Some people are speculating that Opus 4.7 is distilled from Mythos due to the new tokenizer (it means Opus 4.7 is a new base model, not just an improved Opus 4.6)
The new tokenizer is interesting, but it definitely is possible to adapt a base model to a new tokenizer without too much additional training, especially if you're distilling from a model that uses the new tokenizer. (see, e.g., https://openreview.net/pdf?id=DxKP2E0xK2).
Yes, I was thinking that. But it could as well be the other way around. Using the pretrained 4.7 (1T?) to speed up ~70% Mythos (10T?) pretraining.
It's just speculative decoding but for training. If they did at this scale it's quite an achievement because training is very fragile when doing these kinds of tricks.
Reverse distillation. Using small models to bootstrap large models. Get richer signal early in the run when gradients are hectic, get the large model past the early training instability hell. Mad but it does work somewhat.
Not really similar to speculative decoding?
I don't think that's what they've done here though. It's still black magic, I'm not sure if any lab does it for frontier runs, let alone 10T scale runs.
Nah, those are completely different beasts. DeepSeek's MLA solves the KV cache issue via low-rank projection - they literally squeeze the matrix through a latent vector at train time. TurboQuant is just Post-Training Quantization where they mathematically compress existing weights and activations using polar coordinates
Somebody took a deeper look at Claude Code and claims to find evidence of Anthropic's PaaS offering [1]. There's certainly money to be made by offering a nice platform where "citizen developers" can push their code.
From Astral the (fast) linter and type checker are pretty useful companions for agentic development.
I don't think this holds because we're talking about developers who know how to use a package manager, on a piece of software you have to install anyways. The friction of "uv add $other_llm_software" is too low for it to have a real impact.
I think they're more into the extra context they can build for the LLM with ruff/ty.
I don’t think they’re targeting the C suite with it, because they don’t use uv and Microsoft already has Copilot for the “it’s bad but bundled with stuff you’re already paying for” market.
idk, i think it's the other way around. I imagine in 5 years my new laptop setup will look like:
$ curl 'claude.ai/install?key=abcd123' | bash -e
$ claude 'finish laptop setup from http://github.com/justjake/Dotfiles'
claude will be the one to install / set up the system, not the other way around. claude was certainly the one who installed `uv` on my current machine.
The most straightforward one: They run a lot of computational sandboxes that need fast setup. Making sure you can shape the package manager to your needs is a pretty straightforward desire.
I'm not so sure. I sort of wish they hadn't been acquired because these sort of acquihires usually result in stifling the competition while the incumbent stagnates. It definitely is an acquihire given OpenAI explicitly states they'll be joining the Codex team and only that their existing open-source projects will remain "maintained".
The value is to control the tool chain from idea to production so it can be automated by agents. It's no secret that the final goal is to fully replace developers, the flow "idea to production". It's easier to control that flow if you control each tool and every step.
I won't be surprised if the next step is to acquire CI/CD tools.
There is the literal benefit of "we use the hell out of this tool, we need to make sure it stays usable for us" and then there is what they can learn from or coerce the community in to doing.
I don't know about OpenAI using a lot of Python, but Astral builds all their tools in Rust and just exposes Python bindings. Codex is all Rust. It feels like a reasonable acquisition from that perspective. They're banking on at least in part on the Astral team being able to integrate with and supercharge Codex.
Why do you think that uv, etc. will stay maintained? They will for now, but as soon as cash is tight at OpenAI, they'll get culled so fast that you won't see it coming. This is the risk.
I share the feeling; but people using it are mostly non-technicals (despite the 50+ config files lol) and are just runing it constantly to do random things.
But a message bot + Claude Code/Codex would be the better version
I tried it for 2 days and honestly don't see the usefulness either. Although, the big reason is that I paired it with Claude, which only uses the per token billing method. Here are the few improvement on a simple Claude usage:
- As you mentioned, the message bot thing was kind of cool.
- It can browse the internet and act (like posting on MoltBook, which I tried).
- It has a a permanent "memory" (loads of .md files, so nothing fancy).
- It can be schedulded via cron jobs.
Overall, nothing really impressive. It is very gimmicky and it felt very unsafe the whole time (I had already read about the security issues, but sometimes you gotta live dangerously). The most annoying part was the huge token consumption (conversations start at 20k+ because of all the .md files) and it cost me roughly $12 for a few hours of testing.
Non-technical people haven't even heard of OpenClaw or Github, let alone know how to use and deploy them. Non-technical people don't even know what OS their Samsung or iPhone is called.
If you can find something on Github and deploy it on your system, you're part of the technical crowd.
>My hairdresser knew all about it and had ordered a Mac mini.
Your hairdresser can't be a technical person because they're a hairdresser ?? I know a surgeon who writes FOSS software as a hobby. What does profession have to do with being technical or not? Most technical people are self taught anyway.
I know them very well, and they are not a coder, or a 'technical person' by a broad HN definition.
What I'm saying is that we are at the point where technology is so pervasive in our society, and the lure of AI so seductive, that many more people are excited to try things out than I might have expected.
I suppose it has similarities to the early to mid 1980s and the home computing revolution. Where many people thought they should have a computer at home, even if they were not sure what they'd do with it.
It will maybe solved soon if we train yet another neural network on scanning GitHub activities; but by also adding other forges like codeberg, gitlab, self-hosted forgejo, etc... to not lock non-github users out
Yeah, scanning non-GitHub is on the roadmap and really should be done. I expect there would be value in understanding all of the current GitHub competitors. And I think the forecasts of new GH competitors getting launched (likely by AI companies) will become relevant in the near future.
The 1M window might be usable, but it will probably underperform against a smaller window of course.
reply