magnesit

18 Apr 2025

Due to a shocking turn of events, @grapheneos.org is urgently searching for one or more senior developer(s). Do you have extensive experience working with the AOSP? Passionate about cyber security? GrapheneOS might be the perfect flexible, fully-remote workplace for you. See grapheneos.org/hiring.

20 Aug

People who oppose or hinder mRNA research are literally traitors against humanity, and today showed just that again

18 Aug

The Pixel 10 is a surprisingly shitty smartphone. I bought it, coming from my Pixel 7a, and it's genuinely not an upgrade. Can absolutely not recommend. Battery life is, at best, mid for a newly bought phone with almost no apps installed, and 128 GB of storage is genuinely tight for the price.

18 Aug

The CPU is faster, but heck, you don't need that in everyday life. The 7a was more than enough from a performance standpoint already. 12 GB of RAM is nice though. The camera is not that much better. Sure, I can zoom 20x now, but why? I want storage and battery life, not... that.

18 Aug

The display is lower-resolution than my Pixel 7a. It's brighter, but the 7a was already bright enough. I love the design of the Pixel 10. It's much better than this blocky atrocity from before. All of that for a higher price. It's enshittified. Been 3 years and almost no technological development.

1 Aug

Reinforcement Learning From Human Feedback

Digital Brain@yourdigitalbrain.bsky.social

A deaf man's cat figured out that meowing was pointless, so he taught himself sign language to communicate with his owner. No training. Just observation.

14 July

Why does it have that Starmer look on the eyes

This post is unavailable.

Hmm. Reading the numbers, Grok 4.5 seems to be an absurdly good model. And relatively cheap. The high price you pay isn't in dollars, though... Credit where credit is due

Tested it for a bit. Alignment is very pro-Elon, but not as disastrous as I thought. Still problematic.

What the fuck, EU. Losing respect at a record pace. You know you'll never be able to get this through in practice. What the fuck.

For my German readers, SPD and CDU/CSU representatives voted for Chat Control. It passed. Do with that info whatever you want.

23 June

Listening to Ribs by Lorde when a cool breath of air streams through your window while laying on the bed in the heat is corny but peak as fuck and you ain't gonna change my mind

16 June

Can Veritasium talk about a different topic than Prime Numbers for more than 2 consecutive months ffs 😭

16 June

Mistral can do the funniest thing now... Pleeeease let them release Le Chaton Fat

6 June

Gotta love these GitHub accounts that fork every repository in existence, do nothing to it and let leave, refusing to elaborate Absolute cinema

4 June

lol it's happening, GitHub copilot is pulling the rug so hard

16 May

Sorry but can anyone confirm this? They released more models than ever last month???

POLITICO Europe@politico.eu

EXCLUSIVE: Several Chinese tech companies have recently stopped updating previously open AI models, which could indicate they're moving development behind closed doors, Germany’s cyber chief has warned. 🔗 politico.eu/article/chin...

12 May

How does this keep happening? Bluesky is a platform full of very loud anti-AI people, but still, it's the single platform that drowns in slop and and everybody keeps liking these images. Again, what do you want, man? How does this keep happening?

This post is unavailable.

12 May

It's a self-sustaining cycle of rage bait. Many accounts spam AI generated images, and then many other accounts spam how much they hate these images in the comments. So weird

magnesit reposted

Something's odd about the launch of Mistral Medium 3.5, and it's not their fault. Artificial Analysis paid close to $1k to benchmark the model but still defaults to showing Mistral Small 4, a much worse model, on their front page. Medium 3.5 scores way better (Deepseek V3.2 level).

When OpenAI launches "GPT-5.5-Codex-mini (xhigh Special Edition)", support is there within 1/10th of a second, but as soon as a Mistral model is remotely okay-ish, it's professionally ignored. Leaves a very subtle and weird taste. Again, I hope this is just a genuine coincidence.

OpenRouter has literally spent time adopting OpenAI GPT-4o-mini Transcribe and Whisper Large V3 Turbo this (very young) May, as if there's nothing else to do. As you can here, all of that makes me just a little bit salty.

However, they pretend like it's not there. You have to specifically search for it to see it there. I believe that this is not done with malicious intent, but it is weird nonetheless. Another example: OpenRouter. Not providing the new model. Support for Small 4 was there quickly though.

What's going on? Adoption of this model is surprisingly slow. I don't understand why several services pretend like Medium 3.5 does not exist, but Small 4 does. Historically, adoption of new models was much faster. Let's give it time. It's annoying though - this is REAL revenue Mistral is losing.

The new pricing of Medium 3.5 suggests Mistral is in need of money, probably because of failed training runs or low demand. They were able to provide Devstral 2 (123B dense) at a much lower cost than Medium 3.5 (128B dense), even though it should effectively be almost identically heavy on hardware.

Mistral dropped the bomb: a 128B DENSE Medium 3.5. They indeed seemed to take a Devstral-like architecture (same hidden size, attn heads/dims, among other things), maybe even literally a Devstral checkpoint, and added a vision encoder on top. But not just that: They added reasoning 1/5

model Mistral has released in a long time. Now it's a good time to wait for independent benchmarks to come in and listen to the user experience of others. I've not been able to try it out myself for now, but definitely feel like it's going to be fire! Unfortunately hard to run locally. 4/5

mistral.aiRemote agents in Vibe. Powered by Mistral Medium 3.5. | Mistral AIIntroducing Mistral Medium 3.5, remote coding agents in Vibe, plus new Work mode in Le Chat for complex tasks.

Mistral did a good write-up about it here: mistral.ai/news/vibe-re... 5/5

and strong agentic capabilities to it. This might actually be it. This might be a ticket back to the frontier-iiiish (doing a lot of heavy lifting here). It's probably slow and expensive to run at first, but if more inference providers pick it up, hey, why not? 2/5

Benchmarks looking promising, but not SOTA. There's still room for improvement, but for now, it seems like they're doing okay. They've also told us that Medium 3.5 will be the new model powering Le Chat from now on. It's priced at a whopping 7.5$/M output tokens, the most expensive 3/5

Gemma 4 E2B reaches a whopping 620 tok/s prefill speed on the dated Tensor G4 GPU! This is truly impressive. Shows that simply interleaving sliding window attention with full attention is still a worthy technique. Wouldn't have guessed that, to be honest.

New open-weights, likely multimodal+dense Mistral Medium model (128B) inbound for tomorrow, hopes are high as always! It's still unclear whether the model has reasoning baked in, as you can't see that from the architecture alone, but if it has, it could become a real banger!

github.com/vllm-project...

github.com[FEATURE] Add EagleMistralForCausalLM (#41024) · vllm-project/vllm@e9f8f31Signed-off-by: juliendenize <julien.denize@mistral.ai>

It seems like no new PRs have been opened in the meantime. I suspect the model will be released around 15-16:00 MEST (09-10:00 Pacific), that seems like a reasonable time... (this is going to come back and bite me)

Hm. Interesting. So far, there's only code for Eagle speculative decoding models to accelerate the bigger Mistral models on the vLLM repo. There is a certain level of ambiguity to the naming. Large 3 didn't have any active parameter count in its name, so Medium 3.5 might be MoE.

That would be great to see and might make Medium 3.5 actually a deal for local setups as well. Let's see until tomorrow. The rest is pretty standard stuff, and Eagle speculators have been published for Large 3 as well. Let's just hope Medium 3.5 isn't just Devstral + vision encoder.

Essentially an update to the dated Magistral Medium 1.2 (proprietary). Love to see Mistral shifting to open-weights again. I'm going to dig into the merged code now to see whether there's any interesting architectural tweak applied to it!

"merged code", wtf did I write, I meant the code in the merge requests for Transformers and vLLM. On a side note, Mistral seems to be the only lab left that reliably pumps out huge dense models if the prophecy comes true. Lemme check.

Deepseek V4 finally released today, if you haven't noticed yet. I've spent some time reading their technical report and benchmarks and my feelings are extremely mixed. I am full of thoughts, but I feel like it's too early to pinpoint most things for V4 yet.

The competition, however, can happily use Engram now. This is likely the most annoying part for the team. It's such a pity. Had they started training any week later, it might have made it in.

Fuck, my writing already sounds like it's been generated by AI... I'm definitely reading too much LLM-created slop. Rest assured that I actually typed all of that myself, if you wanna believe me.

I like V4. It seems like a solid release, and it's definitely an improvement over V3.2 by far. Chinese AI Labs are still building the best open-weights LLMs, and they'll probably do so for months to come. I wonder when we'll see Engram in a production release next.

It's not worth it to start another training run for V5 already, just to get Engram in. It's also not possible to continue training V4 with Engram added to it as an afterthought, at least not trivially. Thus, I do not believe we'll be able to expect a "V4.1" with Engram anytime soon.

I am aware that Deepseek is primarily focusing on ML research quality. They simply don't aim to push out a state-of-the-art LLM every other month, and that's fine. This is why I'm even more surprised by V4 not really surpassing K2.6 by that much.

You see, they incorporated all of these crazy new steps and architectural innovations like mHC into their new model, but it's not that much better - if at all, depending on the benchmark - than the previous open-weights frontier models. All while having an enormous 1.6T+ parameter count.

People are expecting another "R1 moment" from Deepseek, showing the Western AI companies that SOTA can be achieved with less money and recourses. In a way, Deepseek was capable of doing this again, but the effect is much smaller this time. People are already aware of their capabilities.

Unfortunately though, as it seems like right now, the V4 models fall into the "barely SOTA" category instead. FP4 training was pioneered by Moonshot AI for K2 Think, as well as the enormous parameter counts of 1T+. All of this is also incorporated into V4.

API pricing might develop rapidly (what I mean is, it'll get cheaper) as more compute comes online. Deepseek is still working on improving inference and getting more HUAWEI Ascend NPUs to join their infrastructure.