Your curated collection of saved posts and media
π Live from @NVIDIAGTC, we're releasing Holotron-12B! Developed with @nvidia, it's a high-throughput, open-source, multimodal model engineered specifically for the age of computer-use agents. Get started today! π€Hugging Face: https://t.co/oaSviLi8IN πTechnical Deep Dive: https://t.co/pDItQB1frU πΌWe are hiring: https://t.co/fcNoR9FIYQ #AI #ComputerUse #NVIDIA #OpenSource #ReinforcementLearning #HCompany #GTC @NVIDIAAI @nvidia
Two weeks after sharing @adaption_ai adaptive data, we are excited to share our work on blueprint π blueprint steers data towards your goals, and learns penalties if AI violates any of your rules. Very proud of the team.
Yes! The βare you sure?β problem (link below) is especially pervasive in any complex coding task. Ask Claude or GPT to review a PR then ask it to double check its findings when it finishes - itβll flip on at least 1 of its findings. https://t.co/WUSuOxTuDy
Yes! The βare you sure?β problem (link below) is especially pervasive in any complex coding task. Ask Claude or GPT to review a PR then ask it to double check its findings when it finishes - itβll flip on at least 1 of its findings. https://t.co/WUSuOxTuDy
One thing that makes me feel that code factory has not arrived yet is the following experiment: 1.Ask a LLM to do an in-depth rigorous review of your code 2. In a new thread, as same/different LLM to consider those review comments independently and address issues it agrees with
πA closer look at Omnilingual No Language Left Behind, the encoder-decoder system presented as part of @AIatMeta new Omnilingual Machine Translation work!π Many say encoder-decoder is dead in the age of decoder-only LLMs but we show itβs not! π:https://t.co/isvEzRZbnw π§΅1/n https://t.co/RLs8ncUy0H
In our paper "Towards a Science of AI Agent Reliability" we put numbers on the capability-reliability gap. Now we're showing what's behind them! We conducted an extensive analysis of failures on GAIA across Claude Opus 4.5, Gemini 2.5 Pro, and GPT 5.4. Here's what we found β¬οΈ https://t.co/GkdAxk0wDO
Current vision-language models still struggle with simple diagrams. Feynman is a knowledge-infused diagramming agent that enumerates domain-specific concepts, plans visual representations, and translates them into declarative programs rendered by the Penrose diagramming system. Great insights for those building agents for diagrams and visualizations. One pipeline run produced 10,693 unique programs across math, CS, and science, each rendered into 10 layout variations, yielding over 106k well-aligned diagram-caption pairs. Paper: https://t.co/F4vNS0TII4 Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
One thing that makes me feel that code factory has not arrived yet is the following experiment: 1.Ask a LLM to do an in-depth rigorous review of your code 2. In a new thread, as same/different LLM to consider those review comments independently and address issues it agrees with 3. Keep repeating until no new concerns I find that this loop always goes on for a ridiculously long time, which means that there is a problem with the notion of claude-take-the-wheel. This seems to happen no matter the harness or the specificity of the specs. It works fine for simple applications, but in the limit if the LLMs have this much cognitive dissonance you cannot trust it. Either this, or LLM are RLHFd to always find some kind of issue.
Even the best reasoning models hit an accuracy collapse beyond a certain problem complexity. Giving an LRM the exact solution algorithm doesn't fix it either. This new work, BIGMAS, improves LLM agents by taking inspiration from the human brain. BIGMAS outperforms both ReAct and Tree of Thoughts across all three tasks. It organizes specialized LLM agents as nodes in a dynamically constructed directed graph, coordinated through a centralized shared workspace inspired by global workspace theory. A GraphDesigner builds task-specific agent topologies per problem, and a global Orchestrator routes decisions using the complete shared state, eliminating the local-view bottleneck of reactive approaches. Across Game24, Six Fives, and Tower of London on six frontier LLMs, including GPT-5 and Claude 4.5, BIGMAS consistently improves accuracy. The gains are largest where models struggle most: DeepSeek-V3.2 jumps from 12% to 30% on Six Fives. Paper: https://t.co/sMqUfvHAGp Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c
πA closer look at Omnilingual No Language Left Behind, the encoder-decoder system presented as part of @AIatMeta new Omnilingual Machine Translation work!π Many say encoder-decoder is dead in the age of decoder-only LLMs but we show itβs not! π:https://t.co/isvEzRZbnw π§΅1/n https://t.co/RLs8ncUy0H
Current vision-language models still struggle with simple diagrams. Feynman is a knowledge-infused diagramming agent that enumerates domain-specific concepts, plans visual representations, and translates them into declarative programs rendered by the Penrose diagramming system. Great insights for those building agents for diagrams and visualizations. One pipeline run produced 10,693 unique programs across math, CS, and science, each rendered into 10 layout variations, yielding over 106k well-aligned diagram-caption pairs. Paper: https://t.co/F4vNS0TII4 Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
This is one of the most impressive world model projects I have seen. Very elegant and highly effective combination of an image retrieval mechanism (using 3D locations/views) and otherwise just pure generative modeling. This is the way.
What if a world model could render not an imagined place, but the actual city? We introduce Seoul World Model, the first world simulation model grounded in a real-world metropolis. TL;DR: We made a world model RAG over millions of street-views. proj: https://t.co/Bx4KUAqrRs ht
... and a follow-up chapter about Subagents, now a feature of Codex and Claude Code and Gemini CLI and Mistral Vibe and OpenCode and VS Code and Cursor https://t.co/suGmK4g3Hp
i dj'd a set this weekend and planned half of it with an app i built on @GoogleDeepMind's new multimodal embeddings model. it understands what music actually sounds like - not bpm, not genre tags. raw audio in, vibes out. built with @cursor_ai + @openai gpt 5.4, gemini embeddings and @convex to keep everything running smoothly @ericzakariasson @DynamicWebPaige @waynesutton @gabrielchua here's what it does (demo below) π§΅
i dj'd a set this weekend and planned half of it with an app i built on @GoogleDeepMind's new multimodal embeddings model. it understands what music actually sounds like - not bpm, not genre tags. raw audio in, vibes out. built with @cursor_ai + @openai gpt 5.4, gemini embeddings and @convex to keep everything running smoothly @ericzakariasson @DynamicWebPaige @waynesutton @gabrielchua here's what it does (demo below) π§΅
In our paper "Towards a Science of AI Agent Reliability" we put numbers on the capability-reliability gap. Now we're showing what's behind them! We conducted an extensive analysis of failures on GAIA across Claude Opus 4.5, Gemini 2.5 Pro, and GPT 5.4. Here's what we found β¬οΈ https://t.co/GkdAxk0wDO
Tried the viral poster-skill with Claude Code on the trending Moonshot paper :) Not too bad! https://t.co/I8lb0aUrbT
Tried the viral poster-skill with Claude Code on the trending Moonshot paper :) Not too bad! https://t.co/I8lb0aUrbT
Introducing π¨ππππππππ πΉππππ ππππ: Rethinking depth-wise aggregation. Residual connections have long relied on fixed, uniform accumulation. Inspired by the duality of time and depth, we introduce Attention Residuals, replacing standard depth-wise recurrence with learned, input-dep
Two weeks after sharing @adaption_ai adaptive data, we are excited to share our work on blueprint π blueprint steers data towards your goals, and learns penalties if AI violates any of your rules. Very proud of the team.
Introducing Blueprint, a new capability within Adaptive Data. We firmly believe data that evolves with the world is only useful if it evolves the right way. Blueprint allows you to steer the data space towards any goal you want. https://t.co/8k0WEMYmdd
βEveryone should be a GPU programmer.β @clattner_llvm's goal with @Modular: βWhat Modular is doing is opening up the box. Weβre fixing the language problem and the platform problem. "The goal is to let more developers learn modern compute. And to give developers real choice in the hardware they use.β βThose two things unlock the ecosystem.β
... and a follow-up chapter about Subagents, now a feature of Codex and Claude Code and Gemini CLI and Mistral Vibe and OpenCode and VS Code and Cursor https://t.co/suGmK4g3Hp
Excited to introduce OmniClone, a robust teleoperation system for humanoid mobile manipulation. While systems like TWIST2 and SONIC paved the way, we put efforts into solving the critical stability and scaling gaps. 1/ π Moving past "vibe-based" testing. Weβve built a comprehensive diagnostic benchmark to systematically evaluate whole-body teleoperation. No more trial-and-errorβget the actionable insights needed for true policy optimization. 2/ π€ Universal Human-to-Robot Mapping. Teleop often breaks when switching operators. OmniClone mitigates biases from hardware fluctuations and, crucially, diverse human body shapes, ensuring high-stability control regardless of the person in the suit. 3/ π System Optimizations for Whole-body Manipulation Policy. By optimizing for affordability and reproducibility, OmniClone provides the high-fidelity pipeline necessary to collect data and train humanoid whole-body policies at scale. fully The model checkpoints and deploy code are now fully releasedβwelcome to play with it! π¦ π Paper: https://t.co/kDm60WeuMD π Project: https://t.co/WGcfYridEs π» Code: https://t.co/U1QLgaipcd
@SalajSonar1086 Already done :). The respective tutorial articles are linked via the βView in Articleβ links there
Happy to share πOmnilingual Machine Translationπ In this work @AIatMeta we explore translation systems supporting 1,600+ languages. We show how our models (1B to 8B) can outperform baselines of up to 70B while having much larger language coverage. π:https://t.co/isvEzRZbnw https://t.co/8sdgkQuJ3B
π Live from @NVIDIAGTC, we're releasing Holotron-12B! Developed with @nvidia, it's a high-throughput, open-source, multimodal model engineered specifically for the age of computer-use agents. Get started today! π€Hugging Face: https://t.co/oaSviLi8IN πTechnical Deep Dive: https://t.co/pDItQB1frU πΌWe are hiring: https://t.co/fcNoR9FIYQ #AI #ComputerUse #NVIDIA #OpenSource #ReinforcementLearning #HCompany #GTC @NVIDIAAI @nvidia
VisMatch is on pypi! VisMatch is a wrapper for image matching models, like LightGlue, RoMa-v2, MASt3R, LoFTR, and 50+ more! It's literally as simple as: pip install vismatch vismatch-match --inputs img0 img1 --matcher choose_any To run image matching on any 2 images [1/4] https://t.co/dIr2YapWak
Jensen today announced Alpamayo 1.5 at #NVIDIAGTC! #Alpamayo 1.5 is a major update to Alpamayo 1β@nvidiaβs open 10B-parameter chain-of-thought reasoning VLA model, first introduced at #CES. Built on the #Cosmos-Reason2 VLM backbone and post-trained with RL, it adds support for navigation guidance, flexible multi-camera setups, configurable camera parameters, and user question answering. The result is an interactive, steerable reasoning engine for the AV community. Weβre also releasing post-training scripts to help researchers and developers adapt the model. Additionally, weβve significantly expanded the Alpamayo open platform across data and simulation, including releasing highly requested reasoning labels for the PhysicalAI Autonomous Vehicles dataset (https://t.co/fD9eUcndya), as well as our chain-of-causation auto-labeling pipeline. π Learn more about Alpamayo 1.5 and the latest extensions to the Alpamayo open platform: https://t.co/P0nuqkwBab (please note that most of the links will become active in the next few days.) Happy buildingβand stay tuned for more in the coming months! @NVIDIADRIVE @NVIDIAAI
π οΈ Claude Code "opusplan" λ§ κ·Έλλ‘ νμ΄λΈλ¦¬λ λͺ¨λΈ.. 곡μμ! Claude Codeμλ opusplan λͺ¨λΈμ μ νν μ μμ΄μ. > /model opusplan νμ΄λΈλ¦¬λ λͺ¨λΈ aliasμΈλ°, μμ λ¨κ³μ λ°λΌ μλμΌλ‘ λͺ¨λΈμ μ νν΄μ. 볡μ‘ν μΆλ‘ μ μν νλ λͺ¨λμμλ Opusλ₯Ό μ€ν λ¨κ³μμλ SonnetμΌλ‘ μλ μ νλ©λλ€! Opusλ‘ κ³ννκ³ κ΅¬νκΉμ§ νλ κ²λ λ¬Όλ‘ κ°λ₯ν΄μ. νμ§λ§ μ΄λ―Έ ννν κ³νμ΄ μλ€λ©΄, μ€νμ SonnetμΌλ‘λ μΆ©λΆνκ³ λ μ λ ΄ν μ μμ΄μ. κ° μμ μ λ§λ λͺ¨λΈμ μ°λ κ² = ν¨μ¨ π νλλκ³Ό μ€νμ μꡬλλ μΈμ§ λΆνκ° λ¬λΌμ. Opusμ κΉμ μΆλ‘ λ₯λ ₯μ κ³ν μ립 λ¨κ³μμ κ°μ₯ λΉλκ³ , μΌλ¨ ννν κ³νμ΄ μΈμμ§ μ΄νμ μ€νμ SonnetμΌλ‘ μΆ©λΆν 컀λ²λ μ μμ΄μ. μΈμ μ°λ©΄ μ’λꡬμ? - 볡μ‘ν κΈ°λ₯ μ€κ³κ°μ΄ μν€ν μ² κ²°μ μ΄ μ€μν μμ - 리ν©ν λ§ κ³νκ°μ μν₯ λ²μ λΆμμ΄ νμν κ²½μ° - Opus ννμ μ¬μ© λλΉ λΉμ© μ κ°μ΄ νμν λ μ΄κ±° μ΄μ λ§μ΄ νμ©νμ€ λ―!!!
Did you know about the opusplan model in Claude Code? /model opusplan It's a hybrid alias that automatically uses Opus in plan mode for complex reasoning, then switches to Sonnet for execution. Best of both worlds: Opus thinks, Sonnet builds https://t.co/r7un0X5bVg
Subagents are now supported in Codex. They're very fun and make it possible to get large amounts of work done *quickly*:
Subagents are now available in Codex. You can accelerate your workflow by spinning up specialized agents to: β’ Keep your main context window clean β’ Tackle different parts of a task in parallel β’ Steer individual agents as work unfolds https://t.co/QJC2ZYtYcA
Banger report from the Kimi team: Attention Residuals Residual connections made deep Transformers trainable. But they also force uncontrolled hidden-state growth with depth. This work proposes a cleaner alternative. It introduces Attention Residuals, which replace fixed residual accumulation with softmax attention over previous layer outputs. Instead of blindly summing everything, each layer selectively retrieves the earlier representations it actually needs. To keep this practical at scale, they add a blockwise version that compresses layers into block summaries, recovering most of the gains with minimal systems overhead. Why does it matter? Residual paths have barely changed across modern LLMs, even though they govern how information moves through depth. This paper shows that making the mixing content-dependent improves scaling laws, matches a baseline trained with 1.25x more compute, boosts GPQA-Diamond by +7.5 and HumanEval by +3.1, while keeping inference overhead under 2%. Paper: https://t.co/04IG6FDiVr Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX
The moment he realised that https://t.co/vWmBsnR1nt isn't fully built on transformers and we can run on a single GPU with high accuracy and lower cost https://t.co/ZJYuL62UB8
The moment he realised that https://t.co/vWmBsnR1nt isn't fully built on transformers and we can run on a single GPU with high accuracy and lower cost https://t.co/ZJYuL62UB8
