@TECleveland @cremieuxrecueil Despite some bogus grants, a lot of government is like this cartoon. I bet this "small team" would find a line item for "screwworm sterilization" and delete it, not knowing that they just devastated American farming: https://t.co/OAnAkxDSEK
solving continual learning requires thinking about it clearly
here is my attempt to specify the problem
stay tuned for part 2, where we'll attempt to benchmark it with @PrimeIntellect :)
@Merridew__ 2026 Q2 models burn much more tokens for reasoning and do much more tool-use calls to do the same job compared to 2025 Q4. Given that most of Ant's revenue is enterprise API-priced tokens and token price is stable, revenue increased automatically even if userbase is constant.
@_catwu A graph. A directed graph with one source and one sink. Almost like a chain. A chain with LLM loops as its nodes. May be we can call it a @langchain 🤔 May be we can even design a DSL to express it, a "lang. chain expression language"?
@Merridew__ They can also start hosting big yearly conference in Vegas to teach programmers how to use the most compute possible for any type of problem.
@unixpickle And even if the model generates a wrong answer early, do you think it is less of a problem for a reasoning policy which would verify and correct (compared to chat- or instruct- models)?