And a major open science release from Prime Intellect: they don’t stress it enough but SFT part is beyond post-training. This is a fully documented mid-training with tons of insights/gems on MoE training, asynchronous infra RL, deep research. storage.googleapis.com/intellect-3-...

SAM3 dropped for those who celebrate! Similar to SAM2, it can segment stuff based on points and track stuff, but now it can directly segment stuff based on text and image prompts, too. Webpage: ai.meta.com/sam3/ Repo: github.com/facebookrese...

Me: I want to have more friends Tech companies:

I love widely used high quality datasets.

Just finished Clevatess, first season. Super solid dark fantasy, with an old school vibe, but enough fresh twists to keep you hooked. Strong personal contender for aoty 2025.

Comparatively tiny models, trained on purpose made data, can into reasoning. Beautiful work!

Surprising: Math requires a lot of memorization Goodfire is at it again! They developed a method similar to PCA that measures how much of an LLM’s weights are dedicated to memorization www.goodfire.ai/research/und...

This is the first company that is unironically making actual gynoids and it's not even for porn reasons. They just are.

Cursor made an LLM it’s called Composer, it’s an extremely fast model that was previously available under code name Cheetah it’s an MoE trained in fp8, RL’d on Cursor Agent traces cursor.com/blog/composer

Hundreds of hours of European driving data from NVIDIA! 1700 hours total

New company announced, with the intent of making GPU programming better with Rust: www.vectorware.com/blog/announc... The founding team has impressive Rust credentials. They're targeting a wide range of usecases, not just ML.

codex is definitely "senior engineer" material because it takes forever to think about it before it tells you to fuck off

Sometimes, when writing throwaway UUID v4s, I feel bad about exhausting the global uuid supply. It can't run out, can it...

is fiction a superstimulus?

Incredible that you can... just have Sonnet 4.5 at home. Blog post here: z.ai/blog/glm-4.6

Sonnet 4.5 Better than Opus 4.1 on almost every benchmark Still the classic Sonnet prices, $3/$15

Alibaba released Qwen3-VL The flagship model Qwen3-VL-235B-A22B is released as open-weight and available in both Instruct and Thinking versions ✅ Instruct outperforms Gemini 2.5 Pro on key vision benchmarks ✅ Thinking achieves state-of-the-art (SOTA) performance on multimodal reasoning tasks

Very cool. More stuff on the bad site: x.com/Alibaba_Qwen...

Spectral Labs SGS-1: Generate CAD geometry from description, cool. The output is a format compatible with CAD tools and so this seems quite practical. www.spectrallabs.ai/research/SGS-1

Feels like a JoJo's episode, amazing.

Moondream3 preview dropped, 9B MoE, A2B, visual reasoning, it's beautiful. Claims it's better than big boys opus 4.1, gemini 2.5. pro etc. @vikhyat.net cooked moondream.ai/blog/moondre... x.com/vikhyatk/sta...

Sometimes you bang against the walls of your own flesh. www.youtube.com/watch?v=tzvM...

In addition to some ambient/techno classics pounding the brain, I also find folk music extremely stimulating for doing work. But only on the condition I can't understand the lyrics. What's this about? No clue open.spotify.com/track/5B4lRo...

Us europoors eating good for once. Supported Languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Russian, Ukrainian

recently gave a talk on <Reality Checks> at two venues, and discussed (and rambled) about how leaderboard chasing is awesome (and we want it to continue) but that this isn't easy because everyone (me! me! me!) wants to write more papers. the link to the slide deck in the reply.

Out of all the model drops today, Genie 3 is the most mind boggling. Good job! Shame all the juicy details are locked down tight...

gpt-oss, OpenAI's open weights model 120B & 20B variants, both MoE with 4 experts active openai.com/index/introd...

people in many fields seem to be doing the left part for ai-generated text submissions of some kind and haven't quite made it to the right part yet

The biggest update in 3D reconstruction world - VGGT (new weights) has been released for commercial usage as well. Kudos to @jianyuanwang.bsky.social to make this happen! github.com/facebookrese...

32B apache2 and opus4 in the same image, waow.

K2 writing small bangers

Moonshot AI's Kimi K2 - the best open weight per their eval. huggingface.co/collections/...

it’s new entrant week! today? Kimi-K2 an open weights model that’s competitive with Claude 4 Opus - 1T, 32B active MoE - a true agentic model, hitting all the marks on coding & tool use - no training instability, due to MuonClip optimizer new frontier lab to watch! moonshotai.github.io/Kimi-K2/

her: "what is `**kwargs`?" me: "it's a feature of dynamically typed programming languages, such as Python, that you can use to write horrible code that makes me want to kill myself"

This looks interesting! "Together these results support the idea that the principles of self-attention used in LLMs overlap with the mechanisms of language processing within the human hippocampus, possibly due to similar prediction-oriented computational goals."

Music can only be consumed sequentially, and unlike podcasts/videos and other forms of audio expression which can be conveniently sped up, music... can not be sped up. This is a great tragedy, an underappreciated ceiling of human ability.

Writing a report for a project about VLMs, and the "theory" section feels like trying to catch a speeding train with a fly catcher. Most sources are outdated the moment they've been published. With multi-modal it feels like all variations of it work with enough elbow grease (and data and compute..).

I feel baited.

Wow

Honestly, this photographer was wasted on corporate stock photos.

I didn’t want to learn about Platonism, but I finally caved

The sigil was drawn in salt and ash, the candles lit at the pentagram points, the incantation declaimed. There was a shimmer - a demon appeared. "Curious. What ritual is this?" "I got it from ChatGPT. I included all protections in my prompt!" "I see," the demon said and stepped out of the sigil.

zapping a box of rocks until it summons the omni-comedian to deliver a mediocre joke derived from the thousands of tortured souls it has enslaved

Yeah we finally have a model report with an actual data section. Thanks Qwen 3! github.com/QwenLM/Qwen3...

New blogpost: "Training as we know it might end". It was originally a panorama of the new methods of synthetic generation but the stakes are now much higher and I openly wonder if model training is not soon going to change forever. vintagedata.org/blog/posts/t...

In 1999 philosopher Douglas Lackey polled philosophy professors asking them to rank the 5 most important books and the 5 most important articles in philosophy from the 20th century. Here are the results from the 414 replies:

honestly these are quite helpful. i feel like i’m finally understanding the benchmarks