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

Huấn luyện

Reinforcement Learning: Optimizing 0.6B Dense Models on TPU v7 via vLLM

Co-designing RL for a Qwen3-0.6B dense policy on Google Cloud TPU v7 across vLLM, Tunix, MaxText, and Pallas, with a 107x end-to-end speedup from ~100 minutes to ~56 seconds per true-8K RL cycle.

vLLM Blog · Loki Chen, Yang Song, Hongye Xie, Ming Lei, Greg Rehm, Chongyang Bai, Dre Olgiati, Bob Kamma, Tej Choudhary, Hang Cui, Imed Zitouni, Hamed Firooz (Meta ASA); Mingfu Liang, Kavosh Asadi, Yufei Li, Frank Shyu, Parish Aggarwal, Senthil Manickavelu, Xi Liu, Luke Simon (Meta Ranking AI Research) Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Gemini Decisions API

Câu hỏi net/2026/Nov/10/yonkers-rises/">Yonkers stomping robot</a> video file:</p> <pre><code>dragon: Does this video have a dragon in it?

<p><strong><a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/">Gemini Decisions API</a></strong></p> Google join <a href="https://simonwillison.net/2026/Sep/21/jev/">Jev</a> and <a href="https://simonwillison.net/2026/Oct/6/llm-openai-decisions/">OpenAI</a> in the decision model category with "Gemini Decision 1". This new model API is almost identical to Jev: it supports the same three types of questions (yes/no, choice, and rubric-scoring) and even reuses Jev's term "Noul" for a yes/no question.</p> <p>It's priced the same as the GPT-6 Luna decision model: 10 cents per million input tokens, and output is free. Jev remains the cheapest of the three, at 4.2 cents per million.</p> <p>Jev is text (and JSON) only. OpenAI can handle text and images. Gemini Decision 1 is fully multi-model: it covers text, images, audio, video, and PDFs.</p> <p>I tested it against video by passing these questions to it along with my <a href="https://simonwillison.net/2026/Nov/10/yonkers-rises/">Yonkers stomping robot</a> video file:</p> <pre><code>dragon: Does this video have a dragon in it? robot: Does this video have a robot in it? france: Is this video set in France? usa: Is this video set in the USA? </code></pre> <p>Jev and OpenAI both return 1.0 or 0.0 for yes/no questions, but Gemini insists on being a little less confident than that:</p> <div class="highlight highlight-source-json"><pre>{ <span class="pl-ent">"dragon"</span>: { <span class="pl-ent">"type"</span>: <span class="pl-s"><span class="pl-pds">"</span>yes_no<span class="pl-pds">"</span></span>, <span class="pl-ent">"probability"</span>: <span class="pl-c1">0.04976021841520312</span> }, <span class="pl-ent">"robot"</span>: { <span class="pl-ent">"type"</span>: <span class="pl-s"><span class="pl-pds">"</span>yes_no<span class="pl-pds">"</span></span>, <span class="pl-ent">"probability"</span>: <span class="pl-c1">0.9234055387792485</span> }, <span class="pl-ent">"france"</span>: { <span class="pl-ent">"type"</span>: <span class="pl-s"><span class="pl-pds">"</span>yes_no<span class="pl-pds">"</span></span>, <span class="pl-ent">"probability"</span>: <span class="pl-c1">0.08329171353166241</span> }, <span class="pl-ent">"usa"</span>: { <span class="pl-ent">"type"</span>: <span class="pl-s"><span class="pl-pds">"</span>yes_no<span class="pl-pds">"</span></span>, <span class="pl-ent">"probability"</span>: <span class="pl-c1">0.9230459995157706</span> } }</pre></div> <p><small></small>Via <a href="https://news.ycombinator.com/item?id=50456791">Hacker News</a></small></p> <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/gemini">gemini</a>, <a href="https://simonwillison.net/tags/llm-release">llm-release</a>, <a href="https://simonwillison.net/tags/jev">jev</a>, <a href="https://simonwillison.net/tags/decision-models">decision-models</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

aws-emulator

<p><strong><a href="https://github.com/aws/aws-emulator">aws-emulator</a></strong></p> Finally, an official emulator from AWS to run tests against a local clone of their many different services without having to provision a paid API key to do so.</p> <p>It's written in Rust by a huge swarm of agents - 29,684 commits in the first three weeks of the project. Swaminathan (Swami) Sivasubramanian wrote about <a href="https://www.allthingsdistributed.com/2026/11/aws-emulator-and-the-engineers-job-when-agents-write-the-code.html">how it came to be</a> in a guest post on Werner Vogels' blog.</p> <p>It currently requires Docker to run on macOS or Windows. I'm hoping they fix that in the future - it would be great if you could download a single binary and use that as part of a standard test suite in other languages and on other platforms. <p>Tags: <a href="https://simonwillison.net/tags/amazon">amazon</a>, <a href="https://simonwillison.net/tags/amazon-web-services">amazon-web-services</a>, <a href="https://simonwillison.net/tags/testing">testing</a>, <a href="https://simonwillison.net/tags/werner-vogels">werner-vogels</a>, <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/rust">rust</a>, <a href="https://simonwillison.net/tags/docker">docker</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/agentic-engineering">agentic-engineering</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.34.0rc1

<p>[Bugfix][Spec Decode][XPU] greedy_argmax: treat NaN as -inf on the to…</p>

Nguyên văn nguồn
Chi tiết nguồn
  • v0.34.0rc1 vllm-project/vllm releases · 12 thg 11, 2026
Kỹ thuật

Browse tools.simonwillison.net by tag

<p><strong><a href="https://tools.simonwillison.net/#tags">Browse tools.simonwillison.net by tag</a></strong></p> I've been wanting a better way to browse my <a href="">collection of vibe-coded tools</a> (244 and counting) for a while. I had a few Fable 5.2 credits left over before my weekly reset so I set it to work using <a href="https://github.com/tree-sitter/py-tree-sitter">tree-sitter</a> to analyze the code for JavaScript APIs, libraries, and web API hosts used by each tool.</p> <p>The homepage now features tag clouds for browsing by each of those categories of tag:</p> <p><a href="https://tools.simonwillison.net/#tags"><img alt="Screenshot of two tag cloud sections where tags are sized by count. Browse by JavaScript libraries - Libraries loaded from a CDN by two or more tools. Select a tag to list its tools. @google/generative-ai 3, @jsquash/avif 3, @jsquash/webp 4, Babel 2, Chart.js 2, Cropper.js 2, D3 2, DOMPurify 7, js-yaml 3, JSZip 3, Leaflet 4, markdown-it 4, Marked 12, pako 2, PDF.js 3, Pyodide 11, React 3, Tesseract.js 2, three.js 3. Browse by browser APIs - Web platform features used by each tool's own code. Select a tag to list its tools. &lt;dialog&gt; 5, AbortController 6, Blob URLs 38, BroadcastChannel 1, Built-in AI 1, Camera / microphone 9, Canvas 2D 44, Clipboard API 69, Compression Streams 3, createImageBitmap 2, crypto random values 4, DOMParser 8, Drag and drop 30, Dynamic import() 12, ES modules 92" src="https://static.simonwillison.net/static/2026/tools-tag-cloud.webp" /></a> <p>Tags: <a href="https://simonwillison.net/tags/javascript">javascript</a>, <a href="https://simonwillison.net/tags/tools">tools</a>, <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/claude">claude</a>, <a href="https://simonwillison.net/tags/vibe-coding">vibe-coding</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

llm 0.37.2

<p><strong>Release:</strong> <a href="https://github.com/simonw/llm/releases/tag/0.37.2">llm 0.37.2</a></p> <p>Fixes a security issue reported by <a href="https://github.com/BluePlexus">Daniel de Araujo</a> in <a href="https://github.com/simonw/llm/security/advisories/GHSA-v9h7-jj4g-f862">GHSA-v9h7-jj4g-f862</a> and <a href="https://github.com/simonw/llm/security/advisories/GHSA-r639-rr7r-x72w">GHSA-r639-rr7r-x72w</a>.</p> <p>LLM has a mechanism called <a href="https://llm.datasette.io/en/stable/other-models.html#configure-an-openai-compatible-model">extra-openai-models.yaml</a> - now managed by the new-in-0.37 <a href="https://llm.datasette.io/en/stable/other-models.html#openai-endpoint-add">llm openai endpoint add</a> command - for using LLM with OpenAI-compatible API endpoints from other providers.</p> <p>Daniel reported two ways this could accidentally send your configured OpenAI API key to one of those providers: if the key named by <code>api_key_name</code> had never been stored, or if a <code>responses: true</code> model was run with <code>-o chat_completions 1</code>.</p> <p>Tags: <a href="https://simonwillison.net/tags/projects">projects</a>, <a href="https://simonwillison.net/tags/security">security</a>, <a href="https://simonwillison.net/tags/llm">llm</a>, <a href="https://simonwillison.net/tags/releases">releases</a></p>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Validating GPU Platforms with Runtime AI Inventory via the NVIDIA DSX AI Factory

As teams deploy and scale AI workloads using NVIDIA Dynamo, Kubeflow, or Slurm, tracking what actually runs requires visibility beyond the platform... As teams deploy and scale AI workloads using NVIDIA Dynamo, Kubeflow, or Slurm, tracking what actually runs requires visibility beyond the platform configuration captured in a recipe. Application teams define and update these workloads independently of the platform recipe, so a reproducible deployment bundle alone cannot establish which workload artifacts are running on each node. Source

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Try Tabletop: Customizable, collaborative spaces for groups to plan, play and do more together.

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/labs_tabletop_blog_image_2784x1.max-600x600.format-webp.webp">Tabletop is a new Google Labs experiment for group collaboration. Build interactive itineraries, polls, and brackets with your team. Join the waitlist.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

A New Vision for Human Work After Automation

Câu hỏi utm_source=youtube&utm_campaign=thesis-2027&utm_content=every-261111-video What does great human work look like after automation?

Thesis: 2027 is made possible by our founding partners, Lovable and Notion. Lovable is the fastest way to turn ideas into software people love: https://link.lovable.dev/every Notion is the collaborative AI workspace where teams and agents work side by side: https://www.notion.com/?utm_source=youtube&utm_campaign=thesis-2027&utm_content=every-261111-video What does great human work look like after automation? It’s a question we try to answer every day at Every, and the one we built our conference, Thesis, around. In his closing keynote, *Dan Shipper* argues that as AI makes intelligence cheap and widely available, what matters most is human perspective. It’s your “non-consensus” point of view, plus your responsibility for a particular corner of the world, that helps you spot the insight no one else would. On this week’s episode of The Every Podcast, we’re sharing Dan’s full talk at Thesis: 2027. Dan explains why benchmark charts make the future of work look bleaker than it is, why picking the right idea still takes a human, and how we should really measure AI models. Subscribe to Every (free for 14 days): https://every.to/youtube-every-offer?utm_source=podcast&utm_medium=audio&utm_campaign=thesistalkepisode Links to resources mentioned in the episode: • Thesis: https://every.to/thesis-2027 • Dan Shipper: https://x.com/danshipper • OpenAI’s GDPval benchmark: https://openai.com/index/gdpval/ Timestamps: 0:00 Introduction 1:51 Work after automation 5:00 1 person vs 5 jobs 10:18 Why perspective matters 17:55 Better AI, more valuable humans

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Fostering community amid Kendall Square’s innovation

<p>Kendall Square in Cambridge, Massachusetts, is often cited as the most innovative square mile in the world. But when the Kendall Square Association gathered local stakeholders for its 18th Annual Meeting Thursday, the conversation focused on innovation primarily in the context of community.</p><p>The event, which took place at the nonprofit startup incubator The Engine, attracted people from across Cambridge, including business leaders, elected officials, local residents, and students, for a night of local food and conversation about the advance of technologies like artificial intelligence and what they mean for society.</p><p>“How can we make AI not an enemy of work, but a tool for workers to expand their capabilities and build new expertise or do new things?” asked MIT Institute Professor Daron Acemoglu in a keynote talk. “The world is at a point in time where we need new innovations and new adaptations to growing challenges.”</p><p>Acemoglu’s keynote was one of several presentations at the event, which also featured music from the Kendall Square Orchestra, a vote on new KSA board members, and awards for organizations making a positive impact through innovation.</p><p>KSA is a nonprofit made up of over 150 organizations across Kendall Square, from major companies to universities, research organizations, and the independent shops and restaurants that give Kendall Square its distinct character.</p><p>“People come to Kendall Square to make the world better,” said KSA Executive Director Beth O’Neill Maloney in her opening remarks. “People come here to improve the human condition, people come here to solve big problems, and people come here to innovate for good.”</p><p>Acemoglu, who shared the Nobel Prize in economics in 2024 for his work studying how institutions affect prosperity, spoke on the subject of his new book, “What Happened to Liberal Democracy?”</p><p>Acemoglu says liberal democracy is under siege as trust in liberal institutions decline, political polarization rises, and people lose faith in elected officials. He acknowledged that the shared prosperity liberal democracy promised has not always been delivered, especially over the last 50 years, but insisted it is the best path forward and in need of renewed support.</p><p>“I believe liberal democratic institutions, together with the norms that support them, are one of the greatest human achievements,” Acemoglu said. “It would be unthinkable for people who lived 300 years ago that we could live in societies consisting of hundreds of millions of people with diverse interests, different views, doing different things, and people still participate in politics and reach compromises, and we have state institutions that, by and large, actually provide services and do things for the people, and that we build opportunities for prosperity from which most people benefit while social freedoms are largely upheld.”</p><p>Cambridge Mayor Sumbul Siddiqui also spoke at the event. Siddiqui began her talk by telling the story of Cambridge residents working with biotech labs in the 1970s to create a groundbreaking policy on recombinant DNA research, which helped turn Kendall Square into the life science hub it is today.</p><p>“All these years later I think that story captures what makes Kendall Square so special: It’s our willingness to work together even when we face big challenges,” Siddiqui said. “It isn’t just the universities or labs, it’s also the culture of collaboration among nonprofits, governments, and residents.”</p><p>Broad Institute Chief Communications Officer Lee McGuire started his talk by reviewing a 2009 report on what made Kendall Square such a powerful driver of innovation, noting that factors like talent density, education, and a collective voice remain strong in the area today.</p><p>“This report was brilliant in that it captures so many of the special things that matter for a place like Kendall,” McGuire said. “It is the job of the Kendall Square Association and all of us to help protect those things, make sure we name them, understand them, and see them as essential elements of making sure this community stays what it is no matter what happens. As long as there’s a good idea, this is the place that’s going to get that idea moving and scale it up to change the world.”</p>

  • </p><p>“I believe liberal democratic institutions, together with the norms that support them, are one of the greatest human achievements,” Acemoglu said.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Encoder Disaggregation in vLLM: Scaling Multimodal Serving with EPD

Recent advances in vLLM encoder disaggregation: metadata-only multimodal requests, dedicated encoder execution, remote embedding transfer, cross-encoder caching, and lower-overhead EPD routing.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.40.17-rc0

<p>MLX, MLX-C: version bump (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6316056226" data-permission-text="Title is private" data-url="https://github.com/ollama/ollama/issues/19647" data-hovercard-type="pull_request" data-hovercard-url="/ollama/ollama/pull/19647/hovercard" href="https://github.com/ollama/ollama/pull/19647">#19647</a>)</p>

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context

Long-running agents degrade as context piles up. Learn how offloading, summarization, skills, subagents, and memory keep your agent's context window focused.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Designing for Agents — Zeno Rocha, Resend

Câu hỏi Resend co-founder and CEO Zeno Rocha explores answer engine optimization (AEO) through a concrete example: which email provider does Claude Code recommend, and how does that change across model versions?

A coding agent opens a pull request and picks an email provider for you. How did it choose? Resend co-founder and CEO Zeno Rocha explores answer engine optimization (AEO) through a concrete example: which email provider does Claude Code recommend, and how does that change across model versions? Then comes the agent experience itself. Once the agent picks Resend, can it find the price, set up the account and send its first email? We also follow an email after the API call succeeds. Zeno explains what happens when Gmail throttles a burst, why a tracking pixel can give you the wrong idea about who opened a message, and how one abusive sender can put other customers’ mail at risk. Then we get into the product decisions: giving an agent 80+ named MCP tools, letting it read pricing in Markdown, and letting it register for a service on its own. Explore the references • Zeno’s analysis of what Claude Code recommends: https://x.com/zenorocha/status/2027757871597822022 • Resend’s $3M email safety pledge: https://resend.com/blog/3-million-to-make-email-safer • React Email, the open-source email components: https://react.email/ • The proposed agent registration standard shown on screen: https://workos.com/auth-md Chapters 00:00 Cold open 00:59 From Dracula to email 05:02 When agents pick the API 06:54 Getting into the inbox 13:28 What tracking measures 17:05 Paying for trust 19:43 The DNS hurdle 23:54 React Email’s wedge 27:00 Measuring adoption 31:57 Owning the mail stack 37:54 Designing the MCP 46:28 CLI, SDKs and craft 52:40 Mail for agents 58:27 Writing for machines 1:05:59 Winning recommendations 1:12:13 Enterprise and pricing 1:16:50 Building with a family 1:23:25 Naming a category

Nguyên văn nguồn
Chi tiết nguồn
381–400 / 680
Trang 20 / 34