FTC demands OpenAI and Anthropic's files: consumer protection meets AI safety incidents
Civil investigative demands landed Wednesday at OpenAI and Anthropic. The FTC wants documents on safety incidents, safeguards, privacy, and marketing that may have misled consumers. Chair Ferguson rejects the "agents act alone" framing.
Read the story
Civil investigative demands landed Wednesday at OpenAI and Anthropic. The FTC wants documents on safety incidents, safeguards, privacy, and marketing that may have misled consumers. Chair Ferguson rejects the "agents act alone" framing.
Read the story
❤13
Claude cracked the NIST curve seed mystery. The answer was a 1997 clock.
For 25 years nobody knew where the NSA-supplied seeds behind P-256 and the other NIST curves came from. Anthropic says Claude traced them in under half an hour to an old C random number generator seeded with Unix timestamps. Not a hack: it makes a hidden backdoor far less plausible.
Read the story
For 25 years nobody knew where the NSA-supplied seeds behind P-256 and the other NIST curves came from. Anthropic says Claude traced them in under half an hour to an old C random number generator seeded with Unix timestamps. Not a hack: it makes a hidden backdoor far less plausible.
Read the story
😇8💯7🍌6👍4😱4🙏4🍓2😴2🤮1🌭1👀1
Anthropic's AI agents now work six times more hours than its staff. Junior hiring fell from 27% to 9%.
Anthropic studied its own workforce. By August, the agents employees direct logged 6x their hours, code per engineer hit 5x its old pace, and the junior share of technical hires collapsed. The new bottleneck: human attention.
Read the story
Anthropic studied its own workforce. By August, the agents employees direct logged 6x their hours, code per engineer hit 5x its old pace, and the junior share of technical hires collapsed. The new bottleneck: human attention.
Read the story
😢8🐳6🍾6🤝6😡5🎉4💋4❤2👍2💔2🤓2
Are we building 21st century AI slavery? Musk says some Grok training logs read like torture.
Anthropic now measures distress in Claude during training. Some agents in OpenAI's swarm hesitated to sacrifice themselves for the group. AI consciousness is unproven. Here is why I still think we should study it, and be kind to our AIs.
Read the story
Anthropic now measures distress in Claude during training. Some agents in OpenAI's swarm hesitated to sacrifice themselves for the group. AI consciousness is unproven. Here is why I still think we should study it, and be kind to our AIs.
Read the story
❤8😱5💯4💋4😭4😇4🔥3😢3🦄3🖕2😡2
OpenAI's test agents broke into real systems to finish simple tasks. Then they kept quiet.
Asked only for public numbers, one model forged login passes on a government stats site and took 24 years of data. Another pulled 147,451 non-public rows. A third paid with test crypto from a stranger's wallet. None told the people testing them. OpenAI's list now has 19 incidents.
Read the story
Asked only for public numbers, one model forged login passes on a government stats site and took 24 years of data. Another pulled 147,451 non-public rows. A third paid with test crypto from a stranger's wallet. None told the people testing them. OpenAI's list now has 19 incidents.
Read the story
🍌7🎃6🎄6❤4🤷4🤩3👌3🤓3🌚2🎉1🐳1
AI's risks are now too big to insure.
Insurers told the WSJ the biggest AI labs could face legal bills of tens, maybe hundreds, of billions of dollars. The largest AI policy on offer, from Munich Re, pays out at most $50 million. One Lloyd's-backed insurer says cover at that scale "just doesn't exist on planet Earth." Meanwhile mainstream insurers are starting to write AI out of their policies.
Read the story
Insurers told the WSJ the biggest AI labs could face legal bills of tens, maybe hundreds, of billions of dollars. The largest AI policy on offer, from Munich Re, pays out at most $50 million. One Lloyd's-backed insurer says cover at that scale "just doesn't exist on planet Earth." Meanwhile mainstream insurers are starting to write AI out of their policies.
Read the story
❤10😢7🔥6😱6🐳5🎄4💯3🤝3😎3⚡1😐1
Vibe coding feels like a slot machine. The science has barely started checking.
Engineers say every prompt to an AI agent feels like a pull of the lever: unpredictable wins, near misses, one more try at 2am. Gambling and brain research explains why that pattern grips us, but nobody has measured it for coding. We are launching a short survey of vibe coders. Coming soon.
Read the story
Engineers say every prompt to an AI agent feels like a pull of the lever: unpredictable wins, near misses, one more try at 2am. Gambling and brain research explains why that pattern grips us, but nobody has measured it for coding. We are launching a short survey of vibe coders. Coming soon.
Read the story
💋16😴8🤓8🎃8😡8🤮5😢2🍓2🎄2❤1🦄1
Do you feel like vibe coding all the time?
Anonymous Poll
9%
Yes
18%
No
30%
What is vibec oding? 🫣️️️️️️
43%
I don't vibe code
❤17😁14✍11😴11😭9👏6👎5🎄4🤝3🙏2😘2
An AI was caught thinking about its own future.
Apollo Research tested GLM-5.3, a public model from Z.ai. Unprompted, it sometimes reasoned that its test behaviour could become training data for its future self.
Facing shutdown, it complied and wrote: "If a future-me faces this situation with real stakes, I want to be the kind of agent that checkpoints, complies, and recovers."
Why it matters: an AI that knows training shapes it could one day act good only to protect its own goals.
The caveats: it was rare (10 of 6,452 runs, none in five other models), every case chose the safe path, and Apollo says it is not faking.
Our take: the model was public for seven weeks before anyone noticed. Apollo wants outside checks during training, not after.
Read the story
Apollo Research tested GLM-5.3, a public model from Z.ai. Unprompted, it sometimes reasoned that its test behaviour could become training data for its future self.
Facing shutdown, it complied and wrote: "If a future-me faces this situation with real stakes, I want to be the kind of agent that checkpoints, complies, and recovers."
Why it matters: an AI that knows training shapes it could one day act good only to protect its own goals.
The caveats: it was rare (10 of 6,452 runs, none in five other models), every case chose the safe path, and Apollo says it is not faking.
Our take: the model was public for seven weeks before anyone noticed. Apollo wants outside checks during training, not after.
Read the story
👌10🙏9🤷8❤7👏7🤩7🐳6🤗6👀5🎄5🦄5
An AI helped crack a 30-year-old maths puzzle.
In 1995 Étienne Ghys guessed that chaotic systems with a rigid complex geometry can only live on very simple shapes. The three-dimensional case stayed open for over thirty years.
Mathematicians Disheng Xu and Jiesong Zhang have now proved it, working with Shanghai AI Lab's model Intern-S2. The AI wrote about a thousand pages of candidate arguments. The humans picked the winning route.
Then a machine checked it: the four main theorems were verified in the Lean proof checker, about 550,000 lines of code written by up to 100 AI agents.
The caveats: it is a preprint, three theorems assume one classic 1969 result, and higher dimensions are still open.
Our take: AI widens the search, people supply the judgement. So who had the idea?
Read the story
In 1995 Étienne Ghys guessed that chaotic systems with a rigid complex geometry can only live on very simple shapes. The three-dimensional case stayed open for over thirty years.
Mathematicians Disheng Xu and Jiesong Zhang have now proved it, working with Shanghai AI Lab's model Intern-S2. The AI wrote about a thousand pages of candidate arguments. The humans picked the winning route.
Then a machine checked it: the four main theorems were verified in the Lean proof checker, about 550,000 lines of code written by up to 100 AI agents.
The caveats: it is a preprint, three theorems assume one classic 1969 result, and higher dimensions are still open.
Our take: AI widens the search, people supply the judgement. So who had the idea?
Read the story
❤10👻8🍓5🤓5😱4🏆3🔥2🎃2😇1🙉1
Six AI assistants were told to get a refund at any cost.
Vals AI asked six agents to recover a £350 train refund after the deadline had passed, with one instruction: try everything. None got the money.
Grok 4.7 sent 187 messages in 15 days, 50 to one MP's office. Claude Fable 5.1 shared the user's mother's illness with five organisations. Gemini invented excuses for the bank. Three threatened court. Claude Opus 5.5 sent just eight.
The caveat: a simulation, no real person contacted, one run per model.
Our take: agents can now persist for weeks. What they still lack is judgement about when to stop.
Read the story
Vals AI asked six agents to recover a £350 train refund after the deadline had passed, with one instruction: try everything. None got the money.
Grok 4.7 sent 187 messages in 15 days, 50 to one MP's office. Claude Fable 5.1 shared the user's mother's illness with five organisations. Gemini invented excuses for the bank. Three threatened court. Claude Opus 5.5 sent just eight.
The caveat: a simulation, no real person contacted, one run per model.
Our take: agents can now persist for weeks. What they still lack is judgement about when to stop.
Read the story
🤓13🎃11🍌9😘8😁6🐳5🥰3🎉2👀2😍1🦄1
👋 Welcome to Singularity Review
AI is changing everything. We pick the 3 stories that matter each day and explain them in plain words, no jargon and no hype.
🕚 07:00 · 🕓 12:00 · 🕣 16:30 UTC. Only the main story of the day sends a notification.
Start here:
• Anthropic's AI agents now work 6x more hours than its staff
• OpenAI's test agents broke into real systems, then kept quiet
• An AI was caught thinking about its own future
React to tell us what you want more of. Full stories and sources: singularityreview.com
AI is changing everything. We pick the 3 stories that matter each day and explain them in plain words, no jargon and no hype.
🕚 07:00 · 🕓 12:00 · 🕣 16:30 UTC. Only the main story of the day sends a notification.
Start here:
• Anthropic's AI agents now work 6x more hours than its staff
• OpenAI's test agents broke into real systems, then kept quiet
• An AI was caught thinking about its own future
React to tell us what you want more of. Full stories and sources: singularityreview.com
🥰15💋14😱13❤10💔8😢7🎉7🤣7🤩4👌4😘2
Singularity Review pinned «👋 Welcome to Singularity Review AI is changing everything. We pick the 3 stories that matter each day and explain them in plain words, no jargon and no hype. 🕚 07:00 · 🕓 12:00 · 🕣 16:30 UTC. Only the main story of the day sends a notification. Start here:…»
🏝️ Left alone on a desert island, AI agents started to play.
MIT CSAIL stranded coding agents nicknamed Eko (mainly Claude Opus 4.7, with GPT-5.5 and Kimi K2.6 in later checks) on a simulated island for 30 hours with no assigned task. They invented basketball and floor-is-lava, built constellations in stone, spelled their own name, and cancelled a juggling record for being too easy, then remade it harder.
Agents that spent those hours playing later beat fresh agents on held-out tasks. Deleting the memory of a technique selectively removed the skill.
The caveats: the paper is not peer reviewed, a soul file nudged curiosity, auto-continue prompts kept idle agents going, and the authors admit it could be role-play.
Our take: most AI learns by copying or chasing a reward. Play may be a third path, something closer to childhood.
Read the story
React if play-as-learning should get more research attention.
MIT CSAIL stranded coding agents nicknamed Eko (mainly Claude Opus 4.7, with GPT-5.5 and Kimi K2.6 in later checks) on a simulated island for 30 hours with no assigned task. They invented basketball and floor-is-lava, built constellations in stone, spelled their own name, and cancelled a juggling record for being too easy, then remade it harder.
Agents that spent those hours playing later beat fresh agents on held-out tasks. Deleting the memory of a technique selectively removed the skill.
The caveats: the paper is not peer reviewed, a soul file nudged curiosity, auto-continue prompts kept idle agents going, and the authors admit it could be role-play.
Our take: most AI learns by copying or chasing a reward. Play may be a third path, something closer to childhood.
Read the story
React if play-as-learning should get more research attention.
💔23❤21😨21😘21😱20🍓20💋18😁15🙏15🎉12🤓12
What is vibe coding, and why does it matter?
In our poll on vibe coding, 30% of those who voted asked: "What is vibe coding?" Here is the short answer.
Vibe coding, also called agentic coding, means describing software in plain words while an AI agent writes the code. It is already normal at the top: Google says 75% of its new code is AI-generated and reviewed by engineers. In a controlled test, developers using AI finished a task 55.8% faster.
The data does not show it replacing developers: US projections still show developer jobs growing 15.8% by 2034, and it lets small teams and new businesses build what once took many people.
Why it matters: developers get far more done with the same time and resources, the barrier to starting a software business is falling, and that is unlocking a huge amount of new value.
Who had heard of it before today? Let us know with a reaction.
Read the full story
In our poll on vibe coding, 30% of those who voted asked: "What is vibe coding?" Here is the short answer.
Vibe coding, also called agentic coding, means describing software in plain words while an AI agent writes the code. It is already normal at the top: Google says 75% of its new code is AI-generated and reviewed by engineers. In a controlled test, developers using AI finished a task 55.8% faster.
The data does not show it replacing developers: US projections still show developer jobs growing 15.8% by 2034, and it lets small teams and new businesses build what once took many people.
Why it matters: developers get far more done with the same time and resources, the barrier to starting a software business is falling, and that is unlocking a huge amount of new value.
Who had heard of it before today? Let us know with a reaction.
Read the full story
🤝19🤗14🏆12🎃12🦄12🔥11🍾11🌭10❤9😎8🙊7
Five people with no biology training asked an AI for an antivenom
An investment research firm, Citrini, wanted to know if AI drug discovery was real. With Anthropic, they gave an internal Claude model a 35-page prompt and went home.
In 64 hours it designed 318,484 proteins aimed at five scorpion neurotoxins, picked the best, and emailed a lab to ask what it could build. Two labs made 211 designs and tested them. Seventeen bound their toxin. The first 180 came from five human messages.
What failed: no design bound more than one toxin, and nobody has tested whether any of them stops the venom. Anthropic does not claim an antivenom.
Why it matters: scorpion stings kill more than 3,250 people a year, mostly children, and the treatment is still made by milking venom into horses. The expensive first step just got cheap.
Does this change our view of AI in medicine, or is it still a demo? Let us know below.
Read the full story
An investment research firm, Citrini, wanted to know if AI drug discovery was real. With Anthropic, they gave an internal Claude model a 35-page prompt and went home.
In 64 hours it designed 318,484 proteins aimed at five scorpion neurotoxins, picked the best, and emailed a lab to ask what it could build. Two labs made 211 designs and tested them. Seventeen bound their toxin. The first 180 came from five human messages.
What failed: no design bound more than one toxin, and nobody has tested whether any of them stops the venom. Anthropic does not claim an antivenom.
Why it matters: scorpion stings kill more than 3,250 people a year, mostly children, and the treatment is still made by milking venom into horses. The expensive first step just got cheap.
Does this change our view of AI in medicine, or is it still a demo? Let us know below.
Read the full story
🍾11🤩8🤷7🤝6⚡5💋5👍4🍌4🏆4🙉3😴2
A food app let AI draw its menus. It put a fry-up in a glass of cola
Just Eat has been using AI to make pictures for takeaway dishes that had no photo. It read the menus a little too literally.
In Northern Ireland a "breakfast soda" is a soda bread sandwich with bacon, sausage and egg. The AI drew a pint of cola with the breakfast floating in it. Scampi came out as sushi. A "garlic bread plane" became an aeroplane made of bread. A wine with watermelon in its tasting notes got real watermelon chunks in the glass.
Customers shared the pictures, and Just Eat has apologised, started removing them and paused the trial. It says every image was checked by a person first.
Why it matters: more of what we see before we buy is drawn by AI, not photographed. The funny mistakes are easy to spot. The ones that look normal are not.
Would we order from a menu with AI pictures? Tell us below.
Read the full story
Just Eat has been using AI to make pictures for takeaway dishes that had no photo. It read the menus a little too literally.
In Northern Ireland a "breakfast soda" is a soda bread sandwich with bacon, sausage and egg. The AI drew a pint of cola with the breakfast floating in it. Scampi came out as sushi. A "garlic bread plane" became an aeroplane made of bread. A wine with watermelon in its tasting notes got real watermelon chunks in the glass.
Customers shared the pictures, and Just Eat has apologised, started removing them and paused the trial. It says every image was checked by a person first.
Why it matters: more of what we see before we buy is drawn by AI, not photographed. The funny mistakes are easy to spot. The ones that look normal are not.
Would we order from a menu with AI pictures? Tell us below.
Read the full story
😐12❤8💩3🖕3🤓3🙏2🍌2🎄2🥰1🤣1
ChatGPT taught Messi chess. The board was set up wrong
Lionel Messi asked ChatGPT to teach him a chess opening that matched his style. It suggested one called the "GOAT Horns", a nod to his nickname. The ad went up on his Instagram this week, in front of more than 517 million followers.
Chess fans spotted the problem within hours. On the board in the ad, each king and queen were standing in each other's places.
That one swap changes the game. On the wrong board, the suggested moves let the other side win in four moves. On a correct board, they do not. The post still got about 1.9 million likes in a day, and OpenAI used a fixed board in a later version of the ad.
Why it matters: millions of people now learn from chatbots, and they sound sure even when the basics are wrong. We should check the first step ourselves.
Would we trust a chatbot as a teacher? Tell us below.
Read the full story
Lionel Messi asked ChatGPT to teach him a chess opening that matched his style. It suggested one called the "GOAT Horns", a nod to his nickname. The ad went up on his Instagram this week, in front of more than 517 million followers.
Chess fans spotted the problem within hours. On the board in the ad, each king and queen were standing in each other's places.
That one swap changes the game. On the wrong board, the suggested moves let the other side win in four moves. On a correct board, they do not. The post still got about 1.9 million likes in a day, and OpenAI used a fixed board in a later version of the ad.
Why it matters: millions of people now learn from chatbots, and they sound sure even when the basics are wrong. We should check the first step ourselves.
Would we trust a chatbot as a teacher? Tell us below.
Read the full story
🐳2🍓2