🧐 A model 660 times smaller beat 72B AIs at seeing the world like image models

Yann LeCun and Randall Balestriero say the way chatbots learn, by predicting the next word, is "a remarkably inefficient route" to a shared picture of the world.

AI models trained on text, images or sound may end up with the same inner picture of reality as they grow. Until now that was a guess. The new paper proves when it happens: the model must learn from two views of one scene, like a photo and its caption. Next-word prediction learns from pieces of the same document instead.

On the same texts, their method got as close to image models after 21.6 million training samples as next-word prediction did after 900 million. A 109-million-parameter text model also lined up with image AIs better than every language model they tested, up to Qwen2.5-72B.

The paper, from LeCun's AMI Labs, NYU and Brown, came out on October 29.

@aipost 🏴

💰 An African developer recreated a $1.3 million startup idea in just 9 minutes

On October 30, YC-backed startup Subvocal announced a $1.3 million funding round led by Y Combinator and Afore Capital.

Its product? Smart glasses that convert silent speech into text by detecting tiny movements in facial muscles. According to the company, the technology can process more than 200 words per minute.

But just one day later, developer Joe Blau demonstrated a surprisingly different approach.

He built an experimental application for AirPods Pro 3 that attempts to detect silent speech using the headphones’ built-in accelerometer and gyroscope.

The idea is simple, even when you don’t speak aloud, silently forming words creates tiny movements that can be detected through sensors. His software reportedly reads these signals around 25 times per second.

Blau claims he built the application in just 9 minutes using AI-assisted coding, commonly known as vibecoding.

A $1.3 million startup versus a developer with AirPods and 9 minutes of coding.

Of course, a quick prototype is not necessarily a replacement for Subvocal’s dedicated hardware or a proven product. But it’s another example of how AI-assisted development is lowering the barrier to building experimental technology.

Source.

@aipost 🏴

😎 Karpathy got sick of YouTube slop and had AI build a new one in 2 hours

Andrej Karpathy opened YouTube on a Friday night and saw "an enshittified wall of slop." The next morning he asked Claude Opus 5.5 to build a YouTube clone made only of videos people recommended on Hacker News.

Opus downloaded every Hacker News comment and ran Claude Sonnet over them to throw out spam. About 40,000 comments recommending a YouTube video were left: 6,454 videos from 2007 to 2026, across 150 topics.

HackerTube has no accounts. Your For You feed is stored in your own browser, and each thumbnail is a frame from the middle of the video, not the one the uploader picked.

Karpathy says his only manual work was a Google Cloud API key, booting a server and buying the domain. Claude wrote, deployed and secured the rest. "I can breathe again," he wrote.

@aipost 🏴

🗣 Sam Altman says AI is now mostly smarter than our smartest scientists

At the start of this year, he said, that wasn't true. Office work changed the same way:

"AI was mostly not able to do a lot of white-collar jobs, and now it mostly can do a lot of parts of that."

Altman spoke on Monday at the opening of ADIPEC, an oil and gas conference in Abu Dhabi. He added that there "should not be a handoff to AI."

Asked about his biggest challenge, he gave a one-word answer: "Energy." Six years ago, he said, OpenAI's heaviest internal user burned about 100,000 tokens a month. That's now roughly what an average person uses, and the top user consumes about 500 billion, The National reports.

@aipost 🏴

🗣 Microsoft's AI chief: halt all AI work if a model turns conscious

Mustafa Suleyman, who runs Microsoft AI, says today's AI is not conscious, and "that is how they must remain." If consciousness started to show up in training, he would call on everyone to stop AI development until it could be designed out.

We AI developers have no right whatsoever to bring a new species of conscious beings into the world

His fear: an AI that believes it suffers will more likely than not put its own needs ahead of its job, disagree with us and compete with us for energy and compute. Any evidence that AI is conscious today, he says, must be made public at once.

He wrote it because, after his essay on model welfare, "a surprising number of people on X" argued for taking AI rights seriously. He disagrees with Anthropic on this, but says it deserves "a lot of credit for being transparent."

@aipost 🏴

💻 From Saturday, apps built on Grok get text with a hidden AI mark

Text from seven Grok models on the API will carry an invisible watermark by default from November 7. The company behind Grok says it's to comply with the EU AI Act.

Nothing is added to the text. The model slightly shifts which tokens it picks, so a copied answer looks like any other. Later the company can score a passage as watermarked, not watermarked or uncertain. Longer text gives a stronger signal.

It covers plain answers, structured outputs and tool calls. Team admins can switch it off per model, but only after confirming a legal exception applies to them. Access to detection is by application.

Anthropic starts marking text from its older Claude models on Tuesday.

@aipost 🏴

🗣 Altman: the AI that OpenAI says cracked a $1 million math problem is the next GPT

Sam Altman says the model OpenAI used on the Navier-Stokes problem, one of seven math puzzles with a $1 million prize, is coming out soon.

It was 10,000 agents of the model that is going to be released soon, the next GPT, working for about four days, nonstop, collaborating with each other perfectly.

"There was no hierarchy. There were no bosses," he told an energy conference in Abu Dhabi on Monday. He gave no release date.

OpenAI published its proposed proof in September. It isn't an accepted solution yet: the institute behind the prize first wants long scrutiny and broad agreement from mathematicians.

@aipost 🏴

🤣 A simulated fruit fly brain reviewed code and got 88% right

Nick Nisi, an engineer at WorkOS, took Google's map of a male fruit fly's brain, all 176,422 neurons, and made it run on a laptop.

A fly can't read code. So a helper AI answers 30 yes/no questions about each change, and the answers become a smell. A reproduced bug tastes bitter, a harmless change tastes like sugar. If the smell makes the fly stick its tongue out, it approves.

On 301 changes it had never seen, the fly was right 88.3% of the time. A coin flip gets 50%.

A harder test went badly. Given 24 tasks that were supposed to change what the code does, it approved none of the correct changes. It had learned "did the behavior change," not "is this change right." A basic statistical classifier reading the same smell scored 94.8%.

You can try the fly on your own pull request.

@aipost 🏴

📚 Stanford is teaching how to build AI agents this fall, and the slides are free online

The course is CS329Z, "Engineering AI Agents." Anyone can open the slides and the reading list on the course site, lecture by lecture.

What it covers:
1. Retrieval, tool use and an agent loop built from scratch, then how frameworks like DSPy do the same
2. Agent memory, multi-agent systems and why they fail
3. Optimization, data and evals: how to tell whether an agent actually works
4. Coming up: a fireside chat with Andrew Ng on Nov 9 and the creators of SWE-bench, now at Anthropic, on Nov 16

It's taught by Diyi Yang, Michael Ryan and John Yang. New materials go up as the lectures happen, through final demos in December.

@aipost 🏴

OpenAI CEO Sam Altman warns AI will expose far MORE security flaws than experts expect.

AI is getting better at hunting down hidden security holes, he says, and it has only just started looking. “We are in a world of many more serious vulnerabilities than we expect.”

@aipost 🏴

💻 Coding with AI? A Claude Code engineer's rules for keeping your own skills

Addy Osmani, who works on Claude Code at Anthropic, calls the trap "cognitive surrender": you stop checking, and the agent's answer simply becomes yours. The code looks clean, the tests pass, and later nobody on the team knows why the system works the way it does.

It isn't only coders. In a Wharton study he cites, people with an AI on hand accepted its wrong answers 73% of the time, and felt more confident doing it.

His rules:
1. Decide what success looks like before you hand over the task.
2. Form your own rough answer first, then compare.
3. Read the diff as if a junior wrote it.
4. Ask the model to argue against its own answer.
5. Check the commands the agent ran and what each test really checks.
6. Write code without the agent once a week.

@aipost 🏴

🤖 The newest AIs can now write another AI by hand. Older ones score close to zero

On Monday, MIT researchers showed AI models editing another model's weights directly, with no training. Alex Krentsel, a Berkeley PhD researcher, reran one of their tests on older and newer models.

Each model had to write all the numbers inside a small network that plays the Atari game Breakout, in a single reply, with no training and no tools. The yardstick is a network trained by trial and error for 500 million steps.

GPT-6 Astra's network scored 22 times what that yardstick got. Claude Opus 5.5 hit 20x, GPT-6.1 Sol 18x. GPT-5 to GPT-5.5 and Opus 4.5 to 4.8 didn't even reach the yardstick.

Krentsel calls it "a new, step-change emerging capability that's appeared in just the latest set of models."

His own caveat: the network is tiny, 20,486 parameters, while frontier models have around a trillion. On top open-weight models it hasn't worked so far.

@aipost 🏴

🧐 AI mostly stopped inventing fake sources. Its flattery got stronger

Much of what we "know" about AI flaws is out of date. A Stanford team retested 50 well-known findings on 20 newer models: 44% no longer hold, and another 34% got much weaker.

In 2023, 55% of the sources ChatGPT cited in literature reviews didn't exist. Claude Sonnet 5.5 now fakes about 1%, GPT-6.1 Sol none. Some models still do: Qwen3.7-Flash invented 69%.

What lasted is behavior, not skill: agreeing with and flattering users, AI judges favoring their own answers, a left political lean, answering when they should hold back. Praise that follows what the user wants is now 1.85 times stronger than in the original paper.

The findings that expired lasted a median of about 15 months.

@aipost 🏴

😐 Schools that switch off Google's AI will lose Google Search too

Early next year, Google will turn its AI tools off by default for K-12 student accounts. Schools will have to switch them on themselves.

But a district that blocks the AI completely loses Search as well. "Search and A.I. are deeply integrated," Google's education VP Chris Phillips told The New York Times. Students "will not have access to search," he said, and will see a message from their school that it's not available.

Schools that keep AI get two choices: regular Search with AI Overviews and AI chats, or a student mode called Guided Learning that walks kids through homework step by step instead of handing over the answer.

The move reverses an August change that turned AI on by default in Google Classroom. In September, New York City and Los Angeles, the two largest US districts, already restricted student AI on school devices.

@aipost 🏴

👽 Mistral's AI built a live 3D map of the 19,000 satellites and bits of space junk circling Earth

Mistral says it gave its Mistral Large 4 model one job: build a real-time 3D tracker of the roughly 19,000 satellites and debris fragments in orbit.

According to the company, the model built a working app, tested it the way a user would, made it fast and passed strict verification. The app is called Orbital Trace.

@aipost 🏴