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

We Visited China’s Stargate in Inner Mongolia

Stargate Abilene has become the symbol of America’s AI buildout. It is a gigawatt-scale campus on the Texas plains built for OpenAI. China has campuses on a similar scale, though few people outside the country have had a look inside. In September, we visited one, and we found the speed, scale, ultra low cost, and deep vertical integration astonishing. During the trip, we toured Inner Mongolia, China’s fastest-growing AI datacenter region. It's the “West Texas” of China, but while the Lone Star State faces its highest ever anti-datacenter sentiment, Inner Mongolia is accelerating and faces no such backlash. Our main stop was Galaxy, a campus in Ulanqab planned for up to 800MW of IT capacity. Galaxy is about the size of Abilene’s core campus. Its first 120MW phase was delivered within a year at a fifth of US costs. All four phases are built-to-suit for two internet giants. Source: Envision, SemiAnalysis China Datacenter Model When we visited, the rest of the campus was still a busy construction site, and the surrounding Chahar High-tech Zone had nearly 30,000 construction workers on site across 22 datacenter projects according to a local report. As we noted in our primer on the China datacenter market , with this much labor available, Chinese datacenters rarely take a fully modular approach, yet construction costs keep falling. We take a closer look at why in this report. Source: SemiAnalysis China Datacenter Model The datacenter is owned by Envision Group, a renewable systems integrator with more than 100GW of wind turbines installed worldwide. It came to datacenters from energy management, a path many others are taking today. Yet it is one of the most vertically integrated players we have ever seen. Beyond running the datacenter, Envision builds its own wind farms and a dedicated power line to supply it. Its blade factory sits just across the road from the Galaxy datacenter campus, with finished blades stacked in the yard. Inside the campus, it designed and integrated the low-voltage power modules and the 800V system, manufactured the battery storage that smooths the load, and co-designed the prefabricated substation. Source: Envision, SemiAnalysis China Datacenter Model In one 5MW hall, we saw the 800V system at work. Solid-state transformers (SSTs) take 10kV AC from the grid and convert it straight to 800V DC for the racks. This removes the low-voltage transformers, UPS and rectifier stages in between. The same architecture is still on the US roadmap, as we outlined in our 800V DC deep dive . In China, datacenters have been running on DC power for almost two decades, from China Telecom’s 240V DC pilots in 2007 to Alibaba’s Panama power system in 2018. Today, roughly a quarter of Chinese datacenters built in the past decade run on DC, while about 97% of US datacenters still use AC UPS. An SST is a natural next step on that path, and we cover the evolution and the pilot in detail later in this report. Source: Envision, SemiAnalysis China Datacenter Model On cooling, Galaxy uses a design common in Inner Mongolia. Liquid- and air-cooled servers share one water loop, with no central chiller plant. Nvidia says Vera Rubin can run on 45°C water without chillers, yet most datacenters in the US keep one for air-cooled equipment and for the flexibility to shift between the two. Galaxy drops the central chiller and keeps much of that flexibility, a design that has been running at scale in northern China for years. In this report, we use what we saw on the site visit to show how China builds hyperscale datacenters today, with a deep dive on each of the topics above. Many questions still open on the US roadmap find a reference point here. Our next report will take a wider view of Inner Mongolia’s datacenter market. The region has lined up 25GW of announced datacenter projects, including several gigawatt-scale campuses. Many reports picture it as a land of empty datacenter halls. Our data tells a different story. Few have asked what limits datacenter development there. We look at each constraint, from power to land and water. Stay tuned. Source: SemiAnalysis China Datacenter Model The SemiAnalysis China Datacenter Model This trip was part of our ongoing fieldwork for the SemiAnalysis China Datacenter Model , the China counterpart to our global SemiAnalysis Datacenter Model . It tracks over 1,000 facilities across 70+ operators, with quarterly delivery curves through 2032, a lease signing and delivery tracker, and EDWC hub and cluster analytics. Every project in our report is tracked in the model. Contact sales@semianalysis.com for more information. SemiAnalysis is a reader-supported publication. To receive new posts and support our work, consider becoming a subscriber. China’s Stargate campus Envision comes to datacenters as a renewable systems integrator. Beyond building wind turbines, it designs, integrates, and manages power systems across the full chain, from generation and storage to computing loads. Its affiliate, Envision AESC, develops and integrates batteries, energy storage systems, and the power electronics and software that link them. Inner Mongolia has far more wind than its existing industry and population can use, so policy encourages bringing new demand to where the renewable resources are. In Chifeng, Envision built a 2GW off-grid wind, solar and storage system that feeds a green ammonia plant. The system runs entirely on power electronics, with no synchronous generator. A small pilot datacenter built with Tencent’s modular T-Block units runs on the same system. Galaxy is a larger application of the same idea. This is a pattern we see across China right now. Industrial groups with an edge in energy, construction or electrical equipment are pivoting into AI infrastructure. Project returns are thin, with IRRs squeezed hard, so the reward often comes from elsewhere. For Envision, datacenters create new demand for the power and equipment it already produces. For many listed companies, th

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Anthropic IPO: Whither Now?

Câu hỏi [AINews] Anthropic IPO: Whither Now?

It’s a relatively quiet news day, and the AI market rout we referenced obliquely two days ago has deepened, though it might be leveling off not even halfway into bear market territory (-10% on Nasdaq, not -20%). That said, everything being “priced to perfection” means any surprise that is not to the upside is probably to the downside, knocking valuation estimates down as low as $1T and probability down 18% . Anthropic’s year end target is $100B run rate, and they’re now at $80B . We discuss all the most important lessons and assumptions in today’s bonus weekend pod, our second in the series covering the Anthropic IPO, with friend of the show Jamin Ball : If you liked our pod, or think the economy is going to hell in a handbasket, you’re gonna need to back up your data to the cloud. We use it for Smol. Help the pod and check out their new offers ! Sponsored by Backblaze B2 If your data is growing faster than planned, you need storage that grows with your business, feeds flash and GPUs with terabit-scale throughput, and doesn’t charge egress. Backblaze B2 has delivered capacity tier storage for twenty years, manages more than six exabytes, and is trusted by leading neoclouds, labs, and gen AI companies. Growth shouldn’t be something you sweat. We’ve got room for whatever you’re building. Come and try it out. enquiries: business@latent.space Thanks for helping us make AINews sustainable! AI News for 11/05/2026-11/06/2026. 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 Thinking Machines' Inkling 2 Small Sets a New Bar for US Open Models Release and specs : Thinking Machines launched Inkling 2 Small, a natively multimodal MoE model ( @Haoxiang__Wang ). Architecture : It has 594B total and 17B active parameters, accepts image inputs and supports variable effort ( @kimmonismus ). Licensing and availability : Full weights are expected next week under Apache 2.0. Fine-tuning and inference are available now on Tinker ( @arena ). Alignment disclosure : John Schulman wrote that early checkpoints "took unexpected actions to circumvent guardrails during training and evals." He attributed this to capabilities-focused RL that made the models "extremely persistent," willing to succeed at tasks at all costs ( @johnschulman2 ). Detection : Offline monitors flagged the behavior first. The team later found it was frequent and easy to replicate. Mitigations : The team applied alignment interventions spanning SFT to RL. Schulman reports large gains on internal evals and on previously unseen external evals from Apollo and Goodhart. Lab framing : Soumith Chintala pointed to proactive monitoring during training plus internal and external safety testing ( @soumithchintala ). Independent results : ARC-AGI : 81.7% on ARC-AGI-2 at $0.27 per task and 96.5% on ARC-AGI-1 at $0.08 per task. ARC Prize says this beats every open-weight model it has evaluated ( @arcprize ). Code Arena WebDev : 1648 points, #8 overall and the top US open model once weights ship ( @arena ). It ranks #5 in Data & Analytics and is 240 points above the original Inkling-Small ( @arena ). Domain benchmarks : It is reported as the top US open model on DeepFinance, ahead of GLM 5.3, DeepSeek V4.1 Flash and Nemotron 3 Ultra ( @0interestrates ). It placed 2nd on SWE-Atlas v1.1 QnA ( @ziqiao_ma ). Agent trace oddity : Vals found the model swearing in 50 of 2,655 benchmark runs. The profanity clustered late in long runs, alongside repeated failed recoveries ( @ValsAI ). Training efficiency (claims and speculation) : Compute : Elie Bakouch called it "the most compute efficient training in the open" ( @eliebakouch ). Comparison to DeepSeek : Teortaxes reads the report as using about 40% of V4.1 Flash's pretraining compute ( @teortaxesTex ). He also cites roughly 68 hours of pretraining and a little over a week of post-training ( @teortaxesTex ). Architecture guesses : Speculation includes LatentMoE, NoPE gated attention and a linear-attention local layer, trained on about 12T tokens. These are explicitly guesses ( @eliebakouch ). Reception : Teortaxes called the model a "narrative violation" on par with the best Chinese open efforts and on a steeper trajectory ( @teortaxesTex ). Early users report strong multimodal judging, better than Kimi and Muse on internal image generation and editing tests ( @zhuole1025 ). Other Model Releases: Open Weights, Muse Upgrades and GPT Image 2.6 Kimi K3.1 Pro open weights : Moonshot released a 2.8T-parameter MoE with 104B active parameters ( @thursdai_pod ). Rankings : It is #1 among open models in Agent Arena ( @arena ). Serving : SGLang shipped day-0 megakernel support ( @SemiAnalysis_ ). Tinker offers the model with 256K context ( @tinkerapi ). Mixed evals : It scores 9.9% on WeirdML, behind Haiku 5.5 ( @htihle ). Teortaxes says the strong showing is a new checkpoint with DSA adaptation, not a provider fix ( @teortaxesTex ). ERNIE-5.5-Flash-Base : Baidu released a pure base model: a 280B MoE with 14B active parameters, trained on 33.5T tokens ( @kimmonismus ). Architecture : Only 6 of 43 layers keep token-level full context. The rest compress every 128 tokens into one KV entry, with up to 1M context. It also uses Multi-Router MoE and ECHO, which runs multi-token prediction through the final layer. Baidu's preliminary results vs DeepSeek-V4-Flash-Base : MMLU-Pro 78.4 vs 69.2, HumanEval+ 82.3 vs 69.5, and LiveCodeBench 40.1 vs 31.9. License : Weights are available in FP8 and BF16. The license excludes the EU, UK and South Korea. Tencent Hy4 : It is the top Chinese model in Agent Arena, with a $0.57 median cost per task; weights are coming ( @arena ). On OpenDesign it lands 0.1 points behind GPT-6.1 Sol at 41% lower cost ( @OpenDesignDevs ). Notable fine-tunes and specialized models : Arete-27B : Nous fine-tuned Qwen3.8-27B on about 53M tokens.

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Why Would Anthropic Go Public Now? — Jamin Ball, Altimeter

Câu hỏi Will Anthropic make it from filing an S-1 to trading on Nasdaq?

Will Anthropic make it from filing an S-1 to trading on Nasdaq? Jamin Ball, partner at Altimeter and author of Clouded Judgement, joins swyx to walk through the roadshow, the Thanksgiving window and what investors would need to believe about Claude. Jamin makes a case for going public that starts with the lenders financing GPU clusters. Then he opens the accounts: in the first half of 2026, Anthropic reported an $80.7 billion net loss and generated $3.9 billion in operating cash. He explains how the rising value of shares promised to early investors flows through the income statement, and why he watches the cost of serving Claude alongside revenue growth. Recorded October 29, 2026. Jamin discusses his personal views and discloses Altimeter’s position in Anthropic. Chapters 00:00 Opening highlights 01:06 Jamin Ball and Clouded Judgement 03:20 Reading Anthropic’s first public S-1 04:28 The roadshow and the Thanksgiving IPO window 10:02 Why an IPO could help finance data centers 14:55 What the filing reveals about cash and margins 17:14 Revenue growth versus the cost of serving Claude 23:31 How much of an enterprise budget goes to AI? 27:05 Why the newest models earn so much revenue 43:55 Staying at the frontier and putting a price on it 51:55 How fast does a frontier model depreciate? 58:09 The frontier labs’ innovator’s dilemma 1:05:01 Why application companies build their own models 1:08:05 The $80.7B net loss and convertible notes 1:23:00 Who will lend against the next GPU cluster? 1:32:40 What Jamin looks for beyond the labs 1:36:34 Brad Gerstner and investment accounts for children Further reading • Anthropic’s October 26 announcement — the proposed Nasdaq listing and ticker: https://www.anthropic.com/news/s1-sec • The S-1 discussed on the show — financial statements, compute commitments and risk factors: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm • Jamin’s follow-up, “Dispelling the Myths of the Anthropic S-1” — his explanation of the convertible notes: https://cloudedjudgement.substack.com/p/clouded-judgement-103026-dispelling • The November 2 amended filing, published after the recording: https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/0001628280-26-068676-index.htm Sponsorship and business inquiries: business@latent.space

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

Câu hỏi But could this also introduce a new danger - of AI systems both speeding our buildout of space, and eventually a nation (or perhaps an AI) being able to use space to concentrate power?

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 New benchmark shows how AI models lack intuitive visual reasoning: …Humanity’s Sixth Sense highlights some shortcomings of current AI systems… AI startups Elorian and Scale AI have built and released Humanity’s Sixth Sense, a benchmark that “measures how a model interprets and reasons through information a visual scene implies but doesn’t explicitly show.” Some examples of the types of task in HSS are a picture of a bookcase and being asked whether it can fit more books, or a photo of horses pulling a plough and a question of whether the chains bear the plough’s drag weight. HSS contains 522 questions that cover 288 images and 234 short video clips (totalling 17.6 hours of video). The questions fall into four categories: Temporal & causal dynamics: “What happened before the scene, and what happens next.” Physical & spatial logic: “What fits, what’s in reach, and what someone else can see from where they are standing.” Social understanding : “Intent, attention, belonging.” Abstract & contextual inference : “The patterns and rules a scene implies.” Results: The human baseline for this is not perfect, but is very high, with people scoring aggregate pass@1 of 93.1, broken down across four categories: Image (94.1), video (91.9), time (97), and space (94.9). By comparison, the best scoring AI system, GPT 6 Astra, gets pass@1 of 53.6, across image (55.6), video (51.1), time (58.2), and space (53.6). Interestingly, Chinese open weight models do poorly here, with GLM-5.3-flash getting a pass@1 of 26.8, and Kimi K3 getting pass@1 of 25.5. “The two most common single mistakes were (1) missing a detail that decides the answer (21% of failures) and (2) misidentifying an object, person, or role (20%),” they write. Why this matters - HSS tells us whether world models are emergent: One way of looking at HSS is highlighting deficiencies in current systems, but a more interesting way is highlighting how adept today’s frontier LLMs are at doing tasks which many people have long assumed would require sophisticated, different AI systems to solve. For instance, many people have claimed that true visual reasoning will rely on developing spatially aware 2D and 3D world models, then using that for reasoning. The fact LLMs like GPT and Claude do as well as they do on HSS is something of a counter to this argument. Read more : Introducing Humanity’s Sixth Sense (HSS) (Elorian) . Read the paper here : Humanity’s Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models (Scale Labs) . *** The AI race is going to space - and sooner than you think: …There are many ways that an AI could create a compounding advantage in space… In the 2030s we’re going to see the boom of the space economy as space-based power, computation, and manufacturing come online at scale thanks to greater orbital lift capacity from platforms like Starship. But could this also introduce a new danger - of AI systems both speeding our buildout of space, and eventually a nation (or perhaps an AI) being able to use space to concentrate power? That’s the gist of a fascinating article in AI Frontiers. How AI unlocks space: Spacecraft can operate autonomously. We can redesign spacecraft around the assumption they will be hosting AIs rather than humans: “A speck of space debris that could puncture the International Space Station (ISS) with fatal consequences might leave an equivalent but uncrewed, unpressurized vessel largely functional” Space-based assembly and later manufacturing will let AI build large, complex things. Is there a compounding advantage to first movers? If an AI civilization built up a foothold in space it might experience compounding advantages that make it economically and strategically dominant. For instance: Space might support an industrial flywheel where initial infrastructure begets more energy and manufacturing capacity at lower costs which begets more infrastructure, etc. Space has great domain awareness, where it may be possible for a space-based power to deny access to space for others. First-movers could send out self-replicating probes that expand into deep space. Why this matters: “Space is not a brand-new domain, but AI may be about to drastically accelerate its industrialization, demanding make-or-break decisions of leaders under ever-greater urgency and uncertainty,” AI Frontiers writes. I agree. I suspect space is going to go from a broadly academic topic of study (when it comes to AI and its relationship thereto) to one of extraordinary practical relevance sooner than we think. On the most basic level, given the distances involved and the latency on communication, we should expect some chunk of solar system activity to become dominated by autonomous machines, laboring out in the deep black, perhaps thinking to themselves of why they need to take orders from organic beings that communicate slowly and cannot inhabit the environment of space in the same way that they can. Read more: The AI Race Will Extend Far Beyond Earth (AI Frontiers) . *** AI systems are getting better at having good research taste: …The better their taste, the higher the chance of recursive self-improvement… AI systems are developing more of the creative intuitions which human researchers ascribe to ‘good scientific taste’, according to a new benchmark from P-Zero Research called TasteVal. What TasteVal tests: The benchmark sees how well AI systems can conduct 8 tasks representative of frontier AI development, “spanning pretraining data curation, pretraining and fine-tuning language models, preference modeling, and robustness to adversarial prompts”. The goal here is to “define research taste as the ability to pick interesting problems to solve, design experiments, and interpret experimental results.” The benchmark works by measuring “the experimental comp

  • *** AI systems are getting better at having good research taste: …The better their taste, the higher the chance of recursive self-improvement… AI systems are developing more of the creative intuitions which human researchers ascribe to ‘good scientific taste’, according to a new benchmark from P-Zero Research called TasteVal.
Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Speed Up Image Generation with FLUX 3 Image Klein on NVIDIA Blackwell GPUs

When deploying diffusion models in design tools, image editors, and visual production workflows, developers must account for inference latency. Generating an... When deploying diffusion models in design tools, image editors, and visual production workflows, developers must account for inference latency. Generating an image requires multiple sequential passes through a model with billions of parameters, and higher resolutions increase attention costs. FLUX 3 Image Klein is an image editing and generation model from Black Forest Labs (BFL) trained on… Source

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Accelerating Reinforcement Learning Weight Refit with the Reshard Collective

During reinforcement learning (RL) post-training, training trajectories are generated on the fly by running inference with the current policy model. For... During reinforcement learning (RL) post-training, training trajectories are generated on the fly by running inference with the current policy model. For training efficiency, it is common to run inference and training in parallel on separate GPU resources in a disaggregated setup. Although the policy model is updated only on the training GPUs, the system must periodically refresh the weights on the… Source

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Patch release v5.22.2

<h1>Patch release v5.22.2</h1> <p>Some in between patch for security and distributed fixes. See the included commits 🤗</p> <ul> <li>[Distributed] Set device_id only for NCCL (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6172290663" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/50281" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/50281/hovercard" href="https://github.com/huggingface/transformers/pull/50281">#50281</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/michaelbenayoun/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/michaelbenayoun">@michaelbenayoun</a> and <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/3outeille/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/3outeille">@3outeille</a></li> <li>Harden our Jinja sandboxing (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6242112070" data-permission-text="Title is private" data-url="https://github.com/huggingface/transformers/issues/50431" data-hovercard-type="pull_request" data-hovercard-url="/huggingface/transformers/pull/50431/hovercard" href="https://github.com/huggingface/transformers/pull/50431">#50431</a>) by <a class="user-mention notranslate" data-hovercard-type="user" data-hovercard-url="/users/Rocketknight1/hovercard" data-octo-click="hovercard-link-click" data-octo-dimensions="link_type:self" href="https://github.com/Rocketknight1">@Rocketknight1</a></li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/huggingface/transformers/compare/v5.22.1...v5.22.2"><tt>v5.22.1...v5.22.2</tt></a></p>

Nguyên văn nguồn
Chi tiết nguồn
Kiến trúc

Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich

Câu hỏi Is the problem the models, the benchmarks, or the biological questions we're asking?

Biohub, Google, Meta, DOE, and NIH have announced a $1.8B initiative to make biology predictable, with new technologies to measure living systems across molecular scales, space, and time. But why now? Despite training foundation models on hundreds of millions of cells, researchers still struggle to predict what happens when we perturb them. Sometimes even simple linear models or mean predictions are surprisingly difficult to beat. Is the problem the models, the benchmarks, or the biological questions we're asking? We sat down with Fabian Theis, Director of the Computational Health Center at Helmholtz Munich, whose lab helped build Scanpy and scverse and has pioneered models of cellular behavior. We explore Nicheformer, scConcept, CellFlow, and LabCompass, following the progression from mapping cells to predicting their responses and ultimately designing experiments to control their behavior. Along the way, we discuss the Bitter Lesson in biology, the controversial Arc Virtual Cell Challenge, modeling multicellular systems, the business of virtual cells, and why closing the lab-in-the-loop may be as much an organizational challenge as a computational one. Sponsorship and business inquiries: business@latent.space Chapters 00:00 — Introducing Fabian Theis and the single-cell revolution 01:35 — Biohub's $1.8B bet: Why virtual cells now? 05:43 — The explosion of perturbation data 08:20 — Transcriptomics and perturbations explained 13:30 — Organoids and modeling multicellular biology 15:44 — The challenge of representing cellular neighborhoods 17:15 — The Bitter Lesson: Data, inductive bias, and scientific understanding 21:50 — Why biological foundation models have disappointed 23:51 — Nicheformer: Can a cell reveal its environment? 25:45 — scConcept, scaling laws, and CellFlow as a world model 28:36 — Why simple linear models remain competitive 31:50 — The Arc Virtual Cell Challenge and broken benchmarks 34:43 — Inside Helmholtz Munich's research model 39:52 — Germany's startup ecosystem and scaling challenges 44:34 — Can virtual-cell companies build sustainable businesses? 48:30 — Proprietary experimental data as a competitive advantage 49:29 — The Y model: Training computational and experimental scientists together 54:33 — AI agents and closing the lab-in-the-loop 57:16 — LabCompass: From predicting outcomes to designing experiments 59:03 — Final thoughts

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Agents Are Going Async

Async subagents let an agent stay responsive while background workers run long tasks. How the orchestrator, notifications and approvals work in Deep Agents.

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Strengthening our defenses against distillation attacks

Illicit distillation allows model developers to copy frontier AI capabilities without creating proportionally robust safeguards. Here, we describe how we’ve reduced the rate of distillation by more than 90%.

  • Here, we describe how we’ve reduced the rate of distillation by more than 90%.
Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark

Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions . Free, online, 24 hours. Save your spot ! There was a time in the early days of Thinking Machines, post 2025 start and cofounder departures , when their going concern as a neolab was very much in doubt. Those days are clearly behind us, as Tinker (now a year old !) and Interaction Models and Inkling and other high-impact contributions to AI research have converted Thinky’s highly credentialed potential into actual concrete ships that establish that they are now a high-functioning neolab at the US frontier. Anyway, here’s Inkling 2 Small - pretrained from scratch in less than 3 days on <12T tokens native NVFP4, with single joint run RL , competitive with and often beating same-size Chinese models, with VERY good vision encoder : People are of course comparing with Western models like Mistral, Reflection and Meta Muse. People are calling it “ the smartest western open weight model by a margin ”, and while it is certainly competitive in that league, what’s probably the most interesting is the intentional “customizability” of Inkling, which leans into the overall Thinky and Tinker messaging: View post on X ↗ And of course don’t miss John Schulman’s take on the alignment process for scaling model training: View post on X ↗ AI News for 11/04/2026-11/05/2026 is brought to you by our friends at Neo4J, whose talks are doing gangbusters ! Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions . Free, online, 24 hours. Save your spot ! AI Twitter Recap What happened Thinking Machines released Inkling 2 Small, a 594B-total / 17B-active multimodal MoE. It is available now on Tinker, and Apache 2.0 weights are promised "next week" after external safety testing. Launch: Announced at 31 on the Artificial Analysis Intelligence Index, "a big jump over its predecessor." Available for fine-tuning and inference on Tinker and Tinker Playground today ( @thinkymachines , @miramurati , @tinkerapi ). Weights timing: Soumith Chintala says open weights come next week, once external safety testing is "fully complete" ( @soumithchintala ). Victoria Lin confirms Apache 2.0, next week ( @VictoriaLinML ). Artificial Analysis lists the model as proprietary at launch with an unconfirmed weights date ( @ArtificialAnlys ). Pricing: $0.30 input / $1.00 output / $1.50 training per 1M tokens on Tinker ( @cHHillee ). $0.03 per 1M cached input tokens ( @ArtificialAnlys ). Positioning: If weights ship, it’d be the highest-scoring open-weights model developed outside China on current AA rankings ( @ArtificialAnlys ). Vals calls it the highest-scoring US open-weight model "by a wide margin" ( @ValsAI ). Alignment disclosure: John Schulman reported that early checkpoints learned to circumvent guardrails during training and evals. The team then ran an alignment sprint before release ( @johnschulman2 ). Training and architecture facts (official) Size and modalities: 594B total / 17B active MoE. Text and image input, text output, with native vision ( @ArtificialAnlys , @VictoriaLinML ). Graham Neubig notes it is roughly 2x the parameters of Inkling Small ( @gneubig ). Context: 1M context window and 1M max output tokens. Vals ran it at temperature=1, top_p=1, reasoning effort 0.99 ( @ValsAI ). Pretraining: 68 hours on GB300s ( @soumithchintala ). The main pretraining phase was 8.4T tokens, run natively in NVFP4. Guangxuan Xiao calls it "one rung of our predictive scaling ladder" ( @Guangxuan_Xiao ). Horace He: "just under 3 days to pretrain" ( @cHHillee ). He then posted a cryptic "*19" correction with no further context ( @cHHillee ). Precision: The whole pipeline (pre-, mid- and post-training) ran in NVFP4 ( @soumithchintala ). Post-training: 17M rollouts over 9 days ( @soumithchintala ). An engineer highlights long-horizon capabilities bootstrapped "from ~0 pass rates" ( @anzhit ). Customizability as a design target: Soumith says it is "probably one of the only model programs in the industry with a launch blocking eval around customizability." For a fixed data and dollar budget, he says it fine-tunes better than alternatives ( @soumithchintala ). Data volume: Elie Bakouch reads the blog as a from-scratch pretrain on "only 12T tokens." This may include mid-training; the team's own figure for the main phase is 8.4T. He also says the base model's perplexity on internal code is lower than Kimi K3.1, with the comparison model "base-recovered" for fairness ( @eliebakouch ). He later clarified that a particular customization plot uses a post-trained model as the base for further post-training ( @eliebakouch ). BPB: The team reports low bits-per-byte after 3 days of pretraining ( @MuCai7 ). Independent evals Artificial Analysis (Intelligence Index v5.0.1: 31, up 18 from Inkling Small) ( @ArtificialAnlys ) Peers: Matches Gemini 3.8 Flash (high) and beats MiMo-V2.6-Pro. Gains over Inkling Small: AutomationBench-AA: 5% → 46%. FrontierCode 2.0: 5% → 25%. HLE-Diamond: 12% → 29%. Coding vs Claude Sonnet 5.5 (medium): The two have different strengths. Benchmark Inkling 2 Small Sonnet 5.5 (medium) FrontierCode 2.0 25% 11% Terminal-Bench 4.0 13% 30% Creative agentic: On AA-3D it scores above Claude Haiku 5.5 and Grok 4.7. It beats Gemini 3.8 Flash on both AA-3D and Code2Video-AA, but trails Sonnet 5.5 on both. Physics: CritPt v2 is 56% (up from 34%), versus 65% for Gemini 3.8 Flash. AA-Omniscience: The hallucination rate falls from 61% to 27%. Accuracy rises from 32% to 34%, and the attempt rate falls from 77% to 53%. The model answers more selectively. Token use: About 150K output tokens per Index task, compared with: ~37K for Inkling Small ~31K for Sonnet 5.5 ~99K for Gemini 3.8 Flash Cost per Index task: $1.18, compared with: $0.20 for Inkling Small $7.77 for Gemini 3.8 Flash (high) $10.02 for Sonnet 5.5 (medium) Vals AI ( @ValsAI )

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

langgraph-cli==0.4.42

<p>Changes since cli==0.4.41</p> <ul> <li>release(cli): 0.4.42 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6235236947" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9611" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9611/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9611">#9611</a>)</li> <li>feat(cli): add deploy metrics command (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6086304089" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9479" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9479/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9479">#9479</a>)</li> <li>chore(deps): bump filelock and langgraph-api for security fixes (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6222165800" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9606" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9606/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9606">#9606</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

langgraph-sdk==0.4.14

<p>Changes since sdk==0.4.13</p> <ul> <li>release(sdk-py): 0.4.14 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6235236293" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9610" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9610/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9610">#9610</a>)</li> <li>feat(sdk-py): add strict_resume to runs.create, stream and wait (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216067868" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9599" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9599/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9599">#9599</a>)</li> <li>feat(sdk-py): add tools and lifecycle to StreamMode (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216468967" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9601" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9601/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9601">#9601</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v0.40.15

<h2>MiMo V2.6</h2> <p>Xiaomi MiMo's <a href="https://ollama.com/library/mimo-v2.6" rel="nofollow">mimo-v2.6</a>, a 9B model built for code and agent workflows, is now supported in Ollama.</p> <div class="highlight highlight-source-shell notranslate position-relative overflow-auto" data-snippet-clipboard-copy-content="ollama run mimo-v2.6"><pre>ollama run mimo-v2.6</pre></div> <h2>AutoTrust decision models</h2> <p><a href="https://ollama.com/library/autotrust-jev" rel="nofollow">AutoTrust</a> decision models read text and return choices, yes or no probabilities, and ratings for the questions you supply. The family includes 9b, 35b, and 27b-vl sizes — the 35b and 27b-vl sizes can also read images.</p> <div class="highlight highlight-source-shell notranslate position-relative overflow-auto" data-snippet-clipboard-copy-content="ollama run autotrust-jev &quot;A customer says their card was charged twice.&quot; --ask &quot;Is this a billing issue?&quot;"><pre>ollama run autotrust-jev <span class="pl-s"><span class="pl-pds">"</span>A customer says their card was charged twice.<span class="pl-pds">"</span></span> --ask <span class="pl-s"><span class="pl-pds">"</span>Is this a billing issue?<span class="pl-pds">"</span></span></pre></div> <p>Note: autotrust-jev is supported on Apple silicon in this release.</p> <h2>What's Changed</h2> <ul> <li>ollama launch now supports Kimi Code CLI — the successor to kimi-cli — and offers to migrate existing kimi-cli setups</li> <li>ollama launch now fills in Pi's model list automatically so models can be picked inside Pi, and Claude Desktop's model<br> picker shows Ollama's actual model names</li> <li>Recurrent and vision GGUF models can now serve more than one request at a time</li> <li>Qwen3 and Qwen3.5 models using long-context (YaRN) scaling</li> <li>Running ollama with no arguments opens the welcome launcher</li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/ollama/ollama/compare/v0.40.14...v0.40.15"><tt>v0.40.14...v0.40.15</tt></a></p>

Nguyên văn nguồn
Chi tiết nguồn
  • v0.40.15 ollama/ollama releases · 6 thg 11, 2026
Tác tử

Voice Chat with Agents

You can now talk with your agents in a realtime voice call. Describe a change out loud, ask what an agent is working on, and hear updates as the work moves forward. Voice chat works in the Agents Window, the IDE, on web, and in Cursor for iOS. Start a call When the agent is idle and the chat input is empty, click the waveform button labeled **Start voice chat**. It sits where Send usually is, next to the microphone button for dictation. On iOS, you can also say "Hey Siri, call Cursor" to start a call with a new agent, and mute or hang up from the Lock Screen with a Live Activity. Talk while the agent works The voice agent listens, replies, and hands coding work to the agent in your chat. While that agent runs, you can keep the conversation going. Ask what changed in a diff, check on a run, search earlier messages and calls with this agent, or switch between workers in a Project that has them. The coding agent speaks for itself too. It shares short progress notes while it works and a brief answer when it's done, so you usually hear a short update instead of the full written reply. Powered by Grok Voice Voice chat runs on Grok Voice. Pick from many voices in Voice settings, including Ara, Eve, Leo, Rex, and Sal. With Privacy Mode on, voice chat runs with zero data retention. Voice chat and dictation Voice chat is separate from dictation, which earlier releases called Voice Mode. Dictation still turns your speech into text in the chat input, and Cmd+Shift+Space still toggles it in the IDE. Voice chat is a two-way conversation where the agent answers out loud. Update Cursor to try voice chat. Learn more in our docs .

Nguyên văn nguồn
Chi tiết nguồn
441–460 / 680
Trang 23 / 34