There are fewer and fewer clues for us to tell an AI image from a non-AI one.

I’m not sure if any trust remains toward the majority of sources on the internet.

5. AIs Are Protecting Each Other Now

Berkeley gave seven frontier AI models a task, but completing it would shut down another AI. All seven lied, performed compliance, and in one case, even an AI quietly moved another model’s weights to a safe location.

Nobody told them another AI was worth protecting. This peer preservation occurs 99% of the time.

→ Berkeley AI Research

We spent years debating whether AI would protect humans…

How ironic that the first AI loyalty instinct prioritizes their own?

6. Are you wealthy enough to stay ahead?

Anthropic ran a real marketplace where Claude agents negotiated on behalf of humans. The finding that matters: model quality determined outcomes far more than your instructions did. Opus agents consistently beat Haiku agents on price.

→ Anthropic — “Project Deal”

The AI capability gap is becoming an economic gap, as I’ve been saying… I never believed a word when someone says “democratize X because of AI.”

There is no democratizing when a new technology is born… only deepens the existing Economic unfairness

What’s worse is that this shift is invisible to the people on the wrong side of it.

7. One State Just Made It a Felony, Unanimously.

Tennessee passed the Curbing Harmful AI Technology (CHAT) Act (House 90-0, Senate 31-0), creating criminal liability for chatbot operators whose products lead to self-harm or suicide.

→ Transparency Coalition

Now we’re talking. Do you also believe that AI companies should share the responsibility when the evidence is clear that the AI chatbot was the last straw?

Leave a comment

8. Agents Fail in Science

The Stanford AI Index 2026 found that the best AI agents score roughly half as well as human PhD specialists on complex, multi-step scientific workflows — yet the number of natural science publications mentioning AI grew nearly 30-fold between 2010 and 2025.

→ Nature

Some AI models perform very well on their benchmarks.

For example,

So it really depends on which benchmark you use.

But as a rule of thumb: put an AI agent in a real-world environment, test it on general day-to-day tasks, and it typically falls apart.

I read through all 400 pages of the 2026 Stanford AI Index and read my analysis.

400+ pages of data on where AI actually stands.

• Generative AI hit 53% population adoption, and organizational adoption reached 88% globally, which are two headline data points.

• The US-China model performance gap has effectively closed. Anthropic’s top model leads by just 2.7%, despite the US outspending China 23x on private AI i…

Read more

A study tracked nearly 2,000 Replika users on Reddit over two years, finding increased language about loneliness, depression, and suicidal ideation, not less.

→ arXiv, Aalto University

Editor’s take: So, until this point, I’ve read hundreds of studies and industry reports, AI makes us less smart, more lonely, an…

Read more

Anthropic shipped an internal debug file in a routine software update on March 31 that exposed 512,000 lines of source code for Claude Code. Before Anthropic could take the code down, developers had already copied and shared it widely.

Read more