⚠️ Researchers are opening the black box of AI understanding

Mechanistic interpretability is giving researchers a closer look at what happens inside large language models and how they develop internal representations that go beyond simple word associations.

Studies have found that models can develop internal features representing concepts, track relationships between facts, and sometimes discover general rules that let them solve problems they haven’t seen before.

One intriguing phenomenon is called “grokking”: during training, a model can suddenly shift from memorizing examples to generalizing from underlying patterns. In some experiments, performance on new problems improves dramatically even after a long period of apparent failure.

This suggests that neural networks can discover compact rules from data without anyone explicitly programming those rules into them.

These findings don’t prove that AI understands reality the same way humans do, or that models simply erase memorized facts and replace them with a complete world model.

Source.

@aipost 🏴

This is probably evil but LMAO 🤣🤣

@aipost 🏴

🗣"Labor and capital are about to divorce because capital doesn't does need your labor, it'll just hire GPUs. So you need to own the means of production."

~ Emad Mostaque, founder of Stability AI:

"In 2 years' time, if your job is on the other side of a screen, you will be the dumbest person on the team that drags down everyone else because all your colleagues will be AIs. And so the value of your cognitive labor is negative. They would pay you not to be a part of the team. As of this week, we are no longer the smartest things on this planet. "

@aipost 🏴

⚡️ Dario Amodei: AI could help double the human lifespan

Anthropic CEO Dario Amodei has argued that AI could dramatically accelerate biology, potentially compressing a century of scientific progress into just 5–10 years.

The longevity race is already attracting major tech fortunes. Jeff Bezos has backed Altos Labs, Sam Altman is involved with Retro Biosciences, and Peter Thiel is associated with NewLimit. Their shared ambition is to understand aging and develop ways to extend healthy human life.

Some of this research is moving beyond the laboratory. Retro Biosciences, for example, is pursuing therapies aimed at improving cellular health, while scientists are exploring ways to reverse aspects of cellular aging. But these approaches remain experimental, and doubling human lifespan is far from a proven outcome.

The bigger possibility is that AI could accelerate drug discovery, uncover mechanisms of aging and help researchers develop treatments that would otherwise take decades to discover.

That makes the next decade particularly interesting for longevity science. We don’t know when major breakthroughs will arrive, or whether they will deliver the results people hope for.

Until then, protecting your health today may be the most practical way to give yourself the opportunity to benefit from tomorrow’s medicine.

@aipost 🏴

😐 Jensen Huang just committed $1 BILLION to put AI in the hands of America's scientists.

At the White House science summit, 11 companies pledged $2.4 billion in compute and AI tools to the Genesis Mission, which now covers more than 15 federal agencies.

Nvidia gave the most at $1 billion over 5 years. AMD gave $500 million, OpenAI $200 million, and Anthropic and Google $150 million each, with Micron, AWS and Crusoe at $50 million each.

The government is quickly becoming one of the BIGGEST new buyers of AI compute.

@aipost

❗️The Information just reported a filmmaker-focused startup is negotiating deep access to OpenAI’s shuttered Sora video model.

Jeffrey Katzenberg, former Sora research lead Bill Peebles and ex-Dropbox CFO Sujay Jaswa co-founded that unnamed startup to build AI video for filmmakers. Under the terms being discussed, the startup could heavily customize Sora and sell software built on it, while OpenAI would get an undisclosed equity stake.

Source.

@aipost 🏴

❗️Satya Nadella says Microsoft is making its "BIGGEST BET YET" on its own AI chips.

Starting in 2027, Microsoft plans MULTI-GIGAWATT Maia deployments, potentially exceeding 1 MILLION chips, making them its primary AI inference system. Maia 300 promises 40% more tokens per watt than leading GPU systems.

@aipost 🏴

❗️Mark Zuckerberg tells the FT that Meta has FINALLY caught up with OpenAI and Anthropic at the AI frontier.

He's pledging more spending and still REJECTS calls to slow AI development. Meta is already building its next model, and expects development to get 'quite a bit faster from here'. Meta expects to do 'way more things' in the future, but each task could require 'somewhat fewer people'.

@aipost 🏴

And here I thought I was AI pilled…

@aipost 🏴

❗️The Information reports, the Trump administration intends to transfer 3 top officials from the Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) to the White House.

Its being done so its testers are in the room when policy on new models gets made rather than being consulted after the fact. A White House official framed it as a fix for poor coordination with Commerce.

@aipost 🏴

❗️Some students at Northwestern University are now asking whether they should stay in mathematics at all, after OpenAI posted hundreds of AI-generated papers on open problems in a single release.

Antonio Auffinger, a professor of mathematics at Northwestern University says his own priority has moved to keeping talented people in the field and helping them grow.

Source.

@aipost 🏴