Carsten Kragelund

@JustinGorya @xeophon @koltregaskes How does an unreleased model qualify as "Open"?

@nyxkrage @xeophon @koltregaskes Told ya.

Blimey, Nvidia, "best open base model", this is quite pedantic. 😀 They are comparing to old *base* models, Kimi K2 and "GLM" (probably GLM 4.5, so this should have been labelled), which are both from the middle of 2025. But in comparison, the newer (non-base) versions are hitting 86-87% on MMLU Pro this year. Where Nvidia really does perform, though, is speed and context window.

@JustinGorya @xeophon @koltregaskes GLM-5-Base exists, but wasn't released. arxiv.org/abs/2602.15763

@nyxkrage @xeophon @koltregaskes Give me the source of GLM-5 being not the base model. There is no way you can convince me that GLM-4.5 is the base model, which is literally half the size and has a different frontend dataset. Btw nobody likes a know-it-all guy

@JustinGorya @xeophon @koltregaskes Try again

@nyxkrage @xeophon @koltregaskes They are the base models.

@JustinGorya @xeophon @koltregaskes Where's the GLM5 and Kimi K2.5 base model then?

@xeophon @koltregaskes GLM5 and Kimi 2.5 I would guess.

Blog on methodology and what I believe the results mean: blog.pid1.sh/chastity-bench/ Leaderboard available on HF: huggingface.co/spaces/NyxKrag… Eval code: app.primeintellect.ai/dashboard/envi… Note: Current leaderboard results are from an earlier implementation but all the parts are still equivalent.

Finally getting around to officially publishing ChastityBench! Benchmarking vision models on recognizing chastity cages without being directly prompted for it. Stop testing general vision capable models on more and more graphslop and OCR tasks.

Releasing mlsweep, a sweep scheduler and visualizer for distributed ML training. It aims to make launching runs across groups GPUs frictionless and achieve near feature-parity with wandb. But you can use it with whatever frameworks or loggers you like, wandb included.

NieR Automata cosplay I made a while back 🤖

we are releasing a series of research journal posts at noumena.com/research/ . these are some of the things we are working on at noumena including 2 (very) pre print paper drafts. happy to answer any questions you have . i will push the code and logs to reproduce most of this today in the nmoe repo

the last missing piece. our SFT training dataset is now available at huggingface.co/datasets/stepf…

@xeophon are you sure you want to look at my data?

time to look at your data, anon

We should all just use UTC

RuneBench is out: measuring long horizon goal optimization across 14 AI coding models inside Runescape

shift-enter does not exist as a concept in terminals, most terminals hack around this by sending Esc-Enter/Alt-Enter instead (they are the same thing), or they follow the xterm other modified keys "spec", or they simply just send enter, as it "correct" for the vt100 terminal that most terminal emulators are supposed to be emulating. You can test this for yourself by running the `Profiles: New Window with Temporary Profile` command in VS Code, opening the integrated terminal, and then running `showkey -a` and typing shift-enter, followed by enter. You'll see the terminal receives the same keys in either case.

Shift+Enter works fine in TUIs, it sends a newline. Ctrl+Enter doesn't, because terminals intercept Ctrl key combos at the OS level as control signals. Ctrl+Enter never even reaches the app. Two workarounds that can actually work for everyone: > Triple-tap Enter to submit 3 quick Enters = send. 1 or 2 = newline. > // or /send on a new line = submit Enter gives you a newline freely. // fires the prompt. @theo agree ??

The fact that “shift+enter” is so hard in TUIs is insane. Maybe the terminal isn’t the right place for us to be writing complex multi-line prompts?

@kalomaze probably deserved tbf

My Most Supportive Mutuals Btw

i want a woman to piss in my mouth

I have made the difficult decision that I will be leaving @PrimeIntellect. Working for Prime Intellect has been one of the most fulfilling things I have done, getting to work on many cool projects, even if sometimes moonshots. This pains me even more, given that I deeply value the role of open source in this industry & the ability to openly talk about ideas. However, I’ve recently reached the conclusion that what’s required next is a degree of scale that has been historically difficult for open source when not subsidized by other highly profitable ventures. It’s for this reason that I think it is necessary for me to continue my work elsewhere. I’m looking forward to when I can soon share more as for what’s next. I’ll continue to work on the long-standing dream of mine towards a true multi-modal streaming recurrent model trained on everything from Guitar tutorials to long-running minecraft let’s plays, which I think at this point in time is only bottlenecked by flawless execution of a theoretically sound concept. I’m also extremely grateful to PI for agreeing to allow me to continue to work seamlessly on the repositories I have built up at PI. It is not a given that companies truly live by the principles of openness and a no-zero-sum mindset. The repositories I’ve been working on are now on my GitHub, even if it is about as messy as research code gets. While the current experiments **technically** haven’t left the text domain depending on how you define text, I would guess the next step from here is already somewhat apparent. github.com/mikex86/fbamtr… github.com/mikex86/termst… github.com/mikex86/pyterm… huggingface.co/datasets/mikex… I have high hopes however, that some of the parts I consider most interesting will remain open and will do my best to ensure that is the case going forward. I will leave with nothing but fond memories of my time at PI and truly wish them the best going forward. It’s been a blast working with such a talented team. Best of luck, and godspeed.

@SIGKITTEN Would it be possible to get it running on iOS 18? I really dont want to deal with liquid ass

Litter iOS just got a beta (TestFlight) release so now you can install it and try it out. It's a remote for codex! - Finds all the computers on your local network (or tailscale) that run codex app-server or ssh - Lets you continue any session or start a new one - Experimental local codex with oauth support Free, fully open-source, fully native, with android release coming soon! Thanks to @iwanttobeinmars @dixiidev @ethnologism for the contributions! TestFlight ⬇️

Normally I wouldn't say anything at this stage but I feel that the moment calls for it. This isn't directed at any particular company (I'm serious), but it's something we should all start thinking about now. I'm moving to NYC to join @pangramlabs The midterms are coming up, and I'm worried about AI generated misinformation and other misuses of AI. Even right now the bots are on overdrive bc of the war, this is bad. My old job paid ~210k/yr, more money than I knew what to do with. (that's also why no new blogs/research) I don't expect to stay at pangram long enough to vest, and so in all likelihood I will be taking a 40k/yr pay cut to do this. And of that income I intend to donate a good amount (political) too. I will be leaving all my friends here. I've gotten vastly better at making new ones but is still very uncomfortable. And I will be going in person instead of working online. (None of this is pangram's fault btw I intentionally didn't negotiate, and they're aware of all the details) My questions to you are: 1. How much is your conscience worth? 2. How easily can your conscience bend? 3. Will ambiguity, intentional or not, satisfy you? 4. Will you simultaneously be capable of good compromise when the time demands it? 5. Can you, right now, in good faith, say that what you are doing is right? Our society is falling apart at the seams, trust is degrading, loneliness is at an all time high, antisocial behavior is common, gambling is everywhere, influencers sell out the truth and their viewers for money. We are building an extraordinarily powerful technology while it all happens. Are you prepared to be the people the world needs right now?