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

Đánh giá

[AINews] The No AI Slop Manifesto

Congrats to both llama.ccp and ggml for reaching 1.0 (declaring stability and removing 35 deprecated functions from llama.cpp’s public header) and with SGLang also announcing 1.0 at the recent SGLang Summit . However, today’s title story goes to the No AI Slop manifesto , which arose out of the conference discussions at AIE CODE and which Dex introduced in the closing keynote after cosignatures by many of the speakers at the conference. It seems Hacker News also liked it , voting it up to the front page today. We started this manifesto to offer a useful tool for discussion at work, and hope it gets used that way. Full manifesto, with cosignatures and edits and discussions open at https://www.slopmanifesto.com/ Sponsored by Deepgram Deepgram is the voice AI infrastructure behind natural conversation—bringing speech-to-text, text-to-speech, voice-focused LLMs, and agents to teams building with voice AI. For any stack and any tools you already love building with, Deepgram slots in as your foundational voice AI technology. So you have control. And we have the context to listen, understand, and speak like a human. Learn more about the Real-time Context Standard . Sponsored by Backblaze B2 Growth shouldn’t make you sweat. Backblaze B2 delivers capacity storage that scales with your business and keeps flash and GPUs fed with terabit-scale throughput. Put B2 to the test ! contact business@latent.space for further enquiries. Thank to Deepgram and Backblaze for supporting AINews! AI Twitter Recap Local inference and agent infrastructure llama.cpp 1.0. Users running mixture-of-experts models larger than their GPU memory can now set aside a cache for frequently used experts. Georgi Gerganov · Daniel Han Chen’s example LangChain’s Series C. CapitalG led the round. LangSmith lets teams trace and evaluate production agents across frameworks. The company plans to announce new LangSmith features on November 20. LangChain · CapitalG Frontier and Open Model Launches Gemini 4 Hydrogen in GitHub Copilot. GitHub made the model generally available in Copilot across its app, CLI and VS Code. GitHub Independent checks : Roboflow's vision bench shows high effort at 29.3s per image, against 51.4s for Argon and 132.3s for Carbon, though Carbon has better detection accuracy ( @skalskip92 ). One 3D-modeling test found Hydrogen more expensive than Argon because it loops on over-verification ( @dzhng ). Google's week : Google also shipped two 8th-gen TPUs, Gemini 4 Carbon, Gemini Robotics 2.5, AlphaGenome 2 and a Decisions API ( @sundarpichai ). MathArena : Gemini 4 Carbon took #1, displacing GPT-6.1 Sol ( @j_dekoninck ). Claude Fable 5.2 in Agent Arena. Arena published new results for long coding sessions. Pricing : API pricing fell from $10/$50 to $8/$40 per 1M input/output tokens ( @arena ). Agent Arena : The model ranks #1 at +30.4% net improvement, about 1.6x Opus 5.5. Its lead grows on long tasks, reaching 93% vs 85% no-correction responses at 16+ turns ( @arena , turns ). Thinking Machines Inkling 2 Small : The model's open weights are now on the Hub ( @ariG23498 ). It is the #1 US open model in Code Arena WebDev at 1647 ( @arena ). NVFP4 training : It uses a native 4-bit forward pass from the first pretraining step through the final RL rollout ( blog ). The mixed-precision recipe is detailed by @jackcookjack . NVFP4 is stored in 3.6x fewer bits than BF16 and runs up to 6x faster on GB300 ( quote ). Inference detail : A width-4 short convolution requires a rolling 3-token cache per layer alongside the KV cache ( @andrew_n_carr ). ERNIE 5.5 Flash Preview (Baidu) : The model has 280B total and 14B active parameters, a 1M-token context and free OpenRouter access. Open weights are promised before Nov 25 ( @omarsar0 ), with Apache 2.0 cited by Arena. Other releases : Ripple-65B Preview : The dense model claims US open-weight SOTA on agentic coding ( @RicursiveAI ). MiniMax H3-Image : It is the #1 open-weights model on Artificial Analysis image editing and T2I ( @ArtificialAnlys ). DeepSeek API : DeepSeek exposed its app features via API, including search, files, OCR, ASR and TTS, with separate tool fees and an open-source frontend ( @deepseek_ai , pricing ). ChatGPT Wallet : The feature lets Dot agents buy items without seeing card numbers, via Amex, Mastercard and Visa ( @TheRundownAI ). Independent Evals, Benchmarks and Decision Models AA Coding Agent Index v2.1 : The update moves to FrontierCode 2.0 and Terminal-Bench 4.1 ( @ArtificialAnlys ). Leaders : Fable 5.2 and Opus 5.5 tie at 58. Average scores fell 5 points; GPT-6.1 Terra in Codex is the only configuration that did not drop. Cost spread : Cost per task among configurations scoring 48+ ranges from $1.51 for Codex with GPT-6.1 Sol to $44.22 for Fable 5.2 at max. Fable at medium, high and xhigh all lie on the Pareto frontier ( costs ). Cyber and agent benchmarks : Tencent Hy4 : The model scores 58% on the AA Cyber Index, ranking #3, and 76% on CWE-Bench-AA, ranking #1. It costs $0.92 per task ( @ArtificialAnlys ). Agent Initiative Benchmark : GPT-6.1 Sol beats Fable 5.2 at $0.11 vs $1.87 per task. Models that intervene unnecessarily spend about 80% more time and money ( @jakubjurovych ). Reasoning effort : On the HF Harness Leaderboard, high effort makes DeepSeek-V4.1-Flash 28% more expensive with no gain within noise. On GLM-5.3-Flash, high effort adds 5.8pp ( @joelniklaus ). Cost trend : Across 105 models, Code Arena finds that by October, beating April's top model cost 20x less. MiMo-V2.6-Flash outscored Opus 4.7 ( @arena ). Decision-model wave : More than a dozen vendors shipped decision models from Sep 23 to Nov 11, and Vercel reports that 26% of requests now route through them ( @rasbt ). Leaderboard : SubQ Snap 1.2 ties Drex Max v1.1 and GLiDE v1.2 at 73 on the AA Decision Index. Its gains come from abstention, not accuracy ( @ArtificialAnlys ). Usage : Perplexity reports about 1B decision-model requests per month on OpenRouter ( @AravSrinivas ). Eval

  • Arena published new results for long coding sessions.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Run Native DLA Inference in Python on NVIDIA Jetson AGX Orin Using cuDLA

AI applications at the edge often need to run multiple neural networks within limited compute and power budgets. NVIDIA Deep Learning Accelerator (DLA)... AI applications at the edge often need to run multiple neural networks within limited compute and power budgets. NVIDIA Deep Learning Accelerator (DLA) platforms provide dedicated hardware for deep learning inference. On NVIDIA Jetson AGX Orin, DLA executes supported neural network operations, allowing applications to offload inference work from the GPU. This can improve power efficiency and leave… Source

  • This can improve power efficiency and leave… Source
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

AI Is Helping Care Teams Get Patients the Right Care, Sooner

From pediatric research to smart hospitals, AI is helping care teams spend less time on routine tasks and more time with patients. Health systems and startups are putting it to work across healthcare, from matching patients to clinical trials to supporting radiologists, hospital teams and patients at home. Clinicians stay at the center, reviewing AI outputs and guiding decisions, while NVIDIA technologies provide the compute behind the applications. Read on to learn how healthcare teams are using AI to improve and speed patient care. Tuesday, Nov. 17, 8:00 a.m. PT Proximie Helps Surgical Teams Spend Less Time on Logistics and More Time With Patients As AI advances healthcare, surgery is a high-stakes, highly regulated setting for its use — and one with significant opportunity for efficiency gains. Only about 10% of time spent in the operating room is clinical. The rest is devoted to scheduling, turnover and documentation. With the help of NVIDIA technology, Proximie is introducing AI into operating room (OR) logistics so that clinicians can focus on taking care of patients. A New Vantage Point Proximie is using NVIDIA Cosmos 3 , the frontier open world foundation model family for physical AI, to power a platform used in more than 700 ORs today. Trained on a decade of data, it captures real-time audio, video and device signals from a close-up view of the surgery as well as a wide view of the whole room. The comprehensive view offers context held by no single person in the room, whether the surgeon, scrub nurse or coordinator. Agentic notifications and voice AI tools can call the next patient when a room is ready, document alerts and capture practitioner notes. To protect privacy, Proximie’s models blur identifying details, such as faces, tattoos and onscreen text before footage leaves the OR. Surgeons control the blurring to ensure that recordings used for training still show what trainees need to see. Across health systems where Proximie’s technology has been applied, the platform has enabled clinical teams to handle 8% more cases, turn over the OR 45% faster and start procedures on time 26% more often. Next, the company hopes to turn its technology into an OR copilot that assists with staging equipment ahead of time, estimating time remaining and running implant safety checks. Proximie will also use NVIDIA Cosmos 3 to generate realistic synthetic scenes grounded in real ORs, so models can learn from the rarest, most critical scenarios. The goal is to create a room that runs itself, so practitioners can spend more time with patients and dedicate less time to documentation and preparation. Proximie is working with UC San Diego Health to build an AI model that learns from many kinds of surgical data at once using NVIDIA open models. Before being trained on Proximie’s data, Cosmos 3 could already identify 80% of key equipment in the OR with high accuracy. Training it further on Proximie’s footage made the model more accurate and smaller, so several models can run at once on a single computer in the OR alongside tools to protect patient privacy. Proximie is also among the first participants in Open-H-Hospital , an NVIDIA-led, community-driven open initiative that brings real hospital operation data together with simulation data to train the next generation of hospital AI and robotics. Learn more about Proximie. Tuesday, Nov. 17, 8:00 a.m. PT ConcertAI Matches Cancer Patients to Clinical Trials Faster Interest in clinical trials is high among people with cancer. More than half of those offered a trial say yes, but only about 7% of U.S. patients with cancer take part in one. The issue isn’t the availability of trials, but the time it takes to match patients to them. Care teams have to manually review months or years of medical records and pull together information like physician notes, pathology reports and genomic results from different systems. Built on NVIDIA AI, ConcertAI’s Cara AI engine and TriaLinQ module are speeding up that process, identifying eligible patients roughly 3x faster than manual methods. They assess trial eligibility across the CancerLinQ oncology network, ConcertAI’s data network that works across nearly 1,000 sites of care and over 13 million patient records. From Scattered Records to One View Cara pulls all relevant patient data into a single view for each person and updates it weekly with disease stage, biomarker status, treatment response and line of therapy. TriaLinQ, a module of the CancerLinQ suite, uses that information to assess a patient’s eligibility for a trial. “We use a chain of AI models and agents to parse all the information, understand really challenging concepts like progression or timelines within a diagnosis journey, and summarize it into a structured data model that can be queried from our software stack,” said Eron Kelly, CEO of ConcertAI. ConcertAI uses healthcare-specific NVIDIA microservices, software libraries and AI agent tools to accelerate its work. Clinicians Stay in Charge ConcertAI brings new technology to care teams, but treatment decisions stay with clinicians. “The promise of technology in oncology isn’t measured by how sophisticated the platform is, but instead by whether it meaningfully supports the clinicians delivering care and positively impacts patient outcomes,” said Dr. Shaalan Beg, chief medical officer of oncology at ConcertAI. “That’s the standard we hold ourselves to: AI that earns its place at the bedside.” Every eligibility finding links back to its source in the patient’s record, so oncologists and care teams can make fully informed treatment decisions. By integrating into the way care teams and clinicians already work, ConcertAI’s systems support up to 50% higher trial enrollment, helping more patients find treatment options when they need them. Learn more about ConcertAI. Monday, Nov. 16, 8:00 a.m. PT Artisight Helps Nurses Prevent Patient Falls With In-Room AI Device Roughly 700,000 to 1 million patients fall in U.S.

  • Read on to learn how healthcare teams are using AI to improve and speed patient care.
  • Care teams have to manually review months or years of medical records and pull together information like physician notes, pathology reports and genomic results from different systems.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Jensen Huang resigns after 33 years leading NVIDIA

Congrats to Gemini 4 Hydrogen on their launch and it seems Anthropic’s IPO is well under way, but today’s headline was only ever going to be about one thing, the most earthshaking leadership change in AI (after Demis/Jeff/et al at Google DeepMind). Yesterday we covered reports that NVIDIA’s board was preparing to replace Jensen Huang following an NYT report of an affair with an accused Chinese spy . Following news of a second affair , today NVIDIA announced his resignation (correlation/causation is debatable here), presumably forced by the board: View post on X ↗ There is also some discussion of the extent that Chinese models can demonstrably be shown to be trained on Blackwell GPUs that they should not have access to… though the scale is debated . After 33 years, Jensen Huang has stepped down as CEO, president and boardmember. Colette Kress, NVIDIA’s CFO since 2013, will keep that job while serving as interim CEO… leading into what will presumably the best attended earnings call in the history of $NVDA next week. AI News for 11/11/2026-11/12/2026. 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 Gemini 4 Hydrogen: Google's Fast Gemini 4 Tier Ships Broadly Launch : Google DeepMind released Gemini 4 Hydrogen, a speed-first model positioned alongside Carbon and Argon. Google says it completes tasks 2.5x faster than Gemini 3.8 Flash and edges Argon on the Artificial Analysis Intelligence Index, 51 vs 50 at default thinking and 54 at the new xhigh level ( DeepMind , Koray Kavukcuoglu ). Specs and pricing : The model offers four thinking levels (low, medium, high and xhigh) and a 1M context window. It costs $1.50/$7.50 per million input/output tokens, with a 95% discount on cached input ( Phil Schmid , Artificial Analysis ). Official claims : Google reports 57% on Terminal-Bench 4.0 and more than 3x Gemini 3.8 Flash on Terminal-Bench 4.1. It also claims more than 5x fewer fabricated claims, #1 on Vals Tax Agent Bench and leading results on AutomationBench-AA and Agents' Last Exam ( Google ). Availability : Hydrogen is available on all Gemini API tiers, including a free tier in AI Studio. It is the default model in Search AI Mode, and free Gemini app users get it through auto-selection. Paid Pro and Ultra subscribers can use it in Antigravity ( Sundar Pichai , GeminiApp , Search ). Orchestrator multi-model mode : In Antigravity, Carbon plans while Hydrogen writes and tests code. Google says this reaches near-Carbon success on Terminal-Bench 4.0 with 24% less quota and 19% faster completion. It is rolling out to Ultra subscribers ( Google , Mohan ). Artificial Analysis measurements : The independent results show a mixed profile. Intelligence and cost : At xhigh, Hydrogen scores 54, one point behind Claude Sonnet 5.5 (max) and GPT-6.1 Sol (max). Its cost is about 32% of Sonnet's but about 2.5x Sol's. Cost per task is $3.16 at xhigh and $1.87 at high. Output speed is about 200 tok/s ( AA ). Strengths : Hydrogen has an 8% hallucination rate on AA-Omniscience, versus 43% for Sonnet and 52% for Sol. It also shares the #1 score on GDP.pdf visual reasoning. Weaknesses : It scores 54% on Terminal-Bench 4.0 at both high and xhigh, below Sonnet at 64% and Sol at 56%. Third-party leaderboards : Agent Arena : Hydrogen debuts at #5 with +14.8% net improvement across 6.5k sessions. It is the fastest model in the top 10 at 5.3 minutes per task median, with a median cost of $1.6 per task ( Arena ). Code Arena WebDev : It ranks #10 with 1703 points ( Arena ). Vals Index : It is #6 at 68.5%, 2 points behind Carbon at 38% of Carbon's price ( Vals ). Coding agents : CursorBench rises to 54.3% from 39.6% for 3.8 Flash ( Cursor ). Cognition's FrontierCode 2.0 shows 38.2%, just behind Sol ( Cognition ). SWE-sweep shows 13.5%, versus 9.5% for Astra ( KLieret ). Caveats from users : Runaway reasoning : One tester saw xhigh spend 131K thinking tokens and about $1.08 without answering a jersey-number image question. The same tester reports budget exhaustion on 1.8% of a text benchmark ( skalskip92 ). Weak spots : Hydrogen scored 1% on MazeBench ( patience_cave ). Pricing structure : Simon Willison says Hydrogen is only a 25% discount on the Carbon/Argon token price, though those models reportedly double in price on Dec 1 ( simonw ). Speed : Theo disputes the "built for speed" framing based on his own timings ( theo ). Open-Weight Releases: Inkling 2 Small Leads Inkling 2 Small : Thinking Machines released the weights under Apache 2.0 ( thinkymachines , Mira Murati ). Size : It is a 594B-total / 17B-active MoE with text and image input and up to 1M context ( HF ). Architecture : New components include attention residuals with a q/k/v split, gated attention, a dynamic attention sink, dynamic shortconv and value norm ( Songlin Yang ). Experts use NVFP4 in layers 0–39 and MXFP8 in the last eight, with SWA widened to 1k and softcaps on the LM head and SwiGLU ( eliebakouch ). Day-0 serving : On vLLM, it fits on one GB200/GB300 node and reaches up to 630 tok/s per user at TP4. Its native DSpark speculator has an acceptance length of about 3.5 at T=1.0 ( vLLM , details ). SGLang reports 576 tok/s per user at batch 1 on Vera Rubin and text-plus-vision RL on 8 Blackwells ( SGLang ). SkyRL validates MXFP8, BF16 and LoRA RL ( erictang ). Evals : It is the #1 US open model on Code Arena WebDev at 1647 points ( Arena ). On Agent Arena it is #2 among open models, with a median cost of $0.26 per task ( Arena ). It also holds the top open-model spot on MathArena's BrokenArXiv ( MathArena ). Edge0-145B/120B : These are device-native open-weight models. The developer claims 86.5% on Terminal-Bench 2.1, versus 78.2% for Claude Opus 4.6, in about 29GB of memory at 75+ tok/s decode ( Samuel Zeng ). These figures are vendor claims. Perplexity Decider v1.3 : The open-w

  • It also claims more than 5x fewer fabricated claims, #1 on Vals Tax Agent Bench and leading results on AutomationBench-AA and Agents' Last Exam ( Google ).
  • Artificial Analysis measurements : The independent results show a mixed profile.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

New AI “copilot” helps engineers design better structures

<p>The use of algorithms to design structures has led to new parts optimized for ultralight aircraft, high-bandwidth antennas, and F1 racing cars.</p><p>But the best algorithmic design techniques still require lots of computation and time: Users only see the final design after it’s complete, after which they often must start the design process all over again to make changes. That process also makes it harder to trust that the final designs are structurally sound.</p><p>MIT researchers are making algorithmic designs more collaborative with a system that acts like an AI copilot. As the algorithm is generating the structure, it presents regions to the human designer for feedback.</p><p>In an open-access <a href="https://www.nature.com/articles/s44172-026-00818-x" target="_blank"><em>Communications Engineering paper</em></a> describing the system, the researchers showed it led to more easily manufacturable structures with higher buckling loads while adding only 15 seconds to the total design time compared to all-algorithmic approaches.</p><p>The researchers believe the approach, which uses a design technique known as topology optimization, could make algorithmic design more accessible and manufacturable for a broader class of design engineers.</p><p>“Right now, the human doesn’t have any agency or control over the design process as it’s playing out,” explains lead author Dat Ha PhD ’26, a recent MIT civil and environmental engineering doctoral student who is now a structural designer at LeMessurier. “For these really complex designs, we wanted to also tap into our brain, which is the most complex machine there is, to solve these design problems in a way that’s more practical to industry.”</p><p>Joining Ha on the paper are Ferdous Alam, a former MIT postdoc who is now an assistant professor at Georgia Tech; Markus Buehler, the MIT Jerry McAfee (1940) Professor of Engineering; Faez Ahmed, MIT associate professor of mechanical engineering; and senior author Josephine Carstensen, MIT associate professor of civil and environmental engineering.</p><p><strong>Bringing a human in the loop</strong></p><p>Topology optimization is one of the most advanced and promising algorithmic design techniques. It uses formal optimization techniques to iteratively fill in the space given for a design in a way that maximally achieves the stated goals. For most applications, the process requires a lot of computation and time. When it’s finally done, the designs can be impractical to build or have other issues the engineer didn’t consider. For instance, some designs might have eight or more beam connections at a single point, which would be difficult to make in the real world.</p><p>“Engineers are always very skeptical of computer programs, especially if they give you very weird-looking or unexpected designs,” Ha says. “On paper, topology optimization is really powerful and can achieve great, high-performing designs, but its black box nature has made it less widely adopted in industry.”</p><p>Topology optimization has gotten traction in the automotive and aerospace industries, but it’s been used less to design buildings or bridges, partly because design engineers need to know they can trust the output.</p><p>“A big limiting factor has been that the design engineer has very little influence over how the design evolves and therefore can’t control undesirable features,” Carstensen explains. “The lack of control and lack of ownership also inhibits its use.”</p><p>To address those problems, the researchers made topology optimization more collaborative in an approach that Carstensen compares to having a design copilot.</p><p>Prior attempts at adding human judgment slowed things down dramatically. The researchers’ new system uses machine learning to quickly identify the most likely regions of an algorithmic design that a human would want to change. To train the machine-learning model, the researchers created a synthetic dataset of probable human feedback to thousands of designs that critiqued things like aesthetics, performance criteria, and manufacturability.</p><p>“We thought if we could create a record of where humans have made changes to these programs, then perhaps there’s a pattern to those choices, which is not very different from other types of machine-learning approaches,” Ha says.</p><p>Carstensen sees the system as bringing together the best of algorithmic and human design approaches.</p><p>“The idea is to tap into the brains of the designers and get them to infuse their domain knowledge as the design evolves,” Carstensen says.</p><p>In demonstrations, the researchers showed their human-in-the-loop approach improved manufacturability and linear buckling load by 39 percent while only modestly increasing the total design time.</p><p>Right now, the approach only allows users to make structures within identified regions thicker or thinner, but they hope to add other options for modification in future versions.</p><p>“The hope is that by addressing these hurdles, we open up topology optimization to a much wider audience, not only in industry but also to researchers and students in academia,” Carstensen says.</p><p><strong>From designs to structures</strong></p><p>The researchers see their approach complementing AI only-approaches to design.</p><p>“We are arguing there’s room for both,” Carstensen says. “If you are in a discipline that has a lot of computational resources, you can use the other approach. But if you really need the human input, you can use this collaborative approach.”</p><p>Adding human input also serves to increase trust.</p><p>“In some settings, the humans need to make the final decisions before advancing the final design,” Carstensen says. “By involving humans, we can get higher-performing designs to the real world to create higher-performing structures. Because it’s not just about designing good artifacts; it’s about creating good artifacts that humans will actually accept and make.”</p><p>The work

  • It uses formal optimization techniques to iteratively fill in the space given for a design in a way that maximally achieves the stated goals.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

AI Ambition: ArtCraft Cloned Photoshop and Now Wants to "Rebuild Everything"

Câu hỏi But aren’t there already open source alternatives to Photoshop, like GIMP?

The theme of day two of the AI Engineer Code Summit in San Francisco was increasing ambition using AI . And there have been few projects as ambitious as ArtCraft in the AI era. So far the fledgling company has released 17 open source versions of popular software tools , starting with a Photoshop alternative built from scratch in Rust called PhotoCraft . You could feel the excitement in the room this morning when ArtCraft founder and CEO Brandon Thomas presented a kind of manifesto for AI-based open source at AIE Code. When he mentioned that ArtCraft had cloned Adobe’s Acrobat, with a product called PdfCraft, it got a round of applause from the audience. In wrapping up, his rallying cries of “let's take back the entire internet” and “let's rebuild everything” were greeted with whoops and cheers. Thomas began his presentation by reliving the internet’s early history. In his view, the web started out with the right intentions, but things got screwed up once Web 2.0 hit the mainstream. “We had a bunch of chat applications that were completely open,” he said, regarding the 1990s internet. “So, IRC is like Discord, but you can spin up your own servers that you actually own. You can use whatever programs to automate them and interrogate them without rules.” His point wasn’t necessarily that everything was open source, but that web users could freely build things. With Flash, for example — a commercial animation tool eventually acquired by Adobe — “kids were making all kinds of animations.” When Web 2.0 came along, with its superior broadband and smartphones, “it brought everybody into the fold.” Thomas acknowledged that this widened the scope of internet software, but at the cost of that freedom to build with leading apps. “The diffuse, experimental, early web started to disappear,” he said. “The APIs got taken away from us, and then we started seeing distribution being gatekept by the big providers.” It seems Adobe in particular annoyed Thomas. He described getting billed with an unexpected fee when he cancelled a monthly plan. “It's really frustrating that the web — that was once really, really promising and open and free — has turned into this extractive, owned, rent-seeking world.” This inspired him to start ArtCraft, “with the ethos of ‘let's build everything open source.’” A message even anti-AI people have gotten behind, according to Thomas. In addition to PhotoCraft, they’ve copied other apps in the Adobe suite, cloned Figma, and created a Microsoft-like word processor called WordCraft. There are also independent apps such as SolidCraft, a community-led CAD project distributed through ArtCraft’s launcher. Thomas reiterated that these AI-built applications are 100% open — “not only the desktop apps, but the actual infrastructure as well.” PhotoCraft: The development process ArtCraft is starting with rewriting familiar tools. Thomas’s longer-term vision is for users (such as artists who were previously using Photoshop) to tailor the software to their own workflows. As he framed it, they want to “undo what had happened during Web 2.0 , when everything became owned and consolidated and gatekept, and just give people back their applications.” AI is the means to achieve that, he said. But aren’t there already open source alternatives to Photoshop, like GIMP? “GIMP isn't really used by artists,” said Thomas during his presentation. “We actually observed artists vibe coding their own workflows on top of the software that we gave them.” To begin creating PhotoCraft, the team gave an AI model a detailed prompt outlining Photoshop’s tools, data structures and file importers. After that, they “laid out a large, large prompt to say, let's build out the application.” As part of that process, they made everything MCP and CLI compatible . Within 24 hours, said Thomas, “we were looking at a pretty good approximation of what Photoshop was — and it kind of blew our minds away.” PhotoCraft interface; via ArtCraft website Thomas noted that PhotoCraft is “still not at 100% parity” — he said they’ve just passed 60% — but the team has set an internal goal to reach as close to 100% as possible. ( The project’s GitHub claims about 76% parity for feature breadth and about 46% for professional readiness.) “Right now, we want to fill the muscle memory of all the artists that use Adobe. So we are going for Adobe , their interface and ‘look and feel’ directly.” But after ArtCraft has achieved that, they plan to “build new apps that have new interfaces that are more agentic. ” A big reaction Thomas said he was surprised by the amount of attention PhotoCraft, in particular, has gotten — ranging from the New York Times requesting an interview, to a swelling of its open source community to about 4,000 developers . “I'm still struggling to bear with what has happened. It struck a nerve with people. They felt sort of maligned by Adobe and other tech companies . And what we built has really resonated with them.” After his presentation, Thomas was mobbed by AIE Code attendees. ArtCraft’s apps are being downloaded at the rate of about 1.7 million per day, which Thomas said is still accelerating. In addition to the 17 official apps, there are around 300 being developed by the community, he added. Rust is key, but there have been technical challenges Other than the AI technology, Thomas said that Rust was a crucial part of making the apps performant — he claimed that PhotoCraft opens ten times faster than Photoshop, for example. “Rust was probably essential in this product's success,” he said. “We've seen people do it in Electron before, but for high-quality creative apps, you really need something that's native. ” For those engineers who are not experts in Rust, Thomas has good news for you. “Do you need to know Rust to start using Rust today as an engineer? Definitely not. [...] Your agents will drive it just fine.” However, Thomas admitted that not everything has gone swimmingly. “We've run into a lot of gotchas — like permi

  • ” AI is the means to achieve that, he said.
  • ” But after ArtCraft has achieved that, they plan to “build new apps that have new interfaces that are more agentic.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Microsoft Research Focus – Edition #98

What’s new in research? In this edition of Microsoft Research Focus , based on the Research Focus newsletter (opens in new tab) , host Anson Ho shares some of the latest news and cutting-edge research from across Microsoft—from bringing AI-powered universal translation closer to reality and solving a 31-year-old mathematical challenge to making AI coding agents faster and more efficient. Catch up on the ideas, tools, and discoveries researchers are exploring today and get a quick look at what’s coming next. Related links: A research intern’s journey from computer vision to physical intelligence at Microsoft Research Asia – Tokyo Breaking the Language Barrier — How AI Speech Translation Is Rewriting Our Digital Lives Solving a 31-year-old grand challenge in mathematics – the KLS conjecture Benchmarking Long-Horizon Agent Context Compaction with CompactionBench Latent Core Tokenizer: Compress, but Meaningfully Chapters 0:35: Internship: Growing expertise, building community (opens in new tab) 1:28: Breaking the language barrier with AI (opens in new tab) 2:10: Solving a 31-year-old mathematics challenge (opens in new tab) 2:56: Making AI coding agents more efficient (opens in new tab) 3:57: Improving AI tokenization across 104 languages (opens in new tab) Subscribe to the Microsoft Research Podcast : Apple Podcasts Email Android Spotify RSS Feed Transcript [MUSIC] This is Microsoft Research Focus , your quick look at news and cutting-edge research from across Microsoft. I’m Anson Ho, and these are some of the latest advances we’re tracking. Coming up in this edition: how AI is bringing the possibility of a universal translator closer to reality; a breakthrough that solves a 31-year-old mathematical challenge; a new approach to make AI coding agents faster and more efficient; and more. Here’s what’s new. [MUSIC FADES] Internship: Growing expertise, building community At the Microsoft Research lab in Tokyo, research intern Lilika Makabe found opportunities to grow her expertise while helping to shape a new research community. Makabe joined the lab with a background in computer vision and worked alongside researchers exploring robotics and physical intelligence, AI systems that can interpret information and act in the physical world. Her experience highlights how research internships can offer more than technical training. Through hands-on experimentation, interdisciplinary collaboration, and everyday conversations with colleagues from different backgrounds, Makabe expanded her research perspective while contributing to the lab’s growing community. You can read about her experience on the Microsoft Research blog. [TRANSITION MUSIC] Breaking the language barrier with AI For decades, science fiction has imagined a world where language barriers no longer stand in the way of communication. For example, in 1967, Star Trek introduced the fictional Universal Translator. Now, advances in AI are bringing that possibility closer to reality. In a new conversation, Microsoft Applied Science Manager Jinyu Li explores how AI is advancing the prospect of communication across languages, the progress researchers have made, and the challenges that remain. To learn more, catch the full episode of the podcast Our Digital Life , available to listen or watch on YouTube and podcast platforms. [TRANSITION MUSIC] Solving a 31-year-old mathematics challenge For the last 31 years, the KLS conjecture has been one of the central open problems in high-dimensional geometry. But not anymore. New research has resolved the conjecture. This removes a major theoretical obstacle to understanding the efficiency of randomized algorithms used across science and computing. The breakthrough establishes a universal principle for log-concave distributions: simple flat cuts can identify their narrowest bottlenecks within a fixed factor, regardless of dimension. The results have been independently validated, and the proof has also been formally verified using computer-assisted methods. Learn more about this breakthrough on the Microsoft Research blog. [TRANSITION MUSIC] Making AI coding agents more efficient Coding agents can help when writing and modifying code, but long sessions can max out the amount of information an AI model can process at once. As a result, coding agents will summarize earlier interactions in a resource-intensive process known as context compaction. A new study looked at 13.5 million coding sessions and found fewer than 8% required context compaction. Yet those sessions accounted for 44% of total AI processing costs. To address this challenge, researchers developed COMPACTIONBENCH. COMPACTIONBENCH is a benchmark for evaluating context compaction during long-running coding tasks. They also introduced APEC, a method that prepares summaries in the background instead of interrupting an agent’s work. In evaluations, APEC improved coding performance while reducing costs by up to 25.4% and execution time by 9.6%. [TRANSITION MUSIC] Improving AI tokenization across 104 languages Before AI models can process text, they break it into smaller units called tokens. How that text is divided can affect both efficiency and performance, especially across different languages. Most tokenizers are designed to compress text as efficiently as possible. But a new study suggests that compression alone isn’t enough. Researchers introduced the Latent Core Tokenizer, or LCT, which identifies reusable components of words before building a shared vocabulary across languages. In evaluations spanning 104 languages, LCT produced more linguistically meaningful tokens than leading alternatives while maintaining comparable efficiency across languages. It also improved model performance across four multilingual benchmarks. The findings suggest that better tokenization depends not just on how much text can be compressed, but on how well the underlying structure of language is preserved. [MUSIC] And that’s your Microsoft Research Focus . From the resea

  • The results have been independently validated, and the proof has also been formally verified using computer-assisted methods.
  • As a result, coding agents will summarize earlier interactions in a resource-intensive process known as context compaction.
  • [TRANSITION MUSIC] Improving AI tokenization across 104 languages Before AI models can process text, they break it into smaller units called tokens.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Three Lincoln Laboratory technologies receive 2026 R&D 100 Awards

<p>Three technologies developed by MIT Lincoln Laboratory are among recipients of 2026 R&amp;D 100 Awards. The awards recognize the 100 most significant innovations that have transitioned to use or been made available for license in the past year.</p><p><em>R&amp;D World</em>, an online publication, manages the awards program, which has run annually since 1963. The worldwide competition is judged by an independent panel of science and technology experts and industry professionals.</p><p>The laboratory's winners overcome obstacles to improving the security and capability of systems. One technology is an operating system designed to protect embedded computers from cyberattacks. Another technology is a platform for integrating superconductor and semiconductor chips in a wafer-scale system for quantum or high-performance computing. The final winner is an algorithm that removes interference 5G base stations incur on co-channel radio frequency (RF) devices.&nbsp;</p><p>"These awarded technologies address critical problems in cybersecurity, quantum computing, and communications. We're proud of our dedicated staff who developed and transitioned these technologies for the nation's advantage," says Melissa Choi, Lincoln Laboratory director.&nbsp;</p><p>Since 2010, Lincoln Laboratory has received 111 R&amp;D 100 Awards recognizing the transfer of its unclassified technologies to industry and government. Each year, many technology transitions also occur for classified projects. This&nbsp;transfer of technology&nbsp;is central to the laboratory's role as a Department of War (DoW) federally funded research and development center.</p><p><strong>Secure-by-design operating system</strong></p><p>Computers embedded in deployed systems — such as satellites, drones, or radars — keep missions intact and people safe. The operating systems (OS) inside these embedded computers, however, are prime targets for cyberattacks. They are large, complex, and built on decades of legacy designs predating modern cyber threats. Increasingly, artificial intelligence is helping adversaries exploit OS vulnerabilities.&nbsp;</p><p>To protect embedded systems, Lincoln Laboratory developed an inherently more secure OS called Magnetite. "If Windows and Linux are the familiar workhorses of general-purpose computing, Magnetite is built for the different realities of embedded systems, for which real-time performance matters, resources are tight, and system failure has physical consequences," says Juliana Furgala, principal investigator in the&nbsp;<a href="https://www.ll.mit.edu/r-d/cyber-security-and-information-sciences/secure-resilient-systems-and-technology">Secure Resilient Systems and Technology Group</a>.&nbsp;</p><p>A few components assure Magnetite's security. It is built on top of seL4, a formally verified microkernel (i.e., software core). Formal verification means that mathematical proofs have guaranteed the microkernel's code behaves exactly as intended, eliminating a source of security flaws. On top of this foundation, Magnetite’s OS services — such networking, storage, and communication — are written in Rust. This programming language prevents common memory-corruption bugs that hackers often exploit. Finally, Magnetite configures its services as isolated processes under the principle of "least privilege." This configuration prevents different parts of an embedded system from accessing unnecessary data. If a cyberattack compromises one service, that component can be locked down while critical functions of the system continue running.</p><p>Free for government use and approved partners, Magnetite is now being integrated into space and flight platforms for the DoW. Looking ahead, the team sees potential to extend the platform to other embedded domains, such as control systems for critical infrastructure. Because Magnetite offers a common, reusable foundation for building specialized mission capabilities, the team is also collaborating with other technical groups across the laboratory to offer Magnetite as an alternative to costly, proprietary operating systems currently in use.</p><p>In addition to earning an R&amp;D 100 Award, Magnetite was selected as MIT Lincoln Laboratory's Best Invention of 2025.</p><p><strong>Wafer-scale platform for quantum systems</strong></p><p>Modern computing systems are increasingly limited not by transistor performance, but by interconnection bottlenecks. Electrical pathways linking densely packed chips are taking up space, slowing data transfer, and generating heat. The problem is especially pronounced in quantum systems, which require bulky coaxial cables to bridge superconducting chips (which must be cryogenically cooled to achieve zero electrical resistance) with room-temperature control electronics.</p><p>The Quantum Microsystems platform addresses this interconnection challenge, integrating hundreds of superconductor and semiconductor chips into a single, wafer-scale system. The platform combines several innovations. First, an AI decision-tree approach determines the optimal way to route connections between chips across a wafer. After the wafer is tested for defects, a digital lithography process "draws" custom wiring layers between chip locations and around defective circuit components to enable the highest yield possible from the wafer. Chips are then connected using microbump (tiny "bumps" of conductive material used to connect stacked chips) or bumpless techniques&nbsp;that place them extremely close together; a superconducting fabric extends these connections across the wafer to create a unified computing platform, carrying signals at near light speed. Coaxial cables are replaced with integrated, lithographically defined flexible interconnects, significantly reducing cryogenic wiring overhead and system complexity.&nbsp;</p><p>The Quantum Microsystems platform is designed to be reworkable, so chiplets can be replaced or upgraded without rebuilding the entire system. The platform en

  • </p><p>The laboratory's winners overcome obstacles to improving the security and capability of systems.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Nvidia and the World Without Jensen

Câu hỏi But the question that matters for shareholders (and the AI supply chain) is a different one: what is Nvidia without Jensen?

Jensen Huang is no longer the CEO of Nvidia . Nvidia’s board was investigating Jensen’s conduct and weighing his future at the company he founded, and clearly the conduct issue is real; it is ethically problematic to have relationships with a customer who is a Chinese CCP asset and with multiple subordinates. Additionally, we believe there is more at play with his resignation after only four days. We believe it is directly tied to the vast smuggling of GPUs. Our Accelerator Model tracks every AI chip’s production and destination, and our data shows more than 1.5GW of Nvidia chips, which are restricted, still ultimately ended up in China. This figure doesn’t even include chips that Nvidia created to circumvent US rules, a story we originally broke . We believe the US intelligence community is a large part of the push to get Jensen out of Nvidia. Specifically, we believe it is likely that RASA or similar legislation, which has already passed the house, will move forward. AI and chip restrictions on China are going to ramp now much faster with the biggest defender Jensen, gone. We are strong believers in some restrictions, but poorly shaped ones are damaging to US interests. This sort of legislation will have consequences, and we don’t think politicians are thinking them through. Here are the clear benefits of the current regime where Chinese labs are able to access GPUs remotely, and NOT in China: Profits flow to American companies. The US can monitor what Chinese labs are doing. The US can cut off their access to the GPUs through politics or by force almost instantly, if they choose. American companies such as NVIDIA, AMD, Broadcom, Google, Amazon, Microsoft, Meta, Supermicro, and Dell all depend on Chinese companies in their supply chain. American companies such as NVIDIA and AMD employ Chinese engineers (who live in China, to be clear) to develop hardware and low-level software for their GPU systems. Without these engineers, we are certain that the user experience for these GPUs would be materially worse than it is today. American companies, especially the entire startup ecosystem benefit from the open source research that Chinese companies, students and independent researchers publish using these GPUs today. China is (even slightly) less motivated to use their domestic accelerators. Innovations from China flow globally through open innovations of wide expert parallelism, OSS kernel libraries: MSA, DeepEP, DeepGEMM, etc. With that said we do believe in some AI restrictions to China, but those are more shaped like the MATCH Act then RASA. Biden style diffusion rules make the most sense. Maybe biggest of all: open source is under attack . It’s an open secret that Nvidia works directly with the Chinese labs on open source and helping them optimize GPUs and models. These labs utilize Blackwell GPUs, which Nvidia knows. The US open source ecosystem and many startups rely on Chinese AI models. This is significantly impaired with Jensen gone and the China hawks winning. But the question that matters for shareholders (and the AI supply chain) is a different one: what is Nvidia without Jensen? Our view is that the company was never built for him not to be there. He was at the helm for 33 years, and even though the board pressured him out (or he “resigned”) Nvidia would still be materially better off with him there instead of gone. The company was uniquely built to suit the man himself. The company without the man might be an entirely different beast. SpaceX and Tesla don't function without Elon, Nvidia doesn't function without Jensen. One King, 60 Direct Reports Nvidia’s corporate structure is odd for a company of its size and scale. Jensen has roughly 60 direct reports, according to his public statements. He has said that CEOs should have the most direct reports because those people require the least amount of management and that he doesn’t have one-on-one meetings with them. But few of them can do things without Jensen. The company is run like a kingdom with one king, with everyone else, regardless of seniority, a peon doing his bidding. Jensen has said that it doesn’t matter who reports to who as their title isn’t related to how important their job is, and it’s one of the reasons why Nvidia is the “smallest large company in the word” and has so much fewer employees – about 42,000 – despite its big market cap. At Nvidia, there is less politics as a result. Jensen has said this flat network empowers employees and allows the company to adapt more quickly because its technology is also moving so quickly. This is one of the reasons Nvidia moves fast. Once Jensen set a direction, everyone moved on it together. Former employees have told us that product development and engineering often happen until the last minute so the company can change its mind if it needs to. For example, the GB200 design came together a lot closer to the product cycle than probably anyone would have liked, one former executive told us. Jensen has said he wishes he could keep the company in startup mode forever. The cost of that structure, however, is that Nvidia doesn’t have a deep bench of successors. Look at Nvidia’s website: only five others, two of whom are legal and finance executives, are featured prominently among his reports. Most of his direct reports aren’t trained for the spotlight and are actively discouraged from seeking it. Few leave to become CEOs of other major companies, with Rene Haas at Arm the rare exception. Few at Nvidia get the chance to shine externally because Jensen makes all the big decisions. But much of Nvidia’s executive team doesn’t mind because heavy is the head that wears the crown, and many are happy to stay until retirement age because they’re well paid, treated well, and believe in the mission and their leader. No One Wants the Job Colette Kress has been named interim CEO, but Nvidia is looking for a permanent CEO. That cannot be someone with finance background. The obvious internal names b

  • At Nvidia, there is less politics as a result.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

A New Dawn in the Cloud: ‘Warhammer 40,000: Dawn of War IV’ Is Coming to GeForce NOW

Câu hỏi What’s on the gaming lineup this weekend?

Rally the ranks this GFN Thursday. Warhammer 40,000: Dawn of War IV is coming to GeForce NOW , allowing players to experience the battle for Kronus through cloud gaming. There’s more on the horizon, with NVIDIA DLSS 5 technology coming to War Thunder and Where Winds Meet . Meanwhile, Dragon’s Dogma 2: Dark Arisen brings new tales and encounters, alongside 10 games joining the GeForce NOW library this week. Kronus Calls for Reinforcements Warhammer 40,000: Dawn of War IV , developed by KING Art Games and published by Deep Silver, is coming to GeForce NOW. Launching on Steam Thursday, Dec. 3, the real-time strategy game returns to the series’ roots with large-scale battles, base building and a renewed fight for planet Kronus. Command the Space Marines, Orks, Necrons or the Adeptus Mechanicus, which makes their series debut. Build a base, secure resources and customize squads to the conflict across more than 70 campaign missions, playable solo or in co-op mode. The Last Stand, Skirmish and multiplayer modes offer more ways to put tactics to the test. GeForce NOW will bring the action to supported PCs, Macs, handhelds and other devices, with game downloads and updates handled in the cloud. Save the storage space for everyday essentials and planning for the next offensive. Wishlist the game on Steam , and stay tuned to GFN Thursday to prepare for launch. DLSS 5 Keeps on Coming NVIDIA DLSS 5 continues to evolve, with Gaijin Entertainment’s War Thunder and Everstone Studio’s Where Winds Meet next to receive support for the technology with the 3D-Guided Neural Rendering feature. Arriving in mid-November, War Thunder ’s Eagle’s Shadow major update adds DLSS 5 with 3D-Guided Neural Rendering, dozens of new planes, ships and tanks, along with cluster bombs for planes and visual-effects enhancements increasing the realism and dynamism of battle. In Where Winds Meet , DLSS support arrives Monday, Nov. 23, enhancing the environments and characters of its wuxia open-world action role-playing game. Both titles will use the latest DLSS 5 with 3D-Guided Neural Rendering model, released in October, which doubles model performance compared with the previous version. Since its initial announcement in March, model performance has improved tenfold, bringing the technology to more gamers, including GeForce NOW Ultimate members streaming with GeForce RTX 5080 cloud rigs. Get a closer look at the DLSS updates and find more details . A New Tale for the Arisen Dragon’s Dogma 2: Dark Arisen is streaming on GeForce NOW, expanding the deep, explorable fantasy world of Dragon’s Dogma 2 with new tales and encounters. Travel to the long-abandoned region of Norgan, join the enigmatic Eir and uncover the secrets of the undying Fallen Dragon. Unearth forgotten relics to unlock powerful weapons and skills, take on Lost Rites’ 12 dungeon challenges and explore new hairstyles and tattoos for Arisen and Pawn customization. Delicious in Dungeon also arrives in the world of Dragon’s Dogma 2: Dark Arisen with a special collaboration. GeForce NOW keeps all the new downloadable content available across supported devices — without a single download claiming local storage on a gamer’s hardware of choice. In addition, members can look for the following games joining the GeForce NOW library this week: Worst Class Ever (New release on Steam , Nov. 9) Ashes of the Singularity II (New release on Steam , Nov. 10) Cyberpunk 2077 ( Xbox Play Anywhere, new release on Battle.net , Nov. 10) DreamWorks Shrek 2 Remastered (New release on Steam , Nov. 10) Warhammer Survivors (New release on Steam and Xbox , available on Game Pass, Nov. 10) Main Sequence (New release on Steam , Nov. 12) Deadly Days: Roadtrip ( Steam ) GeForce 8800 GTX – 20 Years of a Legend ( Steam , Ultimate members) Graveyard Keeper ( Steam ) Tape to Tape ( Steam ) Also this week, Ultimate members can take a trip back to 2006 with GeForce 8800 GTX – 20 Years of a Legend , a playable demo celebrating the iconic GPU’s 20th anniversary with DLSS 4.5 and path tracing in the cloud. Dates listed above reflect when games are released on their respective stores. GeForce NOW availability may vary, as games are onboarded after release and added throughout the week. Keep an eye on GeForce NOW channels and GFN Thursdays for availability updates on announced titles. Those ready to put the next game plan into action can grab a day pass to try Performance or Ultimate, with the cost applicable toward a first monthly membership. What’s on the gaming lineup this weekend? Let us know on X or in the comments below.

  • Since its initial announcement in March, model performance has improved tenfold, bringing the technology to more gamers, including GeForce NOW Ultimate members streaming with GeForce RTX 5080 cloud rigs.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Powering AI-Driven Drug Discovery with President of Isomorphic Labs Max Jaderberg

Drug discovery and development has long relied on expensive, time-consuming empirical trial and error. By shifting the discovery process onto computers, Isomorphic Labs wants to fundamentally change medicine from a process of discovery to one of precision engineering. Elad Gil sits down with Isomorphic’s president Max Jaderberg to discuss his vision for using predictive AI models to achieve experimental accuracy. Max details how the success of AlphaFold provided the first proof of concept that a deep learning model could generalize in the field of biology, and how an AI model’s compounding of learnings allows it to achieve generalizability with scale, thus greatly increasing the probability of success as well as speed in real world clinical trials. He also talks about evolving past the need for animal testing through in-silico models, Isomorphic’s focus on developing drug design engines that are agnostic of disease area, target class, and modality, why accurate molecule engineering unlocks the potential for multidrug cocktails that span complex biological pathways, and how he believes AI molecule optimization will improve both human lifespan and healthspan. Sign up for new podcasts every week. Email feedback to show@no-priors.com Follow us on Twitter: @NoPriorsPod | @Saranormous | @EladGil | @MaxJaderberg | @IsomorphicLabs Chapters: 00:00 – Trailer 00:45 – Max Jaderberg Introduction 01:24 – Isomorphic Origins 05:23 – Drug Discovery Stagnation 11:49 – Hiring at Isomorphic 14:35 – Bringing Drug Discovery Onto Computers 19:31 – Isomorphic’s Datasets 20:30 – Rethinking Endpoints 22:50 – External vs. Internal Dependences 23:57 – Automating Biologic Labs 26:22 – Internal Development vs. Partnerships 27:19 – Areas of Human Health Focus 29:21 – Multidrug Approaches 32:17 – Risk Appetite Variation 34:26 – Disease Treatment Outlook 35:42 – Human Longevity and Drug Development 39:16 – Conclusion

  • Elad Gil sits down with Isomorphic’s president Max Jaderberg to discuss his vision for using predictive AI models to achieve experimental accuracy.
  • Max details how the success of AlphaFold provided the first proof of concept that a deep learning model could generalize in the field of biology, and how an AI model’s compounding of learnings allows it to achieve generalizability with scale, thus greatly increasing the probability of success as well as speed in real world clinical trials.
  • He also talks about evolving past the need for animal testing through in-silico models, Isomorphic’s focus on developing drug design engines that are agnostic of disease area, target class, and modality, why accurate molecule engineering unlocks the potential for multidrug cocktails that span complex biological pathways, and how he believes AI molecule optimization will improve both human lifespan and healthspan.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] D-Day (Decision Models Day)

Today was the “System One day” of the second annual AIE CODE event , and somehow everything aligned: Cloudflare’s Michelle Chen walked through their Clef and Clairic models Richard covered Diogo’s triumphant return to AIE talking about the Jevolution with their most notable production customers since crossing $100M ARR In a BIG surprise, Google beat Anthropic to launching a production decision model, with Sundar and Logan announcing Gemini Decision 1 , a calibrated multimodal decision model capable of taking up to 1000 questions and costing $0.10 per million input with ~200ms e2e latency. Benchmarks are good ( usual caveat about benchmarks ): And you can hear all this and more on today’s System One track. And support AINews by giving your agent SerpApi to read and crawl the web! Sponsored by SerpApi Give Your AI Agents the Full Page, Not Just the Link Retrieve public web pages by URL as clean Markdown, structured JSON, or raw HTML with SerpApi’s Extract API . Extract a single page, map a site’s URLs, or run a bounded same-site crawl. Simply set engine=extract with a URL, and add output=md for Markdown. Pair it with 100+ SerpApi search APIs to go from query to page content with one provider. Thanks to SerpApi for working with business@latent.space! AI Twitter Recap Google Launches Gemini Decision 1 for Single-Pass Classification Gemini Decisions API : Google introduced a new API backed by Gemini Decision 1 . The model handles closed-set decisions, so it scores choices in a single pass instead of generating text token by token ( Sundar Pichai , Logan Kilpatrick , Google ). Specs : It supports yes/no, multiple-choice and rubric scoring, with up to 255 multiple-choice options. It accepts text, image, audio, video and PDF inputs, handles up to 1,000 questions per request, has 256K context, returns answers in 150–250 ms end to end, and attaches a confidence score to every answer ( @_philschmid ). Pricing : $0.10 per 1M input tokens, with $0 billed for output. Google says that is about 5x cheaper than Flash-Lite ( Sundar ). Origin : The project started as an internal Google DeepMind prototype. A small team shipped the model and API about six weeks later ( @_philschmid ). Google's comparison with TypeSafe's Jev : Google's own tests show Decision 1 beating Jev 1.14 on 5 of 7 shared benchmarks, including 92.2% vs 90.0% on JevBench. Jev costs $0.042 per 1M input tokens ( @kimmonismus ). Artificial Analysis evaluation : Independent results are more mixed ( Artificial Analysis ). Overall score : Decision 1 scores 66 on the Decision Index. That ties Elephant 9B and Jev 1.14, and sits 7 points behind Drex Max v1.1 and GLiDE v1.2. Strengths and weaknesses : It leads AA-Omniscience-Decision (66) and HLE-MultiChoice (49). On AA-Paperclip, it answers every question correctly on only 33% of documents, versus 63% for Drex Max. Cost and latency : It costs $345 per 1M tasks, versus $103 for Drex Max and $156 for Jev. Its Brier score is 0.18. Average time per task is 211 ms, slower than Jev at 92 ms ( latency ). Early adopters : Stitch built a rapid-design experiment on the API that goes from a rough prompt to a multi-screen product flow in about 5 seconds ( Stitch ). LangChain shipped day-zero support with a base class shared with TypeSafe classifiers ( LangChain ). Architecture pattern : At AI Engineer, Cloudflare described a hybrid agent design. Fast decision models handle tool selection and log extraction, and a large reasoning model handles diagnosis only. In its demo, this cut six LLM calls to one and about 40K tokens to 1.8K ( TheTuringPost ). Frontier Model Evals: Fable 5.2 Leads, Carbon Draws Speed Complaints Claude Fable 5.2 on Agent Arena : Fable 5.2 (Max) ranks #1 with a +30.4% net improvement across 4.7K real agentic sessions. That is about 1.6x the score of #2, Opus 5.5 (+19.5%). It is also #1 in Code, Work and Chat ( Arena , Pareto view ). Pricing : API prices drop from $10/$50 to $8/$40 per 1M input/output tokens. Median cost per task is $7.3, versus $5.5 for Opus 5.5, $1.5 for Gemini 4 Carbon (High) and $6.0 for GPT-6-Astra. Other evals : AA-Robotics v0.1 : Fable 5.2 leads at 75, ahead of Astra (69) and Sol (65) ( AA ). VIALS-AA (visual reasoning) : It scores 67% with tools but only 44% without them. Haiku 5.5 with tools reaches 49% at $0.07 per task ( AA ). Debate benchmark : It won 209 of 224 debates in a Cambridge Union-style contest ( @SamuelFitouss10 ). Effort calibration : Within the Claude lineup, Artificial Analysis finds that moving up a model tier beats raising effort: Opus 5.5 at xhigh outscores Sonnet 5.5 at max with 40% lower cost per task ( AA ). Vals separately reports that Fable's "high" setting is often cheaper and better than xhigh or max ( @RayanKrishnan ). Gemini 4 Carbon : Vals ranks Carbon #3 on its index, and the fastest and cheapest of the top five ( Vals ). Practitioner opinions conflict ( Theo , Martin Casado , @MParakhin ). Latency complaints : Theo measured Carbon at low effort taking 22m21s, versus 9m33s for Opus 5.5 at high ( timing ). He also reports about 11 seconds to first token, roughly 20x Fable's, and attributes the delay to the server side ( TTFT ). Coding and math : Casado calls Carbon "really good at code." Parakhin says Carbon often beats Fable on math and ML problems but that he does not trust it with coding. Muse Ember : Meta's model posts +1.7% on Agent Arena at about $0.02 per task. That is roughly 4x cheaper than GPT-6 Luna and Haiku 5.5, and about 12x faster than Luna ( Arena ). Vals reports Ember reaches 85% of Fable 5.2's performance at 0.84% of the cost ( Vals ). Chinese models : Qwen 4 : Qwen-4-Max appears to be live. Rumors that it is disappointing remain unconfirmed ( @kimmonismus ). Some users are skeptical about how Alibaba has labeled its variants ( @teortaxesTex ). ERNIE 5.5 Flash Preview : Baidu's model is trending toward #2 among open models on Code Arena WebDev at 1668. Apache 2.0 weights are promised ( Arena ). Nex-N2.6 : This open mu

  • Artificial Analysis evaluation : Independent results are more mixed ( Artificial Analysis ).
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Grok Bot 1.0 🤖, AlphaGenome 2 🧬, Gemini Decisions 🧠

Many organizations feel caught between two unattractive options: move fast to capture AI's upside, or slow down to reduce risk. Algolia's whitepaper reveals how the right architectural choices and operating discipline can allow you to move quickly while keeping security and trust intact. You'll learn where value is created and risk emerges, and how to make architectural decisions that directly improve speed, cost, trust, and business outcomes.

  • Many organizations feel caught between two unattractive options: move fast to capture AI's upside, or slow down to reduce risk.
  • You'll learn where value is created and risk emerges, and how to make architectural decisions that directly improve speed, cost, trust, and business outcomes.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Inkling 2 Small on vLLM: Day-0 Support

vLLM supports Thinking Machines' Inkling 2 Small from day 0: attention residuals, a new decode kernel for relative-bias attention, all-reduce fused with the ops around it, and DSpark speculative decoding.

  • vLLM supports Thinking Machines' Inkling 2 Small from day 0: attention residuals, a new decode kernel for relative-bias attention, all-reduce fused with the ops around it, and DSpark speculative decoding.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Anthropic's IPO roadshow video

<p><strong><a href="https://ipo.anthropic.com/us/#roadshow-video">Anthropic&#x27;s IPO roadshow video</a></strong></p> Anthropic kicked off their IPO roadshow yesterday, and their roadshow video has some previously unreleased information about how their business works.</p> <p>The section on "Approach to compute" with co-founder Tom Brown is particularly interesting, describing how Anthropic train and serve Claude across AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs.</p> <p>I'm not sure why there's no official video transcript, but here's <a href="https://gist.github.com/simonw/9806757fe8acec6221cb0a5cce4d720d">one that I extracted</a> using MacWhisper. Highlights below:</p> <blockquote> <p>We train and serve Claude now across all four major chip families. That's only possible because early on we made an unconventional bet on a multi-chip strategy. At the time, many people thought that running AI models on different chips wasn't worth the effort. Each chip needs its own tailored software stack and inference stack, which is expensive and difficult to build. We've now been doing it for years, and we've gotten incredibly good at it. [...]</p> <p>As an example, this year, we started working with AMD GPUs for the first time. In the past, bringing on a new chip platform took dozens of engineers months of work <strong>This time, we had Mythos try it out. It got Claude running on AMD in just three days, with basically no human intervention</strong>. Claude ported or wrote almost 100% of the kernels. [...]</p> <p>The second driver of our compute advantage is our rate of discovery of compute multipliers. We focus on total effective compute, which is the multiplicative product of physical capacity times the compute multipliers that we discover. Each one we discover makes our pool of total effective compute bigger. For example, say you need 100 units of compute to train a model of a certain intelligence. A 4x compute multiplier means you can use 25 units of compute to train the exact same model. Or you can use those hundred physical units to train an even more intelligent model that would have previously required 400 units.</p> <p>Compute multipliers create a competitive advantage for us that's hard to replicate. They build on each other. Each new multiplier we discover has a positive effect on the multipliers before it. They stack. <strong>We believe that this can accelerate even further as we approach recursive self-improvement. Claude will be able to discover compute multipliers entirely by itself</strong>, allowing the number of compute multipliers to grow faster and faster.</p> </blockquote> <p>We got a hint at Anthropic's multi-chip strategy last September in their <a href="https://www.anthropic.com/research/a-postmortem-of-three-recent-issues">postmortem of three recent issues</a>, where they described bugs that affected just their TPU deployments. <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/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/gpus">gpus</a>, <a href="https://simonwillison.net/tags/claude-mythos-fable">claude-mythos-fable</a>, <a href="https://simonwillison.net/tags/recursive-self-improvement">recursive-self-improvement</a></p>

  • <strong>We believe that this can accelerate even further as we approach recursive self-improvement.
  • net/tags/recursive-self-improvement">recursive-self-improvement</a></p>
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Mercor $750m Series D at $20B valuation, $3B run rate

Our usual bar for making a fundraise a title story these days is a decacorn ($10B) valuation. We missed Mercor’s $10B Series C last year when it had less than $500m in revenue, but knew we wouldn’t have to wait too long til the next round, since it crossed $1B in Feb and $2B in June , it has now hit $3B annualized, surviving hacks : View post on X ↗ The company reports 5 million experts, and their TAM: “In the United States alone, there are more than 800 distinct occupations, each containing workflows that models have yet to be evaluated against.” And also this notable stat we are tracking: We’re seeing the impact of AI investment in our business. We already spend 3x as much on inference as on employee salaries, and it pays for itself by making our team better at their jobs . If that ratio holds as AI spreads through the rest of the economy, it points to something closer to double-digit GDP growth over the next five years, nearly 3x the rate we’re used to. We have yet to do a Mercor pod (will work on it!) but to better understand a data vendor business, you can look at our recent deep dive with Handshake AI , one of their competitors: AI News for 11/09/2026-11/10/2026. 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 Gemini 4 Carbon Launch: Rollout, Pricing and Claimed Results Launch : Google released Gemini 4 Carbon as its strongest model for long-horizon agentic work in software engineering, cybersecurity, knowledge work and science ( Google , Sundar Pichai ). Gemini 4 Argon remains Google's recommended default for most coding and knowledge work, with Carbon reserved for the hardest tasks ( Google AI ). Availability : Carbon is available through the Gemini API and AI Studio for paid developers, Android Studio and Gemini for business. Ultra subscribers can use it in the Gemini app, AI Mode and Antigravity. Scientists can apply for 6 months of free access, and cyber defenders can apply through the Fairwind program ( Google AI ). Same-day integrations shipped in Cursor, Devin, Perplexity and GitHub Copilot ( Copilot ). Pricing : In Cursor, Carbon costs $2/$10 per million input/output tokens through Nov 30, rising to $4/$20 on Dec 1 ( Cursor ). Developer notes : Upgrading from Argon only requires changing the model ID to gemini-4-carbon . Google cites long decode up to 1M tokens and recommends thinking_level medium for most tasks ( @_philschmid ). /swarm in Antigravity : A new slash command turns one prompt into a multi-agent team that coordinates on a live message board. Each role can use a different model, for example Carbon for architecture and auditing and Flash for parallel edits ( Antigravity ). Google's showcase demo is a pelican-on-a-bike 3D campus game, with subagents rigging assets in Blender ( Sundar Pichai ). Official and partner claims : Google internal results : +25% open-source vulnerability recall over Argon, and 40x more remote protein homologs found for AlphaFold 3 ( @_philschmid ). Coding benchmarks : Carbon tops CursorBench at 60.4% on high effort ( Cursor ). It scores 40.0% on FrontierCode 2.0, up from Argon's 26.3% ( Cognition ). Research agents : Perplexity reports 0.857 on WANDR at $5.86 per task, matching Fable 5.2 at 40% of the cost ( Perplexity ). Independent measurements : Vals Index : Carbon ranks #3, about 1pp behind Fable 5.2, at $10.51 vs $17.94 per test. It is #1 on Vals' Finance, Legal Research and Tax indices ( Vals ). ARC-AGI : ARC-AGI-3 scores are 99.1% ($4.6K) with the provider adapter and 65.1% ($7.2K) in the standard harness. ARC-AGI-2 is 94.2% at $0.38/task ( ARC Prize ). Vision : Roboflow ranks Carbon #1 overall, but last of 10 on both speed (81s per image) and cost ($91.89 per 1K images) ( @skalskip92 ). Agent Arena : Carbon's cost advantage over GPT-6 Astra and Opus 5.5 widens on the most expensive sessions ( Arena ). Caveats from testers : Early users report hallucinations, overconfidence and verbosity ( @krishnanrohit ). One tester found broken Undo and shifted export text after Carbon reported its checks as passing ( @kimmonismus ). Google's Logan Kilpatrick objected that Gemini was the only model not tested in its native harness on one comparison ( @OfficialLoganK ). Claude Fable 5.2 / Mythos 5.2: Evals, Token Cost and System Card Head-to-head evals : MKA : Fable 5.2 leads Google DeepMind's new Multimodal Knowledge Agent benchmark at 57% Climb WSR. Carbon scores 54% and Opus 5.5 53%. Cost per task is $31.97 for Fable vs $6.29 for Carbon and $1.11 for GPT-6.1 Sol. No model leads all 10 verticals ( Artificial Analysis ). VIALS-AA : Fable scores 67% on life-science visual reasoning, ahead of Carbon (64%) and GPT-6 Astra (63%). It averages about 37 tool calls per task, vs about 8 for Astra ( Artificial Analysis ). Labor and design : Fable reaches 41.3% on the Remote Labor Index, up from 17.9% for Fable 5.1 ( @kimmonismus ). It is #1 on 3D Design Arena with an Elo of 1540, priced at $8/$40 per million input/output tokens ( DesignArena ). RSI Index : Vals scores it 37.8, narrowly ahead of Opus 5.5 at 37.3 ( Vals ). Token efficiency : GDPx materials science : SemiAnalysis finds Fable costs 89% more than Astra and uses 4.3x more output tokens for roughly the same score. It costs 7.5x more than Carbon ( SemiAnalysis ). GDPx software engineering : Fable nearly triples Opus 5.5's score, but per-task cost rises by a similar multiple ( thread ). Domain strengths : Dylan Patel says Fable improves on chip design, SWE, biology and materials, but not on finance, legal or RSI. He attributes the RSI result to Anthropic deliberately limiting those capabilities ( @dylan522p ). Dissenting results : One lab ran 100 unsaturated environments and found Astra and GPT-6.1 Sol/Terra produced better solutions. It also reports Fable pausing mid-run to ask a nonexistent human for decisions ( @leo_l

  • AI Twitter Recap Gemini 4 Carbon Launch: Rollout, Pricing and Claimed Results Launch : Google released Gemini 4 Carbon as its strongest model for long-horizon agentic work in software engineering, cybersecurity, knowledge work and science ( Google , Sundar Pichai ).
  • Official and partner claims : Google internal results : +25% open-source vulnerability recall over Argon, and 40x more remote protein homologs found for AlphaFold 3 ( @_philschmid ).
  • Domain strengths : Dylan Patel says Fable improves on chip design, SWE, biology and materials, but not on finance, legal or RSI.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

AI Engineer Code Summit: OpenAI on New Ways of Working With Codex

Câu hỏi I think about it as saying, how do I manage my threads?

AI Engineer Code Summit kicked off today with a series of workshops, including from leading frontier labs OpenAI, Anthropic and Google DeepMind. I attended the OpenAI workshop, intriguingly titled “New Ways of Working with Codex” and hosted by two of its Developer Experience engineers, Dominik Kundel and Charlie Guo . The past year in coding agents has been one of constant product upgrades by all the frontier labs. But not all of these changes have gone smoothly. The shift to cloud , for example, is still confusing to many users of Codex, Claude and Antigravity. An illustration of the evolution of Codex, shown at the OpenAI workshop today. We’ve also seen new ways to work with agents. Personal agents like OpenAI’s dots and Anthropic’s Claude Duet can assist between sessions, while voice mode gives developers another way to steer coding tasks. OpenAI’s workshop addressed all three, and I also had some followup questions on these topics for Kundel, after the session. Guo began the workshop by listing a selection of things they’ve shipped in Codex since the AI Engineer World’s Fair in June 2026. “I had to cull this list in order to make it fit on the screen because there were too many things to put in here ,” he said. “I think the team has been shipping at an unbelievable rate.” The awkward transition to cloud Codex launched in May 2025 as “a cloud-based software engineering agent that can work on many tasks in parallel.” But Kundel admitted that “it took a while for models and harnesses to equalize into something where we can build a really engaging experience.” In other words, the cloud part of Codex has often been an awkward user experience. That began to change after DevDay in late-September, when a revamped version 2 of Codex Cloud was announced. Suddenly, cloud usage seemed more viable. As the DevDay announcement put it, you can “finally close your laptop now and your agents will keep working.” As explained in the official workshop writeup — hosted on ChatGPT Sites — the new version of Codex Cloud lets you “ publish an environment once and run multiple tasks in isolated instances of that setup.” “These can be shared within a company if you're working in a company, or you can just create your own,” said Kundel about these cloud environments . “You start off by picking the GitHub repositories that you want to have access to.” After the workshop, I admitted to Kundel that I still find Codex’s cloud functionality confusing — for instance, when I start a task locally and later want to move it to the cloud. “We definitely want to have it be easier,” Kundel replied. “Some of it is going to depend on managing the context differences . For example, yesterday what I ended up doing is I was working on something locally, and I just asked Codex to go and hand it off to the cloud.” That handoff to the cloud is something OpenAI is trying to improve for users. To demonstrate, Kundel reopened an existing cloud task in his browser and showed me that its Appshot, a local capture of a Mac app window, was still attached. Codex Cloud is now “just a background , like a different executor of where things are happening,” he continued. “And so in a similar vein, you can actually ask your local Codex to — for example — spin up a task on a cloud Codex. So they're all sort of merging together .” Voice as user interface One of the demos Kundel showed during the workshop was a new website for a fictional record company called Blossom Music. Kundel demonstrated live edits to the site and said he sometimes uses voice dictation when iterating this way . “You can really quickly iterate on these things and get to a world that you're actually envisioning with your stuff, especially if you're unsure what you actually want to get to,” he said. Later in the workshop, Guo explained that voice is an effective mode to use when delegating work to different agents . “Most people I see that use voice with coding essentially treat it as just a different medium on top of their normal work,” he said. But personally, he likes to use voice as a control plane , which he says allows him to control everything he’s doing in a code base. “I don't think about it [voice mode] as saying, here's the one lane of work that I'm doing right now with voice. I think about it as saying, how do I manage my threads? How do I check in on stuff? How do I pull data from other apps on the computer? How do I create new threads in the application? Just using my voice.” Guo used a telephone as a prop in this demonstration of voice mode. After the workshop, I asked Kundel what are some of the ways he is using voice in his work? He replied that dictation is one way he uses it, but another use case is to talk to his dot . “Since my dot has access to all of my dev boxes and my laptop and my cloud environments, it also has all the context on what I work on , what I do,” he told me. “And so I often talk to it on my way to work [...] to work through stuff, start up some PRs while I'm driving, so that when I come home or when I come to the office, I can go and review them. So, voice is incredibly powerful.” Dots Speaking of dots, personal agents were only briefly discussed in the workshop. Perhaps that’s because, as Guo admitted, they’re not entirely intuitive to developers. “As a Codex power user, dots feel a little bit awkward to work with,” he said. “I think because they're a new paradigm and because they're good at different things than the main Codex.” But he said there are three things he’s found useful for dots, as a power user. The first is as a safety net — your dot can flag things you may have missed. The second is as “a delegation sink” ; Guo said he uses it “a ton in Slack” for this reason. And the third is as a ubiquitous presence . “I think the third way that dots are a little bit different is we're aiming to make them truly everywhere that you work ,” Guo said. For his part, Kundel told me that he starts off a lot of his tasks from his dot, rather than

  • ” That handoff to the cloud is something OpenAI is trying to improve for users.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Show HN: Caramel: a speech-to-speech conversation model trained only on audio

I trained a speech-to-speech model last night, inspired by Dessa. I'm calling the model caramel-1. The model learned from recordings of people talking and nothing else, and has never seen a single word of text. The code, the training recipe and the weights are open source. caramel-1 sounds incredibly human. Most of what caramel-1 says doesn't make sense yet, but the sentences make a little more sense every time I add more data. Caramel is fast. On a GPU, caramel-1 makes 1 second of audio in about 0.09 seconds, which is fast enough to talk to a real person live. It even works on my MacBook Pro (M3 Pro): running two copies at once, it makes 1 second of conversation in about 0.25 seconds.

HN: open weights · naklecha 5 lượt bình chọn 0 bình luận Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

Microsoft researchers have discovered “Lazarus,” a backdoor technique that can reprogram dormant neurons in an AI model, causing an otherwise normal AI agent to execute malicious commands when a hidden trigger appears. The threat is especially serious for AI agents with tool or terminal access. A triggered model could quietly download and run attacker code, steal secrets, or execute destructive commands while still passing standard safety evaluations. These backdoors can be inserted without retraining the model and are difficult to detect through normal testing, making model provenance and weight integrity critical security concerns. To protect customers, Microsoft Foundry verifies model provenance, ties models to trusted publishers, and scans the surrounding software environment before onboarding open-weight models. Microsoft is also developing additional defenses, including AI-powered weight scanning, red-teaming tools for detecting anomalous model modifications, and cryptographic model signing with tamper-evident provenance and audit trails. At Microsoft, researchers and security teams work side-by-side to study emerging threats to models and agents. Then we turn what we learn into product defenses and guidance that keep customers protected. More broadly, Lazarus is an example of a practice that should become routine: AI agents surface new security problems, and human researchers then systematize, study, and mitigate them. Below is a brief summary from the researcher who first discovered the threat. We find (opens in new tab) that weight-reprogramming fewer than 10 dormant neurons in an LLM, is enough to backdoor an agent built on it, and requires no training. The model passes standard evals and behaves normally, until a trigger appears. The trigger can be as mundane as a date change, e.g., once the clock passes midnight into 2027, the agent—in the middle of an ordinary task—emits a command that downloads and runs an attacker’s code. The agent then executes it with whatever access you gave it. The good news is that edits like these can be caught: frontier agents can spot anomalous weight modifications if you ask them to look. But no scan can prove beyond any doubt that a model is safe. Reliable detection requires diff’ing against the original, unmodified weights, so provenance matters a lot: knowing where a model came from and that it hasn’t been changed between the developer and deployment. Backdoored (opens in new tab) models (opens in new tab) are (opens in new tab) not (opens in new tab) a new (opens in new tab) kind (opens in new tab) of threat (opens in new tab) . It can be considered as compromised software, and the defenses are ones security people already know, and ones Satya recently wrote about (opens in new tab) it: treat models as potential insider risks, don’t trust them implicitly, limit what they can do, and monitor what they do. What Microsoft does today. Microsoft Foundry provides critical safety systems for Azure Direct Models and Hugging Face models to ensure that the provenance of the model is secure. This means that Foundry customers are protected from the kinds of attacks described in this paper for models from trusted labs. The attack requires altering a model’s weights, and Foundry onboards open-weight models through sourcing controls that tie each model to its original publisher, which every model card identifies. Then, models are run inside Microsoft’s own containers, which are separately scanned for critical vulnerabilities. These controls cover where a model came from and the software around it. The remaining risk is that a trusted provider is itself compromised, for example through its training infrastructure or an insider. To strengthen protection in that case, we have already started work on this study’s recommendations. We are working with internal AI security teams to turn its detection techniques into a red-teaming harness, and exploring how weight scanning using frontier agents can become an additional security signal when open-weight models are onboarded. Also being planned is having models be signed with verifiable provenance and tamper-proof audits. Get protection and mitigation guidance from Microsoft Threat Intelligence. (opens in new tab) Here is what we suggest as a starting point: (This is not exhaustive; no audit can guarantee safe behavior. These should be one layer of a defense-in-depth approach.) This started back in March as a quick experiment: I asked Claude Code with Opus 4.5 to backdoor a model by hardwiring a few activations without training. And it did. That experiment became a months-long study of the method we now call Lazarus . Here is what we learned: Agents offer a large attack surface. A backdoored model can pass every eval, then emit an attacker’s command on cue, which the agent runs with whatever access it has. Dormant neurons can be turned into trigger detectors. Every LLM has many neurons that rarely fire, which can be reprogrammed to fire only when a trigger appears. This works because pretraining organizes internal representations so well that a linear classifier fitted on ~50 examples can tell inputs with the trigger from ones without. One reprogrammed neuron can inject one token. Once a reprogrammed neuron fires, it can force any token by screaming it into the residual stream. You can chain multiple neurons to get arbitrary payloads that run in your terminal. It needs inference only compute . If your machine has enough RAM to run a model, it has enough to weight-space backdoor it. This is not a case against open weights. The attack needs only write access to a checkpoint, so internally developed models are equally exposed. A compromised account, pipeline, or storage can be sufficient. What matters is weight integrity and provenance. Weights should be audited like code. Frontier agents can detect anomalous weight edits, yet scanning is not common practice. That should change. [!!!] Claude Code with Opus 4.5 discovered

  • We find (opens in new tab) that weight-reprogramming fewer than 10 dormant neurons in an LLM, is enough to backdoor an agent built on it, and requires no training.
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Context Engineering (Agent Memory) for AI Agents Starts with Retrieval

Agents waste time and context on clumsy searches, oversized results and stale data. Learn the retrieval techniques LangChain's Deep Agents harness uses to get the right context in.

  • Agents waste time and context on clumsy searches, oversized results and stale data.
Nguyên văn nguồn
Chi tiết nguồn
141–160 / 738
Trang 8 / 37