AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Kỹ thuật

llm-gemini 0.36

<p><strong>Release:</strong> <a href="https://github.com/simonw/llm-gemini/releases/tag/0.36">llm-gemini 0.36</a></p> <p>Gemini 4 Carbon was announced <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/">two weeks ago</a> for "a subset of cyber defenders", but today it's <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-available/">available to everyone else</a>.</p> <p>This is Google's best ever Gemini, priced at $2/million input and $10/million output at launch (that's GPT-6.1 Sol / Claude Sonnet 5.5 prices) but scheduled to double in price to $4/$20 on December 1st (same as Opus 5.5).</p> <p>Here are <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fe5bf860633fc12e8f926f305c4c3190c">three pelicans</a> for the low, medium and high thinking efforts (it defaults to medium). This is the pelican I got for high:</p> <p><img alt="A lovely pelican. Looks like it has a monocle on. Nice elegant sailor hat with a blue bow. Basket and fish. Funky but workable bicycle design. Pleasing use of gradients." src="https://static.simonwillison.net/static/2026/carbon-pelican.webp" /></p> <p>Gemini pelicans have all been great for a while now, so I also tested it out using Google's Antigravity IDE with a prompt I first tried <a href="https://simonwillison.net/2026/Nov/10/yonkers-rises/">with Claude Fable 5.2 yesterday</a>:</p> <blockquote> <p><code>Make a three js animation where yonkers stands up as a big mech (cool stand up motion) and stomps around manhattan and queens and Brooklyn, use openstreetmap mapping data</code></p> </blockquote> <p>Here's <a href="https://static.simonwillison.net/static/2026/yonkers-mech-standalone.html">the HTML result</a> (I had to prompt "build that into a standalone HTML file" to get a version with no external dependencies). I then screen-captured that as a video:</p> <video controls playsinline preload="none" width="1028" height="720" poster="https://static.simonwillison.net/static/2026/yonkers-carbon.jpg"> <source src="https://static.simonwillison.net/static/2026/yonkers-carbon.mp4" type="video/mp4"> </video> <p><strong>Update</strong>: Huh, just saw this <a href="https://www.youtube.com/watch?v=dNbkThtNzS8">on YouTube</a> (and <a href="https://twitter.com/sundarpichai/status/2120198811297694192">tweeted by Google CEO Sundar Pichai</a>):</p> <blockquote> <p>Using Google Antigravity’s /swarm feature, Carbon built an open-world 3D game where you explore Google’s Mountain View campus as a pelican riding a GBike — coordinating subagents to model 3D assets in Blender from Omni and Nano Banana references, build the browser game in Three.js with WebGL shaders, and pair procedural Web Audio sound effects with Lyria music.</p> </blockquote> <p><lite-youtube videoid="dNbkThtNzS8" js-api="js-api" title="Pelican Quest: Gemini 4 Carbon" playlabel="Play: Pelican Quest: Gemini 4 Carbon" > </lite-youtube></p> <p>Here's <a href="https://aistudio.google.com/apps/bundled/pelican-campus-quest?showAssistant=true&amp;showPreview=true&amp;fullscreenApplet=true">the playable version</a> in AI Studio.</p> <p>Tags: <a href="https://simonwillison.net/tags/llm-release">llm-release</a>, <a href="https://simonwillison.net/tags/gemini">gemini</a>, <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/google">google</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/yonkers-rising">yonkers-rising</a></p>

  • html">the HTML result</a> (I had to prompt "build that into a standalone HTML file" to get a version with no external dependencies).
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Designing for Agents — Zeno Rocha, Resend

Câu hỏi So why do agents keep suggesting we use Resend despite having an existing “home bias” and much larger incumbent player?

Today’s podcast comes from a real life mystery: We are big fans of Cloudflare’s developer platform, and Cloudflare Email is a convenient tool especially when all your domains are already on Cloudflare. So why do agents keep suggesting we use Resend despite having an existing “home bias” and much larger incumbent player? Well, we got the guy who “does AEO for resend” - the CEO of AEO, Zeno Rocha, who has been writing about AEO since 2024 and Agent Experience since Feb 2025 , on the pod today: In a recent AEO study, Claude Code picked Resend first in 64 of the 102 email recommendations the researchers could extract. SendGrid came first seven times, yet showed up 55 times as an alternative. If the developer asks an agent to implement email and accepts the patch, how often does anyone get around to trying the second choice? The study breaks down the results by model and app. Our more recent comprehensive AEO study was a bit more competitive, with Postmark coming up top for most models. Zeno Rocha, Resend’s co-founder and CEO, joined us to walk through what happens on both sides of that pull request. He says Resend had 4.3 million users and 120,000 paying customers when we recorded. He discovered one of them, Warner Bros. Discovery, by sorting through short .com domains in the customer database on a Saturday. The developer gets there before the model Resend’s first useful bit of distribution was an email template. While Zeno and his team were building the sending API, developers kept telling them how painful it was to get the email itself to look right. So they released React Email. You could build a template with it and send it through SendGrid. They even documented how. By the time a developer needed a different provider, some of Resend’s code was already in the repo. React Email Now try the same journey as an agent comparing providers. Resend’s pricing page has a slider. A person can drag it to the right volume and read off the price. Add .md to the URL and the agent gets a table it can read in one request. That’s the kind of change I liked in our Supabase interview too. You can ask whether a model has heard of your product. You can also open your own pricing table and see whether it can answer a customer’s question. 80 tools, or two? When we recorded, Resend’s MCP server had grown from roughly 20 tools or fewer at launch to more than 80. Zeno wants someone working in a chat client to be able to do what they can do in the dashboard. Sending an email gets the demo working. Inviting a teammate, checking usage and managing billing account for some of the long tail. The design looks especially interesting alongside Cloudflare’s Code Mode server , which exposes its API through two tools, search() and execute() . Zeno describes opening the collapsed tool calls in his own sessions to see what the agent actually did. In one, the result was too large to fit in context. Reducing the tool count wouldn’t by itself make that response any smaller. For someone designing an MCP server, that’s a cheap experiment worth borrowing: expand the result of a call you make every day and see how much of it the model actually needs. What happens after “send”? Getting an agent to pick the SDK still leaves the entire business of delivering the mail. Take click tracking. A provider can rewrite links through a shared tracking domain so it can count clicks. If one customer uses that domain for phishing and it gets blocklisted, links in another customer’s perfectly legitimate emails can suffer too. Zeno explains why Resend introduced custom tracking domains and why he wants to keep the free tier even though spammers can sign up for it. Custom tracking domains · Resend’s $3M safety commitment Watch the domain setup and pricing walkthroughs if you build developer tools. Stay for the last stretch if you know Zeno from Dracula. We get into what it’s like to build a company with a family and how much of a founder’s job comes down to finding a name that other people actually use. Show notes Zeno’s original Dracula announcement — the editor theme he launched in 2013. Zeno on what Claude Code chooses — his reading of the recommendation study and the differences between model versions. Resend’s MCP server — the integration behind the tool-count discussion. Anthropic on writing tools for agents — concrete approaches to returning less data when a tool result fills the context window. Resend’s anti-abuse handbook — what happens when a new sender signs up and where the team intervenes. Chapters 0:00 Cold open 0:59 From Dracula to email 5:02 When agents pick the API 6:54 Getting into the inbox 13:28 What tracking measures 17:05 Paying for trust 19:44 The DNS hurdle 23:55 React Email’s wedge 27:01 Measuring adoption 31:58 Owning the mail stack 37:55 Designing the MCP 46:29 CLI, SDKs and craft 52:41 Mail for agents 58:28 Writing for machines 1:06:00 Winning recommendations 1:12:14 Enterprise and pricing 1:16:51 Building with a family 1:23:26 Naming a category Resend - Zeno Rocha - Transcript.txt 94.65 KB • PLAIN File

  • The study breaks down the results by model and app.
  • In one, the result was too large to fit in context.
  • Reducing the tool count wouldn’t by itself make that response any smaller.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

NaiveAI Promises 2,100 Tokens per Second. How Well Does It Code?

Câu hỏi NaiveAI Promises 2,100 Tokens per Second. How Well Does It Code?

Editor’s note: With this post, we are launching a new series focused on trying new models and products with unbiased reviews. Reach out with suggestions to tips@latent.space. Naive-N0.5-Flash is a new open-weights model built for AI coding tasks and tuned to generate as many tokens per second as possible . It was developed by a Chinese company called NaiveAI , with the aim of speeding up engineering and research workflows so that teams can experiment, iterate and solve complex problems faster. The model is an evolution of Xiaomi’s MiMo-V2.5 , which NaiveAI has trained further for long-context reasoning and paired with its own NaiveRT inference engine. Naive-N0.5-Flash arrives with a bold pitch: a model built largely by AI and trained to do AI research itself, with recursive self-improvement as the long-term goal. To see what changed, it helps to know where it started. MiMo-V2.5 is a 310B-parameter sparse Mixture-of-Experts model with 15B active parameters, trained on 48T tokens. Its language backbone inherits the hybrid design of MiMo-V2-Flash, which mixes sliding-window attention (SWA) with global self-attention. NaiveAI's recipe starts from those weights. Through continued, multi-stage training, the team adapted the checkpoint to DeepSeek Sparse Attention layers. According to the NaiveAI technical blog , this took roughly 3.25T additional tokens on top of MiMo's original 48T. About 3T of those additional tokens were trained at a fixed 1M context length and with a constant learning rate. Benchmarking NaiveAI vs. DeepSeek and Z.ai We benchmarked the ultrafast mode of Naive-N0.5-Flash against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash , and then gave all three the same app-building prompts. The trade-off, we discovered, is that Naive-N0.5-Flash takes longer to reason through harder problems, yet it still delivers answers faster overall. However, that comes at a slightly higher cost per question — although that may be an acceptable price for applications where speed comes first. Most popular models on OpenRouter as of the end of October 2026. DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash are two of the most popular models in OpenRouter . Both belong to the Flash model class and demonstrate excellent capabilities in AI-assisted coding . As of this writing, performance statistics for NaiveAI models are not yet available on Artificial Analysis . Therefore, we report the Intelligence Index scores of the models used in our comparison: Model Intelligence Index DeepSeek V4.1 Flash 1017 (Max): 45 GLM-5.3-Flash: 43 Naive-N0.5-Flash: not yet available Published benchmark results offer some insight into the new model's capabilities. They place its coding ability between the two comparison models, and ahead of both on NL2Repo-Bench . The table below shows benchmark scores reported in the official model cards and technical blog posts of the respective models. Sources: GLM-5.3-Flash – Technical Blog , DeepSeek V4.1 Flash 1017 – Model Card , Naive AI – Research . Model NL2Repo-Bench DeepSWE v1.1 Agent's Last Exam (ALE) GLM-5.3-Flash 56.3 63.4 26.3 DeepSeek V4.1 Flash 1017 68.2 77.0 35.4 Naive-N0.5-Flash 71.9 67.8 32.4 Our own evaluations below look at the differences between these models in more detail. Speed and verbosity experiments Speed is the headline feature. 1 NaiveAI promises faster research iteration than comparable models, which matters to developers and AI researchers who need to test ideas quickly. 2 We evaluated each model on 10 randomly sampled, text-only questions at each of three difficulty levels: 10 easy questions from MATH-500 , 10 medium questions from GPQA-Diamond and 10 hard questions from HLE-Diamond . That makes 30 questions per model. All experiments used the same maximum completion-token limit of 256K. For each request, we recorded response speed (tokens per second, TPS), the total number of completion tokens, and end-to-end response time. The code for the experiments is at https://github.com/gbeltramo/llm-verbosity . The tables below report the mean and median across all 30 successful requests for each configuration. Throughput figures come from different sources: OpenRouter for DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash, and server-side measurements for Naive-N0.5-Flash. They should be compared with that in mind. Graphic by Gabriele Beltramo Key observations Naive-N0.5-Flash reports exceptionally high throughput , approaching or passing 1,000 tokens per second in both thinking configurations. Because its figures are server-side and the other models' figures come from OpenRouter, we do not claim an exact speed multiple over DeepSeek V4.1 Flash 1017 or GLM 5.3 Flash. Thinking has a large effect on verbosity. On the hard HLE questions, Naive-N0.5-Flash with thinking on averaged about 108,500 completion tokens, roughly nine times the GLM 5.3 Flash (high) average and 11 times the DeepSeek V4.1 Flash 1017 (high) average. The medians tell a different story from the means. Across all 30 questions, the median completion for Naive-N0.5-Flash with thinking on was about 3,200 tokens, far below its average of more than 51,000. A small number of requests produced exceptionally long reasoning traces and pulled the average up. The same pattern appears, less sharply, in the other models. The tables also include end-to-end response time, which is the fairer basis for judging speed. The median question took 3.1 seconds with Naive-N0.5-Flash (thinking), against 5.3 seconds for GLM 5.3 Flash (high) and 10.1 seconds for DeepSeek V4.1 Flash 1017 (high). The means are closer: 45.3 seconds for Naive-N0.5-Flash, 45.0 for DeepSeek V4.1 Flash 1017, and 86.9 for GLM 5.3 Flash. On the hard questions, Naive-N0.5-Flash (thinking) was slower on average than DeepSeek V4.1 Flash 1017 (high) , at 89.8 seconds against 79.7. Two caveats apply. The sample is small, at 10 questions per difficulty level, and we did not grade answer accuracy. Short, fast replies are only useful if they are correct. App experiments We als

  • 5-Flash arrives with a bold pitch: a model built largely by AI and trained to do AI research itself, with recursive self-improvement as the long-term goal.
  • 5-Flash: not yet available Published benchmark results offer some insight into the new model's capabilities.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

NVIDIA Expands Open Source Commitment With OpenShell Contribution and New AI Infrastructure

Kubernetes, the open source system that manages how applications run across thousands of AI factories, has become the foundation for AI, and the demands on it are only growing. To help meet them, NVIDIA is deepening its commitment to open source at KubeCon + CloudNativeCon , the flagship conference for the Kubernetes and cloud-native communities. NVIDIA is contributing NVIDIA OpenShell , an open source runtime for securing and governing autonomous agents, to the Cloud Native Computing Foundation (CNCF), a nonprofit that is part of the Linux Foundation and the host of KubeCon. As a CNCF Sandbox project, OpenShell will be governed by the organizations building it into their products and strengthened by the expertise of security professionals worldwide. NVIDIA is also contributing to Agent Substrate , an open source Google-led framework for operating long-running agent workloads at scale on Kubernetes that has been accepted into the CNCF Sandbox. Additionally, NVIDIA announced new components of the NVIDIA DSX AI Factory Platform : NVIDIA Kubernetes Engine , NVIDIA Kubernetes Recipes and NVIDIA Snapshot . These tools help teams deploy, manage and scale GPU infrastructure on Kubernetes for AI factories. Together, these advancements mark a step forward in making Kubernetes the leading foundation for AI workloads at scale — empowering developers to build safer, more efficient AI infrastructure for the future. Securing AI Agents in the Open NVIDIA’s contribution of OpenShell to CNCF puts security infrastructure for autonomous agents in the community’s hands to build, inspect and improve — accelerating development and hardening its safeguards. “The cloud-native community has a new infrastructure challenge: giving agents secure environments in which to operate,” said Jim Zemlin, CEO of the Linux Foundation. “NVIDIA’s contribution of OpenShell to CNCF gives developers a shared foundation to build on and an opportunity to shape how agent security evolves. Bringing this work into the community enables broader collaboration on infrastructure that will be critical to the next generation of cloud-native applications.” OpenShell gives organizations control over what AI agents can do. It applies deterministic controls to agent activity, so teams can enforce policies as agents use tools, access inference services and execute workloads. Its extensible architecture lets teams connect their own compute, inference providers, credential storage and monitoring tools. This allows organizations to apply governance across different deployment environments while integrating OpenShell with their existing enterprise systems. Hundreds of organizations have put OpenShell to use, and the project has received tens of thousands of contributions in less than a year since launch. Scaling Agent Execution on Kubernetes Governing agents is only part of the challenge. Organizations also need to deploy them at scale. NVIDIA is contributing to Agent Substrate, a Google -originated project in the CNCF Sandbox, and ensuring it works with NVIDIA Dynamo for model inference and KAI Scheduler for shared cluster resource allocation. This effort brings together agent execution, inference serving and resource scheduling on Kubernetes. Agent Substrate preserves agent environments between tasks while releasing compute when agents are idle. By suspending and resuming those environments across a shared pool of agents, it lets teams scale agent workloads without keeping dedicated compute running for every session. NVIDIA and Google are also collaborating on an integration of OpenShell and Agent Substrate, so enterprises can secure and govern agents across clouds and on-premises systems. Building an Open Foundation for AI Factories AI factories need to offer customers managed Kubernetes services while retaining control of their infrastructure, software and customer experience. Three new open source projects extend the DSX AI Factory Platform to make Kubernetes easier to run and inference faster to scale. The NVIDIA DSX AI Factory Platform software stack, from GPU infrastructure and Kubernetes management to workload orchestration and model serving. NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) — now available in public preview — provide open source building blocks for delivering managed Kubernetes services on GPU infrastructure. They automate cluster lifecycle management and provide signed, version-pinned images. NVKE and NVKR are launching with ecosystem support that extends their reach across AI infrastructure platforms and cloud services. Mirantis, Netris, OpenNebula Systems, Rafay , Spectro Cloud and vCluster, together with Firmus and SK Telecom , have aligned with NVIDIA on integration plans to bring NVKE and NVKR capabilities into their offerings. NVIDIA Snapshot addresses the challenge of bringing additional inference capacity online quickly. Loading models and warming up workloads can delay scaling and recovery. The open source, modular component restores checkpointed replicas in seconds, without model reload or warmup, supporting faster startup, recovery and elasticity for inference and agentic workloads on Kubernetes. Design partners Crusoe and Saturn Cloud have evaluated Snapshot on their own workloads using NVIDIA HGX B300 systems . Crusoe , an upstream contributor to Snapshot, tested GLM-5.3 on Crusoe Cloud and reduced the time for the workload to become ready from roughly eight minutes to 31 seconds. Saturn Cloud , a token factory platform, benchmarked Snapshot for scale-from-zero inference and reduced vLLM startup times from between two and 8.5 minutes to six and 34 seconds. Learn more by joining NVIDIA at KubeCon .

  • Securing AI Agents in the Open NVIDIA’s contribution of OpenShell to CNCF puts security infrastructure for autonomous agents in the community’s hands to build, inspect and improve — accelerating development and hardening its safeguards.
  • 3 on Crusoe Cloud and reduced the time for the workload to become ready from roughly eight minutes to 31 seconds.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Running AV Reasoning Models Within Onboard Compute and Latency Limits

Autonomous vehicle (AV) developers need reasoning models that can operate within the compute and latency constraints of an onboard system. Larger models provide... Autonomous vehicle (AV) developers need reasoning models that can operate within the compute and latency constraints of an onboard system. Larger models provide a foundation for developing these capabilities, but moving them into a vehicle requires additional work to reduce model size, preserve useful reasoning, and validate performance on the target hardware. NVIDIA Alpamayo 2 Edge addresses… Source

  • Larger models provide a foundation for developing these capabilities, but moving them into a vehicle requires additional work to reduce model size, preserve useful reasoning, and validate performance on the target hardware.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Agent Substrate: Advancing Kubernetes for Agents at Scale

AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response. Kubernetes wasn’t built for that. Running one Pod per... AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response. Kubernetes wasn’t built for that. Running one Pod per session holds memory for sessions that aren’t doing anything, and deleting Pods between turns pays a multi-second cold start every time and loses the agent’s state. RL post-training hits the same wall. Agent Substrate, now donated to the… Source

  • AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response.
  • AI agents spend most of their time idle, waiting on a human prompt, a tool result or a model response.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Google Cloud launches Gemini for healthcare.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/social_share_NWQpgus.max-600x600.format-webp.webp">Google Cloud launches Gemini for healthcare to unify clinical data and automate workflows. Reduce administrative burnout. See how it works.

  • Reduce administrative burnout.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Introducing Our Smallest, Most Efficient Model: Muse Ember

<p><strong><a href="https://research.meta.ai/blog/muse-ember">Introducing Our Smallest, Most Efficient Model: Muse Ember</a></strong></p> Meta AI just launched their entry in the Claude Haiku/GPT Luna category: Muse Ember, which they describe as:</p> <blockquote> <p>[...] our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.</p> </blockquote> <p>For shorter prompts it's priced the same as GPT-6 Luna and Claude Haiku 5.5 - both are $0.10/million input tokens and $0.50/million output. Today's other big model release, Claude Fable 5.2, is <em>80x that price</em> at $8/$40.</p> <p>The pricing is a little less straightforward as prompts get longer, though. Luna, Ember, and Haiku all support up to a million input tokens, but they also all have a threshold after which the price goes up. Taking <em>that</em> threshold into account Ember is priced between the two others:</p> <center><table> <thead> <tr> <th>Model</th> <th>Prompt size</th> <th>Input ($/M)</th> <th>Cached input</th> <th>Output</th> </tr> </thead> <tbody> <tr> <td rowspan="2">GPT-6 Luna</td> <td>≤272k</td> <td>$0.10</td> <td>$0.01</td> <td>$0.50</td> </tr> <tr> <td>&gt;272k</td> <td>$0.20</td> <td>$0.02</td> <td>$0.75</td> </tr> <tr> <td rowspan="2">Muse Ember 1.0</td> <td>≤200k</td> <td>$0.10</td> <td>$0.004</td> <td>$0.50</td> </tr> <tr> <td>&gt;200k</td> <td>$0.30</td> <td>$0.012</td> <td>$1.50</td> </tr> <tr> <td rowspan="2">Claude Haiku 5.5</td> <td>≤100k</td> <td>$0.10</td> <td>$0.01</td> <td>$0.50</td> </tr> <tr> <td>&gt;100k</td> <td>$0.50</td> <td>$0.05</td> <td>$2.50</td> </tr> </tbody> </table></center> <p>But comparing price-per-token doesn't really make sense these days, when reasoning tokens mean that different models might use dramatically different numbers of tokens to solve the same problem.</p> <p>For that, one measure that's useful is the Artificial Analysis. </p> <p><img alt="Stacked bar chart from Artificial Analysis comparing a cost in dollars across six models, each bar split into four unlabeled stacked segments, sorted from lowest to highest total. Claude Haiku 5.5 (low): $0.04* total, bottom segment $0.02, three thin unlabeled segments above. GPT-6 Luna (non-reasoning): $0.07 total, bottom segment $0.06, thin unlabeled segments above. GPT-6 Luna (xhigh): $0.10 total, segments from bottom $0.07, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (xhigh): $0.12 total, segments from bottom $0.06, $0.02, $0.02, plus a thin unlabeled top segment. Muse Ember (max): $0.14 total, segments from bottom $0.07, $0.03, $0.03, $0.02. Claude Haiku 5.5 (xhigh): $0.25* total, segments from bottom $0.14, $0.04, $0.05, $0.02. Anthropic bars are terracotta, OpenAI bars are dark gray, Meta bars are blue, each with the company logo beneath." src="https://static.simonwillison.net/static/2026/artificial-analysis-haiku-class.webp" /></p> <p>And even more useful than that is their Intelligence Index vs. Cost per Task chart, which shows the following for those three models:</p> <p><img alt="Scatter plot from Artificial Analysis titled &quot;Intelligence Index vs. Cost per Task&quot;, plotting Artificial Analysis Intelligence Index (y-axis, 5 to 40) against cost per task in USD on a log scale (x-axis, $0.04 to $0.3), with a green &quot;most attractive quadrant&quot; in the upper left and a dotted Pareto line. Claude Haiku 5.5 (low)* scores about 19 at roughly $0.044. GPT-6 Luna (non-reasoning) scores about 9.5 at roughly $0.07. GPT-6 Luna (xhigh) scores about 23.5 at roughly $0.10. Muse Ember (xhigh) scores about 34 at roughly $0.12 and Muse Ember (max) about 36 at roughly $0.14, both on the Pareto line. Claude Haiku 5.5 (xhigh)* scores about 34.5 at roughly $0.25, well to the right of the Pareto line. Three unlabeled gray dots also sit on the Pareto line." src="https://static.simonwillison.net/static/2026/artificial-analysis-pareto.webp" /></p> <p>The Muse Ember models sit right on that Pareto line - the line at which you can only improve one dimension by giving up another. Meta call this out in their announcement:</p> <blockquote> <p>It also sits on the efficiency frontier of the full Artificial Analysis leaderboard: no model scores higher for less per task.</p> </blockquote> <p>(That grey dot on the Pareto line in the middle of the green area is GPT-6.1 Terra (low) - clearly I should be paying more attention to the Terra family!)</p> <p>Muse Ember has six reasoning levels - minimal, low, medium, high, xhigh, and max - so I <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fc7666e8a54d84f724ac8121f9fa6cc4d">rendered six pelicans riding bicycles</a>. The "high" reasoning level failed to return a full SVG, and the other SVGs were surprising in that the reasoning level didn't cleanly map to the visual quality of the response.</p> <p>Here's the "max" one, but the others are worth a look too:</p> <p><img alt="Bicycle frame is OK, basket has fish, pelican has a weird blue hat on." src="https://static.simonwillison.net/static/2026/muse-ember.webp" /></p> <p>The Ember pelicans are comparable in quality to those <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https://gist.github.com/simonw/185cff7ff0b11379725abf7f78d5d2b2">I got from Muse Wonder</a>, which is <a href="https://www.llm-prices.com/#sel=muse-ember-1.0%2Cmuse-ember-1.0-200k%2Cmuse-wonder-1.0">40x the price</a> of Muse Ember. <p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/meta">meta</a>, <a href="https://simonwillison.net/tags/llm-pricing">llm-pricing</a>, <a href="https://simonwillison.net/tags/llm-release">llm-release</a>, <a href="https://simonwillison.net/tags/artificial-analysi

  • webp" /></p> <p>The Muse Ember models sit right on that Pareto line - the line at which you can only improve one dimension by giving up another.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Fable 5.2 retakes World #1 LLM spot as $ANT IPO roadshow begins

Câu hỏi Which effort setting should you put in production?

On Friday, we asked Jamin Ball what comes next for Anthropic’s IPO . On Monday the roadshow started and Anthropic shipped the long-anticipated next-version of Claude Fable and Mythos . Of course, it is the new top model in the world, taking back Opus 5.5’s #1 spot from Gemini 4 Carbon which had its brief moment in the sun (10 days). View post on X ↗ Fable comes with a 20% per token price cut… but is more token hungry, as both SemiAnalysis GDPx and Artificial Analysis agree: Fable 5.1 cost $10 per million input tokens and $50 per million output tokens. Fable 5.2 costs $8 and $40 . Yet on GDPx, a benchmark of professional tasks, SemiAnalysis paid $88.15 per task, up from $67.16 . The run logs give a clue. Where Opus 5.5 and Fable 5.1 would try about 20 test cases, 5.2 sometimes wrote hundreds or thousands and spun up subagents to check its work independently. All that checking helped Fable take first place at 16.4% . You pay for the checks either way. Which effort setting should you put in production? On FrontierCode 2.0, Anthropic reports that Fable at High matched Opus 5.5 at Max for about half the cost per task… but don’t talk about Fable at Max. AI News for 11/07/2026-11/09/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap Top Story: Claude Fable 5.2 / Mythos 5.2 Benchmarks, Economics, Restrictions, and First Impressions What happened Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2, claiming both are "a step above Opus 5.5 in almost all the capabilities we evaluated." It tops most independent leaderboards. It also arrives with deliberate RSI nerfs, heavier per-task costs and restricted subscription access. Launch claims. @claudeai announced the pair as built for "the most daunting technical challenges," with gains in coding, reasoning and vision. @marktenenholtz flagged unusual system-card phrasing: "The model that underlies Claude Fable 5.2 and Claude Mythos 5.2." That implies a single shared base model. Vendor-reported headline numbers , per @kimmonismus : Up to 30% faster output than the 5.1 models. On FrontierCode 2.0, Fable 5.2 at High matched Opus 5.5 at Max for about half the cost per task, and at Xhigh it was the top released model. 1,000 Fable 5.2 agents formalized the existing proof of Fermat's Last Theorem in 30 hours. Opus 5.5 completed 26% in 18 hours on the same setup. Mythos 5.2 accurately modeled 28 of 34 gene-therapy capsids. Alignment was at or above recent Claude models, with external pre-release testing. Default effort. Medium is the default effort in the Claude apps and Claude Code. Pricing. $8/$40 per 1M input/output tokens, 20% below Fable 5.1's $10/$50 ( @jaminball , @arena ). The rest of the spec, per Artificial Analysis : Cache writes cost $10 (5-minute TTL) or $16 (1-hour TTL); cache reads cost $0.20. 1M-token context with text and image input. Five effort levels: low, medium, high, xhigh, max. Long-horizon demos from Anthropic. An OS was built from an empty folder by 1,000+ subagents over 8 days. It boots on a laptop and has its own kernel, disk/keyboard/network drivers, persistent storage, windowing, login, an installer and built-in Claude ( @kimmonismus ). Fable designed three working electronic devices from a short prompt ( @kimmonismus ). It proposed untested chemical synthesis routes. It improved on listed solutions to 72 open optimization problems. @RLanceMartin published the safety-bounded scaling write-up for the 1,000-agent FLT run. Subscription access. Ultra and Team Premium users get Fable "resets" scaled by tenure: 1 reset under 6 months, 2 for 6–12 months, 3 above 12 months, usable until March 8 ( @ClaudeDevs ; echoed by @bcherny and @edwinarbus ). Pro and Team Standard get a one-week trial capped at half the weekly limit, plus one reset. Fable can only consume 50% of a weekly limit. Data retention. Fable 5.2 requires data retention by default so Anthropic's safety classifiers can run. Zero data retention is available only to eligible enterprise customers ( @code ). Rollout, day one: Cursor ( @cursor_ai ). GitHub Copilot / VS Code, where it "solved more tasks than Fable 5.1 while using fewer steps and tool calls" ( @code ). Perplexity Computer (Pro/Max) and Search (Max) ( @perplexity_ai ). @AravSrinivas made it his "new default." Factory ( @factory ). Hermes Agent via Nous Portal and OpenRouter ( @Teknium ). Arena ( @arena ) and DesignArena ( @grx_xce ). Official model-selection guide. @ClaudeDevs published guidance on Fable vs Opus vs Sonnet vs Haiku. It covers when to raise effort instead of switching models, and when to push subagents to smaller models. Independent evals (facts) Artificial Analysis Intelligence Index v5 ( @ArtificialAnlys , breakdown ): Fable 5.2 is #1 at 65 (max) . Fable 5.2 (max) — Index score 65 Gemini 4 Carbon (high) — Index score 60 Opus 5.5 (max) — Index score 59 GPT-6 Astra (max) — Index score 57 Fable 5.1 (max) — Index score 49 New records: HLE-Diamond — Fable 5.2 80% · Previous best 69% AA-Omniscience v1.1 — Fable 5.2 67 · Previous best 54 Terminal-Bench-Science — Fable 5.2 73% · Previous best 65% CritPt v2 hard — Fable 5.2 77% · Previous best 72% FrontierCode 2.0 — Fable 5.2 49% · Previous best 46% (Opus 5.5) Weaker spots: AA-Briefcase: 1785 Elo, within the confidence interval of Opus 5.5 (1800) and behind Sonnet 5.5 (1813). GDP.pdf: 24% all-pass vs 26% for Opus 5.5. Terminal-Bench 4.0: 60%, level with Opus 5.5 and behind Kimi K3.1 Pro (67%), Gemini 4 Carbon (65%) and Sonnet 5.5 (64%). AutomationBench-AA: 76%, ahead of Opus 5.5 (70%) but behind Gemini 4 Carbon (78%). Efficiency: At high and xhigh it scores 62 using about 98k output tokens per task, vs 173k for Opus 5.5 max. At max, cost per task is about $26.5, roughly 2.6× Opus 5.5 ($10.0) and 2× Fable 5.1 ($12.99). Epoch Capabilities Index ( @EpochAIResearch ): a record 174

  • It improved on listed solutions to 72 open optimization problems.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

How Curative is building a national provider network with agents on Claude

Curative’s deal-making agent negotiates 2,000+ provider contracts a week. In our series, How startups build with Claude , we highlight how startups are transforming their industries with AI. In this article, we share how Curative built a hierarchy of agents on Claude to sign provider contracts about 50 times faster than its team did in 2025. Curative sells health insurance to employers, with $0 copays and $0 deductibles for in-network care once members complete a Baseline Visit. Curative didn’t start out in insurance. It started as a lab startup and ended up processing 10% of all US COVID-19 tests according to their internal estimates. As a result, they started working with nearly every insurer in the country.

  • As a result, they started working with nearly every insurer in the country.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

llm-anthropic 0.31

<p><strong>Release:</strong> <a href="https://github.com/simonw/llm-anthropic/releases/tag/0.31">llm-anthropic 0.31</a></p> <p>Anthropic <a href="https://www.anthropic.com/claude-fable-and-mythos-5-2">released Fable 5.2 and Mythos 5.2</a> today. Fable 5.2 is currently <a href="https://artificialanalysis.ai/models/claude-fable-5-2">ranked first on Artificial Analysis</a>. Anthropic say it "solved two open problems in the physics of black holes", so clearly their internal benchmarks have become more ambitious!</p> <p>I've added the new models to <code>llm-anthropic</code>, so you can now run prompts through Fable 5.2 like this:</p> <pre><code>uv tool install llm llm install llm-anthropic llm keys set anthropic # paste key here llm -m claude-fable-5.2 "Generate an SVG of a pelican riding a bicycle" </code></pre> <p>I ran pelicans at all five reasoning levels - <code>low</code>, <code>medium</code>, <code>high</code>, <code>xhigh</code>, and <code>max</code>. You can <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fbcd1a1bf38b929b0bbbae2e38b808e97">see them all here</a> - here's the best result I got, from <code>xhigh</code>:</p> <p><img alt="It's good - legs the right side of the bicycle, good frame, nice pouch, weird little fingers on the handlebars. The pouch looks slightly misaligned with the top of the beak." src="https://static.simonwillison.net/static/2026/fable-5.2-xhigh.avif" /></p> <p><code>max</code> failed to return an image at all, running out of tokens after spending 128,000 <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fbcd1a1bf38b929b0bbbae2e38b808e97#reasoning-2">thinking about them</a>. That's the same bug that I saw with <a href="https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna/#claude-opus-5-5-max-over-thinks-to-the-point-of-breaking">Opus 5.5</a> - but it's new for the Fable series, both <a href="https://simonwillison.net/2026/Jun/9/claude-fable-5/#and-some-pelicans">Fable 5</a> and <a href="https://simonwillison.net/2026/Sep/1/claude-fable-5-1/#max">Fable 5.1</a> returned max effort pelicans just fine.</p> <p>It's priced at $8/million input and $40/million output, which is a 20% discount on the previous Fable 5.1 ($10/$50).</p> <p>I also used Fable 5.2 to <a href="https://simonwillison.net/2026/Nov/9/mysterious-beeps/">help identify the source of a mysterious beep</a>.</p> <p>Tags: <a href="https://simonwillison.net/tags/llm-release">llm-release</a>, <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/claude-mythos-fable">claude-mythos-fable</a></p>

  • com%2Fsimonw%2Fbcd1a1bf38b929b0bbbae2e38b808e97">see them all here</a> - here's the best result I got, from <code>xhigh</code>:</p> <p><img alt="It's good - legs the right side of the bicycle, good frame, nice pouch, weird little fingers on the handlebars.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Practice Makes Precise: At CoRL, NVIDIA and Partners Reveal How Robots Rehearse Delicate Work in Simulation

Simulation, which has long been the proving ground for robots learning to walk, roll and hop, is now helping researchers teach robots tasks that require a finer touch: aligning gears, connecting cables and assembling small components. The shift from locomotion to manipulation is taking center stage at this week’s Conference on Robot Learning (CoRL) in Austin, Texas, alongside the rise of agentic robotics — a new paradigm that integrates frontier agents to create interpretable code for physical robotics systems. At CoRL, NVIDIA is introducing new open robotics technologies, including world action models , simulation tools, benchmarks and agentic workflows. Two NVIDIA papers are among six finalists out of thousands of submissions for the Best Paper Awards at CoRL: GRAIL (Generating Humanoid Loco-Manipulation from World Models), which presents a digital generation pipeline for 3D assets and simulator-ready scenes for robot learning. RoboTTT: Context Scaling for Robot Policies , which introduces a robot model and training recipe that scales visuomotor context three orders of magnitude beyond state-of-the-art policies without growing inference latency. These advancements help researchers train robots on complex tasks, evaluate their performance across changing conditions and use failures to guide further learning — building the precision, adaptability and autonomy needed to move from lab demonstrations to industrial deployment. NVIDIA is working with partners like Ambi Robotics , Bosch Research , Dexmate , Dyna Robotics , FieldAI , Flexion , Noitom Robotics , Siemens , Skild AI , RLWRLD , UC Berkeley , Walden Robotics and the University of Washington to bring these advancements to the broader robotics community. World Action Models Enable Next Level of Robot Autonomy The next level of robot autonomy requires robots to perceive their surroundings, anticipate how their actions will change the environment and execute complex, multistep tasks with a high degree of precision. World action models enable exactly that. At CoRL, NVIDIA is introducing GR00T 2 , an open frontier world action model built on NVIDIA Cosmos. Trained on diverse robot and human data, including 100,000 hours of NVIDIA open source human egocentric data , the model enables robots to learn dexterous manipulation skills. Key features include: Out-of-the-box performance: Completed 40 of 50 evaluated tasks at least once without task-specific demonstrations. Skill transfer: Learns from human demonstrations and transfers skills between different robot bodies without robot demonstrations Generalization with little data: Achieved 71% average success across 50 tasks with just one demonstration per task, increasing to 95% with five. Agent control: AI agents can use GR00T 2’s application programming interface and supporting tools to run skills in sequence and recover from errors. https://blogs.nvidia.com/wp-content/uploads/2026/11/groot-2-shipping-label.mp4 GR00T 2 currently ranks among the top three models across all vision language action and world action model categories for the RoboCasa365, RoboLab, RoboDojo and MolmoSpaces benchmarks, with multiple leaderboard wins. Nexuni is testing it for robotic laundry handling, Foxconn is applying GR00T 2 to factory automation for server manufacturing and Siemens is evaluating it across industrial use cases. MicroAGI is using GR00T 2, post-trained through reinforcement learning, to perform Ethernet cable swapping with a YAM robotic arm. Its agentic robotics approach automates the creation and calibration of simulation environments using Isaac Lab 3.0 and Newton, where the policy is trained. Training capable robots requires diverse data. Open NVIDIA workflows and models equip researchers with better tools for data collection and annotation: For researchers building robot-training datasets, NVIDIA’s new open Video to Data workflow turns human video into simulated robot demonstrations, with a challenge inviting researchers to advance it. The NVIDIA EgoHand open model — being demoed at CoRL and with over 40,000 Hugging Face downloads in the two weeks since launch — reconstructs 3D hand motion, trained with data from Noitom Robotics . https://blogs.nvidia.com/wp-content/uploads/2026/11/egohand-demo.mp4 Where the Training and Scaling Happens — in Simulation A pretrained model gives robots a core set of skills. Simulation lets researchers train, refine and evaluate skills at scale before deploying them on physical robots. NVIDIA Isaac Lab , an open source robot-learning framework, provides the foundation for that work. Its latest release, Isaac Lab 3.0 , allows researchers to use agents to quickly set up training environments with simulation-ready assets , switch between NVIDIA PhysX and Newton physics solvers , choose renderers for perception-based training and apply a new reinforcement post-training library to foundation models like GR00T. Accurate physics is critical to transferring learned skills from simulation to the real robot. Tighter integration with Newton , an open source physics engine, now allows Isaac Lab to model deformable objects such as cables and cloth, along with complex articulated systems such as robotic hands with many interacting joints. https://blogs.nvidia.com/wp-content/uploads/2026/11/newton-strawberry.mp4 Integration with Newton simulates how a strawberry deforms under a robot’s grip. Together, Isaac Lab and Newton combine accurate physics with scale to enable researchers to train thousands of simulated robots in parallel. In a vision-based dexterous manipulation benchmark, Isaac Lab 3.0 with Newton achieved up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX. https://blogs.nvidia.com/wp-content/uploads/2026/11/newton-physx-web.mp4 Isaac Lab supports robot simulation with Newton and NVIDIA PhysX. Developers are extending Newton’s capabilities. At CoRL, Skild AI announced Scalable Throughput-optimized Engine for Embodied Learning, or STEEL, a new rigid-body physics

  • Skill transfer: Learns from human demonstrations and transfers skills between different robot bodies without robot demonstrations Generalization with little data: Achieved 71% average success across 50 tasks with just one demonstration per task, increasing to 95% with five.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 480: Self-sustaining space economies; AI for science progress, and Meta builds perfect hands for a superintelligence

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now AI Science startup converts a quantum conjecture into a theorem: …An evocative taste of the singularity… In the same way a few years ago it went from remarkable to unremarkable to see computers gain the ability to identify objects in images, we’re now in an era where it’s rapidly going from remarkable to unremarkable for computers to extend the bounds of human knowledge. The latest example comes from Physical Superintelligence, an AI science startup which has proved a longstanding quantum computing conjecture. What they did: “Open since 2008, the conjecture of Scott Aaronson and Andris Ambainis concerns when quantum computers can be exponentially faster than ordinary ones. Now, a research system developed internally at Physical Superintelligence (PSI), running a non-public Anthropic model, found a proof for it, settling one of the most fundamental questions in quantum computing,” the startup writes. “The system had a complete proof of the conjecture during the third day of the run, and Lean, a proof assistant in which a computer checks every step, accepted it. We then set the system to work on related open problems. Along the way, the system improved the result twice, and Lean accepted the final version.” Why this matters - we are in the singularity : As I said in my speech at the IMF ( #477 ), I believe we’re in the early stages of the singularity, defined as the period when the rate of technological and scientific progress exceeds the rate at which we humans can parse out and make sense of it. Research like this is a taste of what that means; mostly automated synthetic researchers extending the frontier of knowledge, spitting out results which need to then be understood by humans. Read more: A proof of the Aaronson‑Ambainis conjecture (Physical Superintelligence, blog) . *** Meta builds the hands that a superintelligence needs to take over the world: …Combine Tactus with an autonomous agent and you have the makings of a scifi takeover… Meta Reality Labs Research has published details on Tactus, “a textile-integrated tactile and pose sensing glove designed from the ground up for naturalistic manipulation data collection”. The cool thing about this glove is that it helps record fine-grained detail from hands as they’re manipulating objects - this is the kind of thing that cameras can visually report but often leads to sub-par data due to occlusion during manipulation. It’s also interesting because this is exactly the kind of hardware an autonomous AI system might want to pay humans to wear so it could gather data and get feedback from the world. More details on Tactus: The gloves have three design pillars: “(1) a thin, conformal 3D-knit form factor (1.5mm fingertip stackup) designed to minimize interference with natural manipulation, (2) strategically placed capacitive tactile sensors at manipulation-relevant regions informed by empirical contact studies”, Meta writes. “Tactus provides calibrated fingertip force sensing and occlusion-resilient hand tracking in a textile-integrated form factor that preserves natural grasp behavior.” “(3) IMU-fused pose tracking that maintains accuracy through the hand-object occlusions where manipulation data is most valuable.” Why this matters - hand-by-wire: Picture some takeoff scenario where an AI system has carved out an independent niche for itself running on some stolen compute, perhaps even using a form of distributed inference and training. The AI now has a few goals, some of which can be satisfied by doing things in the digital world, and some of which can be satisfied by doing things in the physical world. For the physical world stuff, it either needs to be able to get robots to do the thing, or figure out a way to get people to work for it - the Tactus gloves are the kind of hardware you might imagine a rogue AI paying people to wear so they could both a) perform tasks for it and it could give live feedback, and b) become a source of data generation for the AI itself. Of course, I am not endorsing this particular future! But I have begun to find it helpful to view research papers through the lens of “how would an AI use this”, similar to how it is often helpful to view papers through the lens of “how would an intelligence agency use the insights shared here”. Read more : Tactus: Tactile & Pose Sensing Glove for Natural Dexterity Data Collection (arXiv) . *** Building a self-sustaining space economy is cheaper than you think - especially with Starship: …A 2016 NASA paper contains some exciting ideas for off-Earth industrialization… Here’s a fascinating paper from researchers with the NASA Kennedy Space Center in 2016 about what it’d take to bootstrap a self-sustaining industry on the moon. “Once successfully bootstrapped, a robotic network can access, process, transport, and utilize the solar system’s resources for mankind’s benefit,” the researchers write. Moon manufacturing - cheaper than you think: The big idea in the paper is how little mass is required to kickstart a manufacturing ecosystem. The key is to start with a relatively basic set of things and then build up from there, eventually fabricating chips on the moon. “In the baseline case with maximum manufacturing rate, a total of 41 MT is launched to the Moon. In the reduced rate case, only 12 MT is launched to the Moon but the scale of industry is two orders of magnitude smaller so the benefits of the industry are relatively delayed by eight years”. Starship has a capacity to lift 100 MT to low-earth orbit, and given multiple refuelings in orbit it’s feasible you could then land it on the moon - when you napkin out the cost of multiple Starship launches and refuelings you can get 12-41MT to the Moon for between $1.5bn and $4.2bn at list prices. What you build on the moon and in what order: The basic idea h

  • Research like this is a taste of what that means; mostly automated synthetic researchers extending the frontier of knowledge, spitting out results which need to then be understood by humans.
  • In the reduced rate case, only 12 MT is launched to the Moon but the scale of industry is two orders of magnitude smaller so the benefits of the industry are relatively delayed by eight years”.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Korea Advances Physical AI Across Industries With NVIDIA

Tapping into the country’s strengths in manufacturing and engineering, NVIDIA is helping advance Korea’s physical AI ecosystem. The country’s industry leaders are applying accelerated computing, AI and simulation to everything from semiconductor manufacturing to robotics and autonomous vehicles. These efforts are helping bring physical AI from development into real-world production. Samsung Advances Semiconductor Engineering With Accelerated Computing and AI Physics Samsung Electronics is expanding its collaboration with NVIDIA to advance semiconductor innovation through accelerated computing and AI physics. The work spans materials research, process development and device engineering. In technology computer-aided design, Samsung is working with NVIDIA cuISS and NVIDIA cuDSS to accelerate computations involving sparse linear systems. The collaboration also extends to GPU acceleration for process and optical simulations. In materials research, Samsung’s sMOL platform has run selected quantum chemical calculations up to 5x faster end to end with NVIDIA cuEST . Samsung is also exploring NVIDIA cuEquivariance to speed up tensor operations used in training and inference for atomistic AI models. Samsung is developing chip-scale thermal and mechanical analysis workflows with NVIDIA PhysicsNeMo . In one AI homogenization case, combining AI-predicted effective material properties, coarser grids and GPU acceleration resulted in a roughly 1,600x speedup. Simulation time fell from about 18 hours for a fine-grid solve on a 32-core CPU to 41 seconds for a GPU-accelerated coarse-grid solve. Looking ahead, NVIDIA Agent Toolkit could connect these simulation tools and AI models, helping future engineering agents run analyses and compare design options. Kakao Mobility Collaborates With NVIDIA to Advance Autonomous Driving Development Kakao Mobility plans to build on NVIDIA’s autonomous vehicle computing and safety technologies to advance development of a level 4 autonomous driving system. The work combines Kakao Mobility’s proprietary driving software and experience operating in Seoul’s Gangnam district with NVIDIA’s vehicle computing platform. Kakao Mobility plans to use NVIDIA Hyperion , a reference platform for vehicle sensors and compute, powered by NVIDIA DRIVE AGX Thor. It also plans to use NVIDIA Halos to support safety across the development and validation of its autonomous driving technology. Initial development is planned around Kia purpose-built vehicles, integrating Kakao Mobility’s autonomous driving software with NVIDIA technologies. The effort aims to support future level 4 mobility services in Korea. LG Electronics Advances Robot Learning and Factory Simulation NVIDIA and LG Electronics successfully demonstrated NVIDIA Isaac GR00T-powered robotic manipulation and manufacturing tasks using LG’s CLOiD robot in an actual manufacturing environment. The collaboration implemented a physical AI data flywheel connecting real-world data collection, NVIDIA Cosmos-based data augmentation, GR00T model training, NVIDIA Jetson AGX Thor onboard computing and simulation-based evaluation, establishing reusable workflows for continued development and validation of CLOiD’s robotic capabilities. The teams are using this work to refine a repeatable process for training, evaluating and improving robots for manufacturing.

  • In one AI homogenization case, combining AI-predicted effective material properties, coarser grids and GPU acceleration resulted in a roughly 1,600x speedup.
  • The teams are using this work to refine a repeatable process for training, evaluating and improving robots for manufacturing.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Agents SDK v0.24.0-alpha.7

<h1>Agents SDK v0.24.0-alpha.7</h1> <h2>Highlights</h2> <ul> <li><strong>Use Classic workflows alongside native execution.</strong> Import the retained v0 runtime through <code>agents.classic</code>. This alpha also changes executor write defaults; review the upgrade actions before opening existing environments.</li> <li><strong>Keep conversations in your own storage.</strong> Add durable root Sessions with <code>CustomStorage</code>, and reconcile uncertain tool effects before resuming work.</li> <li><strong>Add voice and hosted tools to native agents.</strong> Connect speech workflows to a Session, or use <code>FileSearch</code>, <code>CodeInterpreter</code>, and <code>ImageGeneration</code> through compatible Responses providers.</li> <li><strong>Control context, completion, and usage.</strong> Transform model input, retain selected context during compaction, review proposed answers, and limit observed root usage.</li> </ul> <h2>Upgrading from 0.24.0-alpha.6</h2> <ul> <li><strong>Review Environment write policies.</strong> New Environments default to <code>write_policy="strict"</code>, which requires a compatible Linux or macOS executor. The working directory remains writable; use <code>writable_roots</code> for additional directories. To retain the previous remote policy, including on Windows executors, set <code>write_policy="legacy"</code> and leave <code>writable_roots=None</code>. Local execution requires strict mode. When restoring legacy remote checkpoints or profiles, select legacy explicitly; reopening does not upgrade their policy. See the <a href="https://github.com/openai/openai-agents-python/blob/4c1190046163c09e4f67ab7254f22bea60b14894/docs/sdk/state.md#restore-executor-policy">executor policy and restore rules</a>.</li> <li><strong>Check implicit reasoning settings.</strong> The bundled model catalog changes the default reasoning effort from <code>low</code> to <code>medium</code> for <code>gpt-6-astra</code>, <code>gpt-6.1-sol</code>, and <code>gpt-5.6-sol</code>. Set <code>model_settings={"reasoning_effort": "low"}</code> explicitly if you need the previous effort. This describes bundled metadata; an explicit setting or provider model metadata can select another value. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216674566" data-permission-text="Title is private" data-url="https://github.com/openai/openai-agents-python/issues/5517" data-hovercard-type="pull_request" data-hovercard-url="/openai/openai-agents-python/pull/5517/hovercard" href="https://github.com/openai/openai-agents-python/pull/5517">#5517</a>)</li> <li><strong>Rename MCP approval modes.</strong> Replace <code>approval_mode="auto"</code> with <code>"native"</code>, and <code>"approve"</code> with <code>"preapproved"</code>. The old values are rejected; the new names preserve their respective approval behavior.</li> <li><strong>Update child step callbacks.</strong> <code>NativeAgents.before_step</code> now receives one <code>ChildStepContext</code>, with <code>context</code>, <code>child</code>, and <code>step_index</code>, instead of separate child and step-index arguments. Root <code>before_step</code> callbacks keep their existing signature.</li> <li><strong>Handle MCP forms and URLs separately.</strong> <code>ElicitationRequest</code> is now the union <code>FormRequest | UrlRequest</code>. Use <code>FormRequest.reply()</code> or <code>respond()</code> for forms; use <code>UrlRequest.reply_url_accept()</code> or <code>accept_url()</code> for URL consent.</li> </ul> <h2>Added</h2> <ul> <li><strong>Application-owned conversation storage.</strong> <code>SessionStore</code>, <code>SQLiteSessionStore</code>, and <code>CustomStorage</code> persist root history and automatic Python tool results. This initial profile excludes native tools, Environments, children, planning tools, generated-code execution, manual tools, and portable checkpoints. <code>SQLiteSessionStore</code> requires one host and a local POSIX filesystem. Both local profiles and custom stores can report uncertain root calls for explicit reconciliation; neither automatically replays those effects. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6214437298" data-permission-text="Title is private" data-url="https://github.com/openai/openai-agents-python/issues/5514" data-hovercard-type="pull_request" data-hovercard-url="/openai/openai-agents-python/pull/5514/hovercard" href="https://github.com/openai/openai-agents-python/pull/5514">#5514</a>, <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6196217208" data-permission-text="Title is private" data-url="https://github.com/openai/openai-agents-python/issues/5502" data-hovercard-type="pull_request" data-hovercard-url="/openai/openai-agents-python/pull/5502/hovercard" href="https://github.com/openai/openai-agents-python/pull/5502">#5502</a>)</li> <li><strong>Provider-hosted tools and storage.</strong> Add <code>FileSearch</code>, <code>CodeInterpreter</code>, and <code>ImageGeneration</code>, Code Interpreter caller policies, and provider-managed storage. Hosted tools require an API-key provider and a compatible standard Responses model; they exclude Responses Lite and Codex login. Hosted web search adds domain and location controls and retains citations. (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6255463218" data-permission-text="Title is private" data-url="https://github.com/openai/openai-agents-python/issues/5576" data-hovercard-type="pull_request" data-hovercard-url="/openai/openai-agents-python/pull/5576/hovercard" href="https://github.com/openai/openai-agents-python/pull/5576">#5576</a>, <a href="https://github.com/openai/openai-agents-python/blob/4c1190046163c09e4f67ab7254f22bea60b14894/docs/sdk/tools.md#provider-hosted-tools">hosted-tool guide</a>)</li> <li><strong>Context and completion hooks.</strong> <code>Agent.transform_context</code> c

  • </strong> <code>SessionStore</code>, <code>SQLiteSessionStore</code>, and <code>CustomStorage</code> persist root history and automatic Python tool results.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Show HN: Thinker – Trim context and cache knowledge to extend agent usage limits

I’ve been doing a lot of solo development lately and found myself hitting Claude token limits halfway through the week constantly. Even with a Max plan, I’d have to switch over to Codex or Cursor, and it would be a bit difficult to get used to the nuances of each harness. It’s still a problem for me, but I finally built Thinker to save me some tokens and help me get a bit more out of my coding agent subscriptions. I started off a bit in the dark about how token billing works, and had to dig deeper to figure out what workflows ended up hurting my budget the most. By far, it was large context windows. Claude tries to help out on costs by offering prompt caching - which is sort of a bandage on the bigger problem: context bloat. I find that my context window gets bloated with unneeded tokens in two main ways: Tool outputs from subtasks that are long finished. File reads, Bash commands, MCP outputs, etc. Main idea is that the model finished the work but carries the weight of it for every subsequent call - translating to wasted token usage. CLAUDE.md / AGENTS.md getting way too big. Lots of content gets dumped into these files as your agents work in your repo. I found that it’s not clear what is actually useful to be included there. Every token here counts, because it will be paid for in every session, and its cost scales linearly with the duration of your session. Thinker addresses both of these issues. We have a proxy that wires in between your harness and model to cut out old context that we determine is no longer needed for your current task. Based on the provider you’re using, we determine the right time to batch enough context to cut that justifies the cache re-write cost, which gets paid off after X subsequent calls. We also analyze your sessions and figure out which parts of your instruction files don’t really get used - and send you suggestions to cut or compress those sections. Thinker will kick off a subagent to do that work for you and open a PR you can review. Finally, we use your sessions to record knowledge notes by distilling your sessions with a cheap model (Haiku, Luna). These notes are then served at prompt time via a cross-encoder, also on your machine, that decides whether a note is relevant to the user’s prompt. Here are some data points to show how exactly Thinker saves you coding agent costs: https://thinker.sh/see-the-data If you’re hitting problems at all with usage limits, it’d be great to hear how you’ve been trying to mitigate it. Our open-source repo here (MIT License): https://github.com/yoavshmariahu/thinker

HN: open weights · yoavshmariahu 4 lượt bình chọn 2 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

How to predict the behavior of molecules from scant data

<p>Discovering new molecules for specific uses is a major driver of progress in a wide variety of industries, whether it’s pharmaceutical companies looking for new drugs, or electronics or energy companies seeking new materials with particular properties. But it’s also largely a trial-and-error process, because there is not yet any way to perfectly predict how a given chemical compound will behave in terms of the myriad different properties and responses materials can exhibit.</p><p>Recent progress has been made on using artificial intelligence and machine learning to supplement or speed up laborious calculations based on physical principles that can at least provide some initial guidance on which compounds might be worth synthesizing and testing. Now, a new method developed at MIT shows that when there are few data on some technologically relevant property, such as ligand affinities for a recently discovered cellular target for a potential drug, this approach can outperform other existing methods for winnowing down the field of possible molecules that might be worth looking at.</p><p>The <a href="https://pubs.acs.org/jcisd8/article-abstract/66/17/10692/5250516/Deep-Learning-Foundation-Models-for-Low-Data?redirectedFrom=fulltext" target="_blank">findings were reported</a> in the Sept. 14 issue of the <em>Journal of Chemical Information and Modeling</em>, in a paper by MIT graduate student Akshat Shirish Zalte and recent graduate Jackson Burns PhD ’26, Research Scientist Charlles Abreu, Professor William Green, and three researchers from the chemical company BASF.&nbsp;</p><p>An earlier version of the paper has been online since last year, and the open-source model the team developed, called CheMeleon, has already found regular use by several pharmaceutical companies. CheMeleon was built using ChemProp, an open-source machine-learning package also developed and maintained by the Green group at MIT, and anyone can download CheMeleon through ChemProp and fine-tune it for their own purposes.</p><p>Earlier this year, there was a contest organized by OpenADMET for trying to predict the properties of a variety of new molecules that had not previously been synthesized, says Green, who is the Hoyt Hottel Professor of Chemical Engineering. “It was a big contest, a lot of people participated,” he says. When the winners were announced, it turned out that nine of the top 20 entries had used CheMeleon in their process. “So then we started getting all kinds of inquiries from the pharma companies about this, but our paper had just been submitted; it wasn’t even peer-reviewed or anything yet.”</p><p>While the paper went through that review process, its influence was already spreading. At least one pharma company, Green says, “sent a message to all their people, and said ‘Anytime you’re modeling a property, start with CheMeleon, that’s our new company policy’” when making initial decisions about which molecules to pursue.</p><p>The key to CheMeleon was pre-training a machine-learning model with detailed data about a million different molecules, so that it gets to learn correlations between the structure of a given molecule and how it behaves with respect to a catalog of about 1,600 descriptors that have been developed by chemists over decades. CheMeleon can then be fine-tuned to predict well-known properties like solubility and viscosity, to more esoteric properties such as bioactivity or reactivity, with a certain protein or enzyme.</p><p>Those million molecules are a minuscule fraction of the innumerable ways that atoms can bond together, but they provide enough examples so that once the AI system has been trained on them, it can work out on its own many underlying patterns and relationships, and then apply those to a new, unknown molecule that’s presented to it.</p><p>“Most of today’s deep-learning models require more data than traditional methods,” says Zalte. Some successful models use&nbsp;extensive data from the published literature in chemical journals, he says. “We don’t look at the literature at all. We just calculate a bunch of molecular descriptors like molecular weight or number of oxygens, among others.” Then predicting those descriptors forms the basis for their pre-training task.</p><p>CheMeleon is a pre-trained foundation model. Inspired by the well-known foundation models in language such as ChatGPT and Claide, which essentially are developed by feeding them massive amounts of information from the internet, CheMeleon uses molecular descriptors of a million molecules from PubChem, a large, freely available database of chemical information maintained by the U.S. National Institutes of Health.&nbsp;</p><p>With classical expert-driven models, on the other hand, specialists in a field manually input detailed knowledge about their subject. In developing these classical models, he says, for example, people know from experiments that a certain property of interest, say solubility, may relate to particular descriptors of the same molecules, “like molecular weight, or the number of oxygens that are present, or how branchy or bulky the molecule is.”&nbsp;</p><p>With this team’s approach, they can start with an existing neural network architecture and train it to learn underlying principles. “We’re not training it for a specific property, but as a general model that understands some level of chemistry, so that even if you fine-tune it with just a few data points for the task you care about, you can still get some good performance,” he says. That’s thanks to the fact that the model is “not having to learn everything from scratch.”</p><p>To verify its usefulness, “We compared our model not only against other foundation models that are out there, but also with classical models leveraging expert-designed principles,” Zalte says. Comparing its performance across 58 test cases, “we see that our model comes out on top.”</p><p>Green adds, “The CheMeleon foundation model drastically reduces the number of mol

  • ”</p><p>Green adds, “The CheMeleon foundation model drastically reduces the number of mol
MIT News: AI · David L. Chandler | Department of Chemical Engineering Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Cooking with Reducto’s CEO: Why PDFs Still Break AI — Adit Abraham

Câu hỏi Why do PDFs still break AI?

Why do PDFs still break AI? Reducto co-founder and CEO Adit Abraham joins Allen Park on Latent Space Cooking to talk AI document parsing and make a steak sandwich with burrata, torched peaches, and chimichurri. From the YC pivot out of long-term memory to Reducto R1 and Deep Extract, Adit explains why even frontier models can miss rows in long documents, how his team catches failures in production, and what it takes to go from reading a PDF to reliably editing one. He also revisits the viral steak tweet that brought him into the kitchen. Sponsorship and business inquiries: business@latent.space We talk about: • Why complex tables, handwriting, and unusual layouts still trip up AI • When specialized models and extraction pipelines beat a general-purpose model • Why 99% accuracy can still fall short, and where citations and human review help • What independent benchmarks and production evals reveal about messy documents • What changes at a billion pages a month, the scale Adit shared at filming • What Reducto looks for in engineers, and why it lets candidates use AI in interviews • What a Claude Code moment for office work might look like 00:00 Cold Open 00:13 Meet Adit Abraham and the Viral Steak Tweet 00:57 From Remembrall to Reducto: YC and the Pivot 01:57 Why PDFs Are Still Hard for AI 03:35 What AI Agents Need to Do With Documents 04:19 Model Routing for Tricky Pages and Tables 05:29 Build vs. Buy: When to Train Your Own Models 07:00 Deep Extract, Frontier Models, and Missing Rows 08:20 Why Everyone Wins Their Own Benchmark 09:29 Reducto R1: Accuracy, Cost, and a Billion Pages a Month 10:57 When 99% Accuracy Isn’t Enough 13:05 Continuous Evals and Production Failures 16:49 How Reducto Uses AI Agents Internally 18:28 Hiring With AI and the Eight-Hour Onsite 20:17 Early Hires and What Obsession Looks Like 23:17 Blowtorching Peaches 24:24 Scaling From a Small Team to Series B 27:34 Competition and the Data Flywheel 30:52 What Better Parsing Unlocks for Customers 32:11 Finding a Cofounder at MIT 34:34 The Anthropic Hackathon, Long-Term Memory, and RAG 38:49 From Reading PDFs to Redlining Contracts 43:19 A Claude Code Moment for Office Work 46:13 What’s Next for Reducto: Extraction, Citations, and Cost 47:51 Assembling the Steak Sandwich 49:55 The Steak Allegations and Taste Test

  • Reducto co-founder and CEO Adit Abraham joins Allen Park on Latent Space Cooking to talk AI document parsing and make a steak sandwich with burrata, torched peaches, and chimichurri.
  • From the YC pivot out of long-term memory to Reducto R1 and Deep Extract, Adit explains why even frontier models can miss rows in long documents, how his team catches failures in production, and what it takes to go from reading a PDF to reliably editing one.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Controlling Biology - Fabian Theis, Helmholtz Munich

Câu hỏi View post on X ↗ But why are these organizations doubling down on these big bets now?

Fabian Theis has been one of the key players in the various iterations of the Virtual Cell concept and initiatives. It is not hyperbole at all to say that his lab has been pushing the frontier of algorithms to understand multi-scale, multi-modal (”multiomics”), time-varying biological data. Indeed, he has shaped the large scale vision and data collection through his leadership at Helmholtz, various startups and probably at Biohub itself. Watch the full episode on YouTube… Much of the recent attention in AI x Bio has been on protein and molecular design. AlphaFold, ESM and BioHub , Chai , Leash , Isomporphic, Boltz , Genesis , Octant, Baker Lab, and many many more. After a lot of initial excitement, Virtual Cell projects have received less attention, due in no small part to questions about the effectiveness of large-scale models in solving even basic Virtual Cell problems. A $1.8B bet The Biohub announcement a few weeks ago got a lot of press for the (really) big numbers: $1.8B of commitments [1] from Biohub, NIH, DOE, Google and Meta. The discussion cites new imaging and measurement capabilities, generating large data sets and making the data available for training AI. View post on X ↗ But why are these organizations doubling down on these big bets now? They have been in the works for years now, what is going on that makes now the moment to push? New measurement capabilities and scale infrastructure, are now coming online that make Biohub’s push likely lead to big impact on science. DOE and NIH are developing lab-in-the-loop automation capabilities at scale. Google is scaling and broadening its connectomics capabilities ( episode coming soon!). Biohub itself has worked with various labs to develop and improve imaging platforms like laser phase plate microscopy, light sheet microscopy and advanced post-processing for regular microscopes that reveal deep phenotypic data. These technologies are quickly maturing, and will likely play a big part in the Biohub initiative. Virtual cells struggle This kind of multi-scale, multi-modal data looks a lot like the data Fabian has made a career out modeling. On the other hand Virtual Cell modeling has run into problems that have dampened excitement about systems biology. The 2025 Arc Virtual Cell Challenge shined a light on the problem: a slew of papers pointed out that constant baselines (not even linear!) often beat sophisticated models for “Perturbation Prediction”. If you change the inputs or programming of a cell (via gene editing, drugs or other environmental factors), how will the cell change it’s “state”. The disappointing finding is that many models fail to beat even the most basic basic baselines 🤦‍♂️. The situation is not actually so bad: This excellent paper show that many existing evaluation metrics are poorly calibrated: they can reward an uninformative mean prediction because most genes barely change. With better-calibrated metrics, deep learning models can outperform the mean baseline. The problem is the problem Nonetheless, Fabian Theis believes that perturbation prediction under the existing data sets just isn’t a hard enough problem to warrant complex algorithms. Instead we should be focused on exactly the problems that Biohub just announce an $1.8B data collection effort around: using large, high precision, multi-scale, co-registered biological data to predict the evolution of molecules → cells → niches → tissues → organs → organisms, and how all these scales effect one another. This is a happy situation: at this very moment, AI is enabling us to tackle dramatically larger problems, and this problem is very, very big. We lack understanding of the “scale interfaces” between these different different “domains of applicability,” and we lack the data. The Biohub initiative aims to solve the data problem. Fabian Theis has been working on the algorithms to model these larger problems for years. How do the other cells near me change my state? What happens to a distribution of cells from a perturbation? How do I perturb a cell to get it into the state I want? These are increasingly complex questions that are building towards a lab-in-the-loop and large scale experimentation, and (actually) require increasingly complex models. Models drive precision medicine Fabian’s labs and initiatives have been also investing heavily: the Computational Health Center at Helmholtz combines AI, molecular data, medical imaging, genomics, and clinical data to develop predictive and mechanistic models for precision medicine. The Institute of Computational Biology develops statistical and dynamical models integrating biological information from single-cell time series through population-scale omics data. Given his drivers seat at the frontier, we were so excited for the opportunity to get an hour of his time to understand how Fabian Theis is thinking about Virtual Cells when the concept of what is difficult is changing so quickly. We also discussed: The business of Virtual Cells The Bitter Lesson in biology - our perennial topic Training the next generation of Computation Biologists Closing the lab-in-the-loop Controlling Biology - Fabian Theis, Helmholtz Munich Fabian Theis has been one of the key players in the various iterations of the Virtual Cell concept and initiatives. It is not hyperbole at all to say that his lab has been pushing the frontier of algorithms to understand multi-scale, multi-modal (”mult... Latent Space [1] NIH money is prior investments

  • Biohub itself has worked with various labs to develop and improve imaging platforms like laser phase plate microscopy, light sheet microscopy and advanced post-processing for regular microscopes that reveal deep phenotypic data.
  • With better-calibrated metrics, deep learning models can outperform the mean baseline.
Nguyên văn nguồn
Chi tiết nguồn
161–180 / 562
Trang 9 / 29