Skip to content
Only Research#Data / training
November 13Friday
  1. Hacker News · AI58

    Study finds AI-style stale version pins in 13.7% of new npm packages in 2026 H1

    Kobi Mozgori sampled 250,000 npm packages and found that new packages whose pinned dependency versions are mostly round x.0.0 numbers already superseded rose from a 1–2.4% baseline in 2019 to early 2024 to 13.7% in 2026 H1, falling to 7.6% in 2026 H2 through early November. The median pinned dependency was about 155 days old at birth in 2021–2023 but 451 days in 2026 Q2, and the share of packages born with essentially all pins on the newest release dropped from about 28% to about 10%. The author frames the pattern as an inference rather than proof, since registry metadata cannot show who or what wrote a manifest, and notes the decline since mid-2026 is unexplained.

  2. Hacker News · AI85

    Forensic analysis links six Chinese frontier models to export-controlled NVIDIA Blackwell GPUs

    Unemployment Labs' forensic analysis of published model weights finds that six models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialized on NVIDIA Blackwell hardware, while earlier models from the same labs match Hopper. The method recovers each model's random seed and thread layout from preserved weights, and the author says the matches were reproduced on physical H100, B300 and GB200 chips. The report estimates each training run needed roughly 500 to 3,000 Blackwell GPUs over two to three months, and notes the evidence does not establish where the chips were located or whether export controls were violated.

    Why it matters: The report derives hardware evidence solely from published model weights, offering a reproducible method for tracing which GPUs initialized a model without relying on leaks.

November 12Thursday
  1. MIT News34

    MIT researchers develop training technique for faster, more trustworthy climate emulators

    MIT researchers have developed a training technique that boosts the performance and adaptability of AI climate emulators while using only a fraction of the data typically required. The method generates optimized synthetic training scenarios with rapid emissions variations, letting an emulator trained on far less data accurately predict unseen scenarios where traditional emulators failed. The work appears in the Proceedings of the National Academy of Sciences.

  2. Google Research58

    Google Research uses wearables to build CKM risk profiles and personalized coaching

    Google Research introduced an investigational framework that turns consumer wearable data into a coachable health age and simulates personalized 12-week what-if health trajectories. The SensorWM world model was pretrained on 2.28 million person-weeks of de-identified Fitbit data from roughly 100,000 users and forecasts 20 physiological markers under different step, sleep and cardio routines. In a blinded study, lifestyle coaches chose the counterfactual-grounded plan as best overall in 71.4% of evaluations, and physicians rated 99.2% of plans within safe exercise intensity limits; the authors note the models are research prototypes, not clinical diagnostic tools.

November 11Wednesday
  1. inclusionAI Models40

    inclusionAI Releases ZoomGain-9B and ZoomGain-4B Models with ZoomGainBench and RL Training Data

    inclusionAI released the ZoomGain-9B and ZoomGain-4B vision-language model checkpoints, alongside the ZoomGainBench evaluation benchmark and ZoomGain RL training data. ZoomGain mines global-to-zoom perception gaps to build fine-grained visual supervision, and ZoomGain-9B reached superior performance on multiple fine-grained and general perception benchmarks. ZoomGainBench contains 1,000 questions over 941 images, while the RL dataset holds 9,000 training records and 7,713 source images.

  2. inclusionAI Models38

    inclusionAI Releases ZoomGain-4B and ZoomGain-9B with ZoomGainBench and RL Training Data

    inclusionAI released the arXiv paper, ZoomGainBench evaluation benchmark, RL training dataset, and ZoomGain-4B/ZoomGain-9B model checkpoints. The best of 28 models scored 61.08 on ZoomGainBench, while ZoomGain-9B reached superior performance on multiple fine-grained and general perception benchmarks. ZoomGain mines global-to-zoom perception gaps to build fine-grained visual supervision for vision-language models.

  3. Microsoft Research75

    Microsoft researchers describe Lazarus, a backdoor that reprograms dormant neurons in AI agents

    Microsoft researchers have described Lazarus, a backdoor technique that reprograms fewer than 10 dormant neurons in an LLM so an agent built on it emits attacker commands when a hidden trigger appears, with no retraining required. The modified model still passes standard evaluations and behaves normally until the trigger, which can be as simple as a date change, and the agent then runs the downloaded code with whatever access it was given. The authors say detection requires diffing against the original weights, and Microsoft Foundry already ties open-weight models to their publishers and scans their containers, while red-teaming harnesses, agent-based weight scanning and signed models with tamper-evident provenance are in development.

    Why it matters: Microsoft researchers show that reprogramming fewer than 10 dormant neurons can backdoor an LLM agent without retraining, which makes weight provenance and integrity a practical security requirement.

November 10Tuesday
  1. Apple Machine Learning Research34

    Weighted State Souping for Efficient Long-Context Retrieval

    Apple researchers and Harvard collaborators propose weighted state souping, a method that learns query-dependent weights for recombining cached SSM state encodings, improving long-context performance on single- and multi-hop document QA, needle-in-a-haystack retrieval, variable tracking, and tool use. A sparse variant achieves near-constant encoding cost with respect to context length while preserving or sometimes improving performance. The paper was accepted at the PALM workshop at NeurIPS 2026.

  2. Apple Machine Learning Research38

    Apple Study: Gating Mechanism Explains Why SSMs Memorize Instead of Learning In-Context

    An Apple and Harvard study shows that the gating mechanism in State Space Models (SSMs) causes them to first learn an in-weights "memorization" solution, delaying or even preventing convergence to correct in-context learning. This occurs even when the architecture and memory capacity impose no fundamental limits. Gating, however, often improves generalization to long sequence lengths.

  3. Apple Machine Learning Research31

    Apple study evaluates LLM judges for personalized writing style assessment

    Apple researchers tested how well LLM judges assess whether generated text matches a user's preferred writing style, generating text with four LLMs across three long-form writing tasks and comparing judge annotations against human pairwise labels. Judge quality correlated only weakly with general LLM ability on MMLU (Pearson's ρ ≤ 0.2) and varied by writing task and evaluation criteria, and paired comparisons proved more consistent than independent scoring. The strongest LLMs did not most reliably reproduce human evaluations, possibly due to exaggerated sensitivity to subtle details.

  4. Goodfire Research62

    Goodfire studies Qwen3.6-35B-A3B neurons as an interpretable basis

    Goodfire Research published a study arguing that individual MLP neurons in Qwen3.6-35B-A3B are more interpretable than assumed, based on sparsification, attribution tools and research agents. Ablating the 92.2% weakest MoE neuron activations per forward pass raised language model loss by only 2.9%, and 69% of neuron poles had a single mode responsible for at least 90% of their causal effect. Natural-language neuron descriptions let a judge model predict firing with 95% accuracy against random text and recover 73% of ablation damage through simulation.

    Why it matters: Goodfire's study of Qwen3.6-35B-A3B neurons argues that strong activations alone carry most of the model's computation, offering a concrete alternative to feature-dictionary interpretability.

November 9Monday
  1. Hugging Face Blog38

    NVIDIA and Elith Release Nemotron-Personas-Finance-Japan Synthetic Dataset

    Elith and NVIDIA released Nemotron-Personas-Finance-Japan, an open synthetic persona dataset for Japan's financial sector built on official government statistics. It contains 9M personas (1M records × 9 personas) across 50 fields and ~1.5B tokens, generated with NeMo SDG-PGMs and NeMo Data Designer, and is available under CC BY 4.0 for commercial use.

  2. inclusionAI Repositories38

    ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision

    ZoomGain constructs questions about a full image from visual details that only become clear when zooming in, using this global-to-zoom perception gap as supervision for vision-language models. The release includes the arXiv paper, the ZoomGainBench benchmark (1,000 questions over 941 images covering attribute recognition, counting, and OCR), 9,000 RL training records over 7,713 source images, and the ZoomGain-4B and ZoomGain-9B checkpoints. The best of 28 models scored 61.08 on ZoomGainBench, while ZoomGain-9B reached superior performance on multiple fine-grained and general perception benchmarks.

November 8Sunday
  1. Artificial Analysis Articles38

    Artificial Analysis Decision Index v2 Released with New Decision Score and Private Benchmarks

    Artificial Analysis released Decision Index v2, with 45% of the Index now built on private datasets, up from 25% in v1. Drex Max v1.1 from Nace AI narrowly leads decision models at 73, within 1 point of GLiDE v1.2 and Mammoth 27B. The new Decision Score rewards both correctness and calibrated confidence, while AA-Paperclip and AA-Symmetry add private document and consistency tests.

November 6Friday
  1. Hugging Face Blog58

    Ai2 releases ToyBlocks, an open simulated block-manipulation world for robot learning

    Ai2 released ToyBlocks, an open simulation framework built around a robot arm and blocks on a tabletop, as the first of its FoundationWorlds for controlled robot learning and evaluation. The framework auto-generates skills and demonstrations via RL and motion planning in Isaac Lab, reaching up to 24x real-time on a single GPU, and ships with 85,336 long-horizon demonstrations covering nearly 1,600 hours and 144 million frames. In its experiments, ManiFlow's atomic-skill success was 80.0% in simulation and 81.1% on a real Franka FR3, while the Molmo 2 plus ManiFlow system succeeded in 55.5% of simulated and 50.0% of real-world episodes.

  2. vLLM Blog77

    REINFORCE Zero: Kernel Alignment Enables Stable Async RL at Policy Lag 64, Extends to DeepSeek-V4.1 and Kimi K3.1

    NVIDIA Research and Inferact's REINFORCE Zero paper shows that aligning rollout and trainer kernels lets unclipped token-level importance sampling correct the remaining weight lag, training stably for 20,000 updates at policy lag 64 and reaching 2.40x the training-step throughput of synchronous training with aligned kernels. With standard kernels, the median rollout-to-learner KL rises 83.33x across the run and online AIME24 pass@1 falls from 34.27% to 13.96%, while a sequence gate stabilizes training but rejects 24.48% of sequences by the last 1,000 updates. The zero-mismatch setup now covers DeepSeek-V4.1-Flash, Kimi K3.1 on release day, and DeepSeek-V4-Flash and Kimi K3.1 on AMD MI355X, with code in Inferact/zero-mismatch-rl and NVIDIA-NeMo/labs-molt.

    Why it matters: NVIDIA Research and Inferact split the async RL rollout-learner gap into weight lag and kernel mismatch, showing that aligning kernels lets unclipped importance sampling hold stable training at policy lag 64.

November 5Thursday
  1. Apple Machine Learning Research38

    Apple's Learned Support Functions Speed Up RAG Retrieval

    Apple researchers presented "Amortizing Maximum Inner Product Search with Learned Support Functions" at ICML 2026, training small neural networks offline to route queries toward their best-matching keys faster. SupportNet learns the support function via input convex neural networks (ICNNs) to improve cluster routing, while KeyNet directly outputs the best-matching key vector, kept geometrically grounded by a consistency loss from Euler's theorem. Both are trained on the actual query distribution, trading an upfront amortization cost for faster, more accurate retrieval at inference time.

  2. Google Research47

    Google Research's Agile Deliberation helps humans define subjective AI categories

    Google Research introduced Agile Deliberation, a framework that turns subjective AI classifier creation into an iterative conversation where the AI helps users sharpen their own definitions. It splits the task into concept scoping and concept iteration, using vision-language models to surface borderline images and route feedback into an automatic prompt optimization loop. In 18 live 1.5-hour user sessions, the approach reduced user fatigue and anxiety and produced more consistent models that outperform standard baselines; the code is open-sourced on GitHub.

  3. Anthropic Research58

    Anthropic studies how researchers use Claude across scientific fields and tasks

    Anthropic analyzed 207,049 Claude and Claude Code research sessions from August 2 to September 30, 2026, using privacy-preserving classifiers, and found scientists use Claude most for data analysis and writing and least for data gathering and ideation. Data analysis accounts for 33.7% of Claude sessions and 48.1% of Claude Code sessions, against 27.6% of researcher time, while ideation takes 4.2% and 2.1% of sessions versus 7.1% of researcher time. The report estimates that wide adoption of current capabilities could roughly double research output in some fields, and full automation of non-physical work could lift fields such as computer science and statistics as much as 15-fold.

  4. Goodfire Research58

    Goodfire Research proposes mechanistic shaping to steer models via internal representations

    Goodfire Research proposes mechanistic shaping, a framework that adds penalties on a model's internal concept directions during training so it learns to rely on or express specific concepts rather than just maximize task reward. In chest X-ray experiments, an LLM agent used sparse autoencoder features to find tube-like shortcut directions, and reliance shaping beat both standard training and a balanced-data baseline.

November 4Wednesday
  1. Anthropic Research66

    Anthropic biology lab reports first prokaryotic vault particle structure, solved with Claude

    Anthropic's biology research lab reports the first 3D structure of a vault particle found in prokaryotic cells, a phage protein called L05 built from 96 copies with a mass of about 9.5 megadaltons. Under the team's direction, Claude searched a database of about 15 billion predicted proteins, designed DNA constructs, optimized a size-exclusion chromatography isolation that cut purification from three days to about 90 minutes, and ran RELION cryo-EM processing that produced a first full model in 34 hours. The preprint has been submitted to bioRxiv, and the team says the vaults' biological function remains unknown.

    Why it matters: Anthropic's biology lab reports the first 3D structure of a vault particle from prokaryotes, with Claude handling database search, construct design, purification and cryo-EM processing.

November 3Tuesday
  1. Ai2 · Allen Institute for AI55

    Ai2 releases StateBench to test how language models track state changes across long documents

    Ai2 introduced StateBench, a benchmark with four difficulty levels for long-context state tracking, from immutable assignment to self-referential and cross-referential updates. In tests, Gemini 3.8 Flash, Claude Opus 5.5, and GPT-6 Sol and Luna handled the easier levels but dropped to 13% or lower on the longest general finite problems, while 275M-parameter Gated DeltaNet and hybrid models trained from scratch generalized almost perfectly on the first three levels and transformers never learned the hardest one.

  2. Apple Machine Learning Research32

    Attention to Mamba: A Recipe for Cross-Architecture Distillation

    Apple researchers propose a two-stage recipe for distilling a Transformer into a pure Mamba architecture without any Attention blocks. The method first distills a pretrained Transformer into a linearized Attention model via an adapted kernel trick, then distills that into an adapted Mamba model. The resulting distilled Mamba retains most of Pythia-1B's performance, reaching 14.01 perplexity versus the teacher's 13.86, with some gaps on tasks relying on in-context recall.

November 2Monday
  1. SemiAnalysis58

    SemiAnalysis releases GDPx benchmark for frontier agentic capabilities across high-value domains

    SemiAnalysis released GDPx, a benchmark covering chip design, cybersecurity, biology, finance, legal, quant trading, RSI, hardware design, materials science, alignment and software engineering, built with 14 specialized RL environment partners. Even the best model scores just 12%, with Opus 5.5 winning quant trading, hardware and chip design, Gemini 4 Carbon leading biology and software engineering, and GPT-6 Astra topping RSI and materials science. Every task runs in a realistic simulated environment, such as a ~4,000-file private equity deal room or live multi-service cyber ranges, unlike OpenAI's GDPval, which SemiAnalysis argues spoon-feeds context and lacks environments.

October 30Friday
  1. Microsoft Research40

    Microsoft builds synthetic enterprise data platform to evaluate Copilot

    Microsoft's Industry Evaluation Data Generation Platform creates full-scale synthetic enterprise environments to evaluate Microsoft Copilot without using customer data. Grounded in public sources such as O*NET occupational data, job postings, and SEC filings, it generates realistic tenants with users, meetings, emails, and documents, plus automated query pipelines and quality frameworks. Microsoft says this establishes the first standardized evaluation baseline for Copilot.

  2. Amazon Science38

    Speeding up LLMs for travel applications: Medusa plus knowledge distillation

    Combining knowledge distillation with the Medusa framework lets a 1.1B-parameter TinyLlama student match the quality of its 70B-parameter Llama-3.1 teacher while running 10 to 20 times as fast at less than one-third the cost. Working with Booking.com on AI Trip Planner, Smart Filters and accommodation description generation, the pipeline closed over 90% of the quality gap and added 2 to 3.6x further speedup from Medusa heads.

October 28Wednesday
  1. Google Research58

    Google Research's sleep framework lets LLMs consolidate new knowledge without catastrophic forgetting

    Google Research introduced a sleep framework for continual learning that alternates active interaction with offline memory consolidation, letting LLMs absorb new facts into expanded Mixture-of-Experts parameters instead of losing prior knowledge. The approach combines parameter expansion, on-policy distillation with reinforcement learning-based imitation, and a dreaming stage that generates and filters synthetic training samples. The authors report near-zero performance drop in continual translation for unseen languages such as Manchu and Kalamang, near-perfect scaling to 10M tokens on BABILong, 48.9% single-task and 46.2% CPT results on SQuAD-based factual incorporation, and 79.2%, 69.0% and 46.1% on AIME 24/25 and HMMT 25.

October 27Tuesday
  1. Ai2 · Allen Institute for AI58

    Ai2 study tests whether LLMs can reliably judge research idea novelty

    Ai2 released Old Ideas, Novel Problems, a study and benchmark testing how reliably LLMs evaluate the novelty of research ideas, along with data, code and a pipeline for building new evaluation sets. Across six judges including GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5 and Claude Opus 4.6, small prompt changes shifted accuracy by more than 50 percentage points on AI-generated ideas, and a simple prompted Claude Opus 4.6 baseline scored 0.854 macro-F1 at $2.59 versus 0.642 at $30.66 for Idea Novelty Checker and 0.772 at $82.05 for Sakana AI's AI Scientist.

  2. Epoch AI · Gradient Updates58

    Epoch AI tests AI on its own research tasks, top models score 65% and fail all

    Epoch AI expanded its earlier effort to automate a senior researcher's job into the Epoch Automation Reports, evaluating AI on real Epoch tasks like graphic design and research planning against the same standards used for human employees. 1 and GPT-6 Astra averaged 65% on the rubric, but a 65% score still counts as a failed task, since every submission would be rejected or revised if a human had made it.

October 23Friday
  1. Ai2 · Allen Institute for AI31

    LitPivot: Ai2's System Helps Researchers Pivot Research Ideas Through Literature

    Ai2 and University of Pennsylvania researchers built LitPivot, a system that organizes retrieved papers around the facet of a research proposal being edited and suggests literature-backed revisions. In a study with 17 researchers, domain experts rated ideas developed with LitPivot as better grounded in the literature than those built with a text editor and an AI chat tool. Participants also reported their revised ideas as more novel and useful, and LitPivot will be presented at UIST 2026.

October 22Thursday
  1. Microsoft Research64

    Microsoft Research and co-author prove the 31-year-old KLS conjecture

    Microsoft researcher Xinzhi Zhang and a co-author have proved the Kannan-Lovász-Simonovits (KLS) conjecture, a 31-year-old open problem in high-dimensional geometry, showing the bound holds with a universal constant independent of dimension. The proof grew out of a human-directed agentic process in which AI agents explored more than 100 candidate approaches while researchers pruned and redirected the search, and a separate unguided attempt with leading models on a comparable token budget fell short. Two other proofs appeared the same day, and a Lean 4 formalization was subsequently released.

    Why it matters: A 31-year-old open problem in high-dimensional geometry is resolved through a human-directed agentic search over roughly 100 candidate approaches, with independent proofs and Lean formalization as confirmation.

October 21Wednesday
  1. Apple Machine Learning Research26

    TEAL: Token-Space Efficient Active Learning for Low-Budget Semantic Segmentation

    Apple researchers (with University of Waterloo) introduce TEAL, the first pixel-level active learning framework built on frozen ViT representations, selecting diverse candidates directly on the native token lattice via a two-level MaxHerding strategy plus margin-based uncertainty refinement. With a single labeled pixel per image, TEAL beats the strongest baseline by up to +24.36 mIoU on CamVid, Cityscapes, ADE20K and Pascal Context, reaching up to 97% of fully supervised performance with just 10 px.

October 17Saturday
  1. Google Research71

    Google Research outlines VERID and closed-loop agent workflows for autonomous wet labs

    Google Research describes three projects aimed at grounding AI agents in real wet-lab experimentation, including the VERID framework, a semantic digital twin of a lab workcell built on the open-source PyLabRobot framework that tracks deck layout, well volumes and tip inventory and dry-runs protocols before execution.

    Why it matters: Google Research outlines three projects that connect AI agents to real wet-lab hardware, showing how simulated dry runs and closed-loop experiments let agents correct their own design rules.

October 15Thursday
  1. Goodfire Research62

    Goodfire research essay argues neural networks are already clean and superposition was misread

    Goodfire researcher argues that the canonical evidence for polysemanticity and superposition in neural networks was misread, and that what looked like arbitrary packing is mostly a motif he calls recursive generalization. Retraining InceptionV1 with ordinary architectural improvements and no change to model size or training data made 99-100% of neurons in its last layer easily interpretable, and preliminary audits of Qwen3-32B estimate roughly 90% of neuron poles have one main job. The essay concludes networks use one learning principle rather than two, so interpretability can proceed by studying neurons and circuits directly instead of first training decomposition models such as sparse autoencoders.

    Why it matters: A first-hand reinterpretation of the field's canonical polysemanticity evidence, arguing that networks generalize recursively rather than pack unrelated concepts, which would make direct mechanistic study viable.

October 13Tuesday
  1. Modal Blog58

    Modal tests mixed-precision, mixed-accelerator RL samplers on Qwen3.6-35B-A3B

    Modal trained Qwen3.6-35B-A3B on agentic coding with samplers spread across six accelerator types (B200, B300, H100, H200, A100, RTX PRO 6000) at NVFP4, FP8 and BF16 precision, feeding one BF16 trainer on B200. Plain GRPO collapsed on both mixed and homogeneous samplers, but adding masked importance sampling plus score centering, with or without top-p mask replay, trained stably to the end and reached SWE-bench Pro V2 pass@4 on par with homogeneous samplers, at roughly 15% lower sampler cost per update. The mixed samplers ran with an average sampler-trainer KL 4.7x that of the homogeneous setup, and the authors argue stability depends on keeping the mismatch bounded rather than low.

  2. Amazon Science75

    Amazon Web Services researchers propose priming to convert pretrained Transformers into hybrid state space models without full retraining

    Amazon Web Services researchers introduce priming, a method that reuses pretrained Transformer weights to cut the pretraining cost of hybrid state space architectures by more than 99%, using under 0.5% of the source model's pretraining token budget. In controlled comparisons, Gated KalmaNet outperformed Gated DeltaNet and Mamba-2, and a 32B GKA-based hybrid matched its source Transformer on reasoning benchmarks while decoding up to twice as quickly at 128,000 tokens of context. The team released 8B and 32B primed hybrid models, training code, Triton kernels, and a vLLM serving plugin under the Apache 2.0 License.

    Why it matters: Amazon Web Services researchers show that reusing pretrained Transformer weights can cut hybrid architecture pretraining cost by over 99%, making controlled comparisons of state space layers practical.

October 10Saturday
  1. Epoch AI · Gradient Updates65

    Epoch AI's InnovationEval finds frontier agents reach only 15% of SDPO's gains on automated AI R&D

    6 Sol and Fable 5 with independently rediscovering a novel post-training technique comparable to Self-Distillation Policy Optimization (SDPO), giving each 3,000 GPU-hours in a sandbox without internet access. 6 Sol's method, largely copied from prior work, reached 15% of SDPO's gains after adjusting for slower training, while Fable 5's technique failed to improve performance.

October 8Thursday
  1. Zyphra Research36

    Zyphra Research: Expert Coupling Cuts MoE Pretraining Communication

    Zyphra Research shows MoE routing is highly structured, with co-selected experts correlated within a layer and a token's expert choices predicting later-layer routing. Exploiting this via correlated expert placement and token shuffling cuts all-to-all communication time up to 2.63× and improves end-to-end training step time up to 1.41× in Megatron-LM experiments on MI300X GPUs, without changing architecture or routing decisions.

  2. Waymo Blog30

    Waymo Study: Sober Drivers Still Face Nearly 4x Nighttime Fatal Crash Risk

    Waymo's peer-reviewed research finds that even fully sober drivers face 3.1 to 3.9 times higher nighttime fatal crash risk than in daytime, and that removing alcohol-impaired drivers cuts average urban fatal crash involvement by 23%, from 1.42 to 1.10 per 100 million miles. A second study of 37,654 fatal cases in the 2023 FARS database shows only 8.2% involved strictly human behavior factors alone. Waymo frames the findings as support for a Safe System Approach and for benchmarking autonomous driving against a zero-BAC human baseline.

October 3Saturday
  1. Epoch AI · Gradient Updates58

    Epoch AI: Compute built over the next two years could run hundreds of millions of AI agents

    Epoch AI estimates that AI compute produced over the next two years could support tens to hundreds of millions of concurrent agents running today's most capable models, or billions running cheaper ones. Using projected HBM shipments from 2025 through 2027, the analysis counts GB300-equivalent GPUs and finds top-tier agents running nonstop could supply the weekly hours of roughly 140 million to 700 million full-time employees, while DeepSeek V4 Pro serving benchmarks imply about 1.9 billion concurrent agents. The report notes the key uncertainty is whether demand grows fast enough to justify the buildout or leaves a glut of agents.