Limits of vector search a new GDM paper shows that embeddings can’t represent combinations of concepts well e.g. Dave likes blue trucks AND Ford trucks even k=2 sub-predicates make SOTA embedding models fall apart www.alphaxiv.org/pdf/2508.21038

Longcat-Flash-Chat (560B) uh, holy shit this one is intriguing. bare minimum they compare themselves to all the (actual) top models and do okay but inside.. damn this one has some cool ideas huggingface.co/meituan-long...

In 2012 when I had to clean data it seemed natural to look for rules I could use to clean it. Now it seems natural to model the noise, find new clean data it can destroy, and then train a model to reverse the process. Machine learning makes you a sicko.

Three things to note about this: 1) AI has obvious utility to many, this is a tremendous amount of use already 2) There is room for multiple frontier model providers, at least for now 3) Any losses from subsidizing cost of AI use (and it is not clear this is happening) are now relatively small

"The Serial Scaling Hypothesis" (arxiv.org/abs/2507.125..., Liu et al) is interesting I think, not as new as it completely looks (autoregressive models are used serially, models have depth,..) but feels like a good formalization and intuition as of where current GPT based LLMs will typically fail

1/ Can open-data models beat DINOv2? Today we release Franca, a fully open-sourced vision foundation model. Franca with ViT-G backbone matches (and often beats) proprietary models like SigLIPv2, CLIP, DINOv2 on various benchmarks setting a new standard for open-source research.

In the coming age of agents, I think vibe coding will die out, same lasting power as prompt engineering. For things LLMs excell at, you might as well stick to higher level directives and let it own the work, Claude Code is a good example. 1/2

this is probably why Meta was able to poach OpenAI ppl aside from the absolute piles of cash, Sama is very SV-minded and can’t imagine building apart from a product a lot of accelerationists see things differently, more broadly, and ids dissatisfying to be forced into a product box

Still not a lot of ML talk on bsky (at least in my feed), hence paper Sunday: my two most interesting recent reads - H Nets arxiv.org/abs/2507.07955 - Energy Based Transformers arxiv.org/abs/2507.02092

Little bit of personal news, shared in other circles already: I'm moving to Mistral in August, after three years at Photoroom. I'm really proud of what we built in the ML team with relatively limited means, lasting SOTA on the existing foundations (saliency segmentation) while growing a lot on genAI

𝗗𝗲𝗽𝘁𝗵 𝗔𝗻𝘆𝘁𝗵𝗶𝗻𝗴 𝗮𝘁 𝗔𝗻𝘆 𝗖𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻 Boyuan Sun, Modi Jin, Bowen Yin, Qibin Hou arxiv.org/abs/2507.01634 Trending on www.scholar-inbox.com

kyutai open sources its TTS model as well as Unmute, a framework for building audio AI apps notable: - high accuracy - actually streaming (can use streaming text input) - serves 32 simultaneous users on a single GPU - voice cloning - supports all 24 official EU languages kyutai.org/next/tts

Alex Nichol is one the rare many-hits researchers of the field, with on top of that a track record of practical models which affect the public/ship. That Meta wouldn't target him is pretty rich

Automatically generate a fused megakernel in triton.. diving in, but if it works half as well as it reads it would already be quite something. Aligns with torch.compile of course github.com/mirage-proje...

Sharing that Photoroom open sourced _Dataroom_, as promised some time ago. Accompanying blog post and mini thread github.com/photoroom/da... www.photoroom.com/inside-photo... 1/N

Still haven't tried Cursor, but I recently moved from Github Copilot to Continue with Codestral (free API), and it's absurd how much better Continue with Codestral is (vs. Copilot with expensive and slow models). Made me realize that there is zero moat in this field, at least for Copilot.

In the last 2 weeks: - Slack locked down its messages data. - X locked down its post data. - Anthropic cut off OpenAI's Windsurf. - Google will stop using Scale. The dream of unfettered MCP interconnects is a mirage. www.dbreunig.com/2025/06/16/d...

While framed as a critique of Apple’s recent paper, I found this article mostly interesting because it made me think about reasoning in general: mikecaulfield.substack.com/p/the-apple-...

Self adapting language models, still early but fascinating prospects. There's a dimensionality curse of course: since the dimensions the LLM can touch per generated token are very small as such, needs a massive lever / dimension reduction to be able to self improve. arxiv.org/pdf/2506.10943

Great write up of AMD new offerings, catching the nvidia train on the software side it seems. 3x speedup on MI300X since release, was required but still great to grab morethanmoore.substack.com/p/amds-ai-fu...

Got nerdsniped into printing this a little while ago.

datago now available with webdataset compatibility (streaming tarballs, so you get the data as it arrives). Just pip install datago and give it a whirl if you'd like ? Speed without the dataloader processes, and typical ViT/DiT pre-processing baked in. example code here github.com/Photoroom/da...

Great link with bsky.app/profile/dbre...

SageAttention3 paper reads great, and looks like B200s just got a good value boost. QAT or PTQ-free use of FP4, I expected this to be much more complicated or come later to be honest. Only at the attention level and LLMs are most often MLP bottlenecked but stil arxiv.org/abs/2505.11594

Big Marigold update! Last year, we showed how to turn Stable Diffusion 2 into a SOTA depth estimator with a few synthetic samples and 2–3 days on just 1 GPU. Today's release features: 🏎️ 1-step inference 🔢 New modalities 🫣 High resolution 🧨 Diffusers support 🕹️ New demos 🧶👇

Striking in retrospect how some leaders position at the time of the first atomic bomb (“other countries won’t get it”), Truman for instance, then space age, then now AI age, rhyme. Same as before, I think a bunch of places are bound to be SOTA AI, ideas and progress cannot be pinned to a wall

Apple readying in house server grade AI processors feels quite bizarre to me: Apple Intelligence has been underwhelming so far, so current status is probably far from end game. But lowering the code now into hardware limits future flexibility, bad timing ? www.tomshardware.com/pc-component...

[1/2] We've released the code for LegoGPT. Our autoregressive model generates physically stable and buildable designs from text prompts by integrating physics laws and assembly constraints into LLM training and inference. Code: github.com/AvaLovelace1... Website: avalovelace1.github.io/LegoGPT/

Some men just want to watch the world burn.

Now (early) supporting webdataset. Draft PR, but runs and ok speed which seems competitive with the webdataset python lib (no extra python process here) github.com/Photoroom/da...

It’s been a couple of years, but I’m still looking forward to these evals somehow

Qwen 3 offers a case study in how to effectively release a model simonwillison.net/2025/Apr/29/...

🚗🌆 We introduce EVolSplat — a feed-forward 3D Gaussian Splatting model that enables real-time, photorealistic rendering without per-scene optimization. Trained on KITTI-360 & Waymo, it sets a new SOTA for autonomous driving applications. #CVPR25. Paper & Code: xdimlab.github.io/EVolSplat/

Test time compute is there, clearly, with small models outperforming bigger ones (not just thanks to this, but counts). qianwen-res.oss-accelerate.aliyuncs.com/assets/blog/...

🔥 Prometheus: 3D-Aware Latent Diffusion for Feed-Forward Text-to-3D Scene Generation. Allows object- and scene-level generation from text in seconds. freemty.github.io/project-prom...

I asked "on the other platform" what were the most important improvements to the original 2017 transformer. That was quite popular and here is a synthesis of the responses:

Some progress in my side project of a rust written webdataset python dataloader (flavor of datago). Proof of concept running, but I’m not actually all that familiar with webdataset: anybody to warn me about gotchas from this file format ?

huge 1T+ models are fascinating bc they’re like tree rings. they take so long to train that several evolutions of LLM architecture happen during the process in this case, DeepSeek in January was unignorable, but Behemoth was likely too deep into training to change course. hence scout & maverick

Really cool idea and implementation, model compression via huffman coding, fast enough to be used live ! Side note is that there are probably model arch questions stemming from the initial redundancy / weights, would be even better if it was not there to begin with arxiv.org/abs/2504.11651

The geoguessing power of o3 is a really good sample of its agentic abilities. Between its smart guessing and its ability to zoom into images, to do web searches, and read text, the results can be very freaky. I stripped location info from the photo & prompted “geoguess this”

TransMamba: Auto switch between transformer and mamba architectures It uses state space modeling (SSM) to decide between two successful AI architectures, based on token position. Unfortunately DOGE banned it last week arxiv.org/abs/2503.24067

All you wanted to know about Ironwood/TPU v7p Chip + interconnect look pretty strong indeed www.nextplatform.com/2025/04/09/w...

Less than two years ago, Google's LLM (then named Bard, now Gemini) was... not good, as you can see in this example from May, 2023. Now, as you can also see, it is much better. Quite a turnaround for a large company that seemed to be struggling with the challenge of shipping LLMs.

Interesting write up from the HFT Guy on a recent stream of conda related lawsuits, the Intel conclusion is quite something thehftguy.com/2025/04/07/a...

Multi Token Attention, another great fundamental paper from FAIR arxiv.org/abs/2504.00927 It focuses on the fact that the attention is single shot, token to token, and softmax will put the emphasis on, well, the maxima or maximum. MTA makes it possible to accumulate "hints" over multiple tokens

Interesting bit, I cannot make complete sense of it but curious to learn more. Really non obvious, nice that Meta reported out and ablations will be a learning

MoE is becoming the norm but it really took a while when compared to other changes in the field. RLHF, reasoning, synthetic, arch tweaks were all ubiquitous within months, Switch is from 2021 arxiv.org/abs/2101.03961 Scout and Maverick seem very impressive here

Part 2 of SLAM handbook is out for public comments! let us know what you think :-) Issue tracker on GitHub awaits! Link: github.com/SLAM-Handboo...