Knowledge catalogue
DISCLAIMER: the post and the model card was made with the assist of AI. Hiya, I’m releasing Recursive BitNet N-Gram 102M, a small experiment combining ternary weights, shared transformer layers, and h
Alright, I learend a few valuable lessons since I posted about a week ago, and I've made more strides with the DDR4x Ascend 96GB 310P cards I purchased, so here we go, meet the new friendly and less v
SAM 3.1 + DINOv3 from @AIatMeta is a crazy combo 🔥 it can even hold onto surgical tools while a forearm sweeps over them. one leaves for 15 seconds, its near-identical twin stays, and it knows exactly
Banger paper from Meta Superintelligence Labs on user simulators for agent training. Agent RL setups usually let an assistant LLM play the user, so the simulated user is too cooperative and too explic
On-device personal AI requires models and agent software that can operate within fixed hardware limits. Developers also need ways to keep those systems improving after deployment. Model builders are r
arXiv:2610.08618v1 Announce Type: cross Abstract: A robot that offloads its control loop to the network carries the receiver with it, so link quality is decided by where it goes. Simulating this requi
Our SynthID technology lets you easily verify if an image, video, or audio file is AI-generated, and today we’re expanding access. Now anyone can go to http://synthid.com to check files for a SynthID
GPU utilization during post-training depends partly on how efficiently infrastructure moves data and loads updated models. As enterprises continually refine AI agents, reducing delays between training
Sabrina Ortiz / The Deep View: Mistral releases Mistral Large 4, dubbed “le Chonk”, a 1T-parameter open-weight model for general agentic capabilities, trained on 4,000 Grace Blackwell GPUs — Most flag
Reflection AI Inc. today introduced Beam, an open-source large language model with 501 billion parameters. The launch comes a few months after the startup raised funding at a 25 billion valuation. Aro
AI data infrastructure must support growing GPU clusters while making enterprise data usable for models and agents. NetApp Inc. is tying intelligent data infrastructure to production AI through a unif
Hi folks, LIVE PROJECT PAGE TLDR: Moral of the story. You need better CPU to do actual agentic coding doing real work... I've been on a mission to make my RTX5090 go brrr for past 2 months so much so
I forked Strata and worked with Opus 5.5 with some heavy changes to it to make it work on an IBM AC922 I have access to. The IBM AC922 is a 2018 era beast with two POWER9 20 core SMT4 CPUs that are co
Banger paper from Google. It's standard practice to sample several agent rollouts and trust the answers they agree on. This Google paper shows that agreement can hide shared errors, while disagreement
From the first LLaMA 33B I knew I wanted that magic-like intelligence locally, mine, so nobody could take it away when I needed it. I bought a 3090 for my home PC. Then LLaMA 65B appeared and I was
Meir Orbach / CTech: Nasdaq-listed neocloud Nebius acquires Inferize, whose tech helps optimize GPU utilization and reduce AI request costs, sources say for 100M to 150M — The Israeli startup, which w
Editor’s note: AI21 Labs is a leading global AI lab with a long track record of building foundation models, most notably the Jamba family, and today focuses on specialized LLMs and agent optimization
arXiv:2609.38216v1 Announce Type: new Abstract: Existing benchmarks evaluate tabletop manipulation, flat-floor household activity, or humanoid locomotion and manipulation as separate task groups; none
This is impressive! Face-to-face conversation is where AI interfaces are heading. @tavus gave me early access to Griffin, a video-to-video model that watches and listens as you talk. It can interrupt
Token economics are becoming a defining measure of artificial intelligence infrastructure efficiency. As AI workloads move beyond training, the economics of producing useful intelligence are starting
I got thousands of DMs and it looks like @bot can't automatically analyze DMs (when @elonmusk?) so it will take me a few days to go through all of them, sorry! If you don't get an answer, don't take i
In light of the latest Anthropic's latest not-particularly-oblique attempt to build a case against open weights, I thought I would write down how I currently think about open weight AI models. (1) Ope
So I would like to share my analysis of PD disagg that I've run by Horace and many folks at Thinky You may have heard that PD disagg helps with per-token 'tail latencies' in ITL/TPOT, as prefill inter
When scaling up agentic reinforcement learning (RL) and evaluation across massive parallel rollouts, frontier AI labs inevitably hit a bottleneck: Expensive GPU clusters sit idle, waiting minutes for
Bloomberg: After lunch with AI leaders, Trump rejects new federal AI rules, says “we automatically have regulation” via DOJ and FBI but “self-regulation is very important” — President Donald Trump sho
We’re collecting real stories of builders, tinkerers, researchers, and creators who are using Codex to do incredible things. If you want to be a part of the next chapter of the Codex Originals program
vulkan: fix wrong results when a mul_mat reads a slice of a larger cache (#28956) vulkan: read the batch stride of an in place src0 from nb[2] A dim01 contiguous tensor can still be a view whose batch
server : support typed content (vision/audio/video) input for /v1/embeddings endpoint (#29556) server : support multimodal input for /v1/embeddings (Qwen3-VL-Embedding) Accept the OpenAI-style wrapped
From what we know (take with a grain of salt, we need much more transparency!), if @OpenAI had been running this on their own agents that attacked us, they would have caught them before we did! Since
arXiv:2609.31383v1 Announce Type: cross Abstract: End-to-end driving policies are commonly trained through open-loop behavior cloning, yet they must ultimately operate in closed-loop when deployed on
I just ran a paper-research agent on an open stack and had this week's top 5 papers on agent memory in my terminal in about 10 seconds. How does it work? I used Tavily to pull the last 7 days of arXiv
arXiv:2609.30854v1 Announce Type: cross Abstract: Autoregressive LLM inference at long context is bounded by memory bandwidth, not arithmetic throughput, and the binding resource shifts from model wei
An interesting idea is to continually train specialized models on skills. This paper explores that idea. They propose SkillGym which turns human-written skills into 2,756 training environments with co
Cool to see classification models resurge in the era of agents. The effects of Jev are insane. Also very cool to see our emotion dataset built at @dair_ai used to test this new model, Julia-1. I spent
About a month ago I was having FOMO and was going to spend coin I don't really have on new graphics cards. Instead of doing that though I decided to spend the money on a new motherboard+cpu and try to
GLiNER2.5-Decide (DeBERTa-v3-large, 340M) converted to LiteRT: fp16 weights, answers identical to the official fp32 model on 361/361 test requests, 66–70 ms per request on a Galaxy S26 GPU at 128 toke
I've been playing around with Ling 3.0 Tiny, which is an 8 billion parameter model (MoE, 1B active). And I've had a lot of poignant thoughts as a result. Just for fun, I got it running with llama.cpp
Decision models should do more than pick a label. To truly act, they need to extract evidence, understand relationships, produce structured records, and keep related decisions consistent. And they sho
Illinois folks (like myself), write your governor, please. And if you live elsewhere, please pay attention to your local/state politics. https://wgntv.com/politics-3/pritzker-establishes-ai-cabinet-to
I gave Jev, Laya, a finetuned ModernCE-base-nli and a finetuned Qwen3.5-4B the controls to Doom. Thanks to TypeSafe AI for Jev access. The video lines up the starts of four separate games with the sam
Don't get me wrong, not hating on Qwen 2.1 but it DOES have all the hallmarks of GPT's weird 'texture' all over it. Is this just a distilled model of one of the GPT Image models? Similar how many LLMs
Master class in spin. Is it really a “positive for the AI infra trade” that OpenAI and Anthropic are steadily losing market share top open models and on a track to have hardly any? What happens to Cor
really great thread/benchmark for common sense in robotics and a reminder that we are not actually close to AGI GPT-6 Astra attempted harmful actions 97% of the time when it was asked to stab a human-
arXiv:2609.20034v1 Announce Type: cross Abstract: We present Astronex-World 1.0, an open controllable video world-model foundation. Given a text prompt (text-to-video) or an initial observation (image
Been building this for a few months, mostly for myself, and it just got a proper release so figured I'd post it. It's a native GGUF inference runtime with OpenAI/Anthropic-compatible APIs and a chat U
Amazon SageMaker HyperPod Inference Gateway is a Kubernetes-native, GPU-aware routing add-on for Amazon EKS. It uses real-time GPU signals to send each inference request to the best-suited pod, cuttin
🔥: “While AI comes with danger, we should be sceptical of any narrative that conveys inevitability around its trajectory. AI systems do not build their own data centres. They do not manufacture their
Basically the title. Training models take time and if a model was already in the process of pre-training/RL then it can't be re-cooked with any new architectural improvement that has been discovered i
Banger report from Salesforce. Pretty interesting to see more of these custom enterprise models. Salesforce trained the enterprise agent model from the same files it uses to configure agents. Koa star
MLPerf Inference v6.1 introduced the first agentic workload run on datacenter hardware and evaluated a model with more than one trillion parameters (Kimi K2), setting a new benchmark for scale. The te
arXiv:2408.15538v4 Announce Type: replace Abstract: While modern Autonomous Vehicle (AV) systems can develop reliable driving policies under regular traffic conditions, they frequently struggle with s
arXiv:2609.13551v1 Announce Type: new Abstract: Static-image benchmarks do not capture the computational and temporal requirements of practical orchard video analytics. This study presents an end-to-e
okay wow. per folks there, Trump called Jensen huang in the middle of his interview and jensen put the phone on speakerphone Jensen thanked trump for his social media post this morning pouring cold wa
Law enforcers already have authority to charge companies and their CEOs for creating and releasing dangerous, unvetted, or defective products. We shouldn’t let discussions about new legal regimes dist
Anthropic CEO Dario Amodei says the time has come to slow down AI development and will give third-party evaluators like METR access to its models to help ensure its 'adherence to safety practices and
Networking technology startup Ayar Labs Inc. said today it has extended its recent late-stage funding round to 650 million after raising an additional 150 million from an unnamed investor. The company
I wish I could spend $15k on my own homelab hardware, but no this is for work lol. Like the title says, we're looking to run DSV4 Flash (and similar tier models) locally at good speeds, both for token
Spoke with many providers who lurk in this sub, plus met folks who work in inference. For a 10k monthly revenue, a provider was able to retain only 200 dollars in profit due to GPU costs. Their custom
This is a brilliant paper. It's of the cleanest long-context agent designs I have seen in the past couple of months. Sequential memory agents read chunks one after another while maintaining a compact
Two big updates 1. I published an @FT op-ed on the OpenAI/HF incident & follow-ups 2. We’re starting an Open Alignment team at @huggingface to work on safety & alignment for open models, incl cybersec