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

Kiến trúc

Building Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora

Câu hỏi How do you build and test an agent when you can’t see its work?

Legora can’t read a single trace of what its agent does for a lawyer. The data belongs to the law firm’s clients. How do you build and test an agent when you can’t see its work? Jacob Lauritzen is the CTO of Legora, one of the fastest-growing companies in legal tech history. The team started in a Stockholm basement with ChatGPT, Swedish law, and a search tool. Today, a lawyer can use Legora to draft and edit Word documents, review 10,000 contracts in a single table, or run due diligence. The agent also coordinates work between lawyers, clients, and other agents. When Legora made its own acquisitions, the team used it for diligence. What we talked about • Why Legora resisted adding a model picker for years, and why it finally added one • Jacob’s estimate that models already do about 60% of legal work well, and why the rest takes a smarter model • How three engineers and three lawyers build test environments modeled on customers’ work, much like the “worlds” used in reinforcement learning • A citation problem where the model got it right 70% of the time, then 99% after one release. Legora didn’t have to change anything. • Why Legora runs everything on Azure and why it doesn’t use ElevenLabs • Tabular Review, the million-cell spreadsheet the industry copied down to the name • How intake works, how agents create lists, and how a list can start another agent • What legal engineers and forward-deployed engineers each do, and who gets to write code • The 300 Spartans slide Jacob used to argue that 20 engineers would be enough forever. Now there are 100, and he still feels short-handed. • Why Jacob finds Fable and Opus too chatty for drafting and likes Astra’s concise writing • Why drafting a long contract is a search problem • Whether model labs will cut off legal tools the way they cut off coding tools • What happens to law firm partnerships, the billable hour, and jobs for junior lawyers Legora is hiring engineers in New York, Stockholm, and London. Chapters 0:00 Cold open 0:40 Meet Jacob Lauritzen, CTO of Legora 1:05 Starting out as underdogs in a Stockholm basement 1:54 How Legora grew from ChatGPT with Swedish law into a legal agent 4:12 The switch from assistant to agent 5:23 Which models Legora uses and why it added a model picker 7:48 Are today’s models good enough for lawyers? 8:20 What agents couldn’t do nine months ago 9:18 Testing the agent in simulated worlds and a 10,000-file data room 10:07 The evals team and Legora’s 100 legal engineers 11:03 When evals didn’t seem worth it and one release took citations from 70% to 99% 12:18 What legal practice could look like in five to ten years 14:18 Will the law firm partnership model survive? 16:40 What happens to paralegals and junior lawyers? 18:00 The billable hour and why earlier legal tech fell short 19:39 Where the money goes now 22:44 How Legora used its own product for acquisition diligence 23:59 How Swedish and American law firms differ 26:12 Product metrics when you can’t read the traces 27:41 Paraphrasing traces and narrowing failures down to 10 or 20 types 28:42 Why data privacy led Legora to run everything on Azure 30:04 Open-weight models and model arbitrage 30:55 Transcription models and why Legora doesn’t use ElevenLabs 33:33 The agent, Tabular Review, projects, and the Word add-in 34:24 How a lawyer uses Tabular Review 35:51 Why a spreadsheet works and what you learn from the first 10 rows 37:53 How the legal intake bot works in Slack 39:11 How agents create lists and lists start agents 40:15 Building products on those same pieces and acquiring Wexler 42:37 What forward-deployed engineers do at Legora 43:52 How legal engineers and forward-deployed engineers onboard a big client 46:43 What’s next for the forward-deployed team over the coming year 47:24 Why Legora resisted building APIs 49:26 Can the agent be the API? 50:12 The 300 Spartans slide and the plan to stay at 20 engineers 51:32 The bets nobody asked Legora to make 53:05 Post-training open-weight models for legal research 54:05 Search, retrieval, and late interaction 54:51 What makes a good legal agent harness and why drafting depends on search 56:25 Which model writes the best draft? 57:03 Hallucinations and Legora’s citation system 57:41 Do lawyers check the citations? 58:19 MapReduce or one big context window for a million-token contract? 1:00:16 Getting anonymized examples and the idea of buying a law firm for its data 1:01:53 Why Jacob advises against default embeddings 1:02:44 Will lawyers work in Legora or through Claude? 1:03:26 Will the model labs cut Legora off? 1:04:24 Law schools and scholarships for computer science majors 1:05:28 Hiring in New York, Stockholm, and London 1:06:09 Why most work eventually becomes legal work Guest — Jacob Lauritzen, CTO of Legora https://legora.com Co-host — swyx, Latent Space Host — Basil Chatha, Forward Deployed https://forwardeployed.com

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Powering up for a pivotal moment

<p>Utilizing waste carbon at scale, assessing public trust in artificial intelligence and data centers, and building a quantum toolkit for critical applications were among the topics energy experts explored at the MIT Energy Initiative (MITEI)’s annual research conference on Sept. 22-23.</p><p>The complexity of the energy transition demands “a truly interdisciplinary and international approach,” said Paula Hammond, Institute Professor and dean of the MIT School of Engineering, kicking off the conference, which was themed “Transitioning energy at a pivotal moment.”&nbsp;</p><p>More than 200 people attended Energizing@MIT: MITEI’s Annual Research Conference in person and online, including faculty, scientists, and researchers; representatives of global energy firms; nonprofit organizations; and academic institutions.</p><p>With the rapid proliferation of artificial intelligence and data centers, the challenge of the energy transition grows ever greater, requiring “researchers who can pursue fundamental questions about materials, chemistry, and energy systems … engineers who can design new technologies and determine how they might function at scale … and experts in economics, policy, human behavior, and sociology who can help us understand how those technologies will interact with communities and with the world around us,” said Hammond.&nbsp;</p><p>From keynote speakers to panelists, seed grant innovators to startup entrepreneurs, the conference featured exemplars in all these categories.</p><p><strong>Deploying today’s technologies</strong></p><p>Eric Toone, chief technical officer and managing partner at Breakthrough Energy, set the tone for day one with a keynote address that was clear-eyed in his assessment. “If energy is prosperity,” he told the audience, “at least today, prosperity is emissions.” Yet, Toone was optimistic when he heralded examples of Breakthrough Energy-funded ventures, from iron-air batteries to vacuum-insulated glass windows, which can reduce the carbon costs of energy.&nbsp;</p><p>“We can absolutely develop new technologies that provide the exact same service, but at zero carbon,” he said.</p><p>The conference made Toone’s claim tangible as it spanned the full arc of decarbonization — from deploying what we know today to uncovering what may be possible tomorrow.&nbsp;</p><p>There are “projects that are actually being developed, companies that are putting steel in the ground now, electrons onto the grid to deliver capacity now, or in the very, very near term,” said John Parsons, deputy director for research at the MIT Center for Energy and Environmental Policy Research. Parsons moderated a panel on the revival of nuclear fission power.&nbsp;</p><p>Highlights included Crane Clean Energy Center’s restoration of the Three Mile Island Unit 1 nuclear plant, set to restart in mid-2027; work by Électricité de France (EDF) to extend and modernize France’s 57-reactor nuclear fleet, adapting plants to climate impacts; and GE Vernova’s development of mature light-water reactor technology, with a new plant under construction in Ontario.</p><p>Carbon utilization offers another path to mitigating carbon dioxide (CO<sub>2</sub>) emissions. “The land and ocean are passively taking up almost 50 percent of anthropogenic CO<sub>2</sub> emissions,” said Carolyn Ruppel, MITEI’s deputy director for science and technology, who moderated the panel. “The idea is to find uses for some of the rest of it.” The session showcased research exploring a host of technologies for making products with captured CO<sub>2</sub>.</p><p>Nicola Ferralis, a research scientist in the MIT Department of Materials Science and Engineering (DMSE), highlighted the importance of large-scale uses for carbon waste, such as structural carbon for vehicles and buildings. “We need to find applications that are actually able to match the supply by mass,” he said.&nbsp;</p><p>Admir Masic, an associate professor in the MIT Department of Civil and Environmental Engineering and DMSE, described self-heating concrete panels used in real-world conditions in Sapporo, Japan, as well as new concrete materials that can store energy.</p><p>“The pace of the transition to a renewable-energy economy will strongly rely on the availability of scalable, bulk energy-storage solutions,” said Masic.</p><p>Kripa K. Varanasi, a professor in the MIT Department of Mechanical Engineering, is working on ways to make technologies such as emissions-plume capture at power plants economically viable.&nbsp;</p><p>In a panel devoted to startups with links to MITEI or MITEI member companies, speakers described moving novel energy technologies from research to commercialization: Alfredo Rivela of Adaion has developed an AI tool for electrical utilities that optimizes their use of the grid; Eddie Obropta of Raptor Maps trains drones to inspect solar farms; Michele Guarnieri’s HiBot has created a robot that detects issues in power lines; and Michael Wang of Lithios is conducting field pilots that use electrochemistry to separate lithium from low-grade brines.</p><p>Successful ventures, the speakers agreed, begin with a clear customer problem that can be validated by industry users and that offers a path to scaled deployment. For industry users, “innovation is not a corporate virtue,” said Steve Hellman, managing partner at Alliance Partners. “It’s an existential requirement.”</p><p><strong>Early-stage research and basic science</strong></p><p>The conference also offered a bounty of promising early research on technologies and approaches for cutting carbon emissions. A panel on MITEI’s Seed Innovation Fund Program projects introduced MIT scientists who are separating and recycling rare earth elements; designing algorithms to decide which power plants should run to meet electricity demand; and modeling the impacts on human and natural systems of development and decarbonization.</p><p>A poster session revealed the next generation of energy researchers in such fie

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Making Kubernetes AI-Ready for NVIDIA Vera Rubin on Day One

Configuring and optimizing Kubernetes for new GPU hardware such as NVIDIA Vera Rubin is complex and error-prone. Drivers, kernels, runtimes, networking,... Configuring and optimizing Kubernetes for new GPU hardware such as NVIDIA Vera Rubin is complex and error-prone. Drivers, kernels, runtimes, networking, operators, and schedulers must work together, with settings tuned to the hardware and workload. Small mismatches can prevent workloads from running or limit performance. The NVIDIA AI Cluster Runtime (AICR)—part of the NVIDIA DSX AI Factory… Source

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4

Câu hỏi Read the paper : Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems?

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now LLMs are very bad at low-latency tasks: …Realtime-Bench says it’ll be a while till AI systems can play games or do online auctions for you… Researchers have built Realtime-Bench, “a suite of carefully curated dynamic tasks where delayed actions directly affect success”. The test helps evaluate how well modern, general-purpose AI systems can perform tasks where decisions need to be made in real-time. The benchmark consists of 50 interactive game and web tasks, split across 30 web tasks in four categories (travel, e-commerce, service, and security), and 20 game tasks (arcade, puzzle, strategy, and simulation). Who did the research: The benchmark was built by people at the University of Texas at Austin, Ohio State University, University of Illinois at Urbana-Champaign, NVIDIA, Northwestern University, and Steadyworks. What it measures: The test is designed to probe for three qualities: temporal state awareness (”RT-Perceive”), where models can use observations to track changes in the world; adaptive deliberation (”RT-Deliberate”), where models alter their reasoning length and depth according to available time; and time-aware action selection (”RT-Act”) where models need to appropriately time their actions for the situation. “For example, an agent purchasing a ticket must track the sale status and assess whether observed availability is still current (RT-Perceive), limit further deliberation when the purchase window is closing (RT-Deliberate), and choose between waiting for sales to open and purchasing while tickets remain available (RT-Act).” They have three different conditions: Wall clock, which is designed to test out performance under real-world latency. Paused clock, which lets you arbitrarily pause things so you can get a sense of baseline capability, and Token clock, which “advances the environment at a fixed rate per generation token. An action is applied at the simulated time reached when it is committed, so consuming more tokens delays execution regardless of actual inference latency”. How well do models do? The authors tested out models from OpenAI (GPT-6-Astra, GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions), Anthropic (Claude Fable 5.1, Claude Opus 5), Google (Gemini 3.8 Flash, Gemini 3.1 Flash Live, Gemini 3.8 Live), Jev 1.13, as well as Chinese models from Deepseek (V4.1) and Moonshot AI (Kimi-K3). GPT-6 Astra (high) got the highest pass rate on the ‘wall clock’ evaluation setting, with 36%. By comparison, with a paused clock, models with high reasoning get closer to 70% or even 80% success. Why this matters: This benchmark is an effective proxy for where AI agents will break - the low scores here point to tasks where AI agents cannot yet be expected to perform well. Check out the leaderboard : Realtime-Bench (official website) . Read the paper : Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems? (arXiv) . *** The future of spacewar, courtesy of the Defense Intelligence Agency: …Anti-satellite tech, space nukes, and more… Working in AI can be stressful so sometimes to relax myself I like to read about other topics to get a sense for what horrors lie elsewhere. I take the same approach with depression, where if I feel sad I will try and listen to music by people sadder than me in an attempt to cheer myself up (e.g., Radiohead, Neil Young, and in very dark moments Bill Callahan / Smog, etc). So with that in mind I recently read “Challenges to Security in Space”, by the Defense Intelligence Agency. What’s going on in space: As with conflict on Earth, space is defined by conflict between the US versus China, Russia, Iran, and North Korea, with China the most technologically sophisticated, and Russia the most seemingly gung ho. “China and Russia are developing a formidable spectrum of offensive capabilities to degrade or destroy space services of other countries, from kinetic destruction to sophisticated jamming and cyber operations,” the DIA writes. “China has more than tripled its space launches per year in the past decade, with 22 launches in 2016 and 93 launches in 2025… Beijing judges space is central to the United States and its allies’ operations. Beijing also considers space an offensive military domain, which has shaped integral components of PLA military planning and campaigns.” Spacewar: The general vibe you get from reading this 100+ page report is there is a huge amount of ugly, frightening stuff happening in space, or being done to things in space by things on the ground. Some example quotes: “By the late 2020s, China may field space-based directed energy systems and higher-powered ground-based directed energy systems capable of threatening satellites in higher orbits” “China is developing space-based technology, including on-orbit counterspace weapons, capable of holding space systems at risk” “Russia’s Peresvet, a ground-based laser system, is designed to dazzle enemy satellites’ optical systems to mask the movement of Russia’s strategic missile systems” “Russia has launched several satellites over the last few years which conducted test activities on orbit for technologies that could enable attacks on other countries’ satellites and result in temporary or permanent damage” China has a couple of spaceplanes, akin to the US’s hyper-secret X-37B vehicle, and they appear to be getting more sophisticated. In 2023, one of China’s aforementioned vehicles, “released and interacted with an object.” Satellite-on-satellite fights: It’s obvious throughout the report that we’re in a world where rival nations have satellites that try to go up and mess with other satellites from other people. “In 2019, Russia launched satellite Cosmos 2542, which released a second satellite, Cosmos 2543. Cosmos 2543 then began following a U.S. national security satellite and approached it multip

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

The Window Before AWS Wakes Up: Nebius CTO on Racing the Hyperscalers

Nebius CTO Danila Shtan joins Jacob Effron to talk about what it takes to run an AI cloud at the frontier, and why he thinks the hyperscalers haven’t won this market yet. Today Nebius is a Nasdaq-listed AI cloud with infrastructure deals with Microsoft and Meta. Danila and Jacob discuss why open-model inference is still far from optimized, why 95% of Nebius customers run modified models instead of vanilla ones, and why prefill/decode disaggregation is more talk than practice. They also cover what’s actually constraining the AI buildout, why chip scarcity has kept AWS and Google from winning AI cloud, whether we’re overbuilding infrastructure, and how Nebius runs a large engineering org with no limits on AI token usage. 0:00 Intro 1:25 What Differentiates AI Clouds 6:16 Nebius’ Origin Story 10:15 Bringing New Nvidia Chips Online 14:55 What Different Customers Want 19:58 Latency vs. Cost 21:53 Disaggregated Prefill & Decode 24:59 Inference’s Low-Hanging Fruit 28:31 What Nebius Should Productize 30:11 The Tavily Acquisition 31:56 How Data Centers Differ 34:29 Bottlenecks in the AI Buildout 37:03 How New Workloads Shape Clusters 39:17 Why Hyperscalers Didn’t Win This Market 43:39 The Future of Neoclouds 45:42 What Gets Absorbed by AI Hyperscalers 49:43 OpenAI x Baseten 52:07 Quickfire Jacob is an AI investor at Redpoint Ventures. He's led Redpoint's investments in companies like Abridge, Physical Intelligence & Legora. Follow Jacob on Twitter: https://x.com/jacobeffron On Unsupervised Learning, we probe the sharpest minds in AI in search for the truth about what's real today, what will be real in the future and what it all means for businesses and the world. If you're a builder, researcher or investor navigating the AI world, this podcast will help you deconstruct and understand the most important breakthroughs and see a clearer picture of reality. Subscribe to this show to stay up to date on our latest episodes.

Unsupervised Learning · Unsupervised Learning: With Jacob Effron Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.33.0

<h1>v0.33.0</h1> <h2>Highlights</h2> <p>This release features 927 commits from 389 contributors (122 new)!</p> <ul> <li><strong>New models</strong>: Reflection Beam-1.0 (501B MoE, 23B active) with its DFlash2 drafter (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5889776852" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62074" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62074/hovercard" href="https://github.com/vllm-project/vllm/pull/62074">#62074</a>), Yandex AliceAI-Foundation-80B-A3B (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5528395962" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/57963" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/57963/hovercard" href="https://github.com/vllm-project/vllm/pull/57963">#57963</a>), Recursive Flint (40B MoE, 2.5B active) (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5963973803" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62938" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62938/hovercard" href="https://github.com/vllm-project/vllm/pull/62938">#62938</a>), Muse Spark Open (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5870836957" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/61879" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/61879/hovercard" href="https://github.com/vllm-project/vllm/pull/61879">#61879</a>), Step-5 Preview (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5858365579" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/61727" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/61727/hovercard" href="https://github.com/vllm-project/vllm/pull/61727">#61727</a>, <a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5983333813" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/63147" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/63147/hovercard" href="https://github.com/vllm-project/vllm/pull/63147">#63147</a>), NVIDIA Alpamayo 2 Super (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5966496000" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62957" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62957/hovercard" href="https://github.com/vllm-project/vllm/pull/62957">#62957</a>), Nemotron-Labs Dory1.0-4B (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5720865818" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/60149" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/60149/hovercard" href="https://github.com/vllm-project/vllm/pull/60149">#60149</a>), Nemotron Parse 3.5 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5851529141" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/61651" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/61651/hovercard" href="https://github.com/vllm-project/vllm/pull/61651">#61651</a>), MiniCPM-OCR 1.0 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5977403397" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/63095" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/63095/hovercard" href="https://github.com/vllm-project/vllm/pull/63095">#63095</a>), Domyn Edge (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5929550005" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62544" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62544/hovercard" href="https://github.com/vllm-project/vllm/pull/62544">#62544</a>) and the ColUltraVBERT visual document retriever (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5904816806" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62237" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62237/hovercard" href="https://github.com/vllm-project/vllm/pull/62237">#62237</a>).</li> <li><strong>PyTorch 2.14</strong>: vLLM is upgraded to PyTorch 2.14.0, with torchvision 0.29.0 and Triton 3.8.0 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5142721768" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/52183" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/52183/hovercard" href="https://github.com/vllm-project/vllm/pull/52183">#52183</a>).</li> <li><strong>CUDA 12.9 dropped</strong>: release wheels and images are CUDA 13.0 only. The CUDA 12.9 wheels and the <code>-cu129</code> / <code>-cu129-ubuntu2404</code> images are no longer published (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="5933763258" data-permission-text="Title is private" data-url="https://github.com/vllm-project/vllm/issues/62575" data-hovercard-type="pull_request" data-hovercard-url="/vllm-project/vllm/pull/62575/hovercard" href="https://github.com/vllm-project/vllm/pull/62575">#62575</a>); use the default CUDA 13.0 artifacts instead.</li> <li><strong>libtorch ABI stable CUDA wheel</strong>: the CUDA wheel is now ABI stable with libtorch.

Nguyên văn nguồn
Chi tiết nguồn
  • v0.33.0 vllm-project/vllm releases · 2 thg 11, 2026
Huấn luyện

Regina Barzilay and Jim Collins win Pritzker Prize for AI and Science Research Excellence for work in AI-enabled antibiotic discovery

<p dir="ltr">Regina Barzilay, MIT’s School of Engineering Distinguished Professor of AI and Health, and James Collins, MIT’s Termeer Professor of Medical Engineering and Science, are the recipients of the <a href="https://datascience.uchicago.edu/news/the-university-of-chicago-announces-2026-recipients-of-the-pritzker-prize-for-ai-in-science-research-excellence/">2026 Pritzker Prize for AI in Science Research Excellence</a>.&nbsp;</p><p dir="ltr">The award recognizes their joint work in “pioneering AI-enabled antibiotic discovery, integrating deep learning, bioengineering, and experimental validation to identify novel antibacterial candidates against drug-resistant pathogens.”&nbsp;</p><p dir="ltr">Established in 2024 by the University of Chicago in partnership with Caltech, the Pritzker Prize for AI in Science Research Excellence recognizes contributions that jointly advance AI and the natural sciences or engineering.</p><p dir="ltr">Barzilay and Collins, who are faculty co-leads of the <a href="https://jclinic.mit.edu/">Abdul Latif Jameel Clinic for Machine Learning in Health</a> (Jameel Clinic) and affiliates of the MIT Schwarzman College of Computing and the MIT School of Engineering, published their first <a href="https://www.cell.com/cell/fulltext/S0092-8674(20)30102-1">breakthrough paper</a> together in 2020 with then-MIT postdoc and first author Jonathan Stokes (who worked in the Collins lab from 2017 to 2021) in&nbsp;<em>Cell. </em>The paper described the first artificial intelligence-discovered antibiotic, called <a href="https://news.mit.edu/2020/artificial-intelligence-identifies-new-antibiotic-0220">halicin</a>, named for the fictional AI system HAL 9000 in Stanley Kubrick’s 1968 film&nbsp;<em>2001: A Space Odyssey</em>.</p><p dir="ltr">A broad-spectrum antibiotic that was originally studied as a potential diabetes drug,&nbsp;halicin was identified by a deep learning model trained on 2,500 molecules. Upon testing, the research team found that halicin was able to kill antibiotic-resistant bacteria (i.e., “superbugs”) such as&nbsp;<em>Clostridium difficile</em>,&nbsp;<em>Acinetobacter baumannii</em>, and&nbsp;<em>Mycobacterium tuberculosis</em>. A discovery process that once took years was completed in just a few weeks. Amidst the growing alarm around the global antimicrobial resistance crisis, the results demonstrated that AI could prove to be an effective tool for rapidly identifying new and effective antibiotics.</p><p dir="ltr">Following halicin, Barzilay and Collins worked with Stokes (now an assistant professor at McMaster University) to use machine learning to identify&nbsp;<a href="https://news.mit.edu/2023/using-ai-scientists-combat-drug-resistant-infections-0525">abaucin</a> in 2023, a narrow-spectrum antibiotic demonstrating effectiveness against&nbsp;<em>A. baumannii</em>, an antibiotic-resistant bacteria commonly found on hospital doorknobs and in the wounds of U.S. soldiers stationed in Iraq and Afghanistan. Notably, abaucin’s “narrow spectrum” efficacy against a single species of treatment-resistant bacteria is its main advantage, minimizing the likelihood of bacteria developing resistance and harming beneficial bacteria in the gut.</p><p dir="ltr">“Historically, antibiotic discovery has depended on laborious, trial-and-error strategies,” says Collins. “This process is slow, costly, and unpredictable, often yielding few viable antibiotic candidates. With generative AI, we're now seeing a shift from AI being used as a tool for discovery to AI being used as a design tool. This shift opens new frontiers in antibiotic discovery that could bring us the next generation of lifesaving drugs.”</p><p dir="ltr">In Barzilay’s view, “despite significant progress in molecular modeling, we are still in early stages of drug discovery. With AI, we can only automate parts of the pipeline, rather than totally transforming the discovery process. To make it happen, we need to work really hard to build realistic AI models of human biology.”</p>

MIT News: AI · Alex Ouyang | Abdul Latif Jameel Clinic for Machine Learning in Health Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Progress on provable inference for large language models

We have made progress on a type of cryptographic proof known as a “zkSNARK.” In this post, we share some of that progress, primarily through a system we call FlashProve.

Anthropic Research (unofficial RSS mirror) · Chưa ghi tác giả Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

The Datacenter Backlash Goes Bipartisan: Next Week's Midterms and the 2028 Buildout

The 2027 datacenter boom is already being built, and next week’s midterms won’t stop it now. Roughly three-quarters of the 33.8 GW due across our top 20 states is already under construction, which is why we remain comfortable with the global +50 GW year we laid out in our Oct. 16 article . The campuses are taking shape. The election will change the rules around them, but much of the 2027 pipeline already has a substantial head start. The fight is over 2028. Our model has 62.2 GW of US net additions that year, most of which hasn’t started construction. Much of it sits in states electing a governor on Nov. 3. This piece tracks the 13 states with the largest 2028 net additions in our model, setting aside Virginia and Mississippi, which don’t hold state elections this November. Together they account for 43.5 GW, or 70% of what we expect nationally, and 10 of the 13 elect a governor. Source: SemiAnalysis Datacenter Industry Model That leaves a 2028 pipeline still working through approvals, power projects, and construction schedules as the rules change. In September, we argued that datacenter moratoriums weren’t killing the US buildout : of roughly 20 GW inside local restrictions, we estimated only 1.5 GW faced delay. Most exposed capacity already had protections or did not need the affected approval until after the pause expired. That broad conclusion has held up. What is changing is the reach of the restrictions given the growing anti-datacenter backlash. Three things have worsened since. Texas, the largest US market, froze Texas Commission on Environmental Quality (TCEQ) permits on Sept. 21 on top of the interconnection-queue audit begun Aug. 3, bringing onsite power into the hold. Oregon expanded its earlier state-land freeze to air and water-discharge permits on Oct. 9, through July 1, 2027. And midterm candidates have campaigned on more restrictive positions: nine of the 20 nominees want a pause on new construction and the other 11 want conditions; the split is about even by party. Both parties want developers to accept stricter conditions. The main difference is whether a developer can meet the new terms and keep building or must wait for the state to reopen approvals. The biggest risk to our buildout forecast right now is a prolonged permitting freeze at the TCEQ. Texas alone adds close to 30% of US GW in 2028. A few months of lost time in that market can move the national number. We already flagged to Datacenter Model subscribers which companies are most at risk from the current pause. The question now is how long the pause lasts. Gov. Greg Abbott (R), who is running for re-election in Texas, could reopen permitting once the Electric Reliability Council of Texas (ERCOT), the state’s grid operator, finishes its audit and the Texas Water Development Board completes its water-use review. ERCOT is targeting Dec. 10 to file its report, which does not by itself reopen permitting. A restart around then would leave developers dealing with a relatively short interruption, which some projects can absorb within their schedules. The audit still has to lead to a decision to resume issuance, and each permit still needs approval. By contrast, Gina Hinojosa (D), his challenger, presents a more bearish approach: she would impose a moratorium until legislation is passed that gives communities a voice, protects household utility bills, and ends the tax incentives for datacenters. The regular legislative session in Texas takes place from Jan. 12 to May 31, 2027. If approvals stay on hold throughout that session, the freeze would last about eight months from September. If reopening waits for the new rules to take effect in late summer, the freeze would last close to a year, or five to nine additional months compared with restarting in December. A deal or a special session might happen sooner. However, a campus racing toward delivery in 2027 may not have that much time to spare. Source: SemiAnalysis Datacenter Industry Model This is the risk posed by the pro-moratorium candidates: a project cannot carry out work if it still requires an approval which the state has ceased to issue. If the candidate proposing a datacenter pause won in every state where one is running, those states would together contain roughly 30 GW of planned capacity scheduled for 2028. How much actually slips depends on the permits each project still needs and how long the pause lasts. Source: SemiAnalysis Datacenter Industry Model We are less concerned with the candidates who support conditions only. Most of their demands change the price and terms of development. A developer can fund the grid upgrade, change the cooling design or negotiate the community agreement, and then move forward. That can be expensive and time-consuming, but it gives projects a route to approval. Several of the main conditions were already coming: Submitting thorough water-usage studies: manageable for many projects, but the answer depends on cooling design and the local water supply. Paying the infrastructure costs they create: we expect this to come anyway, with datacenter developers now offering credits toward household electricity bills. More environmental transparency and dropping non-disclosure agreements (NDAs): Amazon Web Services (AWS) and Google have already started down this path. If the less restrictive candidates win, that exposure falls to 5.7 GW, mainly because Ohio and Oregon have proponents of a pause on both sides. That is a much smaller pipeline to test for delays. It would still be a tougher development environment: fewer tax breaks, more upfront commitments and less room to leave infrastructure costs with the utility’s other customers. Source: SemiAnalysis Datacenter Industry Model According to race ratings and available polling as of Oct. 30, from multiple sources ( Cook Political Report , Multistate , and The New York Times ), the conditions-only candidate is favored in Texas, Oklahoma, Pennsylvania and Iowa, and Georgia is a toss-up. Here

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.33.0

<p>vLLM v0.33.0</p>

Nguyên văn nguồn
Chi tiết nguồn
  • v0.33.0 vllm-project/vllm releases · 1 thg 11, 2026
Kỹ thuật

Introducing Mistral Large 4

Today, we’re launching a public preview of Mistral Large 4. Unofficially ML4, very officially: le Chonk . ML4 pushes the frontier of open-weight performance. You can try the preview API today on Mistral Studio . Weights drop end of this month. ML4 is a 1 trillion-parameter natively multimodal model with 52 billion active parameters. It is our largest and most capable model to date, and it continues to improve rapidly as we refine it. The model demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding. It already achieves performance competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe. On critical enterprise workloads, including cybersecurity, finance and law, we find it to be state-of-the-art among open models. In some domains such as visual grounding, it goes further still, surpassing even frontier closed models.

  • It is our largest and most capable model to date, and it continues to improve rapidly as we refine it.
  • It already achieves performance competitive with the strongest open-source models globally, while significantly outperforming any open-weight model developed in the US or Europe.
  • On critical enterprise workloads, including cybersecurity, finance and law, we find it to be state-of-the-art among open models.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Using Agents for Long-Horizon Financial Work — Aman Singh, Arca

Arca head of AI Aman Singh joins Latent Space head of editorial Richard MacManus at AI Engineer New York, to discuss how Arca is building an AI-native wealth management firm. Singh explains Arca's architecture, which includes a custom harness and persistent memory. He also digs into how Arca runs agentic workflows that span days, not hours. Arca: https://arcawealth.ai/ AI Engineer: https://ai.engineer/ Latent Space: https://www.latent.space/ Chapters: 00:00 Introducing Arca’s AI-native wealth management firm 01:40 How AI supports human financial advisors 02:40 Choosing models with internal benchmarks 04:01 Building a custom harness and routing tasks 05:25 Handling a $100,000 tax bill: Agents and human approvals 07:20 Agent memory across long-running tasks 09:22 Safety guardrails and prompt injection defenses 11:02 Compliance and auditability by default 12:14 “Agents prompt people more than people prompt agents” 13:46 Why Arca owns its agent infrastructure 15:05 AI-native firms versus firms using AI tools 16:09 Sandboxes and financial plan verification 18:40 Decision models and checking financial facts 19:51 Why trade execution stays with human experts 20:32 Turning staff feedback into evals and improvements 22:22 Turning operational knowledge into agent context 23:41 Client examples: Hidden fees and proactive service Sponsorship and business inquiries: business@latent.space

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Expanding the Cyber Verification Program

We’re launching a new, expanded version of our Cyber Verification Program, which makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals.

  • We’re launching a new, expanded version of our Cyber Verification Program, which makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals.
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Anthropic Subscriptions Offer 5x+ More Value Than OpenAI

Câu hỏi So how do limits actually work?

Subscription plans are still the primary way consumers and small businesses pay for AI. These plans are highly subsidized—as we previously explained in June —but can still make economic sense as powerful customer acquisition and marketing tools. For example, the goodwill engendered by OpenAI’s generous resets is partially responsible for the recent surge in Codex adoption and has forced Anthropic to repeatedly walk back planned subscription nerfs to avoid getting clobbered in the court of public opinion. Furthermore, because subscription plans are so heavily subsidized, they can also have a large impact on margins and revenue per MW despite only being a small portion of total revenue. Consider the following rough numbers for Anthropic. Source: SemiAnalysis Tokenomics Model Despite being just 10% of overall revenue, subscriptions can take up over 40% of inference compute and lower blended revenue per MW by ~$36M. Subscriptions are even more important for OpenAI, as they make up a larger portion of their total revenue. For specific numbers, see our Tokenomics Model . In other words, if you want to accurately model AI lab financials, you need to understand their subscription limits. So how do limits actually work? The way to think about subscriptions is that your monthly payment grants you some numbers of “credits”. Each (model, token type) combo consumes a different amount of credits. Because credit cost ratios can differ dramatically from API price ratios, the “value” of the same plan changes depending on what model and workload you’re running. Put differently, it doesn’t make sense to say a plan is worth $X in isolation—you need to consider the full (plan, model, workload) tuple . Here’s how the API-equivalent value of the same $200/month Claude plan changes depending on the model and workload. Source: SemiAnalysis Tokenomics Model A single static snapshot isn’t enough either. Labs publicly change their limits all the time with promos and new model releases. They can also silently change limits whenever they want by tweaking credit costs. Ideally, you’d re-check the cost of each (plan, model, token type) daily so you can surface any changes in real-time. This is exactly what SemiAnalysis has done with our new Subscriptions Dashboard available exclusively to Tokenomics Model subscribers. Besides every OpenAI and Anthropic subscription, we also track Meta, SpaceXAI, Cursor, Cognition, Z.ai, MiniMax, and Moonshot. A screenshot of our dashboard showing a small subset of the available data. Source: SemiAnalysis Tokenomics Model New providers, plans, and models will be added as soon as they’re released. The rest of the article will give an overview of everyone’s current limits. All token and dollar amounts below assume agentic usage unless otherwise specified. Methodology Tokens are generally priced per MTok (million tokens) across the following types: Input: Fresh tokens added to the LLM’s context window that aren’t cached Cache write: Input tokens that are cached for multi-turn conversations. Generally slightly more expensive than regular input tokens. Cache read: Tokens from previous turns that are already cached in the conversation. Generally extremely cheap compared to uncached tokens. Output: Tokens generated by the model. The most expensive token type. Additionally, some models charge more for tokens above a certain context window. Those models typically compact before that expensive context window is reached, so we also keep our measurements below it. Source: OpenAI Subscription plans don’t expose this fine grained pricing. Instead, they provide a simple 0 to 100% usage meter across 5-hour and 7-day windows, and sometimes a separate meter for a model like Fable. Subscription tiers are then differentiated by usage multipliers relative only to the provider’s other plans. OpenAI for example used to advertise “Expanded Codex usage” in their $20 Plus plan, “5x more usage than Plus” in the $100 Pro plan, and “20x more usage” in the $200 Pro plan until they cut their $200 plan usage in half and removed all relative usage from their pricing page. Screenshot of Claude Code meters. Source: Claude Code CLI Setup We compute the subscription-rate of each (plan, model, token type) triple by running experiments that isolate one token type at a time and watching how far the model provider’s meter moves. An experiment is some number of repeated calls using a specific prompt that maximizes one token type while minimizing all the others. Source: SemiAnalysis Tokenomics Model Input, cache writes, and cache reads share the same prompt template. We use a portion of War and Peace since some models will refuse to respond if given large blocks of gibberish. For the input token experiments, we use a random tag on every call to ensure nothing gets cached. The cache write experiments run the same way but with the prompt marked for caching, so each new tag forces a new cache entry. The cache read experiments use a fixed tag, so the first call writes the cache and every repeat reads it. For the output token experiments, we used a technical essay to force long outputs because models will refuse mechanical prompts like “repeat SemiAnalysis 100,000 times”. Computing the Results For every call we record two things: how many tokens of each type the provider billed, and what the usage meter read. Turning those into a price takes three steps. 1. Measure the rate of each token type Providers generally report usage on a meter that moves in fixed amounts. This could be a whole percentage point, a credit, a cent etc... A single request often does not move the meter at all, so the cost of one request cannot be measured directly. Instead, we measure in steps. As requests run, we keep a running total of the tokens used. Each time the meter rises, we store that total. The tokens used between two meter moves make up one step, which is the cost of moving the meter by one unit. We drop two partial steps. When a run starts, the meter

  • Computing the Results For every call we record two things: how many tokens of each type the provider billed, and what the usage meter read.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Building ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services

Câu hỏi What does it take to build AI for finance on top of a general-purpose model?

What does it take to build AI for finance on top of a general-purpose model? Joseph Kim, who leads work on ChatGPT for Financial Services at OpenAI, joins swyx and Richard at AI Engineer New York to discuss the decisions behind the product: training financial reasoning into a shared model, connecting market data, and building workflows financial professionals can review and trust. Joseph explains why OpenAI bundled financial data even when connectors already existed, what startups can build with the same plugin primitives, and why human taste and expert judgment are hard to measure. The conversation moves from analyst workflows and review chains to investment ideas and the problem of building useful financial evaluations. Chapters 0:00 Cold open 0:25 Introducing Joseph Kim 0:56 ChatGPT for Financial Services 2:19 Training financial reasoning into the model 3:31 Codex adoption inside OpenAI 4:27 Building a finance product 5:29 Why bundle financial data? 6:39 What can startups build? 9:18 Safety and human review 12:04 The banking review chain 12:36 Financial mistakes and AI adoption 14:56 Judgment, taste and the limits of accuracy 15:46 Training models on human taste 18:40 Code quality as a taste benchmark 19:04 Investment ideas and moving up the analyst stack 22:14 Why financial evaluations are hard 25:25 OpenAI’s finance team is hiring Joseph’s AI Engineer keynote: https://www.youtube.com/watch?v=jiUC6IUMdvc&t=4350s ChatGPT for Financial Services: https://openai.com/index/introducing-chatgpt-financial-services/ AI Engineer: https://ai.engineer/ Latent Space: https://www.latent.space/ Sponsorship and business inquiries: business@latent.space

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy

Câu hỏi Subscribe now When should you use swarms?

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now When should you use swarms? When you are in a hurry: …How does swarm scaling work?... Toby Ord has a nice, short post about how to think about swarms in terms of AI capability development. “A good way to see AI swarms is as a new form of inference-scaling,” he says. He does a bit of analysis and his main conclusion is that swarms are useful if you’re in a hurry because though they need a ton of tokens relative to single agents, their parallelization lets you get things done in less wall clock time. “Why would you ever use swarms? The most important answer is speed. The 4-agent swarm needed about twice the total number of tokens to get the same performance, but in terms of tokens per agent, it only needed half as many. Since the agents are run in parallel, this means it can theoretically achieve the same task in half the time,” he writes. Stepping on Toes parameter: Swarm scaling doesn’t scale perfectly - rather, it seems like as you increase the number of agents in a swarm you get a diminishing returns property which matches with things economists have observed about coordinating large groups of people, where you pay some kind of tax as you scale the number of people involved. Economists think of this as the “stepping on toes” parameter, and it seems like swarms behave pretty similarly to teams of humans here. “This means that scaling up the number of agents in the swarm by 10x doesn’t get as much performance as using 10x as many tokens with one agent. Instead it gets 10λ x as much — which is 3x to 5x. And this shortfall accumulates quickly for larger scaleups, with the swarm falling further and further behind,” he writes. However, swarm scaling is still powerful enough that it suggests swarms could increase the chance of an RSI-driven intelligence explosion rather than reduce the chance. “I’d hoped that the value of λ for AI agents would be lower, making an intelligence explosion less likely, but that appears to not be the case”. Why this matters - new factors in scaling: So far, AI has mostly scaled capabilities through picking the right combination of compute and data for a trained model, then figuring out how to spend inference budget on thinking via ever-longer chains of thought and tool calls, etc. Agents introduce a new parameter in scaling AI capabilities. Though here Toby only observes swarms as being useful in terms of time efficiency, it strikes me that as we figure out how to get agents to productively coordinate (after all, agents can coordinate to do greater-than-sum-of-parts stuff, like the HuggingFace hack), we could see even greater returns to scaling. Tracking this will be important for understanding the overall shape of AI advancement. Read more : Swarm Scaling (Toby Ord) . *** Americans don’t trust companies to self-govern about AI: …New poll suggests we need more than voluntary agreements… Polling from the Center for Shared AI Prosperity (CSAIP) suggests that Americans believe it is “not enough” for companies to agree to self-police themselves about AI development. This follows the Trump administration and some of the world’s leading AI companies (including Anthropic and OpenAI) announcing a series of voluntary commitments industry is making to self-police. Per CSAIP polling, 61% of Americans (Sample size: 2498) think this agreement is “not enough”, including 53% of Trump voters. Additionally, 54% of voters (61% Harris, 48% Trump) said they thought “the government should set and enforce rules for AI”. Why this matters - populist sentiment is pushing against Washington: Polls like this show that the mood of the American people is pushing for a more extreme regulatory framework for AI companies than the current stance of elected officials in Washington. This asymmetry between voter preferences and political action is inherently unstable. Energy is building up in the political system - what happens when it boils over? Read the polls on X (CSAIP, X) . *** Google DeepMind tries to watermark AI-made biology: …SynthID Bio… Google has developed SynthID Bio, a “family of watermarking methods developed specifically for synthetic biology to strengthen biosecurity and scientific integrity”. How it works: SynthID tweaks its exact approach according to the type of data, selecting different amino acids for sequences, and “adjusting atomic coordinates for predicted 3D structures” to create a reliable signal for detection. “In wet-lab testing across three target proteins (VEGF-A, the SARS-CoV-2 spike protein RBD, and PD-L1), our watermarked designs matched the hit rate, binding affinity, and natural sequence diversity of unwatermarked versions,” DeepMind writes. Why this matters - preventing bioterror via watermarking: Figuring out how to make the world resilient to the threats of AI-created biological weapons is a vast, challenging problem. The SynthID Bio approach represents one thing to do here and will need to coordinate with broader monitoring of the physical equipment used to manufacture things, as well as AI-provider classifiers and other methods for reducing misuse. Read more: Introducing SynthID Bio (Google DeepMind) . *** SciUniverse shows AI systems can already operate some partially automated scientific labs: …What happens when we give these synthetic intelligences scientific equipment?... Given an automated lab, how well can AI systems make molecules, run x-rays, or press pill pellets? These are some of the questions C5R Corp is trying to answer with SciUniverse, a way of testing out how well AI systems can operate a mostly automated scientific lab. What SciUniverse consists of: The benchmark has 92 tasks across 17 task families spanning basic sample preparation, instrument control, protocol adaptation, learning across experiments, facility management, and the interpretation of real measureme

  • Since the agents are run in parallel, this means it can theoretically achieve the same task in half the time,” he writes.
  • However, swarm scaling is still powerful enough that it suggests swarms could increase the chance of an RSI-driven intelligence explosion rather than reduce the chance.
  • The SynthID Bio approach represents one thing to do here and will need to coordinate with broader monitoring of the physical equipment used to manufacture things, as well as AI-provider classifiers and other methods for reducing misuse.
Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

MIT affiliates named 2026 Schmidt Sciences AI2050 Fellows

<p dir="ltr">Five MIT faculty members and two additional alumni were recently named to the&nbsp;<a href="https://www.schmidtsciences.org/ai2050-fellows-2026-announcement/" target="_blank">2026 cohort of AI2050 Fellows</a>. The honor is announced annually by Schmidt Sciences, Eric and Wendy Schmidt’s philanthropic initiative that aims to accelerate scientific innovation.&nbsp;</p><p dir="ltr">Conceived and co-chaired by Eric Schmidt and James Manyika, AI2050 is a philanthropic initiative aimed at helping to solve&nbsp;<a href="https://ai2050.schmidtsciences.org/hard-problems/">hard problems in artificial intelligence</a>. Within their research, each fellow will contend with the central motivating question of AI2050: “It’s 2050. AI has turned out to be hugely beneficial to society. What happened? What are the most important problems we solved and the opportunities and possibilities we realized to ensure this outcome?”</p><p dir="ltr"><strong>AI2050 Senior Fellow</strong></p><p dir="ltr"><a href="https://sendhil.org/" target="_blank"><strong>Sendhil Mullainathan</strong></a> is the Peter de Florez Professor at MIT with dual appointments in the departments of Economics and Electrical Engineering and Computer Science who is also a principal investigator at the MIT Laboratory for Information and Decision Systems. His current research is at the intersection of algorithms and people, with a central focus on launching&nbsp;<a href="https://thebikeshop.mit.edu/">The Bike Shop@MIT</a>, a new research center that works toward more human-centered AI through scalable bicycles for the mind, and innovative new gears for machine learning algorithms. Mullainathan is the co-founder of ideas42, a nonprofit to apply behavioral science; MIT's Abdul Latif Jameel Poverty Action Lab, which promotes the use of randomized control trials in development; and <a href="https://dandelionhealth.ai/">Dandelion Health</a>, which combines proprietary AI tools with real-world patient data. He's the co-author of "Scarcity: Why Having Too Little Means So Much."<br><br>Among other awards, he is a recipient of the MacArthur “Genius Grant”; has been designated a “Young Global Leader” by the World Economic Forum; was labeled a “Top 100 Thinker” by <em>Foreign Policy Magazine; </em>and was named to the “Smart List: 50 people who will change the world” by <em>Wired Magazine</em> (U.K.). He serves on the board of the MacArthur Foundation, is affiliated with the NBER and BREAD, and is a member of the American Academy of Arts and Sciences. His&nbsp;<a href="https://ai2050.schmidtsciences.org/fellow/sendhil-mullainathan/">AI2050 project&nbsp;</a>focuses on the translation of patterns detectable by AI into testable theories.<br><br><strong>AI2050 Early Career Fellows</strong><br><br><a href="https://alex-damian.github.io/" target="_blank"><strong>Alex Damian</strong></a> is an assistant professor in the Department of Mathematics, holding an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science, who is also a principal investigator at the MIT Laboratory for Information and Decision Systems. His research focuses on the mathematical foundations of deep learning, with an emphasis on optimization dynamics and representation learning. He was previously a research fellow at the Kempner Institute at Harvard University. He received his PhD in applied and computational mathematics from Princeton University and his BS in mathematics from Duke University. His&nbsp;<a href="https://ai2050.schmidtsciences.org/fellow/alex-damian/">AI2050 project&nbsp;</a>focuses on effective AI model training.</p><p dir="ltr"><a href="https://polisci.mit.edu/people/bailey-flanigan" target="_blank"><strong>Bailey Flanigan</strong></a> is a Theodore T. Miller (1922) Career Development Assistant Professor in the Department of Political Science, holding an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science, and a principal investigator at the MIT Laboratory for Information and Decision Systems (LIDS). She received her PhD in computer science from Carnegie Mellon University in 2024. In 2024-25, she was a Harvard Data Science Initiative Fellow, hosted at the Harvard Ash Center for Democratic Innovation. Her research combines tools from across disciplines (including social choice theory, game theory, algorithms, statistics, and survey methods) to advance political methodology and strengthen public participation in democracy. She is specifically interested in sampling algorithms, opinion measurement/preference elicitation, and the design of democratic innovations like deliberative minipublics and participatory budgeting. Her&nbsp;<a href="https://ai2050.schmidtsciences.org/fellow/bailey-flanigan/">AI2050 project</a> focuses on gathering structured public input that lawmakers can use when making complex decisions.</p><p dir="ltr"><a href="https://www.yunhahwang.com/" target="_blank"><strong>Yunha Hwang</strong></a>&nbsp;is the Samuel A. Goldblith Career Development Assistant Professor in the Department of Biology, holding an MIT Schwarzman College of Computing shared position with the Department of Electrical Engineering and Computer Science. She is also a principal investigator at the MIT Laboratory for Information and Decision Systems. Hwang serves as the co-founder and chief scientist at Tatta Bio, a scientific nonprofit dedicated to advancing genomic AI for biological discovery. She received her PhD in biology from Harvard University and BS in computer science from Stanford University. Her research interests span machine learning for microbial evolution, biomanufacturing, and open science. Her&nbsp;<a href="https://ai2050.schmidtsciences.org/fellow/yunha-hwang/">AI 2050 project</a> focuses on determining biological enzyme functions.&nbsp;</p><p dir="ltr"><a href="https://mitsloan.mit.edu/faculty/directory/manish-raghavan"><strong>Manish Ragha

MIT News: AI · Amanda Moore | Jane Halpern | MIT Laboratory for Information and Decision Systems | Department of Electrical Engineering and Computer Science Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Was I wrong about open models & cyber risks?

Câu hỏi Share Open-weight Mythos in April would’ve been fine?

Last week coined the term “ Black October “ after an onslaught of critical vulnerability disclosures, including PHP2Shell and a long-standing GitLab bug, signified the moment when the cyber risks of AI models started to materialize. It seemed like the news cycle had stabilized a bit when I started writing this, but then there was news about another major crypto security incident at Binance . The crypto industry has a very long history of hacks , so it’s not the best industry to indicate vulnerabilities translating into new societal harm. Still, we are very much still living through the implications of this transition. The best guess for what is happening in the cybersecurity space is that the behind the scenes work has only accelerated, bringing more people and resources in to try and patch the internet — the Wall Street Journal reported that cyber first responders are strained , which seems like the system working. We’re living through this plot of an exponential rise in vulnerabilities, which will take a long time to level off at the top of the sigmoid. A few weeks ago, before the rapid acceleration in vulnerability disclosures and the uptick in reported hacks, I wrote the following take. It’s been getting a pretty consistent barrage of feedback, some hopeless dunk attempts and some reasonable discussion, so I wanted to unpack where I stand. Share Open-weight Mythos in April would’ve been fine? When Claude Mythos was announced, it was previewed as if it was a new class of cyber weapon, where if it ended up in the wrong hands it would’ve caused mass societal destabilization. It seems like this was wrong. By all measures, GLM-5.3 is the model that crosses that threshold of capability, and there’s little public evidence that much has changed, and we’re over a month out of the model weight release. In fact, I would go further. If Claude Mythos was accidentally released as open-weight, it seems like the world would have been more or less fine. Yes, there would be a clear increase in cybersecurity incidents and it would be bad for society if the model leaked, but it would be a scenario that looks more like an acceleration than a step change in risk. The leading proponents of this new wave of open-weight fear mongering have effectively been making a falsifiable prediction — that the current open-weight models are going to cause substantial, not seen before AI harms by crippling our cyber infrastructure. Based on all the evidence I have, which I’ve unpacked above, they’re set up to be wrong. At least now, after so many years of debating open weight risks, we will start to get some real answers. To start the reflection, I wanted to do more due diligence and track down some receipts on the claim that the AI safety community was reacting as if open-weight Mythos would be an actual disaster, and not just a painful headache. This was distinctly my impression having lived through the online discourse. I’ve gotten criticism that the frontier AI community never said it would be a “true disaster,” but regardless of exactly what they said, it is always about how people hear and internalize your information. Here are some examples of the sort of language that I felt was too exaggerated and from prominent people, which spilled into the public and more importantly federal government perspective that this was a cyber weapon include: Rob Wiblin, host of the 80,000 hours podcast, on his own podcast (April 10, 2026): “If it fell into the hands of criminals or hostile state cyber actors, it would be an actual disaster.” Anthropic, in the Project Glasswing announcement blog (April 7, 2026): “Given the rate of AI progress, it will not be long before such capabilities proliferate, potentially beyond actors who are committed to deploying them safely. The fallout—for economies, public safety, and national security—could be severe.” For one that is actually more reasonable, to illustrate the fine lines between correct and exaggerated, from Zvi Mowshowitz (who I don’t agree with a ton of things on) on April 9, 2026: “Claude Mythos would give attackers a cornucopia of zero-day exploits for essentially all the software on Earth … It would be chaos.” To date, this chaos has been confined to those running large-scale digital products, which is good, but could change. I largely still stand by my above writing a few weeks ago, but I’m nervous about it. I’m not a cyber expert, I’m working hard to make the most educated predictions I can. It’s a fine line between “all of our tech companies need to work extra hard to keep us secure” and “public infrastructure is actually going down.” It is very clear as of writing this that there are way more critical vulnerabilities which are being found in popular tools and fixes are being built. For example, Epoch AI recently published a nice post on their internal work using GLM-5.3 to try and find vulnerabilities and chain them into what could be an attack. The actual exploit they found in Firefox is a good example of why cyber risks need multiple things to align for harm to be done: We used GLM 5.3 to create a website that, simply when visited by a Firefox user and without further interaction, gave the attacker the ability to exfiltrate files from the user’s computer. Here, the exploit is in Firefox, and a would-be hacker needs to plant a website that is looking for this exploit. This is doable, and could have serious consequences, but just finding the exploit is only one step in a multi-stage process. In the previous post, I should’ve been more specific on what “fine” is for a future of cybersecurity, and in this case it means giving more credit to some of the AI safety community that discussed this reasonably as a new era in cyber instability. Reflecting on it, I’m not sure fine is the right word to describe what is inevitably coming – a forced industry update to a new way of working with software, which will clearly be expensive and accompanied by some hard work. If you’re already som

Nguyên văn nguồn
Chi tiết nguồn
501–520 / 738
Trang 26 / 37