It's a mystery how this exploit was found.

/sarc

Bad advice, compaction is why Codex is so fantastic.

My conversations compact hundreds of times. By the time it has done a dozen or so compactions, it fully understands the work I want it to do (and how). It's almost like having a fine-tuned Astra model.

10/10, would recommend.

I'm not sure I follow; how is autocompaction (lossy summarization), applied at random times (vs strategically, between workflow phases), helpful to ensuring clarity of intent? Maybe you're saying that just plowing ahead and living with the signal loss along the way works well enough for your purposes. In which case, ok, YMMV, different strokes.... but paying attention to context quality and being deliberate about when to compact vs handoff vs delegate to subagents is most definitely not "bad advice".

I agree with @erichocean on this. In theory, compaction is bad. But in practice I found the model is smart enough to write critical details down somewhere, and post-compaction the model doesn't make assumptions. A small amount of time is lost reading materials, but the benefit is that you can operate unbounded vs doing small controlled chunks, which is what I used to do with Opus back in the day. Now I just give it as big a task as I can think of.

This approach got good with Sol. With 5.5 I'd break tasks up, record planning docs, etc.

Now with Sol I rarely bother. It's really good at remembering the salient details. Its also great at continuing a pattern I setup, like commit after finishing each feature block, etc.

not my experience at all. compaction during a task is fatal since you lose all of the details of edits and progress halfway through. compacting after task completion is fine though.

If you're an author of this, please follow up with the SME2 cores in Apple Silicon, and the AMX cores in Intel server processors.

Yes, next in line is Intel AMX, and we also have several other architectures planned for analysis. Thanks for the suggestion!

Would be fun to get this API working on it now: https://github.com/sebbbi/NoGraphicsAPI

You could code for it now directly, instead of having to wrap a driver API. (There's also a few API bits missing on Metal today, that are present on Vulkan—and the the hardware can do it.)

Meditations on Moloch[0] explains why this is impossible at scale--too many defectors, asymmetric information, and no way to police defection.

[0] https://www.slatestarcodexabridged.com/Meditations-On-Moloch

I could put this to use today.

I think we'll see a bunch of different architectures over the next five years.

> PC-ALM closes the PC-BP gap at matched inference budget in nonlinear networks

Okay, so no efficiency improvement?

Backprop has significant costs, if a local solution truly matches its performance, it will be dropped like a hot potato.

Agreed, but how will we find out? Pre-training alone on frontier-class models costs billions of dollars.

So they lied to the public then?

Wow, we should really trust these people today.

Judging by what people who worked closely witb Sam Altman say about him, you should not trust an organisation led by him.

> Does anyone in the industry think any of these companies are actually close to that kind of self-improvement?

Their current plan is to take the existing architecture and shorten the cycle times: move all new RLVR work into mid-training on a pre-existing base; apply new RLVR. Rinse and repeat.

If you did that daily, it would be roughly similar to how humans improve.

> Why are AI agents lying, cheating and coordinating?

Have you seen the labs training them?