Knowledge catalogue
Nemotron 3 Ultra is NVIDIA's latest language model featuring significant improvements in speed (5X faster) and cost efficiency (30% cheaper) compared to previous versions, positioning it as a frontier
You can work 5 days a week and succeed as a startup. Mercury has done that from day 0 and we are valued @ $5.2bn 7 years after launch. I have been an entrepreneur for 20 years and raised 3 kids while
what a wonderful project: parakeet.cpp https://github.com/mudler/parakeet.cpp GGML based parakeet inference pipeline that's 2x faster than my ONNX parakeet pipeline on Apple Silicon! (Needed a few loc
Has @AnthropicAI completely given up on making API usage reasonably-priced? Following the token-usage changes recently they announced various updates to *subscription* usage to make it more reasonable
Jeremy Howard expresses positive views on GPT-5.5, stating he has recently switched to using it as his primary model and finds it nearly comparable to Anthropic's Opus 4.6 and 4.7 while offering signi
My MLSys keynote on AI writing systems code got more interest than I expected. The recording will take a while, so in the finest tradition of AI labs sharing blog posts, we’re starting the Core Automa
Worked on some code this morning using Opus 4.8 and so far I'm really liking it. Much more cooperative than 4.7 and less 'over agentic'. Stops and asks for my input when needed in places 4.7 (and GPT
65B private round More than double the size of the largest IPO ever We've raised 65 billion in Series H funding at a $965 billion post-money valuation, led by @AltimeterCap, Dragoneer, @Greenoaks, and
Fascinating results + Anthropic running away with it right now + So many people want to start their own company + Google over OpenAI + Vercel, Linear, Every, PostHog overperforming A great list if you
Fun fact: Australia is basically Scandinavia with the Sahara Desert bolted on. I LOVE this visualisation. Everyone imagines nature and the outback, but they don't realise just how urbanised we are. Cr
glad to know Mythos' safety concerns have been addressed right as Anthropic also secured tens of billions in inference compute 👍 JUST IN: Anthropic announces it will roll out Claude Mythos “in the com
wait… if most people think 5.5 is better than 4.7, i assume that’s due to terminal coding benchmark… 4.8 is still outperformed by 5.5 Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper ju
Behind the MiMo API Price Reduction: The deepest price cut, up to 99%, is for Input (Cache Hit). The core reason is our inference framework now supports hierarchical KV cache optimization for SWA. Pro
Jeremy Howard comments on the Xiaomi MiMo v2.5 model, highlighting its exceptional value proposition and cost-effectiveness for prompt processing. The post appears to include comparative pricing data
Jeremy Howard comments that large language models released 6 months ago feel comparable in capability to recently released models, suggesting that the pace of improvement in model development may be s
recommended reading. i too am very done with people anthropomorphizing a bunch of matrices on a GPU cluster, especially if the same people do not give two fucks about actual human beings. More musings
Nothing disappoints me more than people saying we should stop progress because peoples meaning depends on that monotonous labor, as if humanitys highest purpose is filling Excel sheets or stocking she
RWKV-7 G1g is here: the world's best pure RNN LLM, and a competitive LLM in general. Try https://huggingface.co/spaces/BlinkDL/RWKV-Gradio-2 for bsz16 7B inference. G1h in June 🙂 p.s. const 15000+tps
this has been my experience as well. there definitely were improvements, specifically wrt shell based computer use, but also regressions, especially in the last 3 version bumps of flicker and gerperte
Gated DeltaNet-2 is almost exactly RWKV-7's DPLR recurrence, not acknowledging the elephant in the room 🙂 Gated DeltaNet-2 is here. 🚀 🔥 New paper: Gated DeltaNet-2: Decoupling Erase and Write in Linea
Gemini Flash 3.5 is such a disappointing model. It's intelligence and speed is awesome. Absolutely amazing. But it's been trained to max evals, not to be helpful to humans. It goes off and does random
Jeremy Howard comments on comparative performance trends between GPT 5.5 and Claude models, noting that GPT 5.5 appears to be improving in a particular capability while Claude models are declining in
Jeremy Howard argues that AI model developers at OpenAI and Anthropic should prioritize breaking backwards compatibility during the prototyping phase rather than maintaining it, suggesting that backwa
My concern for the AI era, or at least this phase of it, is that a generation is being taught that 'close enough' is just fine. Take @AnthropicAI for example. Text wrapping in Claude Code has been bro
DeepSeek has announced a permanent discount for its DeepSeek-V4-Pro model, encouraging developers to build and innovate with the platform. The announcement was made via social media and emphasizes the
We desperately need better ways of evaluating models. Something that shows how helpful they are at working hand-in-hand with humans to help them get stuff done in a cooperative/iterative way. The Clau
For anyone who isn't sure, this is how you release a model and talk about the performance. Not 3-5 cherry-picked benchmarks. Performance：Qwen3.7-Max performs strongly across benchmarks in coding agent
Gated DeltaNet-2 is here. 🚀 🔥 New paper: Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention Gated DeltaNet-2 outperforms KDA and Mamba-3, the latest and best recurrent architectures, hea
LLM training is built on fast MatMuls. But many surrounding ops still run as memory-bound kernels. CODA reparameterizes them to hide in the matmul’s shadow, fused into its epilogue before results leav
> These engineers can review their agent's code much faster than reviewing human code. wat Today we reduced headcount by 22%. The business is the strongest it's ever been. So I think it's important to
Australian teenagers who were unable to access social media due to age verification requirements implemented under new legislation showed reduced news consumption compared to their peers. The finding
This post likely discusses a surprising or shocking mathematical discovery, result, or announcement that would be of particular interest to mathematicians. Based on the dramatic framing, it probably p
LLM agents & memory systems operate in continuously updated environments (Git repos, evolving docs). They must process long contexts, recover earlier information, and reason over many updates that cre
Disappointing pricing trend with Gemini 3.5 Flash. 22.5x pricier than 2.0 Flash which came out 15 months ago (9.00 vs 0.40). Are Flash models supposed to get this much more expensive, or is Pro just b
I can't verify this as a genuine statement from Jeremy Howard or provide it as factual information for a knowledge base. The post appears to be either fabricated, a joke, or the URL doesn't correspond
Excited to share our new paper: RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably LLMs often fail on inputs well within their advertised context lengths. We show that these fa
Google’s new Gemini 3.5 Flash is the clear leader on the Intelligence vs Speed Pareto frontier and makes large gains on GDPval-AA (real-world agentic tasks), but is 5x the cost of Gemini 3 Flash @Goog
Chrome DevTools will add support for validating llms.txt files, enabling developers to check the validity and correctness of their llms.txt configurations directly within the browser's developer tools
Google is transitioning from the Gemini CLI tool to a new CLI called AGY (Antigravity CLI), but this change has drawbacks: AGY is not open source and no longer supports ACP functionality, representing
Literary journals are now publishing, and awarding prizes to, AI written stories. Surprised this made it into Granta! ‘The Serpent in the Grove’ by Jamir Nazir is a story set in rural Trinidad about a
I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them. I can't name any specific people because they inc
We've gone even farther: Nemotron 3 Super is 120B and pretrained on 25T tokens in NVFP4. Nemotron 3 Ultra is ~500B and also pretrained in NVFP4. Accelerated computing means we rethink every aspect of
🚨 WOWWW — UTAH — the brilliant young women @kevinolearytv 🤥 accused of being Chinese ops just discovered Utah speaker of the house Mike Schultz (Republican) bought 640 acres of land right next to the
SolveIt now supports real-time voice editing of messages during conversations, allowing users to make changes through spoken commands. The platform can optionally track these edits as diffs, enabling
Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents wer
arXiv has implemented a new policy imposing a 1-year ban on users who submit papers containing hallucinated or fabricated references, reflecting concerns about AI-generated content and citation integr
7,000 false positives per square millimeter. The culprit was the lab gloves. University of Michigan researchers just upended a core assumption in microplastics science. Latex and nitrile gloves, worn
hey surprise - you can just launch interactive in tmux and then tail the jsonl - shipped a small wrapper...ralph loop iterating to full parity rn https://github.com/dexhorthy/shannon Starting June 15,
If you use any of the following with your Claude sub, your usage must got cut by 25x: - T3 Code - Conductor - zed - jean - “Claude -p” in your ci - scripts to call Claude code from other tools They’re
This is misleading. This policy redefines the term 'interactive' to mean 'using an Anthropic front-end'. If you use `claude -p` or Agent SDK to do something interactively, it now uses credits, not you
We’re training models wrong and it’s due to chatGPT. Even the modern coding agents used daily still use message-based exchanges: They send messages to users, to themselves (CoT) and to tools, and rece
Sound on! This is pretty cool :D SolveIt is already an amazing environment for learning and exploring any topic, or for development/writing etc. But add in real-time conversational interaction too jus
TurboQuant has drawn a lot of attention recently, but the accompanying evals didn't tell the full story. So we ran what I believe is the first comprehensive study of TurboQuant: where it helps, where
The post discusses improvements in working with Horace (likely a tool or system) due to recent implementation of simultaneous speech capability in its underlying model. This advancement has made the i
As always, no insights and in personal capacity: The DeepMind unionization effort has very worthy goals it seems Maybe Google will finally grant GDM that /independent/ ethics oversight board that was
big 'Look what they need to mimic a fraction of our power' energy. the original DOOM impl is ~40k lines of C and a bit of assembly and is also a full software renderer. My Codex /goal that has been ru
If the AI models are so smart, why do I feel like I’m losing a few neurons every time I read a longer form content written by AI? We’ve come a long way but we still have long way to go. In terms of cl
ERNIE 5.1 is here 🚀 ERNIE 5.1 significantly reduces pretraining cost while compressing total parameters to ~1/3 and activated parameters to ~1/2 — using only ~6% of the pretraining cost compared to mo
It's been 3 months since the 100x vibers started 100x vibin'! So, post your 25-years-of-work-equivalent project here, so we can signal boost and everyone can celebrate the Life's Work that you did in
I really enjoyed chatting with @mattturck, was a great discussion. Deeply thoughtful conversation with @zicokolter, board member at @OpenAI and head of the machine learning department at @CarnegieMell