Search: “llm”

New work with @AlecRad and @DavidDuvenaud: Have you ever dreamed of talking to someone from the past? Introducing talkie, a 13B model trained only on pre-1931 text. Vintage models should help us to un

What if instead of building one giant AI, we evolved a coordinator to orchestrate a diverse team of specialized AIs? 🐟 Excited to share our new paper: “TRINITY: An Evolved LLM Coordinator”, published

One of my favorite things about Sakana Fugu is the recursive test-time scaling. When allowed to call itself recursively, it reads its own prior output and spins up corrective workflows on the fly. We

// Survey on Multi-Agent Systems // The paper traces the landscape from classical paradigms (consensus, distributed control, swarm intelligence, cooperative learning) to foundation-model-enabled MAS (

Today, Clearwing has replicated Glasswing's discovery of the @FFmpeg vulnerability using Codex-5.4 as the backend. I am going to attempt the same today, using @Alibaba_Qwen 3.6 on @ollama running on m

What you need to know about Opus 4.7 * Takes instructions literally * Better vision means improved computer use and producing slides and other visual artifacts * Optimized for large-scale real-world a

This post by Harrison Chase (or from his X feed) likely discusses the challenges and considerations around implementing proper memory management in AI agents or LLM-based applications. It reflects on

I am not just talking about the em-dashes and phrasing ('doing the heavy lifting'). Those are more fixable than style, which will always default to an LLM-normal even if that normal changes with new L

Gary Marcus will be speaking at the @Bloomberg New Economy Forum in New Delhi tomorrow, October 15, on AI reliability, control and safety. @GaryMarcus Hope you’ll connect all the dots between… LLM bul

Over the last few months at @river_ai_inc we’ve developed what we believe is the best system in the world for building your own internal custom models, driven by 10 years of experience training top mo

Had fun nerding out on the @_odsc AI podcast about: • Building in public as both market diligence and a way to meet founders • Jev experiments: knowledge graphs, relevance filters, and emotional blobs

Hot take: evals are the #1 thing keeping companies from becoming AI-native. If you want to understand evals, read this post & then the accompanying guide. Tweaking prompts and trying a few inputs by h

50K likes on this!? Most of what Karpthy has shared is things we have been doing for over a year. No surprises, as many replies are highlighting. But I still think this raises an interesting discussio

we're seeing our coding agent costs decrease significantly for second month in row longer blog on this later, but three quick steps we took 1/ cost visibility. you cant control what you cant measure.

America is starting to be World Model-pilled Heard a crazy stat the other day. Video generation represents around 70% of the share of AI token consumption in China. A combination of short form content

A little Ember-1 tl;dr. Seems like a great model! (Was recently asked on a podcast, given $ xx million, what's the best way to develop a frontier LLM today? My recommendation was: start with an existi

Cool paper on catching alignment failures with Jev. The ideas is to ask Jev one generic yes/no question about a model's response, and use its probability as a score. With no extra training, that score

i don’t get it fully but this looks cool We achieved a breakthrough at BTL. AI models are very linear. They think in one direction at a time, and when they're wrong they scramble and start again. For

Banger report on building faster memory for AI Agents. If your agent's memory layer is slow, this design is worth a look. Great weekend read. It proposes a new agentic memory architecture, Jev-Mem, in

We announced our RSI Lab earlier this year: https://sakana.ai/rsi-lab/ Over the last two years, we have systematically shipped the foundations for autonomous R&D: ▪ LLM²: AI automating research to inv

Impressive paper showing the impact of a good harness. Improves Qwen3-8B from 41.2% to 91.8% on long-horizon robot tasks without any change to the model. The gain comes from the harness. GAVEL keeps a

LlamaIndex 发布了一个面向 AI Agent 的文档解析评测基准「ParseBench」 @llama_index 100 种解析方法参与评测（vlm - 67、专用解析器 - 33），2000 页真实企业文档（人工两轮校验），169011 条确定性规则，完全不用 LLM 当裁判，可复现。评测五个维度取无加权平均：表格结构、图表抽取、内容忠实度、格式语义、视觉定位。 https://ww

new unredacted filings in the publishers (including my employer) vs OpenAI and Microsoft lawsuit Microsoft's Director of applied science calls LLM training 'an astonishing theft of unprecedented propo

Reasoning from scratch round 3: This time, I cover generating a verifier for... a) ...evaluation (base model versus any future model improvement) b) ...the reinforcement learning with verifiable rewar

I like this framing a lot: agent improvement is harness improvement, not model improvement The interesting interventions are often at the tool boundary: - what context you pass - when tools are provid

// Agent Zero Memory // This work separates three things that agent memory systems usually collapse into one. If you build agents with long-term memory, this memory design is a worth a read. Here is h

A couple of days ago, I did a quick explainer on Claude’s new watermarking process and implementation. Since it’s such a popular topic and sparked such a lively discussion, I thought it might be inter

Nice case study on using optimized functions whenever possible (except for educational purposes, though 😆) By switching from a hand-rolled GELU to PyTorch's built-in one, I improved my LLM training sp

DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro. ⚡ Up to 4.6× the speed of autoregressive decoding, with the same output. This is the next generation of DFlash, seeded at Z Lab and

Computer History builds on the Chronicle research preview with reduced token usage and more privacy controls. A new timeline view gives you the ability to look back on your work and build skills from

We just pushed a big update to Sakana Chat. No login required and free to use: https://chat.sakana.ai/ It is now powered by Fugu and our updated Namazu Japanese LLM, and comes with full code execution

Really important point and nicely explained! Two further points worth noting: (1) if an AI system relies on a harness (which as Gary notes in other writing is increasingly improtant in top models), th

Very cool idea to have agents design complex systems by searching over the model structure itself. New research from Sakana AI introduces CEDAR, which uses LLM agents to write, simulate, and refine sy

Francesco @francedot reported that the first fully offline computer‑based LLM experience was achieved by running Muse Glimmer 30B (from @AIatMeta) locally on macOS, using Cua Driver to control native

In a recent Twitter exchange, Luca Ambrogioni stated that large‑language models (LLMs) likely possess fundamental limitations that may be obscured by progress in reasoning and agentic pipelines, thoug

I think of it like this: If you are building a simple agent with custom tools in your app -> langchain create agent If you are simply doing some LLM analysis or synthetic data creation -> langchain in

love this frame. calls to mind the role of the hippocampus in human navigation (via place cells and grid cells), and how navigation is, in a sense, what makes agents *agents* vs plain old LLM calls in

// Programmatic Memory Enables Long-Horizon Reasoning // Keep the entire interaction log and search it. It works great and beats bespoke memory harnesses on long-horizon tasks. New research introduces

this is important and true. and if you don’t understand it you don’t really understand what’s going on. @GaryMarcus Yes — model makers quietly went from benchmarking LLMs to benchmarking LLM+harness,

New work led by @FlemmingKondrup and @tomjiralerspong highlights an important vulnerability in chain-of-thought monitoring for agents, I highly recommend giving it a read. Link to the paper: https://a

OpenWiki 0.0.3 just launched with tons of bug fixes and improvements! Some notable changes: - 'openai-compatible' provider to support any LLM provider that supports the OpenAI API spec! Thank you @bra

On the pod: 'Constrained Adaptive Rejection Sampling' with @ucsd_cse professor @lorisdanto. Hear how symbolic AI experts have navigated the LLM era and why the future of AI code generation depends on

Fireworks AI is now live on EvoSkill v1.3.0! You can now use @FireworksAI_HQ directly with EvoSkill to run fast inference on open models as both the evolution harness backend and the LLM scorer. Along

team cooked, spicily We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloa

We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT,

This ComfyUI workflow lets you draw lines and annotations to guide motion, define where actions should happen, and influence movement throughout a scene. An integrated LLM translates those annotations