$20 plan + deepseek-v4-flash is the new $200 plan (but much cheaper).
Graham Neubig
@gneubig
Associate professor @LTIatCMU.
Co-founder/chief scientist @OpenHandsDev.
I mostly work on modeling language.
- It's interesting to think of how AI-based software development enables new ways of developing software that would've been infeasible before. The idea of personalizing OSS via a long-running fork is a pretty interesting one IMO. We've always forked before, but maintenance burden"dev tools must be open source" This blog is making the rounds on Hacker News, and we agree wholeheartedly. blog.exe.dev/devtools-must-… One of the nicest tidbits is the idea of a "long-running fork" where you have an agent merge in upstream changes every night. This happens to be
- The big G didn't even make the poll? 🥲The voters love OpenAI. According to @TheArgumentMag polling, none of the major AI discourse (e.g. Anthropic's fight with Trump) seems to have had much of an impact on public opinion of the AI labs. Even among *people worried about catastrophic risks* OpenAI has a +8
- Strong results, and more interestingly you can really see the advantage of a specialized AI research harnesses over very long horizon tasks (4000+ hours)! Does this mean that all of our ML eng experiments cost $10k each now 😅
- OK, I tried deekseek-v4-flash, and initial impression is that this model is *really good* at long horizon tasks (contrast to terra). It usually doesn't lose the plot. I was previously using gpt sol/terra + glm-5.2 as my daily drivers, but I think this may be a new one.gpt-5.6-sol and gpt-5.6-terra are both good models, and terra, to me, doesn't seem a huge amount "dumber". the thing that makes sol so much better though is its thoroughness/persistence. it gets the job done, in contrast to terra, which seems to do a lot of partial work.
