Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: Photorealistic wide shot of six Ionic limestone columns on a university quad with a complete classical entablature carved with ‘TECH’ in Roman serif letters, technical blueprints and circuit diagrams unrolled at the column bases with brass calipers and precision instruments, late afternoon golden hour lighting, red brick buildings in background, sharp architectural photography with warm beige stone texture and long soft shadows.
Anthropic just posted another banger guide. This one is on building more efficient agents to handle more tools and efficient token usage. This is a must-read for AI devs! (bookmark it) It helps with three major issues in AI agent tool calling: token costs, latency, and tool https://x.com/omarsar0/status/1986099467914023194
@Kimi_Moonshot Congratulations to the entire Moonshot team — today is a great day for open source everywhere. We’re excited to continue supporting Kimi models with fast inference on Baseten. https://x.com/basetenco/status/1986494013109903362
@QuixiAI @Kimi_Moonshot a single H200 node is enough😃”” / X https://x.com/vllm_project/status/1986626058897269070
📢 New Model(s) Drop: Kimi K2 Thinking and Kimi K2 Thinking Turbo are now on Yupp! This pair of thinking models from @Kimi_Moonshot specialize in deep reasoning tasks. We explored their capabilities with some prompts on Yupp: https://x.com/yupp_ai/status/1986469027997491422
🚀 Hello, Kimi K2 Thinking! The Open-Source Thinking Agent Model is here. 🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%) 🔹 Executes up to 200 – 300 sequential tool calls without human interference 🔹 Excels in reasoning, agentic search, and coding 🔹 256K context window Built https://x.com/Kimi_Moonshot/status/1986449512538513505
🚨 New Open Source Model Update! Touted for its reasoning and coding strengths, Kimi K2 Thinking by @Kimi_Moonshot is now live for both Text and WebDev in Battle, Side by Side and Direct. Bring your toughest prompts! 💪 The last time Kimi K2 was in the Arena with a new model, https://x.com/arena/status/1986482438768673107
5 Thoughts on Kimi K2 Thinking – by Nathan Lambert https://www.interconnects.ai/p/kimi-k2-thinking-what-it-means
70% on SWE bench verified 30% terminal bench those are two intuitive thresholds for “”actually useful and not frustrating”” coding assistant. Kimi k2 thinking got 71.3% on SWE-Bench Verified 47.1% on Terminal-Bench”” / X https://x.com/andrew_n_carr/status/1986538323876454461
Congrats to the Kimi K2 team on the great numbers on our SWE-bench Verified, SWE-bench Multilingual and SciCode benchmarks!! https://x.com/OfirPress/status/1986475891158040760
Kimi AI – Kimi K2 is Live https://www.kimi.com/
Kimi API is barely alive right now kinda slow ~20tks/s and get quite a few timeouts / network errors when I let the model reason for a long time”” / X https://x.com/scaling01/status/1986476278908920061
Kimi K2 Thinking feels like a big milestone for open-source AI. The first time in a while that open-source gets ahead of proprietary APIs on their big area of focus (agents). Fun to see that it’s happening at a time when the proprietary APIs have the most money/attention”” / X https://x.com/ClementDelangue/status/1986833436607160600
Kimi K2 Thinking https://moonshotai.github.io/Kimi-K2/thinking.html
Kimi K2 Thinking is now available in anycoder https://x.com/_akhaliq/status/1986468663600337125
Kimi K2 Thinking is the new leading open weights model: it demonstrates particular strength in agentic contexts but is very verbose, generating the most tokens of any model in completing our Intelligence Index evals @Kimi_Moonshot’s Kimi K2 Thinking achieves a 67 in the https://x.com/ArtificialAnlys/status/1986911675820446013
Kimi K2 Thinking just launched on Product Hunt! 🥳 Not chasing votes, just using PH as a clean milestone log for our model updates. 🙂 Huge thanks to the helpful team from @ProductHunt https://x.com/crystalsssup/status/1986714377983304137
Kimi-K2 is an exceptional base model GPQA Diamond 77% GPT-4.5 only got 71.4%”” / X https://x.com/scaling01/status/1986112227875954967
Kimi-K2 Reasoning is coming very soon just got merged into VLLM LETS FUCKING GOOOO im so hyped im so hyped im so hyped https://x.com/scaling01/status/1986071916541870399
Kimi-K2 reasoning is landing soon; it just got merged into vLLM https://x.com/cedric_chee/status/1986073808672067725
Kimi-K2 Thinking ranking 19th on SimpleBench improving Kimi-K2s score from 26.3% (rank 33) to 39.6% This makes it the 3rd best open-source model on SimpleBench. Other chinese open-source models like DeepSeek R1 0528 and DeepSeek V3.1 beat it by roughly 1 %. https://x.com/scaling01/status/1986846212050362510
Live in Cline: kimi-k2-thinking https://x.com/cline/status/1986512739490275680
MoonshotAI has released Kimi K2 Thinking, a new reasoning variant of Kimi K2 that achieves #1 in the Tau2 Bench Telecom agentic benchmark and is potentially the new leading open weights model Kimi K2 Thinking is one of the largest open weights models ever, at 1T total parameters https://x.com/ArtificialAnlys/status/1986541785511043536
moonshotai/Kimi-K2-Thinking · Hugging Face https://huggingface.co/moonshotai/Kimi-K2-Thinking
ollama run kimi-k2-thinking:cloud Kimi K2 Thinking is Moonshot AI’s best open-source thinking model. Try it on Ollama’s cloud! https://x.com/ollama/status/1986640693108863271
Our first research paper: custom Mixture-of-Experts (MoE) kernels that make deployment of trillion-parameter models like Kimi K2 viable for the first time on AWS EFA https://x.com/AravSrinivas/status/1986106660386222592
Unsurprisingly, Kimi K2 Thinking is already number one trending on HF. The AI frontier is open-source! https://x.com/ClementDelangue/status/1986827413532057712
🚀 Day 0 support: Kimi K2 Thinking now running on vLLM! In partnership with @Kimi_Moonshot, we’re proud to deliver official support for the state-of-the-art open thinking model with 1T params, 32B active. Easy deploy in vLLM (nightly version) with OpenAI-compatible API: What https://x.com/vllm_project/status/1986455911066706160
It even compares with GPT 5 Pro on some benches. Looks like Kimi’s interpretation of Pro mode is 8 samples + self reflection https://x.com/nrehiew_/status/1986453238552666320
From my tests, Kimi K2 thinking is better than everything Xai, Anthropic, Google has to offer atm. The only thing that is better than this is Gpt 5 codex (at code) and Gpt 5 pro (at high level algorithm design) It beats the SOTA at creative writing by a mile. Good work”” / X https://x.com/karmay007/status/1986454592809529493
IndQA is a new benchmark designed to evaluate how well AI systems understand culture, context and history to answer questions that matter to people in India. With 2278 questions created in partnership with 250+ experts, IndQA dives deep into reasoning about everyday life,”” / X https://x.com/snsf/status/1985719755551158754
“ChatGPT-o1 & DeepSeek-R1, achieved diagnostic accuracy up to 93.75%. For context, this figure approaches the 96% accuracy benchmark reported for primary care physicians on the same vignette set” Except they told folks to get urgent care too often. Not unexpected given alignment”” / X https://x.com/emollick/status/1985164511947682070
Whisper no longer wears the open weights transcription accuracy crown with new entrants achieving better Artificial Analysis Word Error Rate scores Once considered the default choice for open weights transcription, OpenAI’s Whisper has now been surpassed by newer open weights https://x.com/ArtificialAnlys/status/1986100695989145649
Introducing IndQA — a new benchmark that evaluates how well AI systems understand Indian languages and everyday cultural context. https://x.com/OpenAI/status/1985950264525013210
Enabling Trillion-Parameter Models on AWS EFA https://research.perplexity.ai/articles/enabling-trillion-parameter-models-on-aws-efa
Perplexity is the first to develop custom Mixture-of-Experts (MoE) kernels that make trillion-parameter models available with cloud platform portability. Our team has published this work on arXiv as Perplexity’s first research paper. Read more: https://x.com/perplexity_ai/status/1986101355896098836
This was a key figure from the original GPT-3 paper. It proved LLMs could scale. We now believe we’re the first to see this hold true for robotic foundation models as well. It wasn’t obvious this would be true. We believe prior null results were due to insufficient useful data https://x.com/E0M/status/1985766175255773483
Announcing our Frontier Data Centers Hub! The world is about to see multiple 1 GW+ AI data centers. We mapped their construction using satellite imagery, permits & public sources — releasing everything for free, including commissioned satellite images. Highlights in thread! https://x.com/EpochAIResearch/status/1985788184245293153
🚨 WebDev Leaderboard Update MiniMax-M2 from @MiniMax__AI has landed as the #1 open model! A 230B MoE model with 10B-active-parameters, it’s an open source model built for efficient, high-performance coding, reasoning, and agentic-style tasks. It also ranks #4 in WebDev https://x.com/arena/status/1985465603206107318
Introducing SWE-1.5, our fast agent model. It achieves near-SOTA coding performance while setting a new standard for speed. Now available in Windsurf. https://x.com/windsurf/status/1983667319944712460
Thanks @_akhaliq sharing our work! 🚀 Glad to introduce our newest work — VCode! 🎨 VCode: A Multimodal Coding Benchmark with SVG as Symbolic Visual Representation For decades, RGB pixels have been the default medium for representing images. But in the agentic era, how can we”” / X https://x.com/KevinQHLin/status/1986126304316411928
Today we’re releasing SWE-1.5, our fast agent model. It achieves near-SOTA coding performance while setting a new standard for speed. Now available in @windsurf. https://x.com/cognition/status/1983662836896448756
VCode a Multimodal Coding Benchmark with SVG as Symbolic Visual Representation https://x.com/_akhaliq/status/1986073575216824650
Real shortage of good papers, even working papers, testing agentic (post o3) and Deep Research AI outputs in law, medicine, business, coding, etc. Assume when a research paper discusses AI it means GPT-4o (with occasional Gemini 2.5 or o1) for the next year or so.”” / X https://x.com/emollick/status/1985417078749479212
Ant AQ-Team @AQ_MedAI @TheInclusionAI and SGLang RL Team @sgl_project just helped land Kimi-K2-Instruct RL on slime — fully wired up and running on 256× H20 141GB 🚀 Huge shout-out to @yngao016, @menlzy, @Yonah_x from AQ Team and @Ji_Li_233, @Yefei_RL from the SGLang RL Team for”” / X https://x.com/slime_framework/status/1986811354502906304
Fixed the token generation speed on https://x.com/Kimi_Moonshot/status/1986754111992451337
Here’s the command I ran: “` mlx.launch –hosts first.ip,second.ip –env MLX_METAL_FAST_SYNCH=1 mlx-lm/mlx_lm/examples/pipeline_generate.py –model mlx-community/Kimi-K2-Thinking –prompt “”Write an HTML and JavaScript page implementing space invaders”” -m 16384 “` PR here”” / X https://x.com/awnihannun/status/1986602098017116357
The new 1 Trillion parameter Kimi K2 Thinking model runs well on 2 M3 Ultras in its native format – no loss in quality! The model was quantization aware trained (qat) at int4. Here it generated ~3500 tokens at 15 toks/sec using pipeline-parallelism in mlx-lm: https://x.com/awnihannun/status/1986601104130646266
Gemini 3.0 Ultra or Gemini 3.0 Pro? Which is it and why do you think that? It sounds too big for Pro but too small for Ultra, but since models just get sparser and sparser I believe it’s Pro and very similar to Kimi-K2 1.2T@30B. Also as of right now Ultra is still just a”” / X https://x.com/scaling01/status/1986161974883860486
Team from Ant Group @TheInclusionAI helped land Kimi model @Kimi_Moonshot on @Zai_org’s slime framework! Open AIs help Open AIs ♥️ CN AIs help CN AIs ♥️”” / X https://x.com/bigeagle_xd/status/1986815075785879723
Today we’re announcing ARC Prize Verified, a program to increase the rigor of evaluating frontier systems on ARC-AGI This program adds a third-party academic panel to audit our testing process We are also welcoming 5 new AI labs as sponsors of ARC-AGI-3 https://x.com/arcprize/status/1985802145300693140
Context engineering is the art of giving AI systems the right information at the right time. Most engineers focus on initial retrieval, that is, finding relevant documents from a large corpus. But retrieval is just the first step. The real challenge is prioritization. A reranker”” / X https://x.com/douwekiela/status/1985756688000163892
I don’t think how people are tracking how quickly this is happening, for better or worse. https://x.com/emollick/status/1985132904771399899
🚀 Introducing Arena Expert: a new LMArena evaluation framework to identify the toughest, most expert-level prompts from real users, powering a new Expert leaderboard. We also introduce Occupational Categories that underlie eight new leaderboards: 💻 Software & IT Services ✍️ https://x.com/arena/status/1986153162802368555
Congrats on the launch @nebiusai! Live benchmarks are available on Artificial Analysis for all the language model APIs served on Nebius Token Factory: https://x.com/ArtificialAnlys/status/1986174888080789509
Individual results across all evaluations in the Artificial Analysis Intelligence Index: MMLU-Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME 2025, IFBench, AA-LCR, Terminal-Bench Hard, 𝜏²-Bench Telecom https://x.com/ArtificialAnlys/status/1986911685496746404
Interesting new AI benchmark combining elements of game environment testing with world model testing that finds large gaps between human and AI ability (& some behavioral differences as well) We need more grounded, unsaturated & hard benchmarks like this. https://x.com/emollick/status/1983914683976265755
It’s rare nowadays to find something that is intuitively important and not yet done well by any major language models. But *precisely aggregating lots of information over long contexts* is one of those things. Our new benchmark Oolong tests this ability, see the 🧵 for more!”” / X https://x.com/gneubig/status/1986851194862510102
New eval! Code duels for LMs ⚔️ Current evals test LMs on *tasks*: “”fix this bug,”” “”write a test”” But we code to achieve *goals*: maximize revenue, cut costs, win users Meet CodeClash: LMs compete via their codebases across multi-round tournaments to achieve high-level goals https://x.com/jyangballin/status/1986093902122942700
ok we’re at 51% with “”heavy”” mode > Heavy Mode: K2 Thinking Heavy Mode employs an efficient parallel strategy: it first rolls out eight trajectories simultaneously, then reflectively aggregates all outputs to generate the final result. https://x.com/eliebakouch/status/1986447441668022471
🏆NEW LMARENA LEADERBOARDS🏆 🤓Experts 💻 Software & IT Services ✍️ Writing, Literature, & Language 🔬 Life, Physical, & Social Science 🎭 Entertainment, Sports, & Media 📈 Business, Management, & Financial Ops 🧮 Mathematical ⚖️ Legal & Government 🩺 Medicine & Healthcare https://x.com/ml_angelopoulos/status/1986154276499104186
We looked at OSWorld, a popular evaluation of AI computer use capabilities. Our findings: tasks are simple, many don’t require GUIs, and success often hinges on interpreting ambiguous instructions. The benchmark is also not stable over time. See thread for details! https://x.com/EpochAIResearch/status/1985441059032478172
Remote Labor Index: Measuring AI Automation of Remote Work https://arxiv.org/pdf/2510.26787
When Visualizing is the First Step to Reasoning MIRA, a Benchmark for Visual Chain-of-Thought https://x.com/_akhaliq/status/1986075520962793672
If you’d like to win your own Dell Pro Max with GB300 we’re launching a new kernel competition with @NVIDIAAI @sestercegroup @Dell to optimize NVF4 kernels on B200 2025 has seen a tremendous rise of pythonic kernel DSLs, we got on-prem hardware to have reliable ncu benchmarking”” / X https://x.com/GPU_MODE/status/1985436876384453128
We’ve released an early preview of Qwen3-Max-Thinking–an intermediate checkpoint still in training. Even at this stage, when augmented with tool use and scaled test-time compute, it achieves 100% on challenging reasoning benchmarks like AIME 2025 and HMMT. You can try the https://x.com/Alibaba_Qwen/status/1985347830110970027
Super happy to see the next iteration of ViDoRe: ViDoRe v3 is built on human-created examples, covers more realistic RAG scenarios (including open-ended and multi-hop queries), and should be your new default benchmark for multimodal retrieval! 👀 Congrats to the team!”” / X https://x.com/tonywu_71/status/1986047154620633370
🚀 new 🌤️ lighteval release and our biggest yet! • new benchmark finder to explore all available tasks • inspect-ai integration from @AISecurityInst → more stable and easier to add benchmarks • share your evals and insights with the community on the @huggingface hub • new https://x.com/nathanhabib1011/status/1985720151673880923
Amazing work by @RidgerZhu and the ByteDance Seed team — Scaling Latent Reasoning via Looped LMs introduces looped reasoning as a new scaling dimension. 🔥 The Ouro model is now runnable on vLLM (nightly version) — bringing efficient inference to this new paradigm of latent”” / X https://x.com/vllm_project/status/1985695123469209703
Google DeepMind release: Towards Robust Mathematical Reasoning Introduces IMO-Bench, a suite of advanced reasoning benchmarks that played a crucial role in GDM’s IMO-gold journey. Vetted by a panel of IMO medalists and mathematicians. IMO-AnswerBench – a large-scale test on https://x.com/iScienceLuvr/status/1985685404276965481
While human expert evaluation remains the gold standard for mathematical proofs, its cost and time intensity limit scalable research. To address this, we built #ProofAutoGrader, an automatic grader for IMO-ProofBench. The autograder leverages Gemini 2.5 Pro, providing it with a https://x.com/lmthang/status/1985772094085595570
Continuing our IMO-gold journey, I’m delighted to share our #EMNLP2025 paper “Towards Robust Mathematical Reasoning”, which tells some of the key stories behind the success of our advanced Gemini #DeepThink at this year IMO. Finding the right north-star metrics was highly https://x.com/lmthang/status/1985760224612057092
Visualizations from the Smol Playbook pretraining section, explaining MLA, RoPE, chunked attention, multi-step schedulers and more. A thread: https://x.com/LoubnaBenAllal1/status/1986110843600117760
what is ocr? – introspection ft. harsehaj https://harsehaj.substack.com/p/ocr-models-explained
#1 on the MTEB multilingual leaderboard”” / X https://x.com/fdaudens/status/1984541314063446191
> All benchmark results are reported under INT4 precision. Do you understand what a flex this was. They go toe to toe with GPT-5 on the heaviest, longest-range tasks, with hundreds of tool calls. ALL IN INT4. «Convert to fp8 if you need» Frontier lab. https://x.com/teortaxesTex/status/1986612178133123165
Ghosts in the Codex Machine – Google Docs https://docs.google.com/document/d/1fDJc1e0itJdh0MXMFJtkRiBcxGEFtye6Xc6Ui7eMX4o/edit?tab=t.0#heading=h.zdchmr4v1lbh
It’s SOTA, not only open weights SOTA :)”” / X https://x.com/crystalsssup/status/1986627840310452366
Qwen3-VL Accuracy Differences on Ollama vs MLX Video: https://x.com/andrejusb/status/1985612661447331981
Excited to introduce TWIST2, our next-generation humanoid data collection system. TWIST2 is portable (use anywhere, no MoCap), scalable (100+ demos in 15 mins), and holistic (unlock major whole-body human skills). Fully open-sourced: https://x.com/ZeYanjie/status/1986126096480587941
Butter-Bench: Evaluating LLM Controlled Robots for Practical Intelligence | Andon Labs https://andonlabs.com/evals/butter-bench
But robots can not feel… or can they? [Find the paper here ⬇] Scientists have developed a way for robots to sense touch all over their bodies without needing expensive, complicated equipment. ✅ Instead of using lots of sensors, the robot’s internal tech senses touch https://x.com/IlirAliu_/status/1984176634467844371
The “AI will replace radiologists” prediction remains a rich example. A lot of folks have pointed out the problems with confusing a task (“reading a scan”) with a job (“radiologist”) with many tasks. That is true. But there was a human problem. Radiologists rejected (pre-LLM) AI”” / X https://x.com/emollick/status/1984696156140470530
New pathology foundation model from PathAI, PLUTO-4 Interesting FlexiViT arch for their smaller model, and good discussion of some of the training infra/scaling. Trained on 32 H200s with DINOv2 framework, on 551k whole slide images. Great to see EVA (from https://x.com/iScienceLuvr/status/1986031522231865571
Cross-continent teleop is quite insane when you think about it. It’s remote work for atoms, not bits. For the first time, we can apply the cost advantages of outsourced labor to the physical economy. Starlink is only going to accelerate this.”” / X https://x.com/aryxnsharma/status/1985427799541457043
@eliebakouch Interesting that the quantization is applied to the routed experts but not to the shared one. My understanding is that the shared expert has plenty of time to compute (Meituan fit in two whole shared expert layers during one MoE step with ScMoE) which is probably why. https://x.com/timfduffy/status/1986474172860834263
🔍 What makes DreamGym work? – Removing replay grounding → big drop in state consistency, higher hallucination – Removing reasoning traces → states become shallow, reward errors rise – Removing task generation → policy plateaus early In short: ✅ Reasoning = stable https://x.com/jaseweston/status/1986613058035159264
🔥 Learning rate transfer under μP is now proven!”” / X https://x.com/QuanquanGu/status/1985961289882165674
🔥 New Blog: “Disaggregated Inference: 18 Months Later” 18 months in LLM inference feels like a new Moore’s Law cycle – but this time not just 2x per year: 💸 Serving cost ↓10-100x 🚀 Throughput ↑10x ⚡ Latency ↓5x A big reason? Disaggregated Inference. From DistServe, our”” / X https://x.com/haoailab/status/1985753711344316648
🔥Highly requested by the community, PaddleOCR-VL is now officially supported on vLLM! 🚀 Check out our recipe for this model to get started!👇 https://x.com/vllm_project/status/1985589446197330129
another interesting update: for irregular shapes like below that we see in practice, the cutlass kernel that i have actually beats cublas on NVFP4 https://x.com/mrsiipa/status/1986012708433719519
Appreciate @EpochAIResearch ‘s thoughtful review of OSWorld 👇 A few clarifications from our side as maintainers: OSWorld tasks span the full difficulty spectrum — from simple GUI edits to long multi-app workflows. Abhyankar et al.’s “step count” uses merged custom atomic”” / X https://x.com/TianbaoX/status/1985647751468892434
Beyond Standard LLMs – by Sebastian Raschka, PhD https://magazine.sebastianraschka.com/p/beyond-standard-llms
Can LLMs accurately aggregate information over long, information-dense texts? Not yet… We introduce Oolong, a dataset of simple-to-verify information aggregation questions over long inputs. No model achieves >50% accuracy at 128K on Oolong! https://x.com/abertsch72/status/1986842177692180497
Context Engineering 2.0 This report discusses the context of context engineering and examines key design considerations for its practice. Explosion of intelligence will lead to greater context-processing capabilities, so it’s important to build for the future too. This aligns https://x.com/omarsar0/status/1985747789796483109
Continuous Autoregressive Language Models Compresses a fixed number of tokens into a vector using an autoencoder, and directly models prediction of next vector. https://x.com/iScienceLuvr/status/1985317763334967726
Diffusion Language Models are Super Data Learners https://x.com/_akhaliq/status/1986459558114386381
Full prompts for the Orchestra Pattern: https://x.com/code/status/1986622179237085331
Global AI Confessions Report: Data Leaders Edition https://pages.dataiku.com/global-ai-confessions-report-data-leaders-edition
Graphiti GitHub repo: https://x.com/_avichawla/status/1985958022053838924
How I use AI (Oct 2025) – Ben Stolovitz https://ben.stolovitz.com/posts/how_use_ai_oct_2025/
How is memorized data stored in a model? We disentangle MLP weights in LMs and ViTs into rank-1 components based on their curvature in the loss, and find representational signatures of both generalizing structure and memorized training data https://x.com/jack_merullo_/status/1986493462716555726
How to build a thinking AI models – Geeky Gadgets https://www.geeky-gadgets.com/how-to-build-a-thinking-ai-models/
if you’re using SAM2 for tracking, just swap with EdgeTAM drop-in🤝 very few points behind SAM2, efficiency gains in multipliers find transformers model checkpoint here: https://x.com/mervenoyann/status/1986787448890819043
INT v.s. FP A Comprehensive Study of Fine-Grained Low-bit Quantization Formats https://x.com/_akhaliq/status/1985370441465098709
Introducing Terminal-Bench 2.0 and Harbor https://www.tbench.ai/news/announcement-2-0
It is a shame that model makers have to BEG inference providers and aggregators to implement their damn model correctly”” / X https://x.com/xeophon_/status/1985376786402648357
It’s fascinating that FP16 can reduce training-inference mismatch in RL fine-tuning. Out of curiosity, I tried the same precision swap on the Tiny Recursive Model (TRM) @jm_alexia, which iterates hidden states to reason over inputs. The outcome was different: under FP16,”” / X https://x.com/huskydogewoof/status/1985386675263193289
Lexsi Labs unveils Orion-MSP: Multi-Scale Sparse Attention for Tabular Data This new foundation model achieves state-of-the-art in-context learning on tabular data. It efficiently captures complex interactions across multiple scales, handles high-dimensional tables, and requires https://x.com/HuggingPapers/status/1986526105868603666
Long-term memory doesn’t have to suck! Here is the problem: even LLMs with 1M token windows struggle substantially as dialogues lengthen. Raw context capacity alone is insufficient for effective long-term conversational memory. On the other hand, hybrid memory approaches work https://x.com/omarsar0/status/1985348779193860414
looking ahead, we’re prototyping something new — we call it predictive sensing. our paper cited tons of work from cogsci and developmental psychology. the more we read, the more amazed we became by human / animal sensing. the human visual system is super high-bandwidth, yet https://x.com/sainingxie/status/1986685057415680329
Love the retrospective on disaggregated inference. If you wonder where the technique named “”PD”” in vLLM comes from, read on! Thank you @haoailab for pushing the idea forward.”” / X https://x.com/vllm_project/status/1985761953432944893
made Wan 2.2 work with NVFP4 using my cutlass kernel and i do see a decent speedup with good enough accuracy, i bet i’m missing something and it can go even faster https://x.com/mrsiipa/status/1986122938668782002
Moonshot AI https://www.moonshot.cn/
moonshotai (Moonshot AI) https://huggingface.co/moonshotai
New video: What do models learn during fine-tuning? The nascent field of model diffing tries to answer this and more by comparing two versions of a model and studying the diff My MATS scholars @Butanium_ @jkminder and I discuss the field, and their two model diffing papers ⬇️”” / X https://x.com/NeelNanda5/status/1986590658862502380
NVFP4 quantization procedure for those who care: In essence, it is a block-scaling method so a block of 16 consecutive elements gets assigned a scale factor. Given a tensor x each block b (= 16) of consecutive elements in higher precision gets quantized to NVFP4. There are two”” / X https://x.com/mrsiipa/status/1986152319004856491
of course merged by the 500IQ Tsinghua GOAT himself https://x.com/scaling01/status/1986072602306089262
Our adaptive speculator, ATLAS, delivers 400% faster LLM inference by learning from your workloads in real-time ⚡ According to @VentureBeat: “”The shift from static to adaptive optimization represents a fundamental rethinking of how inference platforms should work.”””” / X https://x.com/togethercompute/status/1986543029671637299
PptxGenJS Presentation https://ai.wharton.upenn.edu/wp-content/uploads/2025/10/2025-Wharton-GBK-AI-Adoption-Report_Executive-Summary.pdf
Prompt engineering is worrying what the user should write. Context engineering is worrying what the model should read. These used to be the same thing, and few foresaw their divergence. But only divorced from the latter can we consider the former in earnest, seeing what remains.”” / X https://x.com/goodside/status/1985583995644497931
Run Any LLM from One API: Introducing any-llm 1.0 https://blog.mozilla.ai/run-any-llm-with-a-single-api-introducing-any-llm-v1-0/
Spent some time with @NoahZiems’ arbor over the weekend. It’s basically online RL for @DSPyOSS programs. Plug ArborProvider into your DSPy LM and train with GRPO/mmGRPO to optimize multi‑module pipelines on real rewards (quality, cost, privacy). It’s experimental but I’m a fan ↓ https://x.com/RajaPatnaik/status/1985407144209105026
State of Generative Media Survey Report 2025 | Artificial Analysis https://artificialanalysis.ai/media/survey-2025
The framework shows the largest relative improvements in tasks requiring long-range reasoning: summarization (+160.6%), multi-hop reasoning (+27.2%), and preference following (+76.5%). Ablation studies reveal that all components become increasingly essential as context grows. At https://x.com/omarsar0/status/1985348825197039718
the NVFP4 quant kernel achieves 3.2TB/s memory throughput and has a runtime of 33us which seems to be good enough: NCU complains about non-fused FP32 operations but all i see is FMUL so that might be a false alarm.”” / X https://x.com/mrsiipa/status/1986786214289301593
there’s a subtle but very important issue with this implementation: the scale factor should be stored in a tiled layout, so each 128×4 tile should be laid out as 32×16 in the memory i.e. interleaved pattern but the current implementation does not store the scales this way. https://x.com/mrsiipa/status/1985302904635597238
Thermodynamic Computing: From Zero to One | Extropic https://extropic.ai/writing/thermodynamic-computing-from-zero-to-one
Thoughts by a non-economist on AI and economics – Windows On Theory https://windowsontheory.org/2025/11/04/thoughts-by-a-non-economist-on-ai-and-economics/
turned out to be a RoPE precision issue”” / X https://x.com/vikhyatk/status/1985163608603636195
update: wrote a triton kernel for this – has correct tiled layout for scale factors – uses inline ptx for conversion – 4x faster than torch compiled version triton is absolutely amazing for writing memory bound kernels tbh”” / X https://x.com/mrsiipa/status/1985333503756849326
We just released our complete guide to Context Engineering. (These 6 components are the future of production AI apps) Every developer hits the same wall when building with Large Language Models: the model is brilliant but fundamentally disconnected. It can’t access your private https://x.com/weaviate_io/status/1985741429579170276
We’re excited to see @dify_ai integrating Qdrant into its Knowledge Pipeline. With Qdrant as the vector engine, Dify users get: – Fast, low-latency retrieval for large-scale knowledge bases – Built-in hybrid search combining keyword and semantic matching – Simple configuration”” / X https://x.com/qdrant_engine/status/1986014287718916463
Why do AI models use so many em-dashes? https://www.seangoedecke.com/em-dashes/
YES, int4 QAT training is a bit slower, but evaluation and RL-inference is much faster”” / X https://x.com/bigeagle_xd/status/1986457715913453571
LLMs memorize a lot of training data, but memorization is poorly understood. Where does it live inside models? How is it stored? How much is it involved in different tasks? @jack_merullo_ & @srihita_raju’s new paper examines all of these questions using loss curvature! (1/7) https://x.com/GoodfireAI/status/1986495330201051246
Pro tip to avoid public embarrassment: vibe test your models aggressively for things you can’t easily measure. Exhibit A: during SmolLM3, we accidentally nuked all the system messages in our training data and the model had no clue who it was 🙈 https://x.com/_lewtun/status/1985995034970214676
RayZer: A Self-supervised Large View Synthesis Model TL;DR: Modern self-supervised methods share the same principle of learning by predicting “”missing”” data. GPT predicts missing next token leveraging language sequential prior. Can we do the same for self-supervised 3D model? https://x.com/Almorgand/status/1984295391219617863
Did you know that “”OSWorld”” does not really exist and everyone benches a different set of prompts, so you can’t really compare scores? … yeah I couldn’t believe it either 🙃 Full review: https://x.com/xeophon_/status/1985441764132499883
the score are insane, very cool to see native int4 quantization for the MoE layers > To overcome this challenge, we adopt Quantization-Aware Training (QAT) during the post-training phase, applying INT4 weight-only quantization to the MoE components. It allows K2 Thinking to”” / X https://x.com/eliebakouch/status/1986451219892646124
Reminder that Huawei SuperPoDs with plans to scale to gigawatt-class, 1M-device superclusters by Q4 2027 are dedicated to DeepSeek. Regular thanks to the Whale for inspiring so many. R1 will be one of the most significant events in the history of tech, close to ChatGPT 3.5. https://x.com/teortaxesTex/status/1985567870227460166
This is a great achievement from the SGLang team! I got immediately asked by a few people what vLLM’s plan is. Multimodal generation with omni models is something we have been working on with the community and model vendors. As a sneak peek, Hunyuan-image 3.0 (that beats”” / X https://x.com/rogerw0108/status/1986919399346106490
I am very sympathetic to the delays in publishing papers, but I think we need to be careful with “”AI can’t do this”” claims when our empirical evidence pre-dates even o1 class Reasoners. The strongest model here is GPT-4 (which does better) and the next best is Llama 2 70B(!!)… https://x.com/emollick/status/1985610450709434527
No excuse anymore not to train your own models! This 200+ pages with full transparency. Let’s go open-source AI!”” / X https://x.com/ClementDelangue/status/1985357572300321213
Key Results/Insights: LIGHT consistently outperforms both long-context LLMs and RAG baselines across all conversation lengths, with improvements growing as context increases: +49-60% at 100K-1M tokens, and dramatic +107-156% gains at 10M tokens where no model natively supports”” / X https://x.com/omarsar0/status/1985348807849300249
Someone from xAI reached out and asked me to retest grok-4-fast, because they’ve improved the injected system prompts. Huge improvement! grok-4-fast-reasoning: 77.5% -> 94.1% grok-4-fast-non-reasoning: 77.9 -> 97.9% I really appreciate that xAI takes this topic seriously. https://x.com/xlr8harder/status/1986728144712380682
The results are in. LTX-2 is now ranked #3 video model on @ArtificialAnlys Video Arena. No surprise here. LTX-2 delivers on quality and speed. Huge credit to the LTX team here at @lightricks for making it happen. https://x.com/LTXStudio/status/1986442720534016449
New, extremely challenging visual reasoning benchmark “”MIRA”” where current models fail…great resource to research reasoning with imgs/video🌌 https://x.com/Muennighoff/status/1986519726823211129
here’s the bf16 video for reference, close enough right? i guess this is due to the reduced quantization block size in NVFP4 (16) vs MXFP4 (32) so ideally the former preserves the details more but just a speculation. https://x.com/mrsiipa/status/1986123806357020865
Introducing Cambrian-S it’s a position, a dataset, a benchmark, and a model but above all, it represents our first steps toward exploring spatial supersensing in video. 🧶 https://x.com/sainingxie/status/1986685042332958925
One interesting potential side effect of the nature of AI generated music & video is that it may make good critics good at making stuff. These systems have high variance, so the best approach is making many songs/videos and selecting the best. Critics are practiced at selection”” / X https://x.com/emollick/status/1985139697849413849
We discuss their papers showing that model diffing is unexpectedly easy when fine-tuning in a narrow domain, and on finding and fixing flaws with crosscoders, a sparse autoencoder based approach Video: https://x.com/NeelNanda5/status/1986590670631674217
We present MotionStream — real-time, long-duration video generation that you can interactively control just by dragging your mouse. All videos here are raw, real-time screen captures without any post-processing. Model runs on a single H100 at 29 FPS and 0.4s latency. https://x.com/xxunhuang/status/1985806498811789738
you might ask — isn’t this just a data or scaling problem? partly, yes. that’s why we’re building the new Cambrian-S video MLLM family. we want to push the limits of the current paradigm. we think data and scaling are essential for supersensing (just not sufficient.) the core https://x.com/sainingxie/status/1986685054559342916
Leave a Reply