richardfey

Was it inspired by this?

https://housing.over.pizza/

Tried to use it and the error messages and documentation were overwhelmingly unhelpful. Spent about an hour on it. Maybe it will get better though, good luck!

I wish the useful improvements would be ported to https://github.com/open-annihilation/open-annihilation which seems a better open source project.

Yes; or something which has a similar effect.

It feels lke they replace older models with "optimised" versions which are cheaper to run, but keep the same name.

For pixel owners: the Google photos app cannot be uninstalled, but you can disable it using adb.

> "Instead of improving an agent against a fixed test, we let the evaluation evolve alongside the agent"

This quote should have been highlighted earlier in the article.

It defines and introduces a lot of concepts/acronyms in the thinking blocks which we normally don't read.

> If the model was designed specifically to quantize down to 1.58b, then it's different.

> AFAIK, there's no large models designed for this yet.

Isn't BitNet b1.58 2B4T what you are looking for? (haven't tried it myself though)

> Quis custodiet ipsos custodes? is a question that is roughly 2000 years old. Still no good answer.

I feel like we are moving towards a very specific answer, without much thought: AI.

Did anyone calculate this for OpenAI?

Looking forward to giving this a try with llama.cpp. I’m watching the open-weights competition with high expectations.

I think that would spoil a lot of the fun. I would favour visual and audio feedback loops to interactively learn how things work under the hood.

Sounds like a great way to inflate your success metrics for AI queries

I disagree. Whether something is manipulation depends on whether you are trying to change someone’s opinion or behavior, not on whether the manipulator has “good” or “bad” intentions, since those judgments are not objectively universal.

> Part of this can be good (you talk about what they care about, where 90% of broadcast messaging might not apply) and part of it can be bad (manipulation.)

Side note: it's manipulation either ways because you chose what to talk about, with a goal in mind.

Serious question: how do we verify claims like these on the effectiveness of a harness?

Has anyone tried Kimi K3 against gpt-5.6-sol on real projects?

Exactly; this is a no-go for me, I will wait for an independent provider to sell the service, which is possible thanks to the open weights.

I understand you're trying to be funny, but my point is that with novel technology there are novel ways to claim innocence in courts because of the legislative void.

They could use an agent to summarise the source material, and then train models on those summaries, and claim that some sort of clean-room training has happened?

> Or are you referring to the claims from Meta, Google, and Adobe -- which failed to hold up under independent evaluation.

This.

> However, he demonstrated a clear lack of understanding regarding what makes the bits "independent" or how to resolve the independence problem.

Yes, I didn’t want to call it out explicitly, but this is exactly the kind of thing that would have made my undergraduate statistics professor lose patience.

I can't find on their website some indication of what kind of usage I can get out it, otherwise I'd be interested.

This is a great statistical analysis and it was a pleasure to read, but I wasn't expecting the claims to be so poorly supported. There's also a reply from one of the Meta authors there, worth checking out.

I remember hearing this perspective when I first started in the software industry, and I agreed with it for quite some time. But frankly, we’ve never been further from it.

The more I read into it, the more pain memory flashbacks I got. Bravo

I'm trying it right now for a side project of mine, compared to Opus it is effectively better at following instructions and somehow has better "depth" when reasoning on complex tasks. However, if it will not be part of subscriptions, I will not use it anymore.

64MB was a luxury

I don't know what I am doing right, or wrong, but I have access to claude and codex and I find myself giving the more serious work to codex recently. I tend to trust it more. I might try again Fable when it's back, but this Sonnet 5 didn't work well for my current projects.

How did you give LLMs tool use?

I could spot numerous bugs in code written recently and less recently, by me or colleagues. I was not angry but grateful and I knew there was no way back!