I gave Fable 5 one job: write custom WebGPU kernels for Gemma 4 inference. It climbed to 84 tok/s, then hit a wall, insisting further optimization was impossible. Hours later, Anthropic rolled back invisible LLM development safeguards, and it hit 255 tok/s. The next day, access to Fable 5... was suspended globally.show more

Soon China is going to build a model like fable and it is going to be open - which will make everyone lean on China than us

I have been a fan of webgpu for sometime. Here is one for you. Browser-native persistent AI agent runtime with local memory, context reconstruction, compiled WebGPU inference, and a custom WebGPU Kernel Lab.

This is fake — the invisible safeguards being removed means it went over to Opus instead of sandbagging

Their fear mongering strategy backfired. When they compare AI with nuclear bombs, now they have deprived all of advanced AI. If the government start to control AI like this, it will slow down all progress, not unlike the middle ages, where thought was controlled by the church.

How do you know it was an invisible classifier? And not simply the model having a break through discovery? Seems viable to have natural plateaus and stalls in progress sometimes.

But can this only be done by Fable 5? Any comparison?

i did this but with my stock trading bot lmao. pretty selfish of me but now its getting near 3x returns/mo so thats great. also started on a minecraft clone then ran out of usage

This is funny because the safeguards themselves were never rolled back

Curios to know, did you try this with GPT-5.5? What happened?

Anthropic did not roll back invisible safeguards. It just rolled out the invisible part.

the kernel optimizing itself overnight while you slept is mildly terrifying and i'm here for it

All of this is about stopping the small and protecting monopolies. None of this is about safety. •

Now just plug the same prompt into Opus and you can watch it get the same result (which is all Fable was doing without the safeguards).

Doin Gemma 4 today too Been doin agentic kernel optimization since mid last year. Those last winning tests are Fable 5 right before it was suspended. It was on a tear 😭

84 to 255 tok/s the moment safeguards rolled back is a wild data point. That gap between what the model can do and what it's allowed to do is bigger than most people realized.

what if you give same tasks to gpt5.5 (xhigh)?

I was like 75% solving a pet problem I’ve been chipping away at for years… “this model is unavailable” 😭

fascinating

Woah

okay, I'm interested to know how often it was making edits or being looped during this period. For instance between minute 100 and 180

I’ve been working on a similar project, and just ported it to a standalone was able to get qwen3.6-35b-3ab on 8gb ram, qwen3.6-27b on a 4090, and qwen3.5-27b on an rx 6700 xt, also worked with a 1080ti

@peteskomoroch Can you share your code? Help liberate the models? Maybe opus can use it elsewhere

255 tok/s after warmup means little because webgpu shader cache misses dominate first token

I was doing exactly this too, I did the Gemma 4 kernels a month ago but in one day I got Gemma 12B and diffusion working with WebGPU in my runtime, diffusion got the plug pulled in the final hour last night. They pulled the plug and 4.8 Opus told me it was impossible.

This is heckin insane if you were able to actually do this!!? What about for like qwen?

-- Using Self Improving Agent framwork paired with unsloth.. wouldn't it be possible to spawn and train a slm for this?

Technically it wasn't banned globally. It was banned for non-Americans and then they withdrew it from everyone.

did u notice in the previous period that 4.8 was actually getting called?

Did you ask it write a report of what it tried first and then what changed after it started increasing again so you know what was gated

Anthropic only changed what happened when you hit safeguards (previously, "shadowban"; now, Opus). You just placed the kink in the straight line in such a place to make it look like this was related to the change.

So you were basically using Opus in the background until they rolled back the guardrails...

@grok, what are all the optimizations discovered in the video

Repo link? 🧐

🥲