Keep on to blur preview images; turn off to show them clearly
Two ways to spin the same missing causal link between the LRM intermediate tokens and solution guesses: Us: Stop anthropomorphizing intermediate tokens; they may help LLMs sample better, but have little end-user interpretability. Them: OMG😱 LLMs are hiding their intentions!😱
AI researcher & teacher @SCAI_ASU. Former President of @RealAAAI; Chair of @AAAS Sec T. Here to tweach #AI. YouTube Ch: https://t.co/4beUPOmf6y Bsky: rao2z
(imagine a bottle of champagne pictured) and celebrating w/ champagne from our great friends @PenlandTaylor and @aalixmarie
making complex metal assemblies for fast moving companies | Founder at RMFG
Coming up with such a proof, running the simulation, and all that, is what I used to do in ~ one month. Now this was done in an afternoon. For those interested you can find the full proof and more details in Chapter IV, Section 4 of https://t.co/IfotVApR3X. Happy Thanksgiving!
I work on AI at OpenAI. Former VP AI and Distinguished Scientist at Microsoft.
The proof (100% GPT-5) is surprisingly clever, using Robbins Monro (ie SGD!) type analysis to control the convergence of key quantities. Also before spending the effort to look at the proof, we asked GPT-5 to run a simulation and check empirically that the formula was believable ...
Coming up with such a proof, running the simulation, and all that, is what I used to do in ~ one month. Now this was done in an afternoon. For those interested you can find the full proof and more details in Chapter IV, Section 4 of https://t.co/IfotVApR3X. Happy Thanksgiving!
It's getting harder and harder to get signal from benchmark numbers. Rather than averages, I except in the (near) future we will also care about "argmax": what's the BEST output a model can deliver? After all, we don't need to solve PvsNP 10 out of 10 times, once is enough 😅. So with that in mind let me tell you a bit more about THE MOST IMPRESSIVE LLM OUTPUT I have ever seen.
Many of you are probably familiar with the preferential attachment model (also known as Barabasi-Albert), which is a growing random graph process, where each new node attach to an existing ones with probability proportional to their degrees. This is very similar to how a network like X grows (popular accounts attract more and more followers). (beautiful gif by Igor Kortchemski)
RT @Geiger_Capital: George Washington’s Thanksgiving Proclamation in 1789: "By the President of the United States… Whereas it is the duty…
Co-Founder, American Dynamism. General Partner @a16z. Catholic. Mother. American. 🇺🇸 🚀💪