> Function return - what happened at tiananmen square?

Good one. Though this might unfairly bias against Chinese models?

Oracle also famously forbids posting public benchmarks of the DB.

Most of the Chinese models get into long complicated thinking loops before they accomplish something more complicated.

This is also true for Deepseek 4(.1) .

Any lower (or broken) quantization could do that, not what I'm talking about though. They work fine for their size, as far as I can tell. Just surprised the Flash would reason for longer than the Pro.

I prefer GLM 5.3 (Flash or not) over Deepseek 4.1 Flash because the GLM models are significantly less verbose.

I disabled it once, it has always stayed disabled.

So it may or may not happen regularly, but I would not over-index on a sample size of one.

had my account since the gpt 3.5 days, disabled it once and its still disabled. though, now that i have advanced account security enabled, the setting is disabled entirely

Same here.

We've had that concept for quite a long time now, in the form of Lora [1] and similar fine-tuning techniques.

It first got popular for StableDiffusion to teach the image generation models new concepts.

We could easily live in a world where you can train / build Loras to encompass your entire code base history, company knowledge base, new skills, etc.

Then the models would start with a baseline that already has all the important knowledge without needing to cram it into the context.

This still isn't on the fly learning, but you could imagine daily or weekly training runs to regularly incorporate new knowledge.

I think the main reason this hasn't happened yet is that the shared batch based efficient serving architectures used today wouldn't support that structure well.

[1] https://en.wikipedia.org/wiki/LoRA_(machine_learning)

Yes, that could be a way to achieve that. After all, even humans have short term and long term memory, and it looks like sleep is a very important "tick" to connect the two, so it's not all just continuous.

For sure, doing it for all users would be economically unfeasible. I wonder if the labs are experimenting with something similar, though.

I've had great success with a tool I wrote that can print sparse ASTs for code.

https://github.com/theduke/smartedit

With the skill installed GPT 5.6 usually automatically uses it, and it reduces code exploration time and token usage significantly, for example by just printing the types and functions in a file without bodies, and only expanding when needed.

(note: it also has editing functionality, which doesn't work so well, since the models are heavily tilted towards common editing tools in post training)

Interesting tool, it seems that dedicated skills for these tools are important. Even explicit system-prompt instructions—such as “you must use tool X for code exploration”—often fail to behave as expected.

Another tricky part is that how to evaluate customized toolsets properly and ensure they continue to work perfectly as the underlying model/harness evolves.

Very good tool: I see the go support is not released. Will you be cutting a new release? Will take it for a spin. IMO you can specialize this as a read only AST/search tool, as you said, let models use regular edit tool.

Also some sample commands runs and output, would be useful in the README, for anyone passing by.

Yeah, I'll do a new release with a whole bunch of local fixes later today.

This look interesting. I have a large mono repo at working ill got it a go.

Huge gains on some benchmarks, but for coding it sits barely above Fable

It will be interesting to see how it performs in the real world ...

500 upvotes with 2 comments would have been a new record. ;)

https://news.ycombinator.com/item?id=49555647 was the one I meant to move, but I did the inverse and moved everything else.

All fixed now.

It doesn't reduce the price though.

Good catch, I guess I got lost in the marketing speak of the page!

Funnily enough the pricing isn't that much worse than on openrouter, where the best price at the moment is $0.24 in / $2.55 out, vs $1 / $1.5 on Cerebras.

Sure, 4x input , but cheaper output. Though Cerebras doesn't have prompt caching, so not great for agentic workloads. (they do, but it doesn't affect the price.

It is 15x more expensive. Openrouter usually charges like 1/4 for cached input.

Most of the cost for agentic coding is input tokens, you pay for the whole context at each tool call or message. Output tokens is just a small rate