@_ryu15_ Ah right, I forgot about that. I just really hope Unsloth or someone with lots of $$$ to do this ๐คฃ. There are still so much potentials to be squeezed.
Balless Dario ๐คฃ. Cancer got his balls ๐คฃ and it turned him into this weird annoying control freak too.
All these annoying controls are ๐ง definitely.
If you try do talk about something that is like something unfair against the white, these anthropic and gpt models will do kneejerk hedge first without thinking deeper about nuances, like frequency for example.
I talked to fable about rupertโs campaigner got hassled by antifa and their campaign materials were thrown all over the ground even with police there. Then I told it turningpoint in the US and UK suffer similar things from the left either their materials got damaged or attacked in some other way like throwing milkshake at them.
Fable kneejerk reacting by looking up same incident against the left which was in 2017. It really ignores the frequency and everything. It just stops using nuances just for the sake of sounding balanced. Fucking ๐ง mandated programming.
Idk but.
This sounds great to me.
The reason I stopped using GPT models was that it did too many things I did not ask.
I wrote here on X saying that I should not write a whole list of what it should not do every time I write a prompt.
Maybe I should try Astra more.
Though I already found some flaws already. Astra absolutely dgaf about the resources it used. Even when making doing some 3d stuff I told it to do, it literally used so many cpu cores and made my m5 mac studio fan blowing hard for sustained time until I told it to tone down the test a peg.
Oh it also did not care about disk space, it absolutely generated so many things that now I wrote in Codex to be mindful of disk space.
GPT-6.2 Sol and Astra do the strangest things.
Here's a real example:
A local model I worked on yesterday failed to load, so I thought it's a good time to test the new models. So I opened Codex and told it, "The model worked fine a few hours ago, but now it fails to start."
That was the prompt. What did it return?
"Yes, I checked and confirmed the model loaded successfully before and now fails to load."
That was the whole response!
So it knows there's a problem and confirms it, but in the end doesn't do anything about it and doesn't even offer to fix it? I had to tell it, "ok... fix the issue."
It's like it doesn't fully understand what needs to happen to make the user happy, if that makes sense.
Then it fixed the issue and reported it was fixed. But it didn't load the model or test that it works. Isn't that the obvious thing to do?
So again, I had to tell it to check that its fix works.
Feels like it doesn't do well interacting with humans, and expects a response or a prompt from an agent instead.
They pretty much are and also benchmaxxed not for generalization but instead for climbing the benchmarks.
So yeah it is as you said, once it is told to do something outside its training distribution it flops.
Unlike western models where when a new benchmark shows up the models can still perform, because western models are aiming for actual generalization. Which is better idea in my opinion.
@loktar00 I love the series, people like me who werenโt in game dev or worked in 3d have too many unknown-unknowns that we donโt even know what to tell the model.
I did the exact same when I tried Qwen4-Max keep failing to do simple task like killing itโs own subagent. It keeps trying to call wrong tools instead of looking it up on the opencode instruction.
Then it gave up.
Then I swapped it to GPT-6 Astra medium. I told it do Qwenโs job. And it thought for a very short moment, then it loaded โopencodeโ skill. Then the next step killed the subagent like no problem.
It happened quite consistently on codex and also on deepseek harness.
I think the model really did have too much initiative before. Though I donโt doubt that agent harness can affect how the model behaves.
The one I used for example was GPT-6.1 Sol, I told it to install Cruzโs DeepSeek Flash 1.62 bpw on my dgx spark by following the recipe and register it to my llama-swap.
Sol noticed that my llama-swap has no login, and it added a login because it thought that it is โhelpfulโ to do so.
in the end i couldnโt access my llama-swap on dgx spark because sol added login and never told me anything. then i had to spend time again waiting for it to carefully remove it.
Fable actually like to do this too, I tried to have it become orchestrator for Opus. I already warned it to only do only what I tell it and donโt do anything I donโt ask.
Then I caught Opus doing something it shouldnโt, I told Fable why Opus was doing this and told it not to think and just kill that subagent first.
Then Fable apologized to me it told me that it is because it added something on its own, something I did not ask. The harness for this is Claude Code Desktop. But Fable has similar โdoing too muchโ on the new Opencode Desktop v2 as well.
@observer0334 correct
when i was that young i did not have impulse to do that
much less hitting little girl
itโs just hard to think if we are the same at all
i probably gonna get shadowbanned if i say too hard ๐คฃ
๐จ๐บ๐ธ DISTURBING footage of a 9-year-old girl being brutally beaten by an older boy on a Florida school bus!
A 15-year-old boy was charged with criminal battery after repeatedly punching the child.
How was a 9-year-old left so vulnerable on a school bus?
Source: NBC 6, CBS Miami \ Writer: Naz
Free Kick Practice - Diorama GPT-6.2 Astra Xhigh
Cost: $3.39
This is a redo with the prompt that Fable-5.2 created based on my base prompt.
The scene is now better than before, however there is still this awkwardness from the weird speed and the animation.
Probably can be iterated, if you know all the concepts and vocabs to say to it, but GPT-6.2 Astra cannot match Fable-5.2's awareness on the work. Fable-5.2 just have better awareness and make small things that are harder to vocalize and nitpick to be done by it naturally.
Free Kick Practice - Diorama GPT-6.2 Astra Xhigh
Cost: $2.32
I gave the same base prompt to both and let them enhance that and be as creative as it can before starting the html. I guess Fable just love being creative more than Astra without detailed prompt from the user.
Though, it should not excuse why the animation feels sluggish.
Free Kick Practice - Diorama GPT-6.2 Astra Xhigh
Cost: $2.32
I gave the same base prompt to both and let them enhance that and be as creative as it can before starting the html. I guess Fable just love being creative more than Astra without detailed prompt from the user.
Though, it should not excuse why the animation feels sluggish.
Free Kick Practice - Diorama
Fable-5.2 Xhigh $20.04
Fable absolutely better than GPT-6.2 Astra in animation. Fable's work is so much more natural in term of movement and also speed.
I will upload GPT-6.2 Astra Xhigh's output after it is recorded and the cost accounted.
Free Kick Practice - Diorama
Fable-5.2 Xhigh $20.04
Fable absolutely better than GPT-6.2 Astra in animation. Fable's work is so much more natural in term of movement and also speed.
I will upload GPT-6.2 Astra Xhigh's output after it is recorded and the cost accounted.