Sunday Links: Cartwheels, OpenAI ID, and reasoning

Links on OpenAI's future, Robot Agility and Reasoning.

Sunday Links: Cartwheels, OpenAI ID, and reasoning

Here are this week’s AI stories:

  • Why can’t AI make its own discoveries? [Podcast with Yann LeCun]. I’m generally very much in the LeCun camp in terms of future cognitive architectures: we definitely need to shift to Neurosymbolic, current LLMs definitely aren’t the answer to everything, and learning representations of the world is going to be crucial. This podcast, though (in title and premise), really over-rotates in my opinion. LeCun is probably trying to be provocative (and the host amplifies that), but I think it’s just plain false to say that today’s LLMs “can't make discoveries.” Every second of every day, LLMs are making novel combinations of things now (at human prompting), and some of them are genuinely useful. LeCun’s primary point is that “human thought” doesn’t happen at the language/token level but, instead, uses abstract models of how the world works. I agree with him that if you learn the abstractions (as they are doing with their JEPA research), that’s very powerful, but I think it’s wrong to say that token prediction is useless. It seems highly likely a lot of our human capacity to build abstractions rest on the concrete token level of examples. In reality, there is already a very fuzzy model of the world being learned, even just from words. In the long run, the real power will come from moving between abstractions and concrete representations. The interview is worth a listen, especially for the second half. I’m excited about the approach they are taking, but we really don’t need to be so negative about today’s state of the art. It's very different to say something is "not possible" versus saying something is "being done only very poorly."
  • Artificial intelligence learns to reason. Melanie Mitchell’s article in Science is almost a riff on Yann LeCun’s point of view and a good introduction to reasoning models. In a nutshell, she argues that “reasoning models” like OpenAI’s 4o are likely just pretending to reason. Mitchell doesn’t call it one way or the other but underlines that what’s happening is that the models are navigating correlations between the kind of things they’ve seen before. I think that’s correct, but even given that, one could argue there is a very fuzzy, poorly represented world model sitting underneath all the token generation.
  • Walk, Run, Crawl, RL Fun | Boston Dynamics | Atlas. In a new demo video, Boston Dynamic’s Atlas robot is seen doing everything from impressive yoga poses to cartwheels. Robot agility just reached another level.
  • Gemini Deep Research is now free. I think one of the most important messages from the future on AI, is the power of the deep research tools that Google Gemini, Grok 3, and ChatGPT now have. These tools take complex queries, mine the web for 200-300 relevant sites, and then synthesize a report. Generally, this takes 5-20 minutes. The result, though, is really worth it: an in-depth analysis (often 20-30 pages) rather than just a few bullet points and links. I think this type of research service will have a massive impact on knowledge work. OpenAI’s version definitely feels like it is ahead of the pack, but if you want to try something almost as good for free, try the Gemini Deep Research service. Go to Gemini and select the “Deep Research” model for your query (you will need a Google account, but the free level is sufficient).
  • An Interview with OpenAI CEO Sam Altman About Building a Consumer Tech Company. Lastly, another podcast really worth listening to. This interview of Sam Altman by Ben Thompson hints at much of the future of OpenAI as a consumer company. Thompson pushes Altman on the consumer opportunity, and it is interesting to hear Altman’s response, which I’ll summarise as “UI, Identity and APIs." The key segment about halfway through is where Altman lays out the idea that consumer AI will have 4-5 key services (chat, search, and others) + an identity service (log in with OpenAI) so you can log into other web experiences via your AI which has a trusted memory of your previous interactions. OpenAI might end up trying to be OpenID as well.

Wishing you a great weekend!