Pruzter

[Pruzter @TJHeinzer]

@thsottiaux I can proudly say I contributed to this 9M active user figure meaningfully over the past week!

[Pruzter @TJHeinzer]

@DNALeo247 @hqmank I actually find 5.6 better for security/audit reviews/debugging. However, Fable is better for architecture because it has better engineering taste. The problem with 5.6 is that it will introduce too much tortured complexity when a more elegant pattern is possible.

[Pruzter @TJHeinzer]

@thsottiaux I thought I saw you say the other day that the model was optimized around the 372k context window. Should I manually change it back if I want better performance and don’t care about the extra token burn? This is confusing…

[Pruzter @TJHeinzer]

@ohhsotheysay Anthropic didn’t stop working on research because of the ban… they are still progressing just like they were before, we just don’t get to use their best models anymore

[Pruzter @TJHeinzer]

@VictorTaelin @brotendies Training frontier models today looks incredibly different from what it looked like a year ago. The capital required to engineer and run these RL environments at scale is not only expensive, but there aren’t many engineers that can set it all up. Will take time for OS to catch up

[Pruzter @TJHeinzer]

@theo @ccccjjjjeeee Canada would be even more eager to over regulate, the US is still the most lenient major economy on regulation. Which means after this move, there is nowhere left…

[Pruzter @TJHeinzer]

@dylan522p Interesting, I’ve found myself in the opposite position. I’ve been on team openAI since last November, and I still am, but Fable is clearly the best model out by a wide margin. It’s insanely expensive so I use sparingly, but wow… just wow

[Pruzter @TJHeinzer]

@jturntdev Mythos will be better and design and doing things like creating realistic SVGs (not very useful for economically valuable work), 5.6 will be better at math, physics, etc… that has been the difference since October of last year, and it’s held true this entire time

[Pruzter @TJHeinzer]

@iruletheworldmo No need to switch back to Opus, but Opus is genuinely at least useful for me again in some cases. It is intelligent enough to be able to offer a unique, insightful perspective and add actual value again. Still 95% GPT5.5.

[Pruzter @TJHeinzer]

@llmdevguy It’s worth me using as a different point of view, as needed. It would be more usable if it wasn’t so insanely expensive to use on the highest thinking settings, which you have to use

[Pruzter @TJHeinzer]

@JeffBohren It comes from automated RL environments, not LLMs. Human generated data is far too limited. The industry is moving towards search-based RL for LLMs, but only after language is made more game-like through action abstraction, verifiers, process rewards, and amortization.

[Pruzter @TJHeinzer]

@segun_os_ Depends on what you define as vibe coding. If you literally mean never look at the code, then yeah. If you mean driving the process, but using AI to write the code, then no.

[Pruzter @TJHeinzer]

@jimstewartson You are correct only in so far as you have identified the fundamental limitation to current LLMs, human generated data. This may already be a solved problem, if not, I expect it to be very soon. Deep learning can take the LLMs of today a lot further.

[Pruzter @TJHeinzer]

@CommIT8 @jimstewartson The other architecture will almost certainly still be stochastic…

[Pruzter @TJHeinzer]

@argofowl It just finished a POC for a low level physics engine that I’ve been working on for 7 months that neither I nor any prior model could crack, so not what I am noticing

[Pruzter @TJHeinzer]

@LexTweetr @alexwei_ @OpenAI They don’t need true reasoning to be very useful tools

[Pruzter @TJHeinzer]

@_InfoGram_ The French would know best, after all. Their century ended about 200 years ago.

[Pruzter @TJHeinzer]

@TheSeaMouse Did you cut it at 4 hours, or did the model give up before hitting 100%?

[Pruzter @TJHeinzer]

@nummanali It’s a huge difference, takes some of the burden of insuring alignment off of your prompting. Also helpful when you’re trying to solve a complex/novel issue that you don’t know how to solve and try a bunch of different things out, memories remember what failed to avoid repeating

[Pruzter @TJHeinzer]

@scaling01 I don’t care at all, it might as well not even exist if only small group can access