Image created with gemini-2.5-flash-image with claude-sonnet-4-5. Image prompt: A chess game played on a luminous green circuit board with copper traces forming the squares, pieces crafted from miniature computer components like CPUs and RAM modules, dramatic side lighting with blue and green LED glow, macro photography perspective showing intricate electronic details and solder points, high-tech aesthetic with shallow depth of field.
We launched SWE-Bench Pro last month to incredible feedback, and we’ve now updated the leaderboard with the latest models and no cost caps. SoTA models now break 40% pass rate. Congrats to @Anthropic for sweeping the top spots! 🥇Claude 4.5 Sonnet 🥈Claude 4 Sonnet 🥉Claude 4.5 https://x.com/scale_AI/status/1980685992987431368
AI trading in real markets https://nof1.ai/
They should have broken the 10k to 10-100 stacks of $100-1k and given them to the identical copies of the same model to be able to see anything remotely meaningful. Right now we are looking at noise!”” / X https://x.com/abeirami/status/1980434468398883076
.@karpathy says that LLMs currently lack the cultural accumulation and self-play that propelled humans out of the savannah: Culture: > “Why can’t an LLM write a book for the other LLMs? Why can’t other LLMs read this LLM’s book and be inspired by it, or shocked by it?” Self https://x.com/dwarkesh_sp/status/1980333945385562176
The most interesting part for me is where @karpathy describes why LLMs aren’t able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across https://x.com/dwarkesh_sp/status/1979259041013731752
This AI trading benchmark is interesting. Each model got $10,000 to invest. ~3 days in: ranking atm: – DeepSeek V3.1: +$2,658 – Grok 4: +$2,236 – Claude 4.5 Sonnet: +$1,911 – Qwen 3 Max: −$211 – GPT-5: −$3,139 – Gemini 2.5 Pro: −$3,719 DeepSeek beats all the other models https://x.com/Yuchenj_UW/status/1980318499185823760
TLDR: OpenAI Atlas > Perplexity Comet in an agent mode head to head. Here is my use case: I have a very real, very tedious use case, which is a manual task that I do every day. 1. I go to the school website to look at each of my daughter’s classes 2. I look at her grades 3. I https://x.com/raizamrtn/status/1980695747227210213
Technological development was slow for 9,288 generations not because past humans were dumb, but prior to books & the scientific method, innovation happened at the level of societies, not people. So tech evolved gradually, not through leaps of genius, but slow cultural adaption https://x.com/emollick/status/1979616432355946961
openai’s codex cli with gpt 5 became better than claude code 🤯 it crawls the codebase to a degree i have never seen seen from claude code. Instantly one-shotted a bug i couldn’t solve with claude code for 3 days new $200 per month subscription **check** https://x.com/samuelstroschei/status/1957655293942460670
🚨 New Model Update MiniMax-M2 by @MiniMax_AI is expected to land next week but is already in the Arena for testing as MiniMax-M2-Preview! Let’s see how it stacks up. Early details suggest it’s an advanced agentic model with strong reasoning and long-context capabilities, https://x.com/arena/status/1981850766039187901
I have been waiting for a paper on AI agents and transaction costs and, well, agency problems Agentic work, by its nature, drastically changes how these operate, with huge implications for how we organize markets and firms, which are largely shaped by agency & transaction costs https://x.com/emollick/status/1980299018220753290
If you want to understand AGI, study the humanities (I am only partially trolling – psychology is a young field, computer science is younger. The brightest minds in history spent a lot of time considering what it meant to be a general intelligence, that’s what the humanities is)”” / X https://x.com/emollick/status/1979637468149530660
New research with @AdtRaghunathan, Nicholas Carlini and Anthropic! We built ImpossibleBench to measure reward hacking in LLM coding agents 🤖, by making benchmark tasks impossible and seeing whether models game tests or follow specs. (1/9) https://x.com/fjzzq2002/status/1981745974700581191
Anthropic is catching up with OpenAI – by Alex Wilhelm https://www.cautiousoptimism.news/p/anthropic-is-catching-up-with-openai
🚨 WebDev Arena: Top 15 Disrupted! 4 new models have been added to the WebDev leaderboard: 🔸 #4 Claude Sonnet 4.5 Thinking 32k by @AnthropicAI 🔸 #4 GLM 4.6 (the new #1 open model) by @Zai_org 🔸 #11 Qwen3 235B A22B Instruct (and #7 open model) by @Alibaba_Qwen 🔸 #14 Claude https://x.com/arena/status/1980367208300835328
Mini Models Battle: Claude Haiku 4.5 vs GLM-4.6 vs GPT-5 Mini https://blog.kilocode.ai/p/mini-models-battle-claude-haiku-45
New paper! We reverse engineered the mechanisms underlying Claude Haiku’s ability to perform a simple “perceptual” task. We discover beautiful feature families and manifolds, clean geometric transformations, and distributed attention algorithms! https://x.com/wesg52/status/1980680563582538099
Virtually Being https://eyeline-labs.github.io/Virtually-Being/
OmniVinci: Joint Visual-Audio Understanding https://nvlabs.github.io/OmniVinci/
nanochat d32, i.e. the depth 32 version that I specced for $1000, up from $100 has finished training after ~33 hours, and looks good. All the metrics go up quite a bit across pretraining, SFT and RL. CORE score of 0.31 is now well above GPT-2 at ~0.26. GSM8K went ~8% -> ~20%, https://x.com/karpathy/status/1978615547945521655
Global AI Tracker https://www.similarweb.com/corp/wp-content/uploads/2025/10/attachment-Global-AI-Tracker-1.pdf
Thanks for sharing the internal benchmarks, @rauchg ! We love to see it. 🔥”” / X https://x.com/Kimi_Moonshot/status/1980219115840385349
A lot I like & some I don’t in this paper: Like: Clear definition of AGI, diverse authors, shows jaggedness, tracking metrics over time (huge leap from GPT-4 to GPT-5) Dislike: AGI defined as replicating a model of human cognition, benchmarks are scattershot, narrow view of AI https://x.com/emollick/status/1978874737892667718
Text or Pixels? It Takes Half: On the Token Efficiency of Visual Text Inputs in Multimodal LLMs Looks like this paper is also exploring the direction DeepSeek is interested in: representing text more efficiently as images, observing almost half reduction in number of tokens https://x.com/iScienceLuvr/status/1980942325573648703
> by storing the data representation natively as image tiles This must be obvious but just to clarify: DeepSeek does not propose to store *screenshots* of your chat logs. Pixel representation can be ephemeral; what is stored is still tokens, just not *language* tokens. https://x.com/teortaxesTex/status/1980453820632297900
Karpathy is undoubtedly vision pilled. And thanks to this casual DeepSeek drop — so will you.”” / X https://x.com/bilawalsidhu/status/1980598830916939880
DeepSeek https://github.com/deepseek-ai/
@mervenoyann The good perf of DeepSeek models matches with what we observe on PrediBench! https://x.com/AymericRoucher/status/1980196484617523445
Again, I will reiterate this: DeepSeek was literally built by chinas top quant firm, and chinas TOP quants.”” / X https://x.com/hamptonism/status/1980182896049811780
After DeepSeek-V3.2-Exp added TileLang & CUDA ops, many asked: what exactly is TileLang? 🤔 In his post “”TileLang: 80 lines of Python kernel code to reach 95% of FlashMLA’s performance””, developer & Zhihu contributor ryume gives a full breakdown of this new AI programming https://x.com/ZhihuFrontier/status/1980170674112188440
GLM-4.6 providers overview: we are benchmarking API endpoints offered by Baseten, GMI, Parasail, Novita, Deepinfra GLM-4.6 (Reasoning) from @Zai_org is one of the most intelligent open weights models, with intelligence close to GPT-OSS-120b (high), DeepSeek V3.2 Exp (Reasoning) https://x.com/ArtificialAnlys/status/1980777360724226282
For people thinking that DeepSeek-OCR is the first model to render text as images, the University of Copenhagen already did this in 2023 Paper is called “”Language Modelling with Pixels””. They trained a Masked AutoEncoder (MAE) by rendering text as images and masking patches https://x.com/NielsRogge/status/1980559120760791125
We’re seeing a lot of usage around DeepSeek’s new OCR model. Alex packaged it so you can deploy and test it yourself – prompts and sample images included.”” / X https://x.com/basetenco/status/1980924381217104338
DeepSeek-OCR looks impressive, but its core idea is not new. Input “Text” as “Image” — already explored by: LANGUAGE MODELING WITH PIXELS (Phillip et al., ICLR 2023) CLIPPO: Image-and-Language Understanding from Pixels Only (Michael et al. CVPR 2023) Pix2Struct: Screenshot https://x.com/awinyimgprocess/status/1980506449706119642
A more serious thread on the DeepSeek-OCR hype / serious misinterpretation going on. 1. On token reduction via representing text in images, researchers from Cambridge have previously shown that 500x prompt token compression is possible (ACL’25, Li, Su, and Collier). Without https://x.com/Kangwook_Lee/status/1980709454522744902
DeepSeek finally released a new model and paper. And because this DeepSeek-OCR release is a bit different from what everyone expected, and DeepSeek releases are generally a big deal, I wanted to do a brief explainer of what it is all about. In short, they explore how vision https://x.com/rasbt/status/1980642191950090585
I quite like the new DeepSeek-OCR paper. It’s a good OCR model (maybe a bit worse than dots), and yes data collection etc., but anyway it doesn’t matter. The more interesting part for me (esp as a computer vision at heart who is temporarily masquerading as a natural language”” / X https://x.com/karpathy/status/1980397031542989305
🚨 DeepSeek just did something wild. They built an OCR system that compresses long text into vision tokens literally turning paragraphs into pixels. Their model, DeepSeek-OCR, achieves 97% decoding precision at 10× compression and still manages 60% accuracy even at 20×. That https://x.com/godofprompt/status/1980233080213590326
Letsss gooo! DeepSeek just released a 3B OCR model on Hugging Face 🔥 Optimised to be token efficient AND scale ~200K+ pages/day on A100-40G Same arch as DeepSeek VL2 Use it with Transformers, vLLM and more 🤗 https://x.com/reach_vb/status/1980170192392270227
a bunch of OCR models released in past few weeks: ~ deepseek-ocr-3b ~ olmo-ocr-2-7b ~ chandra-ocr-8b ~ nanonets-ocr2-3b ~ paddleocr-vl-0.9B ~ qwen3-vl-dense/moe (general vlm) ~ dots.ocr-3b Will be dropping a detailed comparison soon”” / X https://x.com/HarveenChadha/status/1981055277408669934
NEW DeepSeek OCR model that outperforms dots ocr while prefilling 3x less tokens https://x.com/casper_hansen_/status/1980166248878203093
DeepSeek-OCR has some weird architectural choices for the LLM decoder: DeepSeek3B-MoE-A570M -> uses MHA, no MLA (not even GQA?) -> 2 shared experts (like DeepSeek V2, but V3 only has 1) -> quite low sparsity, activation ratio is 12.5%. For V3 it’s 3.52%, for V2 it’s 5% -> not https://x.com/eliebakouch/status/1980193125202083951
I think Glyph coming out on the same day a) corroborates the results of DeepSeek OCR b) confirms the “they had it lying around for a while” suspicion. Charitably, they learned of Zhipu’s project retracing their steps and sped up the release. Other possibilities are obvious too.”” / X https://x.com/teortaxesTex/status/1980642000006451348
deepseek-ai/DeepSeek-OCR: Contexts Optical Compression https://github.com/deepseek-ai/DeepSeek-OCR
what happened this week with OCR and VLMs? * deepseek-ocr * chandra-ocr * nanonets-ocr2 * paddleocr-vl * qwen3-vl (2B, 32B, Instruct and Thinking) * dots.ocr * olmOCR 2 (based on Qwen2.5-VL) * LightOnOCR (smallies) top 5 trending models on @huggingface are still OCR/VLM! https://x.com/MaziyarPanahi/status/1981421331053760775
DeepSeek-OCR Contexts Optical Compression https://x.com/_akhaliq/status/1980260630780162505
DeepSeek OCR dropped … but honestly, Glyph [1], released the same day, showed something more interesting: 3–4× context compression and infilling cost reduction, no performance hit on long-context QA and summarization, which is much less trivial than OCR in many cases. If that https://x.com/arankomatsuzaki/status/1980722682246398069
🚀 DeepSeek-OCR — the new frontier of OCR from @deepseek_ai , exploring optical context compression for LLMs, is running blazingly fast on vLLM ⚡ (~2500 tokens/s on A100-40G) — powered by vllm==0.8.5 for day-0 model support. 🧠 Compresses visual contexts up to 20× while keeping https://x.com/vllm_project/status/1980235518706401405
Introducing Coral NPU: A full-stack platform for Edge AI – Google Developers Blog https://developers.googleblog.com/en/introducing-coral-npu-a-full-stack-platform-for-edge-ai/
🚨🎬 Big news from Video Arena! @GoogleDeepMind’s latest Veo 3.1 now ranks #1 in both Text-to-Video and Image-to-Video leaderboards. 🏆 This is a +30-point leap from Veo 3.0 → 3.1, making it the first model to break 1400 in Video Arena history! Huge congrats to the https://x.com/arena/status/1980319296120320243
Now available on @huggingface, the new dataset of 1.5 million task scenarios, field-tested and open-sourced by researchers from @IBM and the @UW, is designed to improve how agents interact with the world and get things done: https://x.com/IBMResearch/status/1981066891062817274
DyPE Dynamic Position Extrapolation for Ultra High Resolution Diffusion https://x.com/_akhaliq/status/1981705074490704366
We’re one step closer to truly predictable reinforcement learning. @Meta’s ScaleRL brings the first solid framework to predict RL scaling in LLMs, estimating large-scale results from small runs. Their study also revealed that: • Not all RL setups scale the same way. • Tuning https://x.com/TheTuringPost/status/1981487666714800356
🧠 How can we equip LLMs with memory that allows them to continually learn new things? In our new paper with @AIatMeta, we show how sparsely finetuning memory layers enables targeted updates for continual learning, w/ minimal interference with existing knowledge. While full https://x.com/realJessyLin/status/1980662516285075762
OpenAI’s ‘embarrassing’ math | TechCrunch https://techcrunch.com/2025/10/19/openais-embarrassing-math/
Open source coding benchmarks are operating in a different reality. They don’t test real world tasks and expect users to come prepared with a detailed page-long spec of exactly what they want to build or fix. But real people don’t use AI this way. They write vague prompts like https://x.com/pashmerepat/status/1981431374386233840
Very cool open-source work from @PyTorch on reinforcement learning environments (we helped a tiny bit)! Feels like early days on the topic with already exciting work from @PrimeIntellect @MechanizeWork @mercor_ai for example but exciting to make this topic as open-source and https://x.com/ClementDelangue/status/1981737560566005950
Choose the “”:exacto”” version of open-source models in Cline automatically route to the best inference provider for models like GLM-4.6, Qwen3-Coder, and Kimi-K2. Provider quality varies wildly, meaning the same model can yield completely different results at different endpoints. https://x.com/cline/status/1981370535176286355
Across most medical benchmarks, including when real cases & human doctors are involved, there is a clear trend of AI models improving over time (and many where today’s AI beats human doctors) But we do not have many studies measuring real-world performance of AI in medicine, yet https://x.com/emollick/status/1980474407656227258
Cool math insight on Weisfeiler–Lehman color refinement and Attention. Really nicely done!”” / X https://x.com/_arohan_/status/1981546840454811747
Because life’s too short for bad research papers. https://x.com/fdaudens/status/1979917544711442917
We are in the “”gentleman scientist”” era of AI research https://www.seangoedecke.com/ai-and-informal-science/
This is an interesting set of academic research papers about the increasingly key debate of when AI should be used to label data (an expensive task we use humans for) Yang et al show that AI answers are quite different than human, but Briggs finds it may be because AI is better https://x.com/emollick/status/1980476586437799990
Anthrogen https://www.anthrogen.com/odyssey-launch
I think the “Erdos problem” blowup obscured the fact that multiple math professors have confirmed recently that yes, AI really can solve some open (but not yet major) problems in mathematics, with guidance The question is whether the ability of these models continues to increase”” / X https://x.com/emollick/status/1980801216729759934
The fallout from the fact that data science/classical machine learning & generative AI are both called “”AI”” has been remarkably broad & persistent. Policy addresses the wrong harms, companies have been confused about who should lead efforts, academic discussion is often muddled.”” / X https://x.com/emollick/status/1981000395137921071
MEG-GPT: A transformer-based foundation model for magnetoencephalography data MEG is another noninvasive neuroimaging technique similar to fMRI. Don’t think I’ve seen a large-scale foundation model for this type of data yet, very interesting! “”we introduce MEG-GPT, a https://x.com/iScienceLuvr/status/1980945270369399234
Nice, short post illustrating how simple text (discrete) diffusion can be. Diffusion (i.e. parallel, iterated denoising, top) is the pervasive generative paradigm in image/video, but autoregression (i.e. go left to right bottom) is the dominant paradigm in text. For audio I’ve https://x.com/karpathy/status/1980347971935068380
Two quick updates: GLM-4.6-Air is still in training. We’re putting in extra effort to make it more solid and reliable before release. Rapid growth in GLM Coding Plan over the past weeks has increased inference demand. Additional compute is now deployed to deliver faster and”” / X https://x.com/Zai_org/status/1981700688401879314
Looking back at an exponentially improving technology & you will see how momentum led to R&D which overcame tech barriers The fact that reasoners were developed at exactly the moment AI pre-training slowed is how Moore’s Law works, too: new technique appear to maintain the trend https://x.com/emollick/status/1980601776710525103
BAPO (Balanced Policy Optimization with Adaptive Clipping) from @FudanUniversity dynamically adjusts PPO’s clipping bounds during training, balancing positive and negative updates. It stabilizes off-policy RL, preserves exploration, and surpasses Gemini-2.5 and o3-mini with a https://x.com/TheTuringPost/status/1981860282629837136
BERT is just a Single Text Diffusion Step | nathan.rs https://nathan.rs/posts/roberta-diffusion/
I have a thing for empirical deep dive into learning dynamics like done in this paper. Sounds like muP mostly helps the early training, while wd affects the long term.”” / X https://x.com/giffmana/status/1981483376604565969
🧠Do reasoning models really follow our instructions? Together AI’s newest paper “”ReasonIF: Large Reasoning Models Fail to Follow Instructions During Reasoning””, studies how well large reasoning models (LRMs) follow user instructions during reasoning. Authors: @ykwon_0407, https://x.com/togethercompute/status/1981441935303975059
Stop what you are doing and try out GEPA now! “”GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning”” presents such elegant ideas by a collection of amazing researchers! Here is a tldr of how it works: GEPA (Genetic-Pareto) is a sample-efficient prompt”” / X https://x.com/joelniklaus/status/1980651047720001884
Super grateful to the @thinkymachines team for their recent post showing the importance of targeting the expert layers for LoRA training of MoEs. This was a huge PR and I doubt it would have gotten into vLLM so quickly without that impetus!”” / X https://x.com/corbtt/status/1980678250608443467
One sign of this being a really cool idea is that while reading, I had tons of follow-up ideas immediately come to mind, and only few “”hmm but””s. Plz read thread and paper, but TLDR: add layer of input independent kv, and fine-tune only the high tfidf kvs for continual learning.”” / X https://x.com/giffmana/status/1980869216149619009
You spend $1B training a model A. Someone on your team leaves and launches their own model API B. You’re suspicious. Was B was derived (e.g., fine-tuned) from A? But you only have blackbox access to B… With our paper, you can still tell with strong statistical guarantees”” / X https://x.com/percyliang/status/1981612361309098383
Automating Algorithm Discovery: A Case Study in MoE Load Balancing https://adrs-ucb.notion.site/moe-load-balancing
Language Models are Provably Injective and Invertible! A groundbreaking paper challenges the long-held belief that LLMs lose information. They prove mathematically and show empirically across billions of tests that inputs map uniquely to representations, making them lossless. https://x.com/HuggingPapers/status/1981452722495787286
GPTQ is a post-training, weights-only quantization method that compresses a model to int4 layer by layer. For each layer, it uses a second-order method to update weights while minimizing the error on a calibration dataset. It comes built-in in Keras 3 (works with JAX, TF, torch) https://x.com/fchollet/status/1980343806265552918
Here is what the future looks like for training at scale: 1) Pre training: torchtitan -> for coms: torchcoms (new) -> low precision: torchao -> framework: monarch -> decentralize: torchft 2) Post training: torchforge (new) -> training: torchtitan -> inference: vllm -> RL env:”” / X https://x.com/eliebakouch/status/1980642834404319388
Prompt-MII: Meta-Learning Instruction Induction for LLMs https://x.com/gneubig/status/1980646347188707787
Notes on this Continual Learning paper. https://x.com/nrehiew_/status/1981714450089676877
LLMs Can Get “”Brain Rot””! https://llm-brain-rot.github.io/
The coverage principle: How pre-training enables post-training New preprint where we look at the mechanisms through which next-token prediction produces models that succeed at downstream tasks. The answer involves a metric we call the “”coverage profile””, not cross-entropy. https://x.com/canondetortugas/status/1981481591177105740
Another interesting paper about how to scale weight decay with muP for AdamW, from a different perspective. They argue for «independent weight decay scaling» as a way to maintain relative update sizes and ensure hyperparameter transfer. https://x.com/tonysilveti/status/1981406663086391588
[2507.02151] Non-exchangeable Conformal Prediction for Temporal Graph Neural Networks https://arxiv.org/abs/2507.02151
🤔 Can AI optimize the systems it runs on? 🚀 Introducing FlashInfer-Bench, a workflow that makes AI systems self-improving with agents: – Standardized signature for LLM serving kernels – Implement kernels with your preferred language – Benchmark them against real-world serving https://x.com/shanli_xing/status/1980705452699926851
🚨 RL Collapse Explained! 🚨 Anyone doing RL has seen it: reward curves suddenly crash—aka “”training collapse.”” 🤯 Ling Team dev & Zhihu contributor 萧轩 reveal it’s due to train-inference mismatch, a core RL issue, and share practical fixes. 🤔 Why RL collapses: Unlike https://x.com/ZhihuFrontier/status/1981337266523164694
To learn more about temporal difference learning, you could read the original paper ( https://x.com/RichardSSutton/status/1980150877177688544
A fun collaborative project! We leverage TunedLens (~linear decoding of tokens) to explore how LLMs’ internal representations change from layer to layer. 1/”” / X https://x.com/neuranna/status/1981357907170959799
torch.OutOfMemoryError: CUDA out of memory.”” *sigh* One always remember the awesomeness of flash attention when it is missing. @tri_dao”” / X https://x.com/francoisfleuret/status/1981487811489317175
@_fqnn_ You can 𝚊𝚜𝚢𝚗𝚌 𝚏𝚞𝚗𝚌𝚝𝚒𝚘𝚗 𝚍𝚊𝚒𝚕𝚢() { “”𝚞𝚜𝚎 𝚠𝚘𝚛𝚔𝚏𝚕𝚘𝚠””; 𝚠𝚑𝚒𝚕𝚎 (𝚝𝚛𝚞𝚎) { 𝚊𝚠𝚊𝚒𝚝 𝚜𝚕𝚎𝚎𝚙(“”1 𝚍𝚊𝚢””); 𝚊𝚠𝚊𝚒𝚝 𝚢𝚘𝚞𝚛𝚃𝚑𝚒𝚗𝚐(); } } It’s wild how robust the model is. It seems magical but it’s ackchually super consistent”” / X https://x.com/rauchg/status/1981426366982824387
Inspect is my favourite evaluation library and now you can easily evaluate huge models directly from your laptop :)”” / X https://x.com/_lewtun/status/1981692392295276885
use workflow”” And your async await calls become durable. Supported everywhere TypeScript runs https://x.com/cramforce/status/1981399119559348290
With project-scoped Memory, I’m now completely project-pilled. It’s the best way to avoid cross-pollination of memories into conversations where they’re not applicable.”” / X https://x.com/alexalbert__/status/1981421146886328778
Re memory layers: Shouldn’t memory layers include a sink (ie. one memory slot with score 0 which is always included that can be used as none-of-the-above)? Performing softmax on the top-k means that the memory layer can never defer, so you will always grab some random memory https://x.com/BlackHC/status/1981022197415068129
Last night I taught nanochat d32 how to count ‘r’ in strawberry (or similar variations). I thought this would be a good/fun example of how to add capabilities to nanochat and I wrote up a full guide here: https://x.com/karpathy/status/1981746327995465816
BOLT – How Mura wrote an in-house LLM Eval Framework https://mackey.substack.com/p/bolt-how-mura-wrote-an-in-house-llm
Check out deep dive into all our perf updates: https://x.com/basetenco/status/1981757380816748757
Bubble, Bubble, Toil and Trouble – by Zvi Mowshowitz https://thezvi.substack.com/p/bubble-bubble-toil-and-trouble
LoRA finetuned experts in MoE now runs properly in vLLM https://x.com/casper_hansen_/status/1980525929026973904
bring-your-own-inference (BYOI) strikes again”” / X https://x.com/cline/status/1980311303001633125
Context engineering is sleeping on the humble hyperlink • mbleigh.dev https://mbleigh.dev/posts/context-engineering-with-links/
Kimi K2 is up to 5x faster and 50% more accurate ：）”” / X https://x.com/crystalsssup/status/1980147163629047854
Excited to partner with @PyTorch on bringing over 2,000+ RL environments from OpenEnv to Unsloth! All Atari games + tonnes others now work out of the box in Unsloth! Check out our free Colab notebook on RLing 2048: https://x.com/danielhanchen/status/1981428184215363956
Wanted to get better intuitions for how RL works on LLMs. So I wrote a simple script to teach Nanochat to add 5 digit numbers. I was surprised at how fast it learned. Until I looked at the model’s generations and realized that it had just learned to always call the built-in https://x.com/dwarkesh_sp/status/1980427914639524278
Papers like this show that are a lot of potential pathways forward on some of the hardest outstanding problems in AI. The amount of low-hanging fruit suggests that AI lab R&D might continue to find ways around barriers to continual improvement of AI models.”” / X https://x.com/emollick/status/1980704687377486182
Some of the most interesting experimentation in AI are people trying to chart a way forward by developing new approaches to doing work based on the strengths and weaknesses of the models we already have. Example: @danshapiro on slot machine deployment: https://x.com/emollick/status/1978883777117008116
There is a lot of analytical clarity in thinking of AI as a large-scale industry, not just a technology. The comparison is not only to consumer technologies but to industrial policy decisions: where should resources go? What trade-offs does it have relative to other industries?”” / X https://x.com/emollick/status/1980687729232105472
This thread backs up my belief that fine-tuning is mostly useful in narrow situations (where it can be very valuable) but I remain skeptical that fine-tuning is the right solution for many problems for which it is being proposed. Often prompting is all you need, try it first.”” / X https://x.com/emollick/status/1979291926772724017
New method for automatic prompt optimization! TL;DR: We do reinforcement learning to train an LM that can take in a lot of task examples and generate a prompt that describes the task. Trained/tested on 3000+ classification datasets on Hugging Face!”” / X https://x.com/gneubig/status/1980644772902789603
Codex IDE extension is growing incredibly quickly. How to get the most out of it:”” / X https://x.com/gdb/status/1979268596267438588
Helion: A High-Level DSL for Performant and Portable ML Kernels – PyTorch https://pytorch.org/blog/helion/
How Well Does RL Scale? — LessWrong https://www.lesswrong.com/posts/xpj6KhDM9bJybdnEe/how-well-does-rl-scale
Import AI 431: Technological Optimism and Appropriate Fear | Import AI https://jack-clark.net/2025/10/13/import-ai-431-technological-optimism-and-appropriate-fear/
As part of our recent work on memory layer architectures, I wrote up some of my thoughts on the continual learning problem broadly: Blog post: https://x.com/realJessyLin/status/1980697898141774017
LLM Exchange Rates Updated – by Arctotherium https://arctotherium.substack.com/p/llm-exchange-rates-updated
LOLMIL: Living Off the Land Models and Inference Libraries https://dreadnode.io/blog/lolmil-living-off-the-land-models-and-inference-libraries
Neuro SAN is All You Need https://www.cognizant.com/us/en/ai-lab/blog/neuro-san-is-all-you-need0
Below is a deep dive into why self play works for two-player zero-sum (2p0s) games like Go/Poker/Starcraft but is so much harder to use in “”real world”” domains. tl;dr: self play converges to minimax in 2p0s games, and minimax is really useful in those games. Every finite 2p0s https://x.com/polynoamial/status/1980697004658556972
Origin | Nucleus Labs https://mynucleus.com/labs/origin
Spread your workloads across the globe with many regions and clouds, to get the best availability, cost-efficiency, and runtime. Check out this example for running large-scale batch inference across the globe: https://x.com/skypilot_org/status/1980307993842622471
Statement on Superintelligence https://superintelligence-statement.org/
The Continual Learning Problem https://jessylin.com/2025/10/20/continual-learning/
The Not-so Bitter Lesson – Marius Vach Blog https://blog.mariusvach.com/posts/bitter-lesson
Thoughts on the AI buildout https://www.dwarkesh.com/p/thoughts-on-the-ai-buildout
Really like the way the insights engine automatically breaks down the categories and scores for our projects. LangSmith has always been a powerful tool for helping you look at your data, but we want it to require even less elbow grease.”” / X https://x.com/WHinthorn/status/1981403256598192451
📢 New Model Drop: MiniMax M2 is live on Yupp! This new model from @MiniMax__AI is built for strong reasoning and efficiency. Let’s see what it can do: https://x.com/yupp_ai/status/1981887934812082564
Welcome, to our new work, BAPO, which tackles the challenge of Off-Policy Reinforcement Learning for Large Language Models across settings like Partial Rollout and Experience Reuse. Code: https://x.com/Be1ong1/status/1981297924564046007
We discovered GLM-4.6 was failing in Cline not because the model was flawed, but because inference providers were silently corrupting it. Using the same weights across different endpoints produced completely different behaviors. The variance wasn’t minor, it determined whether”” / X https://x.com/canvrno/status/1981403534471119330
Looking back, Llama4 MoE architecture was very different from recent open models 1) alternating dense/MoE layers with a 1:1 ratio 2) top-k=1 with 1 shared expert, but with 128 total experts for Maverick, which leads to a sparsity of 64. > for comparison, the sparsity of https://x.com/eliebakouch/status/1981747185373827079
Lookahead Routing for LLMs Proposes Lookahead, a routing framework to enable more informed routing without full inference. Achieves an average performance gain of 7.7% over the state-of-the-art. Here is why it works: Lookahead is a new framework for routing in multi-LLM https://x.com/omarsar0/status/1981360482813710384
Monarch seems very nice: https://x.com/finbarrtimbers/status/1980681034359533861
@BlackHC also memory layers seem like they’d be extremely slow? or at least, totally unable to mfu-maxx. those independent random accesses per token are going to kill performance.”” / X https://x.com/gallabytes/status/1981038852539371969
Welcome Ray from @anyscalecompute to join the PyTorch foundation family 🎉”” / X https://x.com/vllm_project/status/1981045521671393441
🚀 Excited to share our work on batch-invariant inference in vLLM! Now you can get identical results regardless of batch size with just one flag: VLLM_BATCH_INVARIANT=1 No more subtle differences between bs=1 and bs=N (including prefill!). Let’s dive into how we built this 🧵👇 https://x.com/vllm_project/status/1981088861506982041
our login broke. went to sst to check why, theirs was also broken. turns out us-east-1 is down https://x.com/vikhyatk/status/1980171953614012448
Continual Learning via Sparse Memory Finetuning Jessy Lin, Luke Zettlemoyer, Gargi Ghosh, Wen-Tau Yih, Aram Markosyan, Vincent-Pierre Berges, Barlas Oğuz https://x.com/nrehiew_/status/1981714473560801446
GitHub repo: https://x.com/_avichawla/status/1981246746497077492
HF Datasets: built for audio, images, videos… And now, PDFs 📕 Still loadable in one line of code: >>> load_dataset(“”username/my_dataset””) What should we do next for OCR datasets ? 🤗 https://x.com/lhoestq/status/1981720383620358449
We’re kicking off an early beta of the new Sesame iOS app, which includes the ability to search, text and think. https://x.com/brendaniribe/status/1980677775058162077
Cool work!”” / X https://x.com/OfirPress/status/1980319814481817901
World-in-World: World Models in a Closed-Loop World https://world-in-world.github.io/
Finally, researchers have open-sourced a new reasoning approach that actually prevents hallucinations in LLMs. It beats popular techniques like Chain-of-Thought and has a SOTA success rate of 90.2%. Here’s the core problem with current techniques that this new approach solves: https://x.com/_avichawla/status/1980159925109309799
New User Trends on Wikipedia – Diff https://diff.wikimedia.org/2025/10/17/new-user-trends-on-wikipedia/
.@romeovdean and I wrote a blog post to teach ourselves about the AI buildout. We were surprised by some of the things we learned: 1. There’s a huge fab CapEx overhang – with a single year of earnings in 2025, Nvidia could cover the last 3 years of TSMC’s ENTIRE CapEx. In 2025, https://x.com/dwarkesh_sp/status/1981074799758921843
kvcached works directly with vLLM and you can use it to serve multiple models on the same GPU. They will share unused KV cache blocks. Check it out!”” / X https://x.com/vllm_project/status/1980776841129701411
I wrote this piece in the Harvard Business Review in December, 2022, two weeks after ChatGPT was released (its seven pages, so these are the first two and last two). I think my predictions all played out, though I underestimated how good they would get at accurate math. https://x.com/emollick/status/1979573319121916037
Remarkable: if someone steals your model, you can catch them using only your training data order and their model’s predictions, even if they do a bunch of fine-tuning. This power traces to subtle, pervasive memorization patterns: the palimpsest.”” / X https://x.com/ChrisGPotts/status/1981739673077657832
OpenEnvs for Reinforcement Learning! 🙏 We are launching a universal RL Environment interface today, teaming up with @huggingface and @UnslothAI Let’s take a trip down memory lane: It’s 2016, you read some papers. RL looks promising. But the reality? Cartpole is best we https://x.com/bhutanisanyam1/status/1981377720157351938
14B, 11FPS on B200, Real-time -Apache 2.0 licensed on Hugging Face 📹”” / X https://x.com/reach_vb/status/1980376352726610342
Do AIs think differently in different languages? https://www.theargumentmag.com/p/do-ais-think-differently-in-different
Tomorrow https://x.com/dwarkesh_sp/status/1978885626255839419
Until recently, datacenters were able to simply repurpose the carcass left over from America’s deindustrialization. Honestly a compelling ode to capitalism. https://x.com/dwarkesh_sp/status/1981113646836306030
Building an infinitely learning geolocation model with online RL – sdan blog https://blog.sdan.io/geospot-infinity/
MoGA Mixture-of-Groups Attention for End-to-End Long Video Generation https://x.com/_akhaliq/status/1980952993563349127
Awesome to see Veo 3.1 top the LMArena video leaderboards by a large distance with big improvements over Veo 3.0 for text-to-video (+30) and image-to-video (+70)! 🔥Huge congrats to the team! Try it for yourself in https://x.com/demishassabis/status/1980397419658645708
Open-o3 Video Grounded Video Reasoning with Explicit Spatio-Temporal Evidence https://x.com/_akhaliq/status/1981564465897509333
Ditto Scaling Instruction-Based Video Editing with a High-Quality Synthetic Dataset https://x.com/_akhaliq/status/1980265202500116525
UltraGen High-Resolution Video Generation with Hierarchical Attention https://x.com/_akhaliq/status/1980952631544799705
