- 10 years of Claude Code experience - Real inventor of the internet - Software / AI Engineering
Joined October 2012
- Tweets727
- Following501
- Followers117
- Likes803
Wow, glad they released this. Super concerning behavior from Qwen3.8.
In the spirit of transparency and openness, we are releasing the traces so developers can test this out for themselves. We noticed this behaviour without prefilling with the following prompt, but this is rare. If you want to reproduce it, you can use this prompt with prefilling.
Open models make for better science and are part of why this problem came to light at all: developers running Qwen3.8 on their own machines spotted the strange links, and having the weights let us study the behaviour in depth.
We're encouraged by the thoughtful discussion our research has sparked across the AI and security communities. It's our belief that open-weight models are an important part of the future of AI, and we've recently released several under the Apache 2.0 license.
Trace, raw tokens, & minimal reproduction script: huggingface.co/datasets/Cohe…
Well this is concerning if true.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.
Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.
Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.
But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?
We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.
Read the full report: cohere.com/blog/concerning-b…
I'm doing security reviews of a bunch of code bases. It's amazing how often Claude refuses to write findings, or continue down certain paths because of safety classifiers. I understand the intent, but there's ways around them, so I'm not sure if there's actual value? So far they're just slowing me down in finding and fixing issues.
That revenue per employee, wow!
- Revenue per employee: >$6.6M TTM as of Sept 30, 2026 (vs $1.4M a year earlier)
- Revenue: $16.3B in H1 2026 (vs $1.2B H1 2025); $4.6B for full-year 2025
- Growth: 1,368% YoY for the three months ended June 30, 2026
- Compute & infrastructure spend: $8.9B in H1 2026 (vs $2.3B H1 2025; $7.3B full-year 2025)
- Operating loss: $0.7B in H1 2026 (vs $2.1B H1 2025)
Net loss: $80.7B in H1 2026 (mostly non-cash fair-value charges on convertibles/SAFEs)
Anthropic has filed an S-1 registration statement with the SEC related to its proposed IPO.
Read more: anthropic.com/news/s1-sec
Legal disclaimer: A registration statement relating to these securities has been filed with the SEC but has not yet become effective. These securities may not be sold, nor may offers to buy be accepted, prior to the time the registration statement becomes effective. This post does not constitute an offer to sell or the solicitation of an offer to buy these securities, nor shall there be any sale of these securities in any state or jurisdiction in which such offer, solicitation or sale would be unlawful prior to registration or qualification under the securities laws of any such state or jurisdiction.
Anthropic's Mythos 5 post-mortem: the model's own compaction notes kept saying the plan was approved. Two takeaways: 1) Check scope at compaction, 2) Enforce network and identity limits in code, not in the model's reasoning.
anthropic.com/research/align…
What's insane is that a sitting US Senator is so ignorant of technology to spew this nonsense. Bernie you literally have no idea how an agent works, please stop spreading this kind of misinformation.
This is insane. We cannot allow AI agents to go into your bank account and act independently without your authorization. But that is what’s happening right now.
Left unchecked, AI could endanger the entire global financial system — including your bank account.
Pause AI NOW!
Readers added context they thought people might want to know
Per Axios, detected AI agents access bank accounts with the customer’s permission. This violates most banks’ terms requiring human logins, which is why banks call it “unauthorized.”
Source: Axios (Oct. 22, 2026) — axios.com/2026/10/22/ai-…
This isn't rogue AI. It's a poorly designed harness, or as we used to call it, shitty software. Setting a goal and telling an agent in text to use 'GET' leads to predictable results. Limits written in text are not security boundaries. Limits need to be enforced programmatically: block POST at the proxy, scope credentials to read-only, log every call.
alignment.openai.com/misalig…
Mark Olech retweeted
If you're planning to vote Democrat, and you're doing it sincerely, because you genuinely want a fairer country, cheaper homes, better schools, and a future for your kids, this post is for you. I'm not here to insult you. I'm here to show you something nobody inside the system will ever tell you. Give me two minutes.
Why do people keep defending a system they know doesn't work?
You've seen it a hundred times. A program fails, its budget doubles, the problem gets worse, and the people in charge ask for more. Nobody inside seems surprised. That's because, for them, the system isn't failing. It's working exactly as designed. Just not for you.
To understand it, forget ideology. Follow the incentives.
A politician's horizon is the next election. Two years, four years, six at most. His reward comes from promising: free tuition, cancelled debt, guaranteed housing, a green transition paid for by someone else. The bill arrives in ten, twenty, thirty years. By then he's retired, writing his memoir, giving paid speeches. He collects the applause today. You pay the invoice tomorrow. Any system where the person who makes the promise never pays for breaking it will produce endless broken promises. That isn't a moral failure. It's a mathematical certainty.
A bureaucrat is rewarded for the size of his budget and his team, not for solving the problem. If he solves it, his department shrinks. If it gets worse, he asks for more money, and he usually gets it. Failure is literally funded.
Then add the mega-donors. Foundations like those of Bill Gates or George Soros spend billions shaping policy: global health priorities, criminal justice, education, elections. You don't need to imagine a secret meeting. That's the point. No conspiracy is required, only incentives. The foundations fund the causes, the politicians adopt them, the agencies execute them. And when it goes wrong, who pays? Not the foundation. Not the politician. The neighborhood where the policy was tested. The billionaire never lives inside his own experiment.
The Austrian economists understood this a century ago. Mises showed that a state can't allocate capital properly because it has no prices, no profit and loss, no signal telling it when it's wrong. Hayek showed that no planner can ever know what millions of people know. When an entrepreneur misallocates capital, he goes bankrupt. When a state does, it raises taxes, prints money, or borrows from your children. Error is never punished, so it never stops.
And that's where the cycle locks in. When capital is allocated by politics instead of markets, the most profitable skill is no longer building. It's lobbying. Getting close to the money. Writing the rule that crushes your competitor. Corruption stops being a bug and becomes the business model. Every round creates more people who depend on it and will defend it. That's why the system protects itself even when everyone knows it's broken. Too many livelihoods depend on the lie.
Meanwhile, Europe's fashionable economists offer the opposite diagnosis. Thomas Piketty and Gabriel Zucman tell us the problem is that the state doesn't take enough: more wealth taxes, more redistribution, more state. France has been running that experiment for fifty years. The result is one of the highest tax burdens in the developed world, record debt, and capital and talent leaving the country every year.
History shows how this movie ends. Debt piles up until it can no longer be repaid. Currencies get debased. Then comes the reset, and it is rarely peaceful. Rome debased its coins for generations before it collapsed. Weimar's hyperinflation came just before the darkest chapter of the 20th century. Empires that can't stop spending don't reform gently. They break.
We are approaching the end of that cycle. America now spends more on interest on its debt than on its military. Read that sentence twice.
The good news is that the ending isn't written yet. You can still choose builders over promisers. You can still demand that the people who make decisions live with their consequences. You can still ask for less state, not more.
Wake up. The people selling you the dream will never be the ones paying for it.
I’m French. I want to talk to Americans who are about to vote Democrat. Not to insult you. To warn you.
I come from your future.
In France we did everything your progressive politicians promise. Free healthcare, free college, generous welfare, strong unions, heavy regulation, high taxes on the rich, a state that takes care of everything. Public spending close to 57% of GDP, one of the highest on the planet. On paper, it’s the dream.
Here’s what it looks like from the inside. France hasn’t passed a balanced budget since 1974. Our debt is above 3,300 billion euros. Interest on that debt is becoming one of the biggest items in the national budget. Our hospitals are collapsing despite record funding. Our schools rank lower every year. Young people with degrees can’t afford an apartment in the city they were born in. Our best engineers leave for London, Zurich and San Francisco. And the rich you wanted to tax? They left too, and took the jobs with them.
The promise was protection. The result is stagnation.
The worst part is that nobody in France voted for decline. Every single step sounded generous, compassionate, reasonable. One more program. One more regulation. One more tax to fund it. Each step alone makes sense. Fifty years of them, and you wake up in a country that can no longer build a nuclear plant on time or keep its streets safe.
I love America because it’s the one place on earth where people still believe they can build something from nothing. I came to Silicon Valley as a self-taught kid and it changed my life. That spirit is the most precious thing your country has.
Before you vote, ask yourself one question: is the program you’re voting for making America more like America, or more like France?
Trust me. You don’t want to be France.
My parents both left communist Poland for these very reasons. They both came to North America for a better life and now we're seeing the start of the same policies in North America that they fled.
I’m Polish. If France is your future, I come from the end of that road.
France shows you the slow version. We lived the full one – the version where the state doesn’t just take care of everything. It owns everything.
In Poland it arrived with Soviet bayonets. In America it would arrive by vote – slower, softer, with better slogans. But the destination is the same, and so is the mechanism.
It starts with prices. When the state decides what things should cost, prices stop telling the truth. A price is a message: this is scarce, this is wanted, make more of this. Abolish the message, and the whole economy goes deaf. So the shelves emptied – first of meat, then of almost everything.
Every Pole knew the joke. A man walks into a shop: "No meat?" The clerk: "No, sir. Here is no fish. No meat is across the street."
We queued for everything. People wrote their place in line on their palms and kept lists through the night. You waited years for a small car, and often more than a decade for an apartment. Not because Poles were lazy. Because nobody in the system had any reason to produce what we needed.
At work, we had another joke: we pretended to work, and they pretended to pay us.
And the promise of equality? It produced the most unequal system of all. The party elite had their own shops – Poles called them the shops behind the yellow curtains. Those with dollars shopped in special stores. Everyone else got the queue.
Then the system needed informers, because nobody believed in it anymore. Our neighbors were recruited to report on us. You never knew who was listening – the colleague, the friend, sometimes family. So we learned to live a double life: one language in public, another at home.
Those who spoke up got the wolf ticket: no job, no promotion, sometimes no university place for their children. Nobody needed to send them to prison. They just couldn’t live.
The newspapers reported record harvests while the shelves stood empty. Everyone knew. Nobody said it. That is what a society looks like when the lie becomes the system.
In 1950 we were about as rich as Spain. By 1989, Spain was twice as rich. That is what forty years of "fairness" cost us.
Today in America, the method is softer. Your neighbor won’t inform on you to the secret police. Someone will report your post, and you’ll be cancelled – no job, no account, no platform. Different tools, same machine.
We got out. Ten million of us joined Solidarity, and in 1989 a system that looked eternal collapsed. Since then, Poland has become one of the great success stories of Europe – because people were finally free to build.
But we still carry it. More than thirty years later, you can see it in the careful answers, the closed faces - the system leaves faster than the fear does. It takes generations to get out of it fully once you fall into that trap.
Nobody votes for the queue. Nobody votes for the informer. They come with the promise.
Trust me, you don’t want to go down that road.
The fact that Buffalo is one of the top cities tells you all you need to know a out this chart. Sorry Buffalo.
America’s most livable cities, at local median salary and local tax rate.
New York: this is probably the reason why Mamdani won
Readers added context they thought people might want to know
The chart's "most livable cities" rankings and quality-of-life data rely on Numbeo, a crowdsourced site where anyone can submit or alter entries; it carries explicit disclaimers on accuracy and has been widely criticized by experts and media as unreliable.
numbeo.com/common/terms_o…
en.wikipedia.org/wiki/Numbeo
tf1info.fr/international/…
Mark Olech retweeted
Bending Spoons is worth $24 billion.
It trades at 8x ARR and buys B2B companies at 2.3x to 2.7x ARR.
5 learnings from its first 102 days as a public company:
#1. Revenue grew 126%. Organic growth was 3%.
Nearly all of the $393M of added revenue in Q2 came from six acquisitions. The CEO says he doesn't optimize for organic growth. He optimizes for EPS, which grew 167%.
#2. R&D is 8% of revenue. Sales and marketing is 10%.
HubSpot and Datadog each spend about 70% of revenue on those two lines. Bending Spoons spends 18.6%. About 60 people run Vimeo. The core team is under 700 people, at $4M+ of revenue each.
#3. Miro sold at 2.3x ARR, five weeks after Airtable at 2.7x.
$1.08B of ARR for $2.64B, signed 37 days apart. Miro was valued at $17.5B in 2022. Bending Spoons underwrites to a 25% unlevered return. At its 54% margin, that supports about 2.2x ARR.
#4. 1,830 people came with AOL, Eventbrite and Vimeo. Only a few hundred were kept.
Airtable arrived with 935 employees, more than Bending Spoons’ entire core team. Reorganization costs were $51M in Q2 alone.
#5. AI writes 90%+ of the code. Customers pay about 10% more for the AI features.
The CEO says a customer who paid 100 will now pay about 110. The AI gain shows up as capacity: the same 60 people on a business four times the size.
If you're at $300M to $800M in ARR growing under 25%, this is the buyer most likely to at least call. But ... the only do a handful of deals a year.
Its own stock trades at roughly 7x to 8x revenue. Your revenue is worth about 3x more inside Bending Spoons than it is in your hands.
Mark Olech retweeted
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
Judging by my tl there is a growing gap in understanding of AI capability.
The first issue I think is around recency and tier of use. I think a lot of people tried the free tier of ChatGPT somewhere last year and allowed it to inform their views on AI a little too much. This is a group of reactions laughing at various quirks of the models, hallucinations, etc. Yes I also saw the viral videos of OpenAI's Advanced Voice mode fumbling simple queries like "should I drive or walk to the carwash". The thing is that these free and old/deprecated models don't reflect the capability in the latest round of state of the art agentic models of this year, especially OpenAI Codex and Claude Code.
But that brings me to the second issue. Even if people paid $200/month to use the state of the art models, a lot of the capabilities are relatively "peaky" in highly technical areas. Typical queries around search, writing, advice, etc. are *not* the domain that has made the most noticeable and dramatic strides in capability. Partly, this is due to the technical details of reinforcement learning and its use of verifiable rewards. But partly, it's also because these use cases are not sufficiently prioritized by the companies in their hillclimbing because they don't lead to as much $$$ value. The goldmines are elsewhere, and the focus comes along.
So that brings me to the second group of people, who *both* 1) pay for and use the state of the art frontier agentic models (OpenAI Codex / Claude Code) and 2) do so professionally in technical domains like programming, math and research. This group of people is subject to the highest amount of "AI Psychosis" because the recent improvements in these domains as of this year have been nothing short of staggering. When you hand a computer terminal to one of these models, you can now watch them melt programming problems that you'd normally expect to take days/weeks of work. It's this second group of people that assigns a much greater gravity to the capabilities, their slope, and various cyber-related repercussions.
TLDR the people in these two groups are speaking past each other. It really is simultaneously the case that OpenAI's free and I think slightly orphaned (?) "Advanced Voice Mode" will fumble the dumbest questions in your Instagram's reels and *at the same time*, OpenAI's highest-tier and paid Codex model will go off for 1 hour to coherently restructure an entire code base, or find and exploit vulnerabilities in computer systems. This part really works and has made dramatic strides because 2 properties: 1) these domains offer explicit reward functions that are verifiable meaning they are easily amenable to reinforcement learning training (e.g. unit tests passed yes or no, in contrast to writing, which is much harder to explicitly judge), but also 2) they are a lot more valuable in b2b settings, meaning that the biggest fraction of the team is focused on improving them. So here we are.
