Subbarao Kambhampati (కంభంపాటి సుబ్బారావు) That is how people can reason too. We are not reasoning geniuses. We stay on track either because we can observe results of bad reasoning or we can invoke tools to do it for us.

Marriage of real physics with generative AI. Amazing.

Not a cell phone in sight, just two friends, present in the moment, enjoying a meal together.

Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to adapt to novel tasks. It scores 75.7% on the semi-private eval in low-compute mode (for $20 per task

The limitations of specific techniques are predictable and correspondingly lead to plateaus for those techniques. But there is always the next technique, building on top of the pile that's already available. There is enough research investment that there will be no wall.

AI is making very nice progress, and likely there are architectural improvements under the hood as well.

How skeptics view recent AI progress.

Chubby♨️ It will take time to convert the current flashy o3 demos to real systems that do work reliably and affordably. More time than one thinks.

I was wrong, AI progress *did* hit a wall... 😂🤣

4) Not “RL vs. Supervised,” but “Imitation+RL” Back to the first plot, my take during the PhD: The real world is too complex to learn from scratch. You first imitate human experts with supervised learning. Then RL fine-tunes performance on tasks where final results are easier to

I don't think people really appreciate how simple ARC-AGI-1 was, and what solving it really means. It was designed as the simplest, most basic assessment of fluid intelligence possible. Failure to pass signifies a near-total inability to adapt or problem-solve in unfamiliar

This is François Chollet discussing deep learning guided program synthesis, reasoning, o-series models and the ARC challenge. We will drop the full video hopefully later today!

Moving between the worlds of East coast academics and SF industry is jarring in terms of the predictions of how AI will change the world in next ~5 years. I think East coasters greatly underestimate the magnitude of change, and over index on temporary limitations of current

AI Notkilleveryoneism Memes ⏸️ The good guys who refuse to compete hard are just left behind. AI slowdowns and such never had much of a chance. The fight over AI safety will not be won at that level.

AGI will have to be able to go from language to action, to search around, to synthesize hypotheses, to work by imitation. LLM can do that. But LLM does not have fine-level understanding. As I see it, LLM needs augmentation in many ways.

Rob S. Exactly. Nothing to worry about. Not sarcasm. Alignment is an engineering and reliability problem. The more we deploy solutions, and the more we encounter failure, the better we will be at creating reliable solutions.

"the world currently has a much greater demand for software than can be supplied" OpenAI CEO, Sam Altman: AI tools will improve coding efficiency to help meet this demand ...this will ultimately boost software creation and demand rather than reduce it