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

Tác tử

v18.20.0

<h2>@oh-my-pi/pi-agent-core</h2> <h3>Changed</h3> <ul> <li><code>prepareCompaction()</code> runs as the <code>compaction.prepare</code> loop phase, so a <code>ui.loop-blocked</code> stall from a cold token walk over a long branch names compaction instead of "unknown" (<a href="https://github.com/can1357/oh-my-pi/issues/15027" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/15027/hovercard">#15027</a>, <a href="https://github.com/can1357/oh-my-pi/pull/18867" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/18867/hovercard">#18867</a> by <a href="https://github.com/wolfiesch">@wolfiesch</a>)</li> </ul> <h2>@oh-my-pi/pi-ai</h2> <h3>Added</h3> <ul> <li>Added the <code>google-decisions</code> judgment API for Google's Gemini Decisions API (<code>gemini-decision-1</code>), with text and image input (<a href="https://github.com/can1357/oh-my-pi/pull/19060" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19060/hovercard">#19060</a> by <a href="https://github.com/kevcube">@kevcube</a>)</li> <li>Added <code>apiKey</code> to successful <code>OAuthAccessResolution</code> results from <code>oauth.accessById()</code> and <code>oauth.accessAll()</code>: the key normal requests send for that account, which differs from <code>accessToken</code> for providers with structured credentials (<a href="https://github.com/can1357/oh-my-pi/pull/19030" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19030/hovercard">#19030</a> by <a href="https://github.com/Xytronix">@Xytronix</a>)</li> <li>Added an optional <code>exemptionAlias</code> to <code>ToolCallLoopGuard</code> so transport calls can match the dispatched tool's name in <code>exemptTools</code> (<a href="https://github.com/can1357/oh-my-pi/issues/18860" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/18860/hovercard">#18860</a>)</li> <li>Added the <code>liquid-systemone</code> judgment API for Liquid AI's decision models, which rejects image sets over Liquid's 10,000 32×32-patch request cap before sending (<a href="https://github.com/can1357/oh-my-pi/issues/18076" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/18076/hovercard">#18076</a>, <a href="https://github.com/can1357/oh-my-pi/pull/18100" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/18100/hovercard">#18100</a> by <a href="https://github.com/SuperUserA">@SuperUserA</a>)</li> <li>Added the <code>nace-systemone</code> judgment API for Nace's Drex decision models, which sends images as Drex's top-level <code>images</code> data URLs and replaces the base64 image, audio, video and PDF data URLs that Drex rejects in judged text with a description (<a href="https://github.com/can1357/oh-my-pi/pull/18950" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/18950/hovercard">#18950</a> by <a href="https://github.com/alphastorm">@alphastorm</a>)</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed a spent Codex account retrying every minute for days when its siblings were only refused the requested model: their short model-policy blocks no longer count as a sibling becoming available, so the usage-limit wait follows the spent account's reset and model fallback can take over (<a href="https://github.com/can1357/oh-my-pi/issues/17637" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/17637/hovercard">#17637</a>, <a href="https://github.com/can1357/oh-my-pi/pull/19019" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19019/hovercard">#19019</a> by <a href="https://github.com/wolfiesch">@wolfiesch</a>)</li> <li>Auth-gateway client tokens (<code>serve --clients</code>) can no longer set who a request is from, who pays, or how the gateway reaches the provider: chat routes drop their request headers, <code>userProfileId</code>, <code>initiatorOverride</code>, <code>metadata</code>, <code>serviceTier</code>, <code>cacheRetention</code> and the other gateway-only options, while the gateway token keeps them (<a href="https://github.com/can1357/oh-my-pi/pull/19067" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19067/hovercard">#19067</a> by <a href="https://github.com/alphastorm">@alphastorm</a>)</li> <li>Fixed the auth gateway's <code>/v1/pi/stream</code> dropping <code>taskBudget</code>, <code>include</code> and <code>anthropicPrefixMismatchBehavior</code>, which its other chat routes accept (<a href="https://github.com/can1357/oh-my-pi/pull/19067" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19067/hovercard">#19067</a> by <a href="https://github.com/alphastorm">@alphastorm</a>)</li> <li>Fixed requests to Vertex AI Mistral partner models (<code>mistral-medium-3</code>, <code>mistral-small-2503</code>, <code>codestral-2</code>) failing because they were sent in Anthropic's request format (<a href="https://github.com/can1357/oh-my-pi/issues/19010" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/19010/hovercard">#19010</a>, <a href="https://github.com/can1357/oh-my-pi/pull/19022" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19022/hovercard">#19022</a> by <a href="https://github.com/yuzu-octopus">@yuzu-octopus</a>)</li> <li>Added Gemini's <code>xhigh</code> thinking level, sent on the Gemini API and Vertex for models whose ladder exposes it (Gemini 4 Hydrogen).</li> <li>Fixed omp failing to start with <code>ENOTDIR: not a directory, stat '~/.aws/credentials'</code> when <code>~/.aws</code> is a file instead of a directory (<a href="https://github.com/can1357/oh-my-pi/pull/19083" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/19083/hovercard">#19083</a> by <a href="https://github.com/iamshatter">@iamshatter</a>)</li> </ul> <h2>@oh-my-pi/pi-catalog</h2> <h3>Added</h3> <ul> <li>Added the Gemini Decision 1 judge model (

Nguyên văn nguồn
Chi tiết nguồn
  • v18.20.0 Can Bölük — Oh My Pi releases · 13 thg 11, 2026
Kỹ thuật

llm-gemini 0.37

Câu hỏi If they're the same price, why have two different models?

<p><strong>Release:</strong> <a href="https://github.com/simonw/llm-gemini/releases/tag/0.37">llm-gemini 0.37</a></p> <p>Google launched <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/">Gemini 4 Hydrogen</a> today - the lowest cost member of the Gemini 4 family.</p> <p>It's actually not too far off the price of its much bigger siblings, Carbon and Argon. I'm finding Gemini 4's pricing <em>really</em> confusing. Here's the pricing scheme:</p> <center><table> <thead> <tr> <th>Model</th> <th>Dates</th> <th>Input ($/M)</th> <th>Cached input</th> <th>Output</th> </tr> </thead> <tbody> <tr> <td rowspan="2">Gemini 4 Carbon</td> <td>Through Nov 30, 2026</td> <td>$2.00</td> <td>$0.10</td> <td>$10.00</td> </tr> <tr> <td>From Dec 1, 2026</td> <td>$4.00</td> <td>$0.20</td> <td>$20.00</td> </tr> <tr> <td rowspan="2">Gemini 4 Argon</td> <td>Through Nov 30, 2026</td> <td>$2.00</td> <td>$0.10</td> <td>$10.00</td> </tr> <tr> <td>From Dec 1, 2026</td> <td>$4.00</td> <td>$0.20</td> <td>$20.00</td> </tr> <tr> <td>Gemini 4 Hydrogen</td> <td>&nbsp;</td> <td>$1.50</td> <td>$0.075</td> <td>$7.50</td> </tr> </tbody> </table></center> <p>Carbon and Argon are <em>the same price</em>, and both are scheduled to double in price on December 1st. If they're the same price, why have two different models? <a href="https://aistudio.google.com/docs/pricing">Google says</a> that Argon "delivers Gemini 4 intelligence with faster responses and lower token spend per turn compared to Gemini 4 Carbon" - so it may be priced the same, but it will use less tokens per task.</p> <p>Hydrogen, supposedly the cheap model, is currently only a 25% discount on its more capable peers. The Hydrogen price will make more sense after December 1st, when it will represent a 75% discount. It's still considerably more expensive than Claude Haiku 5.5 and GPT-6 Luna - they're both $0.10/$0.50, so Hydrogen is 15x that.</p> <p>I had Hydrogen <a href="https://tools.simonwillison.net/markdown-svg-renderer?url=https%3A%2F%2Fgist.github.com%2Fsimonw%2Fce357e95149cd16f65ec7e7106ed4fed">generate some pelicans</a>. Here's a <a href="https://static.simonwillison.net/static/2026/gemini-4-pelicans.html">full comparison grid</a> showing the pelicans for all three members of the Gemini 4 family at their three supported effort levels:</p> <p><img alt="Comparison grid of nine SVG-style illustrations of a pelican riding a bicycle, three models across by three thinking levels down. Column headers: Thinking, gemini-4-carbon, gemini-4-argon, gemini-4-hydrogen. Row HIGH: gemini-4-carbon draws a detailed pelican in a sailor cap riding a red cruiser bicycle with fenders and a wicker basket holding a fish, along a beach road with sea, sun, clouds and birds, captioned 40 input tokens, 20,305 output = 20.31 cents · 4m 11s; gemini-4-argon draws a pelican in a red cap and a streaming teal striped scarf on a red bicycle with a blue fish peeking out of the basket, on a seaside promenade with a pale sun, captioned 40 input tokens, 17,158 output = 17.17 cents · 2m 29s; gemini-4-hydrogen draws a much simpler plain white pelican on a basic red bicycle against a pale blue sky with two clouds, speed lines and flat green ground, captioned 40 input tokens, 5,311 output = 3.99 cents · 19s. Row MEDIUM: gemini-4-carbon draws a pelican in a sailor cap and red striped scarf on a red bicycle with cream fenders and a basket, on a sandy beach with a rope fence and sea, captioned 40 input tokens, 11,020 output = 11.03 cents · 1m 53s; gemini-4-argon draws a pelican with an enormous orange beak, sailor cap and flowing red scarf on a teal bicycle with a basket, on a boardwalk under a purple and pink sunset sky with a large pink sun, captioned 40 input tokens, 11,090 output = 11.10 cents · 1m 32s; gemini-4-hydrogen draws a simple cream pelican on a plain red bicycle with sparse spokes, two clouds and flat green ground, captioned 40 input tokens, 3,964 output = 2.98 cents · 17s. Row LOW: gemini-4-carbon draws a pelican in a sailor cap and red and white striped scarf on a red bicycle with fenders and a basket, passing a white seaside railing on a grey road, captioned 40 input tokens, 9,153 output = 9.16 cents · 1m 34s; gemini-4-argon draws a large, chubby pelican with a sailor cap and huge orange beak on a red bicycle with a basket, gripping the handlebars with a dark blue wing, in front of an orange setting sun and a grey road, captioned 40 input tokens, 10,177 output = 10.18 cents · 1m 21s; gemini-4-hydrogen draws the simplest image, a plain white pelican on a thin red bicycle over a single green line with no scenery, captioned 40 input tokens, 2,996 output = 2.25 cents · 10s. Overall, carbon and argon produce rich, accessorised seaside scenes at every thinking level, while hydrogen produces far plainer pelicans at a fraction of the cost and time, and detail decreases for all three models as thinking drops from high to low." src="https://static.simonwillison.net/static/2026/gemini-4-pelicans.webp" /></p> <p>Considering how close they are in price, the difference in pelican quality for Hydrogen is quite stark.</p> <p>Here's <a href="https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/">how Google describe</a> Hydrogen's intended position in their mix of models:</p> <blockquote> <p>In real-world software development, while teams rely on frontier models like Carbon and Argon for deep architectural judgment when designing systems or untangling subtle bugs, they still need a fast, disciplined execution model for writing code, running tests, and verifying changes. Hydrogen is built to stand on its own as a model for everyday work — or to pair with Carbon and Argon as the high-efficiency execution engine in multi-model workflows.</p> </blockquote> <p>Most notably, Hydrogen is the new free tier model, and is the model that will power AI features in Google Search.</p> <p>Tags: <a href="https://sim

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

langgraph==1.2.21

<p>Changes since 1.2.20</p> <ul> <li>release(langgraph): 1.2.21 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6353233846" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9683" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9683/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9683">#9683</a>)</li> <li>chore(deps): bump the minor-and-patch group across 1 directory with 9 updates (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6143893583" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9530" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9530/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9530">#9530</a>)</li> <li>fix(langgraph): match whole namespace segments in get_subgraphs (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6245081883" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9615" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9615/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9615">#9615</a>)</li> <li>fix(langgraph): read an undeclared subgraph namespace with the parent graph (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6242638688" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9614" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9614/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9614">#9614</a>)</li> <li>docs(langgraph): say each interrupt() call has its own id in UnknownInterruptError (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216068500" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9600" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9600/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9600">#9600</a>)</li> <li>chore(deps): bump filelock and langgraph-api for security fixes (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6222165800" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9606" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9606/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9606">#9606</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Google Search gets a major upgrade with Gemini 4 Hydrogen

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/HydrogenUI_socialshare.max-600x600.format-webp.webp">We are bringing Gemini 4 Hydrogen to Google Search globally to provide you with more helpful, visual, and intuitive responses.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Why Drug Approvals Haven’t Grown

Medicine has been picking low-hanging fruit. What’s left is harder to reach. Isomorphic Labs President Max Jaderberg on why drug approvals are roughly flat. New No Priors episode out now!

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Giving AI a “crystal clear” view of the materials world

<p>In principle, artificial intelligence could make finding the right material for an application as simple as ordering dinner: Scientists would tell a large language model what they want, and it would serve them the perfect material for the job.</p><p>Unfortunately, that approach doesn’t work well yet, and the problem comes down to data: Most materials consist of 3D arrangements of crystals, not the text streams that today’s large language models are trained on. The experimental literature is also heavily biased toward a few popular classes of crystals, making AI outputs skew toward those materials.</p><p>Now MIT researchers have created a multimodal AI model that tackles these problems. The model combines 3D atomic structures with natural language descriptions to create rich linguistic descriptions about the structures from experimental research papers and first-principles computational databases. The unified description outlines not just the chemical composition and structure, but also its properties, processing considerations, and potential applications, helping AI learn more about what makes materials impactful.</p><p>In a <a href="https://www.nature.com/articles/s41467-026-76378-7" target="_blank">new paper in <em>Nature Communications</em></a>, the researchers showed their system could identify better material designs than models trained on text streams alone by allowing the model to “see” the atomic arrangements and bonding patterns. For example, when the researchers asked the model to identify materials suited for solid-state lithium-ion battery electrolytes, its top results included materials that had been independently identified as promising candidates. The researchers also showed their model learned geometric representations of material structures in ways text-only models could not.</p><p>“Structure is really at the heart of the materials science field,” says Ju Li, the Carl Richard Soderberg Professor in Power Engineering in the departments of Materials Science and Engineering and Nuclear Science and Engineering. “If you understand the structure, you can infer quite a lot about the properties and performance of materials. We wanted to build the connection between a material’s atomic structure and its properties and processing for the model.”</p><p>Joining Li on a paper about the model are Yang Jeong Park, first author and former MIT postdoc; Mayank Kumaran, a recent University of Illinois Urbana-Champaign computer engineering graduate; Chia-Wei Hsu SM ’26; and Elsa Olivetti, the Jerry McAfee (1940) Professor in Engineering in DMSE.</p><p><strong>Getting models to see in 3D</strong></p><p>Materials scientists are always on the hunt for new crystals that will lead to the products of the future. Over the last decade, AI has reshaped that process, proposing and designing new crystalline structures, which have highly ordered atomic structures and make up more than 90 percent of solid materials on Earth.<strong>&nbsp;</strong>Most of those models treat solid crystalline materials as 3D graphs.</p><p>The rise of large language models has given researchers another tool for innovation. But existing large language models rely heavily on text from published papers and books, limiting their knowledge of the physical world, since many potential crystal structures are stored in computational databases. Text-based models also lack understanding of visual geometry.</p><p>To overcome those challenges, the researchers decided to create a vision-language model for crystalline materials.</p><p>They trained a language model to create textual descriptions of crystal structures using an approach known as contrastive learning, in which models compare similar and dissimilar examples of structures and textual descriptions to achieve a unified understanding. This is similar to aligning the visual and language regions of the brain.</p><p>Following training, the researchers compared the predictions of two versions of the model: one trained on a dataset containing the AI-generated descriptions and the&nbsp;other trained on a dataset containing text from papers only. They found that the model trained on AI-generated&nbsp;descriptions successfully learned how to align geometric structures with text descriptions while the other did not.</p><p>“We wanted to create an AI model that works broadly across materials,” Li says. “That’s why we decided to synthesize narratives rather than rely solely on existing papers. We found when we used that narrative dataset, the models are better at predicting the function.”</p><p><strong>Accelerating materials science</strong></p><p>The researchers then showed their model could be given a natural language query describing a property and identify crystalline materials with that property, outperforming a baseline.</p><p>“There are sometimes these step changes within materials science around algorithmic integration toward useful tools, which the community uses to further advance the field,” Olivetti says. “This approach helps us see more deeply into three-dimensional structures and link structures to property and processing. I hope this can be one of those step changes.”</p><p>Li compares the model to AlphaFold, an AI model that links protein structures with function that has helped advance biological research and drug development.</p><p>The researchers believe their new approach can work with many crystalline materials, and they hope other material scientists will use it to accelerate their work.</p><p>“Being able to see a three-dimensional crystal and how its atoms are connected is super foundational,” Li says. “I think this will be useful to anyone who practices materials science.”</p><p>The work was supported, in part, by the U.S. Defense Threat Reduction Agency through the University Research Alliance; the U.S. Defense Advanced Research Projects Agency; and the National Research Foundation of Korea.</p>

  • For example, when the researchers asked the model to identify materials suited for solid-state lithium-ion battery electrolytes, its top results included materials that had been independently identified as promising candidates.
  • </p><p>They trained a language model to create textual descriptions of crystal structures using an approach known as contrastive learning, in which models compare similar and dissimilar examples of structures and textual descriptions to achieve a unified understanding.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

langgraph-cli==0.4.43

<p>Changes since cli==0.4.42</p> <ul> <li>release(cli): 0.4.43 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6352546338" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9682" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9682/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9682">#9682</a>)</li> <li>feat(cli): use Agent Server flags with hidden deployment aliases (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6319418360" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9666" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9666/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9666">#9666</a>)</li> <li>fix(cli): remove discontinued local debugger image (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5076278238" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/8543" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/8543/hovercard" href="https://github.com/langchain-ai/langgraph/pull/8543">#8543</a>)</li> <li>chore(deps): bump the minor-and-patch group across 1 directory with 9 updates (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6143896181" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9532" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9532/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9532">#9532</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.34.0rc2

<p>[Doc][Tests] Stop setting <code>gpu_memory_utilization</code> in most extensible…</p>

Nguyên văn nguồn
Chi tiết nguồn
  • v0.34.0rc2 vllm-project/vllm releases · 12 thg 11, 2026
Kiến trúc

The Jev Revolution: What Engineers Are Already Building with System One Models

Today at the AI Engineer Code Summit , the main theme was system one models — non-generative models that return structured choices, scores, or probabilities. They’re also typically much faster and cheaper than reasoning LLMs (a.k.a. system two models). And who better to give an inspirational rallying cry for system one models than Diogo Almeida , co-founder and CEO of TypeSafe AI . Almeida was lead creator of Jev , the product that kicked off this trend less than two months ago. Jev was an instant revelation for AI engineers when it was released in September. In a podcast with Latent Space on launch (which quickly became our most popular episode of the year), Almeida explained the core problem with making LLMs the heart of software: reliability . In his keynote at AIE Code today, he took it a step further by claiming that Jev is “reliable intelligence” that should be used for “just automating everything.” Integrating system one models into code Of particular interest to AI engineers is Almeida’s point that system one models should become integrated with code. As he put it in his keynote, “code is the language of automation , therefore you need to integrate with code.” Latent Space founder swyx interprets this as a call to “liberally sprinkle ‘System One’ model calls everywhere INSIDE application code to make smarter software instead of increasingly convoluted forms of ‘concat strings together and hope for the best" AI engineering.’” The term “smart software” was used throughout Almeida’s presentation. It sounds like the kind of term used by companies like Microsoft or IBM during Web 2.0, but in the AI era it does actually signify a harsh reality about LLM-driven software: sometimes it’s just not very smart. As Almeida quipped during his presentation, “you get Einstein 95% of the time, but Mr. Bean 5% of the time.” “And if you can't tell which is which, how the fuck are you going to automate anything,” he added. Almeida ended with an impassioned plea to “automate everything.” Like an AI-pilled, pink-haired Steve Jobs, he called out to “the dreamers and the people who don't want to be limited by everything that has happened so far,” promising that TypeSafe AI wants to “really fight for you, and we want this to be as reliable as a fourth logic gate. ” How Glean uses Jev With the unenviable task of immediately following Almeida on stage, Chinmay Goyal from Glean presented a more prosaic version of Jev usage. He explained that system one models are used at Glean alongside LLMs and other types of classifiers, and that Jev is especially suited for “zero-shot flexibility.” “As the taxonomy of the decision space changes, we don't have to fine-tune these models again and again,” he said. In a short interview after his presentation, Goyal told Latent Space that their customers don’t actually need to know which type of model they’re using. “Jev powers a significant number of flows [but] the user just needs to have a product that works; that is always our aim with building anything. So we want to hide away all the complexity of whether we are using Jev or we are using a full-blown LLM.” In the conclusion of his presentation, Goyal offered some rules on when to use Jev: Jev patterns and prototypes Next up at AIE Code were Chibi Vikram and Kevin Bernstein from UiPath , a business orchestration and automation platform. One of their examples showed a test of Jev ranking document chunks . They held a “tournament” where Jev picked the most relevant chunks from small groups, then the winners advanced to another round, and so on. This whole process took under a second , even with thousands of Jev calls. “We realized that Jev is a building block that we can adjust ,” said Vikram about this and other experiments, “and we have to find these new strategies where we will be able to apply this building block. It might look less intelligent on its own, but with these strategies and patterns, we can get more intelligent — with even the current Jev model that we have.” Scott Roberts , a Principal Engineer at Workday , finished the ‘system one’ session at AIE Code with some prototypes his team has run with Jev. In one, an interface starts to fill in a time-off card while the user is still typing. It was an interesting showcase of Almeida’s idea that a fast model call can change what the software itself does. Roberts described it as going “from predicting what you want to say to predicting what you want to do.” Like the early internet The biggest learning I took from the System One presentations today is that a model like Jev can open up surprising new ways to build software . I was sitting next to a conference attendee who heads up AI at his company. He told me that in his spare time, he used Jev to create a website that sorts and ranks news about fantasy football. But the final word has to go to Almeida, who likened the new opportunities opened up by system one models to the experimentation of the early internet. “I think that there's going to be an explosion of what software can do, and we will want people to be exploring those boundaries of software. Not necessarily because it's valuable, but because they can, because we have new frontiers to explore , and this is what the early internet was like.”

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

New research could lead to faster, trustworthy AI models for climate prediction

<p>Accurate, up-to-date climate projections are essential for planning decisions that enhance the resilience of power grids, protect public health, maximize agricultural yields, and more.</p><p>Now, MIT researchers have developed a technique that streamlines the production of fast, trustworthy, and adaptable artificial intelligence models for climate prediction.</p><p>These specialized AI models, known as climate emulators, can rapidly simulate the effects of human activities onto a future climate, predicting impacts on temperatures or sea level, for example.&nbsp;</p><p>Because emulators run in a matter of seconds, scientists use them to efficiently test thousands of varied climate scenarios. This provides more reliable estimates and can identify unexpected outcomes of regional interventions.</p><p>The MIT researchers developed a training technique that boosts the performance and adaptability of a climate emulator while using only a fraction of the data typically required. This significantly cuts production time for these practical scientific tools.</p><p>They show that an emulator trained with a much smaller amount of carefully selected synthetic data accurately predicts outcomes of scenarios it did not see during training, while traditional emulators failed to adapt to new situations.</p><p>“Climate emulators are especially well-suited for modeling policy-relevant scenarios. By using our method to generate optimized training data for an emulator, we can speed up the production of much more reliable emulators, so we can better trust their outputs for making policy decisions,” says Christoher Womack PhD ’26, who is now a postdoc at ETH Zurich and lead author of a <a href="https://www.pnas.org/doi/10.1073/pnas.2620731123" target="_blank">paper on this technique</a>.</p><p>He is joined on the paper by Shahine Bouabid, a postdoc in the MIT Department of Earth, Atmospheric and Planetary Sciences (EAPS); Andrei Sokolov, a research scientist in MIT’s Center for Sustainability Science and Strategy (CS3); Popat Salunke, a postdoc in CS3; Glenn Flierl, a professor of oceanography in EAPS; Sebastian D. Eastham, an associate professor at Imperial College London; and senior author Noelle E. Selin,&nbsp;a professor in the MIT Institute for Data, Systems, and Society (IDSS) and EAPS, and director of CS3. The research <a href="https://www.pnas.org/doi/10.1073/pnas.2620731123" target="_blank">appears this week</a> in the <em>Proceedings of the National Academy of Sciences.&nbsp;</em></p><p><strong>Speedier surrogates</strong></p><p>Climate scientists typically use earth system models (ESMs) to make long-term projections about how global temperature or precipitation could change in response to varying greenhouse gas emissions.</p><p>These enormous, physics-based models capture the major components and interactions of the earth’s climate in extreme detail. While powerful, ESMs can take months to run on a cluster of supercomputers — and that’s only to project one possible future climate scenario.&nbsp;</p><p>Selin and her collaborators design climate emulators, which are fast AI surrogate models that mimic the behavior of full-scale ESMs. Emulators help scientists test many scenarios, make more accurate forecasts, and explore specific pollutants and regional effects.</p><p>“Climate emulators can run in just a few seconds, so it really expands what we can do in terms of exploring these future climate scenarios,” Womack says.</p><p>When scientists design a climate emulator, they often train the AI model using a set of six to eight future scenarios based on realistic future trajectories of emissions. They must run each scenario through a full-scale ESM, making these datasets costly and time-consuming to generate.&nbsp;</p><p>The MIT researchers recently found that using realistic future scenarios as training data can lead to underperforming emulators that fail to predict unfamiliar or novel scenarios.&nbsp;</p><p>In this work, they developed an algorithm that generates a synthetic dataset of optimized climate scenarios tailored to boost one emulator’s predictive performance.</p><p>“Our algorithm might generate some weird scenarios which look nothing like a realistic future, but we don’t care so much if the training scenarios are realistic. We are just trying to make the best possible dataset to train an emulator,” Womack says.</p><p>Importantly, rather than needing to run a full ESM six or eight times to generate training data, a user would only need to run that model once.</p><p><strong>Tuning the training data&nbsp;</strong></p><p>The researchers begin by generating an arbitrary climate scenario using a fast and simple climate model. They test how well their emulator can reproduce the results of that ESM, which gives them a measure of error.</p><p>Because of the programming technique they used to build the training pipeline, they can iteratively update the training data to increase the emulator’s predictive skill. For instance, the algorithm could increase CO<sub>2</sub> emissions dramatically in the year 2050 or rapidly decrease methane emissions in 2037.</p><p>Then the optimization updates the training data until the emulator’s error level reaches a certain threshold.</p><p>In the end, rather than resembling realistic scenarios in which emissions follow smooth trajectories, the synthetic data contain rapid variations.</p><p>“By introducing these high-frequency wiggles and oscillations, our data enable the emulator to extract a lot more information about the overall system,” Womack explains.</p><p>Their training algorithm outputs these emissions scenarios, which they can use to train an emulator that is more accurate and adaptable.</p><p>“In the future, we are definitely going to see climates and scenarios we haven’t seen before. If we can simulate these different scenarios more robustly, we can provide more reliable information to the stakeholders who need it,” Selin says.</p><p>For instance, an emulator

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Meta shakes up the frontiers, open and closed

Câu hỏi Is Meta a tier 1 (OpenAI/Anthropic) or tier 2 (Gemini/xAI) frontier lab?

This has been a week where the balance of American open-weight models took a step change back toward the frontier. Meta took the lead of this charge with Muse Spark Open , which is now the strongest American-made open model by quite some margin, but the addition of Reflection’s first model Beam is also relevant. Without these two models, there would be 20+ open-weight models built and released by Chinese companies better than the best Western models. At the time of writing this, Mistral has also announced a strong model, Mistral Large 4 , with an index score of 38 (see below), but its weights are not yet released. There are rumors of a few more strong American open-weight models coming from other organizations too! In many ways, the gap between the best American and Chinese models has been reduced roughly by half. We’ve gone from the best American open models being largely irrelevant to consumers who require the cutting edge of performance, to them being behind but relevant. This paints the picture of the cut-throat, competitive nature of the frontier model business. Reflection and Mistral announced their model scores and Meta dropped the weights of a smarter model before these other two labs could get their own weights out. All of these models are competing with Chinese models released months ago, where the industry is collectively asking — where have Kimi and Zhipu AI been since the releases of GLM-5.3 and Kimi K3 on Aug. 18th and Jul. 16th, respectively? That is a very nuanced question, which I’ll dig into another time. Today’s piece is focused on Meta’s turnaround, what Muse Spark Open means for the ecosystem, and the balance of power across Western frontier labs. Share Why Muse Spark Open should take off — and why it may not One thing we see again and again at Interconnects through our open model analysis is that Western open-weight models get astonishingly more adoption than Chinese models with the same performance-to-size footprint (even though the Chinese models usually come out first)! Muse Spark Open is landing into this environment, at the same time as when open-ended LLM agents are finally real (~ GLM 5.2 capabilities and on ). Riding off a rapid increase in performance of Chinese models, the open ecosystem has been living through an acceleration in demand. The inference providers are rapidly scaling, the open post-training economy is forming, and countless enterprises are looking to own their intelligence. Muse Spark Open’s performance falls somewhere above GLM 5.2, near DeepSeek V4.1 Flash, and behind GLM 5.3 or Kimi K3. It should be squarely “good enough” to slot right into this. For some background, from the Muse Spark Open model card on Hugging Face: Muse Spark Open is an 816B total/42B active parameter open weights model with a context window of 1M tokens. Built for coding and agentic work, Muse Spark Open is natively multimodal. This open weights version has additional safety post-training to make it more robust against misuse as a standalone model without system guardrails. We recommend deploying Muse Spark Open as part of a broader AI system, with additional guardrails suited to the specific use case. Here’s the distribution of the various Muse models on the Artificial Analysis Intelligence Index. This is included to loop back to the original model card snippet. It was shared that this open-weight model underwent additional safety training to prepare for this open release. Many observers saw this release as potentially a model that was “broken” during the safety training, as it came out after Muse Spark 1.3. In reality, this was much more likely a safety tune of Muse Spark 1.2, which would only be a 1 point drop on Artificial Analysis to shore up safety behaviors. When Meta released Muse Glimmer , the 30B agentic model, Alex Wang said that they were preparing to release the Muse Spark 1.2 model as open weights, which fits with the above. If anything will hold Muse Spark Open back, it’ll be the specific license they designed for it. In summary, this custom license has pretty much all of the restrictive license behaviors we’ve seen (and criticized here) in the last few years of building with open models. They are as follows: Users of the models in products must display “Built with Muse” on their products. Meta did this with Llama, and Kimi has tried it, but almost no one complies. There is a revenue-gating factor, where companies that make more than $20M per month across their entire business deploying the model need a separate license with Meta. Other prominent open model providers are doing this too, in order to set up revenue share agreements for re-selling inference. Muse Spark Open makes you agree to an acceptable use policy, which in the true “open source software” corner of the world makes this model extremely not open-source. All of the above applies to outputs from the model as well, and derivatives, so there is legal ambiguity if you can safely use the model for synthetic data. The license does permit using Muse Spark Open for internal research and applications. We’re at a weird equilibrium where this license is complex and unproven in court, but many players will likely ignore it. Many, many people will build on this model from various inference providers, and Meta will only likely go after egregious violations. For now, licenses haven’t been the determining factor of AI adoption, but this could change in the future. This is why we think it’s an unnecessary strategic risk for labs trying to compete and regain adoption to have complex, non-standard open licenses for their models. Interconnects AI is a reader-supported publication. Consider becoming a subscriber. Is Meta a tier 1 (OpenAI/Anthropic) or tier 2 (Gemini/xAI) frontier lab? Just a few days before Muse Spark Open, Meta unveiled their closed frontier models, Muse Wonder and Muse Wonder Full . 1 On traditional benchmarks, these are very strong models only behind Anthropic’s Opus 5.5 and similar to Open

  • In many ways, the gap between the best American and Chinese models has been reduced roughly by half.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v18.19.0

<h2>@oh-my-pi/pi-ai</h2> <h3>Added</h3> <ul> <li>Added the <code>liquid-systemone</code> judgment API for Liquid AI decision models, including validation that rejects image sets exceeding Liquid’s 10,000 32×32-patch request limit before submission.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed Gemini CLI and Antigravity responses losing or misordering thought signatures and reply text, including responses that follow tool calls.</li> <li>Fixed Gemini CLI and Antigravity merging reply text around a thought signature, dropping signatures at the start of a reply or after a tool call, and placing a thought ahead of reply text that streamed before it (<a href="https://github.com/can1357/oh-my-pi/pull/18942" data-hovercard-type="pull_request" data-hovercard-url="/can1357/oh-my-pi/pull/18942/hovercard">#18942</a> by <a href="https://github.com/JustasMasiulis">@JustasMasiulis</a>)</li> <li>Fixed Google Antigravity judging Gemini models by quota buckets they do not consume: Gemini 3.x requests ignore the Gemini 4 Argon weekly bucket, and Argon requests ignore <code>gemini-weekly</code>, so an exhausted or low bucket no longer blocks, reserves, or caps an account for the other models (<a href="https://github.com/can1357/oh-my-pi/issues/16104" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/16104/hovercard">#16104</a>).</li> </ul> <h2>@oh-my-pi/pi-catalog</h2> <h3>Added</h3> <ul> <li>Added Liquid AI provider support with the d1.1 and d1-ultra judge models and Liquid authentication via <code>/login liquid</code> using <code>LIQUID_API_KEY</code>.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed Google Antigravity Gemini 4 Argon models sharing the general Gemini quota scope instead of their dedicated weekly bucket (<a href="https://github.com/can1357/oh-my-pi/issues/16104" data-hovercard-type="issue" data-hovercard-url="/can1357/oh-my-pi/issues/16104/hovercard">#16104</a>).</li> </ul> <h2>@oh-my-pi/pi-coding-agent</h2> <h3>Breaking Changes</h3> <ul> <li>Changed the browser <code>open</code> API: <code>ignore_https_errors</code> is now <code>insecure</code>, <code>allow_file_access</code> has been removed, and <code>file:</code> pages can access other local files.</li> <li>Changed the browser <code>open</code> and <code>targets</code> <code>app</code> option from an object to a string or CDP endpoint/executable path. Use <code>argv</code> instead of <code>app.args</code> and the top-level <code>target</code> instead of <code>app.target</code>.</li> <li>Removed <code>task.file</code> and the task tool's <code>taskFile</code> field; subagent instructions must now be provided inline.</li> <li>Removed blocked todo tasks and the <code>blocked</code> status/blocker metadata. Existing blocked tasks are restored as pending.</li> </ul> <h3>Added</h3> <ul> <li>Added <code>systemPromptTemplates</code>, which maps model-selector patterns to Handlebars template files for customizing system prompts. Templates re-render when switching models, with warnings and fallback to the bundled prompt for missing or invalid templates.</li> <li>Added ordered multi-step tasks through <code>task.steps</code>. Subagents receive steps sequentially, and the final step determines the task result; interrupted tasks resume with their remaining steps.</li> <li>Added Liquid AI judge models <code>liquid/d1.1</code> and <code>liquid/d1-ultra</code>, including image resizing for Liquid's request limits.</li> <li>Added per-step progress, state, and duration reporting for GitHub Actions watches, with expandable details and native Tern Surface Protocol views.</li> <li>Added <code>task.autoParallel</code> (<code>off</code>, <code>suggest</code>, or <code>on</code>) to identify subtasks that can run concurrently and optionally delegate them in a single parallel task call.</li> <li>Added portable <code>awk</code>, <code>tar</code>, <code>lsof</code>, <code>shasum</code>, <code>md5</code>, and <code>strings</code> support to the <code>bash</code> tool, including Windows support and internal-URI paths.</li> </ul> <h3>Changed</h3> <ul> <li>Browser sessions now prefer the bundled ABKit WebKit engine when available, with automatic installation and SHA-256 verification. WebKit provides headed, user-visible tabs and broader browser compatibility; Chromium remains the fallback and can be selected explicitly with <code>app: "chromium"</code>, while <code>app: "webkit"</code> requires WebKit. Configure this behavior with <code>browser.webkit</code> and <code>browser.webkitPath</code> or the corresponding environment variables.</li> <li>Browser visibility controls now show or hide WebKit tab windows immediately. Headed WebKit tabs can be interacted with directly, while Chromium tabs retain their existing mode-switching behavior.</li> <li>In Tern environments with native DOM access, browser observation and interaction use trusted native input across frames and closed shadow roots, with caret-aware typing and case-insensitive text selectors.</li> <li><code>edit.modelVariants</code> now uses the longest matching model pattern, so specific patterns take precedence over generic ones regardless of configuration order.</li> <li>The <code>ask</code> tool now submits the form's current values when its timeout expires; updated settings and documentation reflect this behavior.</li> <li>The <code>read</code> tool loads detailed format selectors and URI sub-path syntax on demand, reducing the default request description while retaining inline details when <code>read</code> is disabled.</li> <li>The default system prompt and tool descriptions have been streamlined, with guidance moved closer to the tools and detailed task/procedure references provided when relevant.</li> <li>The <code>bash</code> pseudo view renders words built from literals and expansions as f-strings (<code>f"repos/org/{r}/rulesets"</code>) instead of <code>concat(...)</code> calls.</li> </ul> <h3>Fixed</h3> <ul> <li>Fixed <code>computer.close()</code> so subsequent computer-control calls s

Nguyên văn nguồn
Chi tiết nguồn
  • v18.19.0 Can Bölük — Oh My Pi releases · 12 thg 11, 2026
Kỹ thuật

DeepSeek-V4.1-Flash on Vera Rubin NVL72: 58x Throughput vs. GB300 NVL72 at 387 tokens/s

On one NVIDIA Vera Rubin NVL72 rack, vLLM serves DeepSeek-V4.1-Flash on SemiAnalysis AgentX at 1.7M tokens/s per GPU at a P90 of 387 tokens/s per user, 58x GB300 NVL72 at the same interactivity.

Nguyên văn nguồn
Chi tiết nguồn
361–380 / 680
Trang 19 / 34