Joonas Uusitalo

Astronomer. Where did everything come from? Where are we heading?

[Joonas Uusitalo @JoonasUusi]

@adam_dorr Sounds like some kind of problem. I was able to use it for hours last night and burned like 30% of fable usage. Although I used high, not max.

[Joonas Uusitalo @JoonasUusi]

@MikaelJungner 🤣

[Joonas Uusitalo @JoonasUusi]

@MikaelJungner Palataanko asiaan kahden vuoden päästä? 😊🤗

[Joonas Uusitalo @JoonasUusi]

@Lumichaoss Dumb & Dumber.

[Joonas Uusitalo @JoonasUusi]

@itsnwts Frontier AI models. 🤗🤗🤗

[Joonas Uusitalo @JoonasUusi]

@asaio87 For me it was able to one shot things I had to fine tune and tweak for a day with Opus 4.8.

[Joonas Uusitalo @JoonasUusi]

@LLMJunky LOL, indeed!

https://t.co/sEw2euf8t4

[Joonas Uusitalo @JoonasUusi]

@DavidSacks @wholemars Sorry, but this is bullshit! Everyone knows that all the models can be jailbraked, just see @elder_plinius. It doesn't appear to be easily fixable. You have been against these kind of restrictions and now you support them??

[Joonas Uusitalo @JoonasUusi]

@NickADobos I don't even want to continue my projects without it. Feels waste of time to spend one day fine tuning something with Opus 4.8 that Fable was able to oneshot. This literally happened.

[Joonas Uusitalo @JoonasUusi]

@DeryaTR_ BS, it's possible to jailbreak any model.

[Joonas Uusitalo @JoonasUusi]

@mark_k BS explanation.

[Joonas Uusitalo @JoonasUusi]

@DavidSacks @wholemars Sorry, but this is bullshit! Everyone knows that all the models can be jailbraked, just see @elder_plinius. It doesn't appear to be easily fixable. You have been against these kind of restrictions and now you support them??

[Joonas Uusitalo @JoonasUusi]

I do not understand the negativity around Opus 4.8. I wonder if there's some kind of campaign against it. It is clearly better than before. With Opus 4.8 I can again continue a project I got stuck with using Opus 4.7 and GPT 5.5.

[Joonas Uusitalo @JoonasUusi]

@markkurajala2 Great catch!

[markku.rajala @markkurajala2]

Claude quietly showing Opus 4.8 in the model picker 👀

Looks like Anthropic is shipping fast. If 4.7 was already strong, this could get interesting.

[Joonas Uusitalo @JoonasUusi]

@JPKivisto @StochasticStat1 ... been solved.
4) Solve the new problem.
5) etc.

Maybe 2) is too hard, I don't know. Or 3) is not possible, which I doubt.

Anyway, I wouldn't bet against our ability to improve these systems. And not saying LLMs are some holy grail, but even they can be powerful.

[Joonas Uusitalo @JoonasUusi]

@JPKivisto @StochasticStat1 ... in the new space etc. etc. Or do think the LLMs ability to solve things get exhausted when all "obvious" problems have been solved?

So I see it something like this:

1) Solve a "known" problem
2) Retrain autonomously with the solution
3) Figure out new problem as 2) has ...

[Joonas Uusitalo @JoonasUusi]

@JPKivisto I don't care whether something "thinks" or not if it can beat any human in chess or go, or in any game whatsoever, and solve mathematical proofs in split second that no human can do. At that point, whether it thinks or not is just a play on words. That's the trajectory we are on.

[Joonas Uusitalo @JoonasUusi]

@JPKivisto Where and how has it been shown? I use LLMs pretty aggressively and I'm not sure I agree with that. I think it may be possible to form chains, where the LLM "thinks", tries and verifies ane eventually memorizes novel stuff.