Covering some of the papers this week with just the interesting bits (or just things I didn't know)
Starting with the BLT paper, we've now learned that @aiatmeta has some kind of food obsession
The most interesting one was the synchronous LLMs paper (at the end)
It's been fun to see Malayalam show up in multilingual testing a lot as a 'lower-resource language'. If anyone needs translation or labelling help lmk haha
What's crazy about the theory of mind paper (ai.meta.com/research/publications/explore-theory-of-mind-program-guided-adversarial-data-generation-for-theory-of-mind-reasoning/) is that you could "finetune" theory of mind understanding (at least going by the numbers). Is this changing how the model functions at a deeper level or is it just bringing out latent capability?
Mathematicians trying to write a novel
Has anyone studied how finetuning is affected when the responses are short (single answer, high entropy decision) vs long (CoT, SQL, etc)?
The syncllms paper (ai.meta.com/research/publications/beyond-turn-based-interfaces-synchronous-llms-as-full-duplex-dialogue-agents/) has a number of useful references about human speech which was good to put numbers to
Once again with the finetuning - are these guys doing something completely different to me?
Very, very likely
Brb trying to tell my girlfriend that interrupting "signifies engaged listening"