Jev is just a newer version of BERT. Only those who don’t know what decoder-only and encoder-only models are get hyped about it. It’s unreliable, won’t work for any serious tasks, and can’t be trusted. Anyone using this shit in a serious scenario like prod is a moron.
It’s being forgotten because it was overhyped trash with no real value to offer.
@sama In short: Ram Altman is calling for an industry-wide brain-size limiter on frontier AI models so we don't have to worry about supersupersuperintelligence going rogue and killing us all before we build the cage.
@merettm For 1 second I thought you were talking about how confusing Astra's responses can be lol. Then I realized you were just talking about perplexity as in CE.
So basically this is just scaling laws 1.01.
Well done tho.🫡🫡🫡
GPT also has a habit of saying things like “I didn’t do A or B, and there’s no need to do C.” Half the time I never asked about any of those things. They came up in its own CoT, or they’re terms it made up along the way.
It’s frustrating to spend so much time and effort and tokens just to figure out what GPT is trying to say. To me, that’s a alignment problem. It doesn’t know what matters to the person reading the answer. It’s almost like Astra was trained to talk to another Astra.
My wild guess is that this is an RL problem. On tasks that are hard for humans to grade, OpenAI might rely more on another model to judge the answers. If that model gives Astra full credit for a technically correct answer, RL can reward answers that make perfect sense to another model but are a pain for a human to read. That would explain why it gets so much worse on complex tasks.
One of them is a constrained combinatorial optimization problem with a nonlinear, reverse-engineered objective function. Nothing conceptually crazy, but I guess GPT got bogged down by endless self-imposed restrictions. And btw, Fable 5.2 SABOTAGED the codebase, so I have to stick with GPT.
GPT is extremely capable, but it’s a pain to communicate with. Like when Theo asked why creating a worktree took so long. Opus just said the disk was slow. GPT went off about pnpm. GPT was absolutely correct, but still extremely confusing.
@takayama1227 Don't pursue women your age. They're delusional and ask for even more while their actual value continues to decrease. Try meeting younger girls (adults only please). Things will get much easier.
@Cloudier0177@thsottiaux If you don’t have my problem, you are lucky that your work can be easily accomplished by Astra with few shots. Once you see Astra fails its task you’ll see what I was talking about. Total misalignment disaster.