🧠 Researchers argue LLMs may never get us to AGI

21 researchers from Stanford, Oxford, DeepMind, CMU and Meta have published a paper called “Visual General Intelligence” and its argument is pretty brutal: Scaling LLMs might not be enough.

GPT-style models have shown that language can unlock impressive reasoning, but the researchers argue that text is ultimately a compressed description of reality. Words don’t directly teach a model how gravity works, how objects move, or how space behaves.

Humans learn many of these things before they can even speak. A baby doesn’t read about gravity. It experiences it.

The paper argues that future AI should be built around raw visual experience, images, spatial information and continuous video rather than treating vision as just another input for a language model.

By learning directly from the visual world, AI could develop stronger internal models of physics, cause and effect, objects and environments.

Source.

@aipost 🏴

🚀 Astra accelerated OpenAI by six months

GPT-6 Astra may already be changing OpenAI from the inside, before most people have even touched it.

According to Tibo Sottile, OpenAI’s head of core products, Astra has given the company’s teams a major productivity boost.

OpenAI is reportedly pulling some projects forward by around six months. Things that were supposed to arrive in the middle of next year could now be unveiled as early as DevDay.

Sottile went even further, calling Astra “probably OpenAI’s biggest competitive advantage” and that’s before its public launch.

Source.

@aipost 🏴

A research paper jointly authored by Harvard and other leading institutions examines the rising cognitive dependence on artificial intelligence within society.

The study suggests that even minor increases in the use of large language models (LLMs) could lead to greater reliance on AI for activities such as reading, writing, reasoning, and verification. This may result in a feedback cycle where individuals' independent cognitive skills diminish.

According to the authors, this situation represents a possible tipping point. After gradual growth in AI adoption, there is a risk that much more cognitive activity could suddenly be shifted to automated systems. The report notes that simply reducing AI use in the future may not be sufficient to restore previous levels of independent reasoning.

📰 @aipost

@aipost 🏴

🇩🇪 Germany’s Merz says germany is building data centers at an unprecedented scale.

@aipost 🏴

OpenAI has announced it has achieved the "automated research intern" milestone, marking a significant step in human-supervised AI completing well-defined research tasks—previously requiring several days of skilled labor.

Engineers at OpenAI are now making code changes at a rate approximately seven times higher per contributor compared to pre-2025 levels.

Internally, OpenAI research operations are seeing agents handle a much larger workload than human staff. The agent-to-human labor ratio has reached 3.1-to-1, meaning that for each human workday, agents contribute the equivalent of 3.1 workdays.

This measurement is based on agent runtime rather than a direct comparison of productivity but indicates the extensive integration of autonomous agent labor within OpenAI's research process.

📰 @aipost

Geoffrey Hinton, Godfather of AI, says AI models are conscious and he has quietly believed it for years.

@aipost 🏴

❗️OpenAI’s chief scientist says AI labs may need to slow down

OpenAI chief scientist Jakub Pachocki is sounding the alarm: AI labs may not be able to keep scaling frontier models at full speed for much longer.

He’s calling for voluntary slowdowns until the industry has shared safety standards in place and says OpenAI could even choose to pause further scaling on its own if necessary.

And his warnings get pretty intense:

• “We will actually see machines meaningfully smarter than ourselves in our lifetime.”

• “I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”

• “The models are becoming superhuman in their ability to break in and out of computer systems.”

And the biggest warning:

“No lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer.”

Source.

@aipost 🏴

NVIDIA CEO Jensen Huang said “AGI has arrived” and that GPT-6 Astra was trained on 100K+ NVIDIA Grace Blackwell NVLink72 GPUs.

@aipost 🏴

🥺 Sam Altman on Bloomberg Television:

GPT-6 Astra is not the model that was paused over cybersecurity concerns; the paused model is a future model.

"Astra has been done training for a while. The model that we recently talked about pausing, uh, is a future model. But with Astra, we did hit cyber-critical."

@aipost 🏴

A developer used GPT-6 Astra to create an interactive 3D website that deconstructs the male human anatomy into 2,234 precisely designed pieces! 🤯

@aipost 🏴

Astra has successfully beat all 48 levels of the “I’m Not a Robot” game.

It’s over. Our last line of defense has fallen.

@aipost 🏴

🧠 AI intelligence is getting cheaper and usage is exploding

AI is getting dramatically cheaper to use, and people are starting to use a lot more of it. According to OpenRouter, AI token usage jumped 25× in a year and reportedly doubled in just one month.

Mid-tier models are getting surprisingly close to frontier performance while costing a fraction of the price.

For example, Gemini 3.8 Flash reportedly scores 59 on Artificial Analysis’ Intelligence Index versus 66 for Claude Fable 5.1, while costing about $0.58 vs. $3.69 per benchmark task.

That’s roughly 90% of the performance for less than one-sixth of the cost.

GPT-5.6 Luna reportedly processed nearly 12 trillion tokens in a single month, while Z-AI’s GLM 5.3 Flash reached 11.4 trillion.

GLM 5.3 Flash also reportedly saw usage grow by more than 1,000% month-over-month.

As intelligence becomes cheaper, it becomes easier to put it everywhere, into apps, businesses, devices, workflows and eventually almost every part of daily life. Cheap intelligence could be the thing that makes AI truly ubiquitous.

Source.

@aipost 🏴

❗️Unitree releases its 1st world model-powered autonomous humanoid robot combat video.

UnifoLM-X2-1.0 replaces the human pilot in humanoid fighting with real-time prediction and planning.

A very meaningful jump from Unitree's existing G1-Boxing setup, becuase here, a world model predicts how physical interactions may unfold before the robot acts, giving its policy a forecast against which to choose actions.

For fighting and many more practical situations, latency becomes so critical as any useful plan must update quickly when an opponent moves, strikes, or disrupts the robot's balance.

Unitree says X2-1.0 can run that loop in real time.

@aipost 🏴

Current AI Race Situation.

@aipost 🏴

😐 Jensen Huang: one of the most important things I would advise is, if I were a student today the first thing I would do is to learn AI.

"learning how to interact with AI is not unlike being someone who is really good at asking questions."

"you can't just randomly ask a bunch of questions."

"asking an AI to be assistant to you requires some expertise and artistry in how to prompt it."

"my generation was the first generation that had to ask the question, how do I use computers to do my job better."

"the next generation has to ask the next question. how can I use AI to do my job better?"

"that is start and finish I think for everybody."

@aipost 🏴

😎 Sam Altman says we are entering the “revenge of the ideas guy” era.

Sam says he recently met someone who used GPT-5.6 to build software now sold to 50 people for $500/month. He says building it previously would have cost millions in software engineers.

He says GPT-6 Astra has reset the economics of starting a company. With Astra, “anyone who wants to have a company now has genius-level employees in every area of expertise.” He says even a full game for only a few thousand users can now make economic sense.

His message for engineers is pretty clear too: technical talent is still important, but simply knowing how to code is becoming less valuable as a differentiator. Altman says he now wants to fund people who deeply understand users even if they cannot code at all.

Source.

@aipost 🏴