New TPUs make JAX & Keras go brrr

Google I/O is today and 'Large language models with Keras' is at 4:30PM 🔥with Samaneh Saadat and Gabriel Rasskin !

Since 2016 I've been told that 'self-play' was going to be a huge thing in AI. With the rise of AI girlfriends, it might finally become true.

People really want to believe in near-term AGI, and will take all intellectual shortcuts available in order to feel justified in this belief.

I also want to believe in near-term AGI, but not to a level of religious faith that would make me renounce basic critical thinking.

The always insightful Melanie Mitchell @melmitchell1 on what LLMs are actually doing (really good retrieval) and not doing (reasoning). aiguide.substack.com/p/evaluating-l…

If intelligence is the ability to deal with what you weren't prepared for, then the modern AI strategy is to prepare for everything, so you never need intelligence.

This is of course a terrible strategy, because it is impossible to prepare for everything. The problem isn't just…

Meanwhile what the AI of today does is to combine extremely weak generalization power (i.e. ability to deal with novelty and uncertainty) with a dense sampling of everything it might ever be faced with -- essentially, use brute-force scale to *by-pass* the problem of intelligence…

The downside of hyping your future V5 so much is that you have to release all of your new models under the V4 brand in order to avoid disappointment -- in perpetuity