iCreatAI News
Back to main site
Skip to content
#Data / training
Nov 4Wed
  1. TechCrunch · AI66

    Anthropic is still grappling with the economics of compute

    Anthropic's updated S-1 filing disclosed results through September 30, showing revenue of $17 billion, up from $11.5 billion the prior quarter, while compute costs rose to 58% of revenue from 48%. The company's quarterly compute bill reached $10 billion, including $5.0 billion under short-term commitments that cost significantly more than multi-year deals, and it said compute and infrastructure expenses will remain its largest operating expense for the foreseeable future.

  2. Meta Engineering34

    F3 Lightning: How Meta Accelerates Ads ML Features to Production at Scale

    Meta introduces F3 Lightning, an infrastructure layer on top of its Facebook Feature Framework (F3) that decouples feature releases from shared execution so each of the 20,000+ features in an ads ranking request can be compiled, validated, deployed, and rolled back independently. To preserve F3's efficiency, it uses stable computation identifiers, a single shared execution plan, one shared memory buffer per request, and dense index mapping for direct array access. Meta reports comparable extraction features reach production about 3.6x faster than prior F3 with no increase in infrastructure compute cost.

  3. Allen Institute for AI Blog55

    StateBench: Testing how language models track state changes over long documents

    Ai2 introduces StateBench, a long-context state-tracking benchmark with four difficulty levels (r-trivial, aperiodic, soluble, and general finite) grounded in Krohn–Rhodes theory. Testing open pretrained models and frontier models including Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna, they found accuracy fell sharply on soluble and general finite problems as inputs grew longer, dropping to 13% or lower on the longest general finite problems. In from-scratch training of 275-million-parameter models, Gated DeltaNet and hybrid models generalized almost perfectly on the first three levels out to tens of thousands of assignments, while transformers never learned the hardest level.

  4. Simon Willison50

    Qwen 3.8 27B on my laptop can interpret model weights, too

    Simon Willison tested the "Sparks of Weight-Space Programming" paper's game-identification prompts, which ask a model to name which of four Atari-style games a policy network was trained on using only its raw weights and no tools. Running Qwen 3.8 27B locally via LM Studio, he got correct answers for Pac-Man (46m 12s, 43,506 reasoning tokens), Pong (1h 4m 54s) and Breakout (52m 25s), but it failed on Freeway. He notes critics like Matthew Honnibal and Sean Goedecke dispute the paper's framing, and he focuses on the no-tools question rather than claims that source code is obsolete.

  5. Hugging Face Blog58

    Palette: teaching a small model to build decks in your design language

    Palette fine-tunes a 20B open-weight model with LoRA to build slide decks in two roles—a designer that writes a structured JSON brief and a coder that writes pptxgenjs code per slide—trained on roughly 1,000 synthetic decks (~10,000 slides) generated by a large open teacher and filtered by a checker. On PresentBench it ranks first on Economics and Education, third on Talk, and second across the three-domain mean (59.7 vs. NotebookLM's 60.8). A check-and-repair harness that keeps a rewrite only when the checker confirms improvement adds 3.5 points on average across 131 decks.

  6. Google DeepMind73

    Agentic cheminformatics with Google Antigravity

    Google describes Cheminformatics Skills for Google Antigravity that pair a language model for workflow orchestration and scientific reasoning with deterministic RDKit-based tools and ChEMBL/PubChem connectors. In a lung cancer walkthrough, an agent retrieved thousands of compounds, ran matched molecular pair analysis, built potency models, and proposed rigidified macrocycles with synthesis routes in about an hour. The authors note that sparse data and assay noise limit predictive precision, so computational designs remain testable hypotheses requiring experimental validation.

    Why it matters: It offers a concrete, technically detailed example of agent-assisted drug discovery, including both demonstrated capabilities and stated limitations.

  7. NVIDIA Newsroom38

    AI Is Changing How Semiconductor Fabs Work — and Engineers Are Moving Faster Because of It

    NVIDIA describes an "AI-defined fab" in which accelerated computing, physics-based AI, AI agents and digital twins help semiconductor engineers shorten simulation cycles, predict process behavior and test changes virtually before physical trials. The article cites up to 10x acceleration for COMSOL multiphysics simulations with cuDSS, up to 17x for Siemens Calibre lithography workloads with cuLitho, and Lam Research predicting plasma behavior in milliseconds versus days for full simulation. It also covers Advantest using NeMo Switchyard agents to cut model-selection response times from hours to minutes and Applied Materials building fab digital twins on Omniverse libraries.

  8. Hacker News AI33

    Show HN: 8 years of "Who's Hiring?" in charts (hiring boom, bust, and AI trends)

    The author used Codex to run semantic analysis on eight years of Hacker News "Who's hiring?" posts, from pre-COVID through October 2026, retrieved via the HN API. Five findings: AI or machine learning appeared in 50% of 2026 posts (up from 17% in 2019); 53% of identified employers appeared in only one of the 94 monthly threads; frontend roles fell from 16% to 3% of posts; staff-level posts rose from 77 in 2019 to 277 in 2026; and only 41 of 2,932 posts in 2026 explicitly named junior roles, versus 163 in 2019. The post also lists 20 other trends.

  9. Hacker News AI31

    Quili AI

    Quilicura, Chile, located in the Maipo River Basin, has become a major hub for AI data centers that use large amounts of water to cool servers, and combined with an ongoing climate crisis, Chile faces a severe drought expected to last until 2040. To make the impact visible, residents of Quilicura became "human A.I." for a day, answering prompts, creating content with local artists, and completing other tasks manually. The initiative aimed to dimensionalize a complex problem and show the real people affected by the irresponsible use of AI and its hidden cost on local water resources.

Nov 3Tue
  1. Hacker News AI38

    2M databases and counting: the architecture behind Sauna's AI teammate

    Sauna, a proactive AI work assistant launched in April 2026, gives every session, agent, connected inbox and calendar its own database, totaling over 2 million databases with about 90,000 created daily. The piece details how Sauna uses Turso Sync for chat histories and synced copies of inboxes and calendars, after earlier attempts with Cloudflare Durable Objects and Litestream, and explains how per-database isolation supports sharing, fast local search and retention. It also covers Sauna's plans to consider moving its in-house file system to Turso and to experiment with MVCC.

  2. Hacker News AI40

    An Endless Stream of AI Slop: The Burden of AI-Assisted Software Development

    This arXiv paper qualitatively analyzes 1,154 Reddit and Hacker News posts to build a 15-code codebook on how developers perceive and respond to "AI slop" in software development, organized into Review Friction, Quality Degradation, and Forces and Consequences. The authors frame AI slop as a tragedy of the commons, where individual productivity gains externalize costs onto reviewers, maintainers, and the broader community. They report developers' concerns and proposed mitigation strategies, with implications for tool developers, team leads, and educators.

  3. Hacker News AI20

    Consumer Appetite for Personal AI Goes Mainstream

    A Qualcomm Technologies-commissioned Personal AI report, based on a survey of more than 12,000 people across the U.S., U.K., Germany, India, Brazil, and China, finds consumers shifting from single devices toward connected ecosystems coordinated by a personal AI agent, with 70 percent expressing strong interest in an AI assistant that interacts naturally through voice and manages devices, apps, and everyday tasks. The report says 60 percent of wearable owners use AI at least weekly, while privacy, security, and user control are the top adoption concerns, with almost 60 percent wanting wearable data kept private or shared only with approval. Purchase drivers vary by category, from audio quality for earbuds and headphones to display quality for smartwatches, materials and finish for rings and pendants, and control type for smart glasses.

  4. Hacker News AI40

    A Human Corrected the AI. Then the Data Changed

    Zingg describes how adding a simple steward approval flag to its incremental entity resolution and MDM platform raised harder questions about what a human decision locks in, how long it holds, and who may reopen it when underlying data changes. The post lays out design principles (members stay put, no automatic merges, growth allowed), walks through concrete edge cases such as a father-and-son household split, and argues that the system being corrected should not control when a human decision expires. It closes with general guidance for any system where agents and people share decisions over time.

  5. The Decoder69

    Huawei tells customers to find their own AI chip memory, effectively outsourcing procurement to the black market

    Huawei is telling some customers to source memory for its AI chips themselves, effectively pushing buyers onto the black market, The Information reports, as Chinese suppliers cannot produce enough High-Bandwidth Memory and U.S. export rules block direct purchases from Samsung and SK Hynix. Smugglers bundle HBM with cheap server processors below the export threshold and ship them through Hong Kong, where workshops separate the chips, with smuggled HBM3E prices rising from a few hundred dollars to around $4,000 per unit in a year, according to a Shenzhen broker. South Korea's customs authority had uncovered trade violations worth $560 million by May, already surpassing the previous full-year record, a third involving strategic goods like chips.

  6. 404 Media58

    Vibe Coded AO3 Clone Site Down After Fanfic Authors Buried It With 2,000 Takedown Notices in a Single Day

    A vibe-coded site called ao3.life that scraped Archive of Our Own fanfiction, reuploaded the works and charged for AI-generated audiobooks went down Monday evening after its creator, who identified himself as a 17-year-old from India named Ganesh, said he received over 2,000 takedown requests from authors in the single day it was up. The site used AI agents controlling 40 virtual machines to scrape works, offered EPUBs and audiobooks paid after the first three minutes, and claimed authors would receive 91% of profits. AO3 has been down since Thursday following a security incident and is expected to be offline for as long as four weeks, according to the Organization for Transformative Works, which says no data has been lost.

  7. Latent Space64

    [AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop

    AINews reports a paper in which a smart model directly reads and edits the weights of a smaller model, described as one-shot fixing bugs from failed traces, improving in a general direction over about 50 steps, and writing entire network weights from scratch that beat some independently trained models. The post frames this as a new recursive-self-improvement capability that emerged suddenly, notes an independent reproduction appears to have already found it, and cautions it is unlikely to scale beyond simple games and RL policies for now. The same issue also covers DeepSeek's open-sourced "super-agent" harness prototype, the GDPx benchmark where the best model scores 12%, and October model releases.

  8. MarkTechPost49

    Zyphra AI Introduces Zyphra Hierarchical Attention (ZHA): A Path to True Subquadratic Attention

    Zyphra released Zyphra Hierarchical Attention (ZHA), a sparse attention method that uses a tree-structured indexer to search the KV cache instead of scanning the full context, and its paper shows it can be distilled into frozen pretrained models with about 4B tokens. On Qwen3.6-35B-A3B, ZHA adds 81,920 trainable parameters and stays within a few points of dense attention on seven reasoning, coding and retrieval benchmarks, with the wide tree reaching 96.2% on RULER-13 at 256K versus 96.0% dense, while MSA fell to 77.1%. Zyphra released no optimized kernels, so speedups such as 373 to 417x over dense at 1B tokens come from an H100 BF16 roofline model rather than measurements.

  9. Apple Machine Learning Research30

    Attention to Mamba: A Recipe for Cross-Architecture Distillation

    The authors propose a two-stage cross-architecture distillation recipe: first distilling a Transformer into a linearized version of Attention using an adaptation of the kernel trick, then distilling that linearized version into an adapted Mamba model with no Attention block. The distilled Mamba model retains most of the original Pythia-1B Transformer's performance, reaching a perplexity of 14.01 versus the teacher's 13.86, with results largely carrying over to downstream tasks though some gaps appear on tasks relying more on in-context recall. They conduct ablations at 1B scale with 10B tokens, scaling analysis on model sizes and distillation tokens, and a sensitivity analysis on token allocation between stages.

  10. Latent Space40

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

    In this podcast episode, Legora CTO Jacob Lauritzen discusses how the legal-tech company builds and tests agents whose traces it cannot read because the data belongs to law firm clients, using simulated test environments modeled on customer work. He covers model choices and the addition of a model picker, an estimate that models already do about 60% of legal work well, and a citation fix that took accuracy from 70% to 99% in one release. The conversation also touches on Tabular Review, running everything on Azure, forward-deployed and legal engineers, and the future of law firm partnerships, the billable hour, and junior lawyer roles.

  11. Latent Space60

    Building AI for Lawyers - Jacob Lauritzen, CTO of Legora

    Jacob Lauritzen, CTO of Legora, discusses building AI agents for lawyers after the company's $550M Series E at an $8.8B valuation, covering how its evals team simulates customer work with three engineers and three lawyers because client data is confidential, and how Legora runs everything on Azure for law-firm approval. He says models already handle about 60% of legal work well, that complex analysis and editing across million-token contracts still need more intelligence, and that Legora has begun post-training open-weight models for narrow tasks like legal research. The company passed $100 million in ARR in April, less than 18 months after launch, and now has more than 1,800 customers.

  12. TechCrunch · AI55

    Trajectory raises $45M to make AI model training accessible to nontechnical experts

    Trajectory, an 18-person startup founded by three Stanford classmates with experience at Google DeepMind and Apple, raised $45 million in a Series A round at a $300 million valuation led by Sequoia, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. The company uses continual learning to let businesses connect an open-source model to their applications, collect feedback on its performance, and retrain it so specialists like lawyers and doctors can improve models themselves. Customers including Harvey, Rogo, Decagon, Mercor, and Nvidia are already using the platform, which is now publicly available.

  13. NVIDIA Technical Blog25

    Keep Distributed Training Running with Automated Workload Recovery

    As distributed AI training scales, failures across hardware and software dependencies increasingly disrupt training progress, making fast, reliable recovery essential. The material notes that a stalled training rank can trigger an NCCL timeout, a GPU node can stop responding to the scheduler, or a straggler can slow the entire job, and that each disruption is manageable on its own.

  14. Hacker News AI34

    AI × Hiring

    The authors interviewed 30 recruiting and engineering leaders in New York and San Francisco in August and September 2026 about how AI is changing hiring, covering inbound, sourcing, referrals, evaluation and winning candidates. They report that fraud can make up 25% to 80% of applications for remote roles, that no team trusts AI to rank applicants, and that agentic sourcing output degrades after the first two to three days unless someone keeps steering it. The report also describes new evaluation formats such as system design, building with AI and reviewing AI-written code, and argues that effective filters should be harder for low-quality candidates to fake than for strong ones to emit.

  15. Allen Institute for AI Blog34

    How Olmo 3 earned the highest ethical rating from Nextcloud

    Nextcloud awarded Olmo 3 its highest Ethical AI rating, green, which requires open-source training and inference software, a self-hostable trained model, and publicly available training data. Nextcloud bundles Olmo 3 with its llm2 local LLM backend, citing the 7B model's ability to run on consumer hardware without high-end GPUs and its fit for customers seeking independence from big tech. Nextcloud engineer Marcel Klehr said fully open models make AI transparent, explainable, auditable and controllable, and that the company is working on expanding its AI features.

  16. TechCrunch · AI58

    New York Times and other outlets sue web browser Brave over alleged content scraping

    The New York Times sued Brave on Monday in federal court in Manhattan, alleging the privacy-focused browser company scraped its copyrighted content and distributed it to AI companies including Anthropic, Perplexity, Mistral, xAI, and Cohere. The Daily News, Chicago Tribune, and Orlando Sentinel filed a similar lawsuit the same day. The Times, which has also sued OpenAI and Perplexity, seeks an injunction and damages, alleging Brave built an index of more than 30 billion webpages and offers a feature that bypasses paywalls.

  17. Hacker News AI38

    AI adoption in 2026: use, spending, and returns

    Nathan Benaich compares three AI adoption estimates with different populations and definitions: McKinsey's 2026 survey (nearly nine in ten respondents report regular AI use in at least one business function), Ramp's AI Index (55.9% in September 2026, based on observed payments among US businesses in its transaction sample), and the US Census Bureau's BTOS (19.8% for the period ending May 3, 2026). The piece argues these figures cannot be averaged into a single adoption rate because they measure different things over different periods, and it notes that McKinsey reports 37% of respondents attribute some positive EBIT impact to AI, which is a reported financial outcome rather than an adoption rate or causal estimate.

  18. Hacker News AI62

    Oracle files to sell $20B in new stock, bonds to finance AI buildout

    Oracle filed with the SEC to sell $10 billion in common stock and $10 billion in senior notes to help finance its AI buildout, a plan it had already outlined at last week's investor day. Shares fell more than 6% Monday despite the filing matching prior guidance, and Oracle said proceeds will go toward general corporate purposes that may include capital expenditures, debt repayment and future investments. Oracle shares are down 33% in 2026 and 50% over the past 12 months, and S&P downgraded the company to BBB- in July.

  19. Hacker News AI38

    AI assistants read a Shopify store 10x more often than people click through

    A 30-day analysis of a headless Shopify fashion storefront found that ChatGPT and Claude fetched pages 1,264 times for users while GA4 recorded only 122 visits from AI answers, with ChatGPT alone reading pages about nine times per visit. The logs also showed assistants repeatedly requesting discontinued products that returned soft 404s, Claude probing standard Shopify paths the storefront doesn't serve, and roughly 10,600 requests from fake AI crawlers that failed IP verification and mostly probed for secrets. The authors note their own dashboard initially overstated ChatGPT arrivals by more than tenfold due to utm tag handling in referer headers, and recommend checking 404 status codes, adding redirects for standard Shopify paths, and verifying bots by IP.

  20. MarkTechPost71

    Deep Principle Releases Hydrogen-1.0: A 753B Open-Weight Model for Scientific Discovery

    Deep Principle has released Hydrogen-1.0, an MIT-licensed 753B-parameter open-weight model built by post-training Z.ai's GLM-5.3 on chemistry, materials science and life sciences data, laboratory records and agent discovery trajectories, with a 1M-token context and 128K max output. Deep Principle reports it scores 38.3% on FrontierScience-Research versus 33.3% for GPT-6-Astra and 93.0% on GPQA-Diamond, while trailing GPT-6-Astra by 3.33 points on NMR structure elucidation. Weights are on Hugging Face and ModelScope, with BF16 serving tested via SGLang on 8 GPUs, NVFP4 builds for NVIDIA B200/B300 and MXFP4 for AMD MI350X/MI355X, though full datasets and recipes are not released.

  21. Hacker News AI58

    South Korea targets ' highest performance' sovereign AI model

    South Korean President Lee Jae Myung unveiled the country's largest ever budget, with 2027 spending rising 12.8% to 820.9 trillion won ($610.14 billion), including 21.3 trillion won for semiconductors, physical AI and AI data centers, nearly double the 2026 level. A separate 4.7 trillion won will go toward developing a sovereign AI model that Lee described as having the "world's highest performance," alongside training 2,000 semiconductor specialists and expanding AI education. Lee said the higher spending would result in "virtually no net increase in government debt," with defense expenditure rising 8.2% to 73.3 trillion won.

  22. PyTorch Blog28

    Announcing the Special Interest Group (SIG) for Ray on Accelerators

    The Ray community announced a Special Interest Group for hardware accelerators, aiming to bring together hardware vendors, Ray contributors, and application developers around a vendor-neutral roadmap for accelerator support. Its initial roadmap covers development and testing infrastructure, hardware-aware scheduling, and networking/transport, including Ray Direct Transport for direct device-to-device data movement. Ray currently supports more than a dozen kinds of accelerators, from NVIDIA and AMD GPUs to TPUs, Trainium, and Gaudi.

  23. AWS Machine Learning Blog36

    Build efficient agents with knowledge distillation on Amazon SageMaker AI

    This post explains how to use sequence-level knowledge distillation (SeqKD) to fine-tune a smaller shopping agent on Amazon SageMaker AI, using a frontier teacher (GLM-5.3 on Amazon Bedrock) to generate multi-step tool-use trajectories that are filtered by ShoppingBench's programmatic scorer and used for serverless supervised fine-tuning of a Qwen3.5-4B student. The distilled student reached an average Absolute Success Rate of 0.541 on 324 held-out ShoppingBench problems, versus 0.297 for the base model and 0.540 for the greedy teacher, while IFEval general instruction-following dropped roughly 6 to 8 points. The post also covers per-step trajectory encoding, LoRA training configuration, a low-cost model-evaluation pre-check, and serving the base model and student adapter as two inference components on one real-time endpoint.

  24. MIT Technology Review · AI22

    Raising the quality of insights in agentic analytics

    An MIT Technology Review Insights report finds that on average only 48% of insights generated by analytics agents at surveyed companies lead to actions that improve the business, up from 37% a year earlier, with a small group of "insight leaders" exceeding 70%. Respondents cite lack of business context (55%) and fragmented or inaccessible data (45%) as the main causes of substandard insights, and 52% of organizations say agents generate too many insights to evaluate properly. Companies are prioritizing stronger AI and data governance (60%) and adding a semantic layer to their data architecture (45%).

  25. TechCrunch · AI52

    42AI comes out of stealth with $62.5M and a world model that can ‘predict human behavior’

    Israeli startup 42AI has emerged from stealth with $62.5 million in seed funding for a "behavioral world model" that predicts how individuals will act, initially deployed at financial brokerages to forecast trader responses to volatile conditions. The company says it studies reactions to seemingly unrelated events and converts them into mathematical trait representations fed into a transformer-based model, claiming 98% accuracy on whether trading flows fall within a predicted range and an F1 score of 0.8 to 0.85 for trading direction and volume. 42AI says it is working with 10 large brokerages and generating more than "seven figures" in revenue, and is exploring gaming, fraud prevention, cybersecurity, and eventually robots and AI agents.

  27. Hacker News AI60

    Anthropic's internal AI use, at retail prices, now exceeds its employee pay

    Anthropic's IPO filing shows that in April the value of Claude AI used internally, calculated at the prices Anthropic charges customers, exceeded its monthly employee compensation costs excluding stock awards, and by September internal AI usage had surged more than 250-fold since January. The company is not actually spending more cash on AI than on salaries, but more than 90% of the code added to its software in September was written by Claude, while revenue has grown far faster than headcount.

  28. Tomasz Tunguz62

    The Four Questions of AI

    The piece lays out four coupled questions across the AI stack: how application companies build moats when models can absorb their features, who owns the harness and context layer given evidence that third-party harnesses outperform native ones, when foundation-model training costs stop paying back, and whether credit and power can keep pace with GPU demand. It cites figures including 98% input-token share, a 71% cost cut from smarter harnesses, roughly 83% of agent tasks routing to smaller models, $4t in projected five-year AI data center debt, and a B300 index at $12.29 per GPU-hour. It concludes that the layers behave like a double pendulum—simple individually but chaotic when coupled.

  29. Hacker News AI50

    Can AI Spot a Bad Business Plan?

    The author tested Claude Fable 5.1 and DeepSeek 4.1 on business-plan review using AI-generated plans with planted mistakes and 31 anonymized real Regulation Crowdfunding companies scored against later outcomes, where both models caught about four in five problems (81% and 82%). The piece also reports that a model identified the real company behind 37 of 68 disguised plans, that both models flagged nearly half of real red herrings as problems, and that the tests are small, single-run signals rather than a ranking.

Nov 2Mon
  1. TechCrunch · AI29

    Omen AI adds liquid know-how in $16 million Aquatic Labs acquisition

    Omen AI, founded in 2024 to build real-time sensors that monitor data center cooling fluids, has acquired competitor Aquatic Labs in a $16 million cash-and-stock deal after raising a $31 million Series A in June. Omen's sensors detect elements in the water while Aquatic Labs' sensor tests alkalinity, and the combined teams aim to integrate both into one sensor expected on the market by early next year, with Aquatic Labs CEO Allan Adams becoming chief science officer. Omen's initial sensors will monitor chips drawing 250 to 300 MW of power by the end of the year, with a goal of installation on the equivalent of 1 GW of chips by the end of 2027.

  2. Hacker News AI55

    AI cut data warehouse migration cost from $70M to $35M

    Westpac says AI-accelerated migration of its legacy St.George data warehouse cut costs from a budgeted $70 million to around $35 million and finished more than two years early, with CEO Anthony Miller citing AI tools for faster code conversion, testing and documentation. The bank expects about $15 million a year in direct benefits, plans to decommission the old warehouse in FY27, and says the project produced a reusable AI data migration capability for other migrations.