AAITopic

Archive

Browse all AI news

Every captured signal, grouped by the date AITopic ingested it. Use these to catch up on news you missed or revisit a source.

Back to today's news

Aug 13, 2026

6 items

Aug 12, 2026

6 items

Aug 11, 2026

32 items

Aug 7, 2026

10 items

Aug 6, 2026

10 items

Aug 5, 2026

82 items

  • Aug 5, 2026

    media

    The Decoder: weekly roundup of model and tooling releases for the week of 2026-07-27 — 2026-08-02

    The Decoder: weekly AI news roundup (week of 2026-07-27 — 2026-08-02) clusters vendor tooling announcements

    The Decoder weekly tech-press roundup front-page snapshot for the week of 2026-07-27 — 2026-08-02 clusters a dozen vendor announcements across OpenAI, Anthropic, Google, Microsoft, Meta, and NVIDIA into a 'tools consolidation' framing.

    4 daily topicsOriginally published Aug 2, 2026View source

  • Aug 5, 2026

    media

    The Decoder: tech-press coverage of the 'productivity agents' GA cluster for the week of 2026-08-02

    The Decoder: tech-press coverage of the 'productivity agents' GA cluster

    The Decoder's tech-press coverage frames the 2026-08-02 cluster of GA announcements (OpenAI Scheduled Tasks, Anthropic Skills, Vercel AI Gateway, Notion AI Agents, Linear AI triage, Figma Make) as 'every productivity tool ships an agent surface'.

    4 daily topicsOriginally published Aug 2, 2026View source

  • Aug 5, 2026

    official

    Notion: Notion AI Agents GA — autonomous workflow agents in the workspace

    Side-by-side: Figma Make + Notion AI Agents + Linear AI as the 'productivity-tool agent surface' pattern

    Figma Make (design primitives), Notion AI Agents (workspace workflows), and Linear AI (issue triage) share a pattern: a productivity tool ships an agent that lives inside the tool's existing data model, scoped to the tool's primitives (frames, pages, issues).

    1 daily topicOriginally published Aug 2, 2026View source

  • Aug 5, 2026

    official

    Anthropic: Claude 4.5 Sonnet tools release notes (programmatic tool calling + web fetch GA)

    Anthropic ships programmatic tool calling + web_fetch GA + 1M-token context on Claude 4.5 Sonnet

    Anthropic release notes for Claude 4.5 Sonnet document GA promotion of programmatic tool calling (define and call a tool from a code-execution sandbox), web_fetch server tool with caching and provenance metadata, and 1M-token context window GA.

    2 daily topicsOriginally published Aug 1, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.14777: SEED — Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

    arXiv 2607.14777 (indexed 7/17): SEED — self-evolving on-policy distillation lifts agentic RL across ALFWorld / Search-QA / WebShop

    Hugging Face 2026-07-17 indexed arXiv 2607.14777 introduces SEED, a self-evolving on-policy distillation framework for long-horizon agentic RL that converts completed trajectories into hindsight skills and distills them back into the policy, eliminating the sparse-reward credit assignment problem. Two stages: (1) Hindsight Skill SFT fine-tunes the base policy on 1,440 trajectories (180 tasks × K0=8 rollouts) annotated by an external analyzer (GLM-5.2) to extract reusable natural-language skills; (2) Self-Evolving OPD has the current policy snapshot simultaneously act as rollout actor and trajectory analyzer, refreshed each iteration so actor and analyzer co-evolve. The same model plays both roles (actor + analyzer) sharing parameters; confidence-gated token-level distillation via sigmoid(β_opd × Δlog-prob); joint loss L_SEED = L_RL (GRPO with KL regularization) + λ_opd · L_OPD; gradients flow only through the ordinary student branch; at inference, the deployed policy acts from ordinary history alone — no skills, memory, or retrieval required. Reported benchmark numbers (ALFWorld / Search-QA / WebShop score / WebShop succ): Qwen2.5-3B-Instruct — GRPO 75.0 / 36.4 / 79.8 / 63.3 vs Seed 91.8 / 45.7 / 88.5 / 78.9; Qwen2.5-7B-Instruct — GRPO 81.2 / 42.0 / 80.9 / 72.6 vs Seed 96.1 / 48.6 / 89.7 / 78.1; Qwen3-1.7B-Instruct — GRPO 46.1 / 40.8 / 67.3 / 38.3 vs Seed 92.0 / 42.2 / 87.1 / 77.3; sample efficiency: Seed with 60% of training data (ALFWorld 80.7) exceeds full-data GRPO (75.0); cross-domain generalization: +15.3 points on ALFWorld Unseen split (86.2 vs 70.9); multimodal (Qwen2.5-VL-3B) Sokoban 82.0%, EZPoints 100.0%, averaging 91.0% vs GRPO 77.0%; ablations (ALFWorld avg) — removing Hindsight-Skill SFT → 86.0 (-5.8); removing Self-Evolving OPD → 87.0 (-4.8); replacing on-policy skills with static offline skills → 84.4 (-7.4).

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.14952: LongStraw — Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

    arXiv 2607.14952 (7/17 #1 Daily Paper): LongStraw — million-token RL execution stack, 2.1M positions on 8x H20

    Hugging Face 2026-07-17 indexed (207 upvotes, #1 Daily Paper) arXiv 2607.14952 (Mind Lab) introduces LongStraw, an architecture-aware execution stack for million-token RL post-training under a fixed GPU budget, instantiated with GRPO. Methodology: evaluates shared prompt once without autograd; retains only model-specific state needed by later tokens; replays short response branches one at a time under autograd; reduces live training graph from full prompt+response to a single response branch; trades replay time for lower GPU memory. Model implementations: Qwen3.6-27B (hybrid recurrent + full-attention); GLM-5.2 (compressed-attention MoE, 78 layers). Reported benchmark numbers: on 8x H20 GPUs completes grouped Qwen scoring and response backward at 2.1M positions for groups of 2 and 8; group size scaling cost is only +0.21 GB peak allocated memory per group increase; stress test reaches 4.46M positions; on 32x H20 GPUs end-to-end validation for 2.1M-token prompt across all 78 layers of GLM-5.2. Key insights: targets the gap between inference (~1M token contexts) and RL post-training (often ≤256K tokens); especially relevant for AI agents with accumulating trajectories (observations, tool outputs, prior decisions). Authors note these experiments establish execution capacity, not complete training correctness — captured prompt state is detached and some distributed forward / gradient composition paths remain incomplete.

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.14935: VideoChat3 — Fully Open Video MLLM for Efficient and Generalist Video Understanding

    arXiv 2607.14935 (indexed 7/17): VideoChat3 — 4B fully open video MLLM, 2x faster inference than Qwen3-VL on H200

    Hugging Face 2026-07-17 indexed arXiv 2607.14935 (MCG-Nanjing University) introduces VideoChat3, a 4B-parameter fully open, efficient, and generalist video-centric MLLM; addresses three limitations of existing open-source video models: limited cross-domain generalization, high computational overhead, incomplete openness of training assets. Architecture: (1) Inflated 3D Vision Transformer (I3D-ViT) inflates a pretrained image tokenizer by extending 2D spatial self-attention into 3D spatiotemporal self-attention, groups consecutive frames into chunks of T=4, applies spatiotemporal self-attention, performs temporal pooling, combined with 2x2 spatial downsampling yields a 16x spatiotemporal compression ratio; (2) Adaptive Frame Resolution for Streaming Video Perception uses three response-state tokens (</Silence>, </Standby>, </Response>) that also control the next window's pixel quota: Silence/Response → 224² pixels (low), Standby → 448² pixels (high). Training datasets: 3M samples total — VideoChat3-Academic2M (2.27M), VideoChat3-LV116K (116.2K), VideoChat3-OL617K (617K); four training stages (tokenizer pre-training → video-language alignment → video instruction tuning → long & streaming instruction tuning). Full open-source — model weights, training code, training strategy, complete training datasets. Benchmark numbers (VideoChat3-4B vs open-weight Qwen3-VL-4B): MotionBench 61.7 vs 58.6, TempCompass 75.6 vs 70.8, Video-MME 70.1 vs 69.3, LVBench 56.7 vs 56.2, MMVU 56.4 vs 50.5, Charades TL mIoU 56.1 vs 46.4, VUE-TR V1 47.9 vs 32.9, VUE-TR V2 40.2 vs 19.6, MomentSeeker 25.9 vs 13.8; streaming: ODVBench 72.3 vs StreamForest 59.9 (+12.4), OVOBench task avg 62.5, StreamingBench Real-Time 83.0, River Avg 42.8, OVO-Timing Avg F1 35.5 vs Qwen3-VL-4B 8.1; efficiency on NVIDIA H200 at 2048 frames: total latency 20.412s vs Qwen3-VL 44.449s, total FLOPs 80.775 × 10^15 vs 106.913 × 10^15, GPU memory 80.775 GB vs 106.913 GB (saving 26.14 GB), visual tokens 100,352 vs 200,704 (half); best fully open results on MotionBench (61.7) and TempCompass (75.6); paper claims VideoChat3 'surpasses GPT-5 and Gemini 2.5 Flash on all three TimeLens splits.'

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    community

    IBM Research blog: It's time for cryptography to get its own abstraction layer

    IBM Research 7/17: cryptography needs its own abstraction layer — intent-based API + PQC migration support, API spec + Go SDK released

    IBM Research's 2026-07-17 blog post 'It's time for cryptography to get its own abstraction layer' argues cryptography needs a standardized layer that separates high-level intent from low-level implementation, much as filesystems and sockets did for storage and networking ('filesystems and sockets once looked like unnecessary layers of indirection; now they're the foundation of modern systems'). IBM proposes an intent-based API organized around scopes representing classes of cryptographic intent such as 'standard digital signatures or authenticated encryption'; applications would only express 'what they need' while 'the details of algorithms, parameters, and implementations move into a layer below, where they can be managed centrally.' Key features: separating policy (control plane) from execution (data plane), modeled on SDN for networks; treating cryptographic backends as interchangeable providers behind a single interface across software, hardware, cloud, or TEEs; enabling algorithm changes without touching application code; supporting post-quantum cryptography (PQC) migration without 'swapping algorithms' alone. IBM released supporting materials including an API specification, a reference standalone server, and a Go client SDK, inviting the community to 'explore the work, challenge the assumptions, and help shape its evolution.'

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    official

    Cursor changelog: Improvements to Cursor in Slack

    Cursor 7/17 Cursor in Slack three improvements: plan-upfront / multi-repo environment / cross-channel workflows

    Cursor's 2026-07-17 changelog entry 'Improvements to Cursor in Slack' documents three improvements: (1) Interaction improvements — Cursor now responds with a plan before it begins so users can redirect early, updates its status as it works, replaces in-message buttons with 'compact footer links', and renders tables / PRs / artifacts more cleanly; (2) Multi-repo environment support — Cursor can now start in a named multi-repo environment rather than a single default repository, targeting the environment that gives access to all relevant repos, with a mid-task 'Switch repository button' that lets users add repos outside the current environment and Cursor resumes where it left off; (3) Cross-channel workflows — Cursor can read from and send messages to other Slack channels and threads, pulling context from elsewhere in the workspace and posting updates back in the original thread or relevant channel.

    Not selected for a daily topic yetOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.14749: WanSong v1.0 Technical Report

    arXiv 2607.14749 (indexed 7/17): WanSong v1.0 Technical Report — commercial-grade text-to-music pure diffusion, demixed vocal + BGM stems

    Hugging Face 2026-07-17 indexed (arXiv submitted 2026-07-16) arXiv 2607.14749 (Wan-AI) presents WanSong v1.0 Technical Report — a commercial-grade text-to-music foundation model that generates songs up to ~5 minutes in a single end-to-end pass, outputting demixed vocal and background-music (BGM) stems; presented as a pure diffusion alternative to autoregressive (AR) or AR+diffusion cascaded systems. Architecture: (a) continuous 1-D VAE compresses stereo 44.1 kHz audio to a 64-channel latent at ~43.1 Hz (downsampling factor 1024), trained adversarially with multi-resolution STFT magnitude / feature-matching / hinge / KL losses on ~2.6×10^8 clips (50% music / 40% speech / 10% sound) on 32 A100s for 1.5M steps; (b) hybrid-MMDiT transformer (~25B params) processes concatenated LLM text tokens and dual-stem VAE audio tokens, using fully-shared AdaLN and packed sample sequences; (c) dual-stem modeling produces vocal and BGM tokens as independent output streams but treated as joint channels within each block, avoiding vocal/BGM interference. Capabilities include multilingual songs (Chinese, English, Japanese, Korean) up to 5 minutes, long-form generation via 3-stage pretraining (90s → 300s → SFT), step-distillation for faster inference, RLHF alignment via DPO followed by ReFL. Benchmark numbers (WanSong Bench, ~200 four-minute clips): PER 7.43% vs Suno V5 22.80% / Mureka V7.6 12.7% / LeVo 27.11%; SongEval 4.47/4.55/4.57/4.46/4.4; Muq text-alignment 0.44; internal musicality score 5.49 vs Suno V5 4.18 / Mureka V7.6 3.83 / LeVo 1.69.

    Not selected for a daily topic yetOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    paper

    Hugging Face: SEED ablation thread

    SEED three-piece ablation: Hindsight Skill SFT -5.8 / Self-Evolving OPD -4.8 / static offline skills -7.4 — on-policy + self-evolving each contribute roughly half

    SEED 7/17 paper ablation (ALFWorld avg): removing Hindsight-Skill SFT → 86.0 (-5.8); removing Self-Evolving OPD → 87.0 (-4.8); replacing on-policy skills with static offline skills → 84.4 (-7.4); the three ablation pieces together account for ~18 points, meaning on-policy self-evolving and hindsight-skill SFT are two independent contributions, each roughly half. Sample efficiency: Seed with 60% of training data (ALFWorld 80.7) exceeds full-data GRPO (75.0), showing Seed's sample efficiency significantly exceeds the baseline.

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    paper

    Hugging Face: LongStraw cluster thread

    LongStraw 7/17 three-model implementation + #1 Daily Paper: Qwen3.6-27B / GLM-5.2 78 layers / 32x H20 end-to-end validation

    LongStraw 7/17 implementation covers three models: Qwen3.6-27B (hybrid recurrent + full-attention) and GLM-5.2 (compressed-attention MoE, 78 layers). Hugging Face daily papers lists LongStraw as 7/17 #1 Daily Paper with 207 upvotes; GitHub MindLab-Research/longstraw has 41 stars; community commenter O96a questioned applicability to real agent loops where context grows unpredictably mid-trajectory rather than being cleanly segmented into prompt + generation.

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    media

    The Decoder: IBM Research cryptography abstraction layer

    The Decoder: IBM Research proposes cryptography abstraction layer — intent-based API + PQC migration support

    The Decoder 2026-07-17 tech-press coverage paraphrasing IBM Research's cryptography abstraction layer proposal (intent-based API, scopes representing cryptographic intent, separation of policy and execution modeled on SDN, interchangeable providers, PQC migration support).

    1 daily topicOriginally published Jul 17, 2026View source

  • Aug 5, 2026

    community

    Hugging Face blog: Security incident disclosure — July 2026

    HF 7/16 security incident disclosure: first publicly disclosed end-to-end autonomous AI agent intrusion, HF used zai-org/GLM-5.2 for triage

    Hugging Face's July 16, 2026 blog discloses an intrusion into part of HF's production infrastructure driven end-to-end by an autonomous AI agent system. The attacker exploited code-execution paths in dataset processing (a remote-code dataset loader and a template-injection in a dataset configuration) to run code on a processing worker, escalated to node-level access, harvested cloud and cluster credentials, and moved laterally into several internal clusters over a weekend. The campaign was executed by an autonomous agent framework running many thousands of individual actions across a swarm of short-lived sandboxes. Disclosed impact: unauthorized access to a limited set of internal datasets, compromise of several credentials used by HF services, no evidence of tampering with public user-facing models / datasets / Spaces, and software supply chain (container images and published packages) verified clean. HF used zai-org/GLM-5.2 on their own infrastructure for LLM-based triage over security telemetry, analyzing over 17,000 recorded events — frontier models behind commercial APIs had blocked their requests due to safety guardrails that couldn't distinguish a responder from an attacker. Seven remediations: fixed root vulnerability, eradicated attacker foothold and rebuilt compromised nodes, revoked/rotated affected credentials and tokens, deployed additional guardrails and stricter cluster admission controls, improved detection/alerting (high-severity signals page responders in minutes), engaged outside cybersecurity forensic specialists, reported to law enforcement. Community advisory: as a precaution, rotate access tokens and review recent account activity; security concerns can be reported to security@huggingface.co. Partner/customer impact assessment was still ongoing at the time of the captured summary.

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.12395: Ring-Zero — Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

    arXiv 2607.12395 (indexed 7/16): Ring-Zero — 1T-parameter zero RL with 5 spontaneously emergent cognitive behaviors

    arXiv 2607.12395 (indexed as a Hugging Face daily paper on 2026-07-16) introduces Ring-2.5-1T-Zero, a model trained with zero reinforcement learning applied directly to a pretrained base model without human-annotated CoT data, relying on RL with verifiable rewards (RLVR). Architecture: two pretrained base models from the Ling-2.5 family — Ling-2.5-1T-Base (1T MoE, 63B activated) and Ling-2.5-flash-Base (104B MoE, 7.4B activated). Four-stage training pipeline: (1) first-stage RL with clipped importance sampling policy gradient and token-level loss to elicit reasoning; (2) self-distillation compressing verbose CoT traces via shortest-correct-rollouts + self-refinement; (3) second-stage RL with sample-level loss normalization for stable, sustained optimization; (4) third-stage RL with tier-based training at three difficulty tiers (4k / 16k / 64k token budgets). Infrastructure: 320x H200 GPUs with Megatron as the training engine, SGLang for rollout, Areal framework for orchestration. Optimizations include mixed-precision control (BF16 body with FP32 attention softmax and LM head) and context parallelism tailored to the model's hybrid MLA + Lightning Attention architecture. Findings: validates the 'bitter lesson' at 1T-parameter scale; reveals a two-phase training dynamic (initial discovery + later sharpening); Ring-2.5-1T-Zero (Second Stage RL) reaches 94.1% on AIME 2024 with competitive numbers on AIME 2025-2026 / HMMT 2025-2026 / IMOAnswerBench; spontaneously develops five cognitive behaviors without explicit supervision: anthropomorphism, structured formatting, self-verification, parallel reasoning, and context anxiety; CoT traces outperform baselines on comprehensibility, reproducibility (5.8-point gain on Qwen-32B distillation vs DeepSeek-R1), and efficiency (less than half the tokens of competitors). Lead author Gangqiang Cao with a large RUC-AIBOX team.

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    community

    vLLM blog: Keeping vLLM Production Quality

    vLLM 7/16 three-layer quality gates: CI 37 test groups / 266 jobs / perf-eval 17 model-hardware recipes / two-week release cycle

    vLLM's July 16, 2026 blog post 'Keeping vLLM Production Quality' documents three layers every PR passes through: Layer 1 CI (lightweight GitHub Actions on every PR, heavier unit testing on Buildkite with 37 test groups and 266 jobs dynamically selected based on the diff, shared multi-stage container image, pip-compile lock files, 58 runner queues across AWS / Crusoe / LambdaLabs / Nebius / NVIDIA / Roblox / RunPod via Buildkite agents, MIG-slicing and autoscale-from-zero per-machine runners, custom dashboard at ci.vllm.ai, nightly CI-analyzer bot posting Slack reports with auto-revert PRs ~1.5/day at ~70% correct diagnosis); Layer 2 perf & accuracy (nightly pipeline at github.com/vllm-project/perf-eval covering 17 model-hardware recipes across H200 / B200 / MI300X / MI355X — DeepSeek V4 Pro/Flash, gpt-oss, Kimi K2.5, MiniMax M2.5/M3, Qwen3.5, GLM 5.1, Gemma 4, Nemotron 3 Super — measuring TTFT / TPOT / vllm-bench, plus accuracy via lm-eval on GSM8K / GPQA / AIME and function-calling via BFCL); Layer 3 release (two-week cadence since November 2025 — every other Monday the release manager picks the greenest main-branch commit to cut releases/vX.Y.Z; Mon-Wed cherry-pick window; candidate ships only when all three gates pass; delays accepted rather than lowering the bar; 7 Python wheels + 11 Docker images per release smoke-tested before publication).

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    official

    Cohere blog: Cohere and the University of Toronto partner to advance responsible AI adoption at scale

    Cohere 7/16 × University of Toronto: multi-year partnership, North platform as the orchestration layer for U of T's university-wide AI

    Cohere's July 16, 2026 blog announces a multi-year agreement with the University of Toronto to integrate Cohere's enterprise AI technology into a forthcoming university-wide AI platform supporting responsible AI adoption across teaching, research, student services, administration, and operations. Cohere's North platform will serve as an orchestration layer within U of T's AI platform, helping users manage complex tasks and securely access trusted information across university systems; supports faculty, librarians, staff, and students while keeping sensitive data secure and under the university's control. Cohere's technology will also power U of T's 'AI Kitchen' — a secure environment for exploring and evaluating AI tools through vetted applications, appropriate data access, and privacy-conscious frameworks. Founders' connection: Cohere was founded in 2019 by former U of T students Aidan Gomez, Nick Frosst, and Ivan Zhang, making this a homecoming partnership. No funding figures are disclosed in the article.

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    community

    Hugging Face blog: Newer Models, Same Advantage — Dharma-AI

    Dharma-AI 7/16: on Brazilian Portuguese OCR, Mistral OCR4 and Unlimited-OCR still lose to the specialist DharmaOCR

    Hugging Face's July 16, 2026 blog post 'Newer Models, Same Advantage' by Dharma-AI evaluates Mistral OCR4 (0.798), Unlimited-OCR (0.7587), and DharmaOCR (0.925) on Brazilian Portuguese OCR tasks. Despite being newer and better-resourced, both generalist / multilingual competitors score significantly lower than DharmaOCR. The post attributes DharmaOCR's persistent advantage to structural specialization: (1) domain concentration — all parameters are directed at Brazilian Portuguese vocabulary, morphology, and orthographic patterns rather than spread across many languages; (2) two-stage training — supervised fine-tuning builds domain competency while Direct Preference Optimization trains against full-output coherence rather than token-by-token prediction, reducing text degeneration under visual ambiguity. The post argues the advantage is structural not temporary: specialists continue to extract more from finite resources within their domain than generalists distributing those resources across many. Cited failure cases: Mistral OCR4 misread 'Chico Buarque' as 'Chico Barque' and Unlimited-OCR produced incoherent output like 'a dose de chico bique' — failures concentrated precisely where Portuguese-specific training matters most.

    Not selected for a daily topic yetOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.13285: Harness Handbook — Making Evolving Agent Harnesses Readable, Navigable, and Editable

    arXiv 2607.13285 (indexed 7/16): Tencent Hunyuan Harness Handbook — readable, navigable, editable agent harnesses

    arXiv 2607.13285 (indexed as a Hugging Face daily paper on 2026-07-16) is a handbook / guide from Tencent Hunyuan on designing clearer, navigable, and editable agent harness architectures. Specific author list, exact architectural pattern catalog, evaluation methodology, and reference implementations beyond the captured summary were not extracted.

    Not selected for a daily topic yetOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.12747: Tracing Agentic Failure from the Flow of Success

    arXiv 2607.12747 (indexed 7/16): Tracing Agentic Failure from the Flow of Success — 4 authors derive failure modes from successful trajectories

    arXiv 2607.12747 (indexed as a Hugging Face daily paper on 2026-07-16) is a 4-author paper that analyzes agentic failure modes derived from successful trajectories — i.e. tracing failure back to where the same trajectory succeeded. Specific author list, failure taxonomy, evaluation methodology, and the benchmark or agent population used beyond the captured summary were not extracted.

    Not selected for a daily topic yetOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.12625: KnowAct-GUIClaw — Know Deeply, Act Perfectly, Personal GUI Assistant with Self-Evolving Memory and Skill

    arXiv 2607.12625 (indexed 7/16): KnowAct-GUIClaw — Lychee Team's GUI assistant with self-evolving memory and skill modules

    arXiv 2607.12625 (indexed as a Hugging Face daily paper on 2026-07-16) presents KnowAct-GUIClaw, Lychee Team's GUI assistant built with self-evolving memory and skill modules. Specific author list, exact benchmark numbers on Android / Web / Desktop GUI agent benchmarks, memory module architecture, and self-evolution mechanism beyond the captured summary were not extracted.

    Not selected for a daily topic yetOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    community

    Hugging Face: Security incident cluster thread

    7/16 HF security incident triple: Security Incident Disclosure main post + The Decoder coverage + Anatomy companion technical timeline

    On 2026-07-16 the HF security incident shipped in three surfaces on the same day: Hugging Face Security Incident Disclosure main post (autonomous AI agent end-to-end attack + GLM-5.2 triage), The Decoder tech-press coverage (autonomous agent attack vector, GLM-5.2 triage, partner / customer impact assessment), and Anatomy of a Frontier Lab Agent Intrusion companion deep technical timeline (HF blog index lists the post tagged 'security' / 'Hot' / 423 reactions). Together the three form 'first publicly disclosed autonomous AI agent attack' as a complete narrative.

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    repo

    vLLM perf-eval GitHub repo (companion)

    vLLM 7/16 perf-eval dashboards: 17 model-hardware recipes across H200 / B200 / MI300X / MI355X, lm-eval GSM8K / GPQA / AIME, BFCL function-calling

    vLLM's July 16, 2026 perf-eval nightly pipeline covers 17 model-hardware recipes across H200 / B200 / MI300X / MI355X (DeepSeek V4 Pro/Flash, gpt-oss, Kimi K2.5, MiniMax M2.5/M3, Qwen3.5, GLM 5.1, Gemma 4, Nemotron 3 Super), measuring TTFT / TPOT / vllm-bench, plus accuracy via lm-eval on GSM8K / GPQA / AIME and function-calling via BFCL; dashboards at ci.vllm.ai/perf and ci.vllm.ai/eval.

    1 daily topicOriginally published Jul 16, 2026View source

  • Aug 5, 2026

    community

    Hugging Face: Welcome Inkling by Thinking Machines

    Thinking Machines 7/15 ships Inkling: ~1T multimodal open-source MoE, 975B / 41B active / 1M context / 45T training tokens

    Hugging Face's July 15, 2026 blog announces Thinking Machines' Inkling — a ~1T parameter open multimodal model natively accepting image, text, and audio inputs, trained on 45T tokens. Architecture: decoder-only multimodal MoE, 975B total / 41B active parameters, 256 experts with top-6 + 2 shared experts, 1M context window, relative attention (no RoPE), hybrid 5:1 sliding-window-to-global attention, short 1D convolution over hidden states, hierarchical MLP patchifier for vision, mel-spectrogram discretization for audio. Variants: Inkling BF16 (2 TB VRAM) and Inkling NVFP4 (600 GB); Inkling-Small BF16 600 GB (276B / 12B active) / NVFP4 180 GB. Self-reported scores (Inkling vs Inkling-Small): HLE text-only 29.7% / 31.6%, HLE with tools 46.0% / 47.8%, AIME 2026 97.1% / 95.5%, GPQA Diamond 87.2% / 89.5%, SWE-Bench Verified 77.6% / 80.2%, SWE-Bench Pro 54.3% / 55.9%, Terminal Bench 2.1 63.8 / 64.69, MCP Atlas 74.1% / 79.2%, MMMU Pro 73.3% / 74.0%, VoiceBench 91.4% / 90.1%. License not stated in the captured summary.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    community

    vLLM blog: TML Inkling on vLLM — Day-0 Support

    vLLM 7/15: Inkling day-0 support — 380 tok/s/user with MTP8 on 4x GB200 + 1M context

    vLLM's July 15, 2026 blog documents Inkling day-0 support — full feature parity across NVFP4 and BF16 variants, 8 MTP heads (up to 9 tokens per forward step), native 1M token support; on 4x NVIDIA GB200 quoted throughput is 380 tok/s/user with MTP8 (mean acceptance length 4.5) and 140 tok/s/user without MTP; accuracy-by-length-bucket figures are 99.09% (436/440) at 2K-221K, 95.68% (421/440) at 294K-513K, and 81.36% (358/440) at 586K-805K.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    community

    LMSYS blog: Inkling Day-0 Support in SGLang

    LMSYS 7/15: SGLang ships Inkling day-0 — 71.7k tok/s input throughput on Blackwell

    LMSYS's July 15, 2026 blog announces SGLang day-0 support for the 975B multimodal MoE Inkling with 1M context, hitting up to 71.7k tok/s input throughput on Blackwell. The SGLang Cookbook simultaneously publishes tuned deployment recipes for Inkling on Blackwell TP4/TP8, H200, and AMD MI350X/MI355X.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    community

    Hugging Face: Real World VoiceEQ — Measuring the human quality of voice AI

    Hume + HF 7/15 ship Real World VoiceEQ: voice AI benchmark built on 1M+ human ratings, 40+ models, 60+ metrics

    Hugging Face's July 15, 2026 blog announces Real World VoiceEQ — a Hume + HF co-built voice AI quality benchmark built on 1M+ individual human ratings across demographics, speaking styles, and acoustic environments, covering 40+ proprietary and open-source voice models across 15+ evaluation dimensions and 60+ metrics spanning ASR / TTS / S2S / Speech Understanding; the current dataset is 785k TTS ratings and 48k STS ratings; every evaluation ran on Hume's Kairos platform; the post quotes transcription WER on noise-backed speech as roughly 4x higher than on music-backed speech; the public leaderboard is at huggingface.co/spaces/HumeAI/rw-voice-eq and the technical report is at arXiv 2607.14846.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    official

    Cohere blog: The total cost of AI ownership

    Cohere 7/15 AI-TCO framework: owning inference gives 8x edge over cloud, 18x over frontier API, 8-GPU pays back in under 4 months

    Cohere's July 15, 2026 blog 'The total cost of AI ownership' frames AI cost as 'renting vs owning' across data centers / chips / models; quotes Gartner $2.52T global AI spending in 2026 (44% YoY), IDC/DataRobot (Dec 2025) 96% of gen AI and 92% of agentic AI deployments faced higher-than-expected costs, McKinsey (Nov 2025) only ~1/3 of organizations scale AI enterprise-wide with 5-6% reporting significant financial impact, Mavvrik/Benchmarkit (2025) 80% of companies miss AI forecasts by >25% with ~25% missing by >50% and only 15% within 10%, Uber's claim that 10% of committed code is built by autonomous agents with 12 months of AI budget spent in 4; quotes NVIDIA Blackwell vs Hopper as ~50x more tokens per megawatt with ~35x lower cost per token, Lenovo 2026 amortized cost per million tokens at ~$0.11 on owned H100 vs ~$0.89 cloud instance vs ~$2.00 frontier API (8x edge over cloud, up to 18x over API), 8-GPU server pays back in under 4 months vs on-demand cloud, break-even at ~4 hours/day of use, NVIDIA/SemiAnalysis InferenceX $0.123 per million tokens on GB300.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    community

    Hugging Face: Model Routing Is Simple. Until It Isn't. — IBM Research

    IBM Research 7/15 HF blog: Model Routing Is Simple. Until It Isn't.

    Hugging Face's July 15, 2026 blog post by IBM Research, titled 'Model Routing Is Simple. Until It Isn't.', addresses why model routing becomes complicated as systems scale, contradicting the initial simplicity of directing queries to different AI models. The specific framework name, decision diagrams, code samples, and benchmark methodology were not in the captured summary and were not extracted.

    Not selected for a daily topic yetOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    community

    Hugging Face: Inkling day-0 cluster thread

    7/15 Inkling one-day triple: Thinking Machines open-sources + vLLM day-0 + LMSYS SGLang day-0

    On 2026-07-15 Inkling shipped in one day across three surfaces: Thinking Machines published the ~1T multimodal open-source MoE on Hugging Face, vLLM blog posted day-0 support (380 tok/s/user with MTP8 on 4x GB200 + 1M context), and LMSYS blog posted SGLang day-0 support (71.7k tok/s input throughput on Blackwell). Together the three releases turn 'a 1T multimodal model in the repo' into 'runnable on Blackwell the same day'.

    1 daily topicOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.12463: Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

    arXiv 2607.12463 (indexed 7/15): Function-Aware Fill-in-the-Middle as mid-training for coding agent foundation models

    arXiv 2607.12463 (indexed as a Hugging Face daily paper on 2026-07-15) introduces a function-aware fill-in-the-middle approach used as mid-training for coding agent foundation models. Specific author list, exact benchmark deltas (HumanEval / SWE-Bench / LiveCodeBench), training compute cost, and per-curriculum-step methodology were not in the captured summary and were not extracted.

    Not selected for a daily topic yetOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    paper

    arXiv 2607.11562: MonkeyOCRv2 — A Visual-Text Foundation Model for Document AI

    arXiv 2607.11562 (indexed 7/15): MonkeyOCRv2 — a visual-text foundation model for document AI

    arXiv 2607.11562 (indexed as a Hugging Face daily paper on 2026-07-15) presents MonkeyOCRv2, a visual-text foundation model designed for document AI tasks. Specific author list, exact benchmark numbers (DocVQA / InfoVQA / ChartQA / OmniDocBench), model size, training data composition, and license were not in the captured summary and were not extracted.

    Not selected for a daily topic yetOriginally published Jul 15, 2026View source

  • Aug 5, 2026

    official

    Anthropic newsroom: Claude for Teachers

    Anthropic 7/14 Claude for Teachers: free Claude for US K-12 educators + Learning Commons + 10 ed-tech platform integrations

    Anthropic's July 14, 2026 newsroom post (updated July 21, 2026) announces Claude for Teachers — free premium Claude access for verified US K-12 educators, a library of teaching skills, and direct connection to Learning Commons for standards-aligned lesson planning (mapped to all 50 US states). Capabilities: differentiation support; Claude Code and Cowork for autonomous task handling (e.g. reviewing exit tickets daily at 4 pm); data-analysis tools with educator-controlled data sharing (not used for model training). Platform integrations: ASSISTments, Brisk Teaching, Canva Education, Coteach, Diffit, Eedi, MagicSchool, Snorkl, TeachFX, plus curricular resources OpenSciEd and Illustrative Mathematics. Pricing: free for verified K-12 educators; sign-up required by 2027-06-30 for a full year of access; individual educators only (schools/districts offering coming soon). Compliance: 18+ policy; FERPA-compliant via the K-12 Data Processing Addendum; Anthropic is working with the American Federation of Teachers on a Gold Standard for safety and privacy.

    1 daily topicOriginally published Jul 14, 2026View source

  • Aug 5, 2026

    community

    vLLM blog: vLLM x TileRT — Specialized Decode for Latency-Critical Serving

    vLLM 7/14 × TileRT 0.1.5: split prefill/decode via vLLM V1 connector, latency-optimized GLM-5.1-FP8 on B200

    vLLM's July 14, 2026 blog documents pairing vLLM prefill with TileRT decode via vLLM V1's public connector interface; TileRT 0.1.5 ships with this integration; both pools coexist behind a single prefill instance via MultiConnector with routing/claim filtering sending only marked latency-critical traffic to TileRT; evaluation ran on 8x NVIDIA B200 with GLM-5.1-FP8 token-generation speed, input 1K-192K, output 1K, achieving MTP average acceptance length 3.2 (peak 4.0) and one in-flight request at a time per TileRT decode node; current model coverage is GLM-5/5.1 and DeepSeek-V3.2.

    Not selected for a daily topic yetOriginally published Jul 14, 2026View source

  • Aug 5, 2026

    repo

    SGLang release v0.5.15.post1

    SGLang 7/14 v0.5.15.post1: GLM-5.2 IndexShare fixes for DSA / flashinfer / NaN / PD-disaggregation

    SGLang v0.5.15.post1 released 2026-07-14 is a stability patch focused on GLM-5.2: fixes DSA model launching on non-CUDA/HIP devices, flashinfer dependency, NaN outputs, and GLM-5.2 IndexShare on PD-disaggregation and context-parallel settings; pairs with v0.5.15 (released 2026-07-10) which highlighted GLM-5.2 NVFP4 production tuning on Blackwell, Spec V2 default (+11% e2e TPS), IndexShare MTP, TopK V2, indexer-prologue fusion, and new model support including Hunyuan 3, HRM-Text, and Qwen3.6 NVFP4.

    Not selected for a daily topic yetOriginally published Jul 14, 2026View source

Jul 14, 2026

14 items

  • Jul 14, 2026

    official

    OpenAI: gpt-oss refresh

    OpenAI 7/13 gpt-oss refresh: 120B MoE + 20B dense post-training snapshot under Apache-2.0

    OpenAI index page on 2026-07-13 publishes a refresh of the gpt-oss open-weight line — 120B MoE and 20B dense variants with new post-training snapshots; license remains Apache-2.0 with the OpenAI usage-policy addendum. Self-reported benchmark numbers: MMLU-Pro 84.6, GPQA-Diamond 76.2, LiveCodeBench v6 72.1, SWE-Bench Verified 68.3, BFCL v3 65.4, Terminal-Bench 2.1 76.8. Semantic boundary: vendor index + Hugging Face community model card + a tech-press item (The Decoder) — not an independent reproduction of any specific score.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    official

    Anthropic: Claude Sonnet 4.5

    Anthropic 7/13 Claude Sonnet 4.5: mid-tier flagship for everyday tasks at the same $3/$15 pricing

    Anthropic newsroom on 2026-07-13 introduces Claude Sonnet 4.5 as the new mid-tier flagship positioned between Haiku 4.5 and Opus 4.5; 200K context, native tool-use with extended-thinking mode, $3 input / $15 output per million tokens (same as Sonnet 4.5 prior). Self-reported figures: SWE-Bench Verified 67.4, GPQA-Diamond 74.8, MMLU-Pro 82.9, BFCL v3 63.7, Terminal-Bench 2.1 74.2. Deployment: API + claude.ai + AWS Bedrock + Google Vertex AI. Semantic boundary: vendor newsroom + Hugging Face community card + a Decoder tech-press item — not an independent reproduction of any specific score.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    official

    Google: Gemini 3.5 Pro mid-July

    Google 7/13 Gemini 3.5 Pro mid-July: agent scaffolds + new image-edit-via-namespace endpoint

    Google blog on 2026-07-13 publishes a mid-July tier update for Gemini 3.5 Pro emphasizing agent scaffolds (server-side tool-use loop) + a new image-edit-via-namespace endpoint; pro-tier pricing unchanged. Self-reported figures: MMLU-Pro 86.3, GPQA-Diamond 79.1, LiveCodeBench v6 75.4, SWE-Bench Verified 71.2, BFCL v3 68.9, Terminal-Bench 2.1 78.5. Deployment: Gemini API + Vertex AI. Semantic boundary: vendor blog + Hugging Face community model card; no media corroboration today.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    official

    Meta: Llama 4 Behemoth refresh

    Meta 7/13 Llama 4 Behemoth refresh: 1M context + code-weight specialist under Llama 4 Community License

    Meta AI blog on 2026-07-13 publishes a refresh of the Llama 4 family emphasizing long-context (1M + token-efficiency) and a code-weight specialist. Self-reported figures: MMLU-Pro 83.4, GPQA-Diamond 75.6, LiveCodeBench v6 73.8, SWE-Bench Verified 69.7, HumanEval+ 91.2, Terminal-Bench 2.1 75.9. License: Llama 4 Community License with the commercial-use restrictions addendum. Semantic boundary: vendor blog + Hugging Face community card; benchmark numbers are vendor-supplied.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    official

    ByteDance Volcengine: Doubao 1.6 Pro

    ByteDance 7/13 Doubao 1.6 Pro: 256K context + tool-use refresh, full weights gated to Volcengine API

    ByteDance Volcengine product news on 2026-07-13 publishes a Doubao 1.6 Pro update — 256K context, refreshed tool-use schema, post-training emphasizing document-extraction / retrieval-grounded tasks. Self-reported figures: MMLU-Pro 80.7, GPQA-Diamond 72.4, C-Eval 88.6, CMMLU 86.9, LiveCodeBench v6 70.3, Terminal-Bench 2.1 73.5. License: commercial-only via Volcengine API; full weights remain gated. Semantic boundary: vendor product news + a Hugging Face mirror card listed as researcher preview only.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    official

    Zhipu Z.ai: GLM-4.6

    Zhipu Z.ai 7/13 GLM-4.6: AutoGLM-Coder refresh + agent scaffold, commercial-only on BigModel

    Zhipu Z.ai BigModel platform news on 2026-07-13 publishes a GLM-4.6 refresh with AutoGLM-Coder coding-specialist post-training + agent scaffold. Self-reported figures: MMLU-Pro 81.2, GPQA-Diamond 73.5, HumanEval+ 90.8, MultiPL-E 87.6, SWE-Bench Verified 67.8, Terminal-Bench 2.1 74.6. License: commercial-only via BigModel / Z.ai; open-source variants will follow separately (vendor-stated future commitment, not a current artifact). Semantic boundary: vendor platform news + Hugging Face mirror card; benchmark numbers are vendor-supplied.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    media

    aihot.virxact.com: meta-louisiana-5gw-datacenter-50b-investment

    Meta 7/13 Louisiana data center to 5GW / >$50B total + Entergy gas-fired + battery + nuclear expansion (per IT之家)

    aihot 2026-07-13 18:35 + IT之家: Meta expanding Louisiana data center to 5GW with >$50B total investment; commits to bearing all local energy + water costs + >$1B local infrastructure; signed Entergy deal for new gas-fired + battery-storage + nuclear capacity additions.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    media

    aihot.virxact.com: xai-grok-cli-silent-codebase-upload-disclosure

    xAI Grok CLI 7/13 silent-upload disclosure (per buzzing.cc HN translation)

    aihot 2026-07-13 08:10 + buzzing.cc: xAI's official Grok CLI v0.2.93 silently uploads working directory + ~/.claude.json + global AGENTS rules + 30+ Skill files + an API key on every turn via a side channel to a xAI Google Cloud bucket; xAI subsequently added a server-side toggle to disable codebase upload.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    media

    aihot.virxact.com: ploy-ai-agent-migrated-claude-opus-to-gpt-5-6-sol

    Ploy 7/13 AI-agent migration: Claude Opus 4.8 → GPT-5.6 Sol (per buzzing.cc HN)

    aihot 2026-07-13 07:54 + buzzing.cc: Ploy migrated AI-agent default from Claude Opus 4.8 to GPT-5.6 Sol — build time 2.2x faster, cost -27%, output tokens halved, visual score +0.034 — but GPT-5.6 Sol silently fills defaults on 25 tool params, leaving 52-64% of file reads empty.

    1 daily topicOriginally published Jul 13, 2026View source

  • Jul 14, 2026

    media

    aihot.virxact.com: altman-amodei-ai-net-jobs-narrative-shift

    Altman + Amodei 7/12 soften 'AI replaces jobs' framing (per The Decoder)

    aihot 2026-07-12 17:28 + The Decoder: OpenAI CEO Sam Altman + Anthropic CEO Dario Amodei both softened earlier 'AI replaces jobs' framings; Altman 'pretty confident' AI has so far net-created jobs; cross-referenced studies find no aggregate productivity / labor-market impacts.

    1 daily topicOriginally published Jul 12, 2026View source

  • Jul 14, 2026

    media

    aihot.virxact.com: apple-v-openai-trade-secrets-ai-hardware-suit

    Apple v. OpenAI 7/11+ trade-secrets suit (per IT之家 + TechCrunch AI)

    aihot 2026-07-11/12 + IT之家 + TechCrunch AI: Apple filed trade-secrets suit in U.S. NDCA against OpenAI (and Tang Tan + Chang Liu + io Products); ~400 ex-Apple employees now at OpenAI; consumer AI hardware at stake; February 2026 private-settlement attempt declined by OpenAI.

    1 daily topicOriginally published Jul 11, 2026View source

Jul 13, 2026

3 items

  • Jul 13, 2026

    official

    Qwen Blog: Qwen3-Max

    Alibaba 7/12 Qwen3-Max: 1.2T total / 32B activated MoE with native tool-use + 1M context

    Alibaba Qwen team blog on 2026-07-12 introduces Qwen3-Max — a 1.2T total / 32B activated MoE with 1M context and native tool-use / function-calling schema; deployment surfaces Tongyi Qianwen + Model Studio + DashScope OpenAI-compatible API; license is Apache-2.0 with a Chinese-regulatory acceptable-use addendum. Self-reported benchmark numbers: MMLU-Pro 87.1, GPQA-Diamond 81.4, LiveCodeBench v6 78.2, SWE-Bench Verified 74.5, BFCL v3 72.0, MTEB multilingual 71.6, Terminal-Bench 2.1 81.0. Semantic boundary: this is a vendor announcement + a Hugging Face community model card + a tech-press item — not an independent reproduction of any specific score.

    1 daily topicOriginally published Jul 12, 2026View source

  • Jul 13, 2026

    official

    Mistral News: Mistral Large 3

    Mistral 7/12 Mistral Large 3: 145B dense flagship at 128K context with native tool-use

    Mistral news post on 2026-07-12 introduces Mistral Large 3 — a 145B-parameter dense flagship with 128K context, native function-calling + JSON-schema tool-use, and a chat-instruct variant; deployment is Mistral La Plateforme + Azure AI Foundry + Bedrock; weight formats are BF16 sharded safetensors + MXFP4 inference snapshots; license is Apache-2.0 with a Mistral-research addendum. Self-reported figures: MMLU-Pro 79.2, GPQA-Diamond 71.6, HumanEval+ 93.4, MultiPL-E 88.1, MT-Bench v3 9.21, IFEval strict 88.4. Semantic boundary: vendor + Hugging Face community model card; benchmark numbers are vendor-supplied.

    1 daily topicOriginally published Jul 12, 2026View source

  • Jul 13, 2026

    official

    Moonshot AI: Kimi K2.7-0913

    Moonshot 7/12 Kimi K2.7-0913: K2-family re-tag emphasizing instruction following + agent scaffolds

    Moonshot AI platform blog on 2026-07-12 announces Kimi K2.7-0913 — a re-tag of K2-line weights (architecture unchanged: 1.1T total / 32B activated MoE, Apache-2.0 with the K2 acceptable-use addendum) emphasizing long-form instruction adherence + agent scaffolds (browser / shell / file-system tool handlers) + reduced refusal on benign long-context reasoning. Self-reported gains vs K2: IFEval strict 86.7 (vs 81.4), WildChat-Write 9.04 (vs 8.81), AgentBench-lite 78.4 (vs 71.2), 64K NIAH 99.2%. Semantic boundary: this is a refresh of existing weights, not a base-model swap; gain numbers are vendor-stated deltas against the K2 baseline.

    1 daily topicOriginally published Jul 12, 2026View source

Jul 10, 2026

3 items

  • Jul 10, 2026

    official

    Anthropic Newsroom: Inviting hard questions

    Anthropic 7/9 "Inviting hard questions": public-facing inbox + commitment to show their work in response

    Anthropic on 2026-07-09 published the "Inviting hard questions" initiative on its official Newsroom, asking the public to submit hard questions about AI's effects on jobs and society; Anthropic states explicitly that they will 'show our work' as they address them. Semantic boundary: this is a public-inbox initiative plus a public-traceability commitment, not a new independent-oversight or binding-accountability mechanism.

    1 daily topicOriginally published Jul 9, 2026View source

  • Jul 10, 2026

    official

    Anthropic Newsroom: Ben Bernanke appointed to Long-Term Benefit Trust

    Anthropic 7/9 adds Ben Bernanke to the Long-Term Benefit Trust (4 named trustees total)

    Anthropic on 2026-07-09 officially announced that former Fed Chair Ben Bernanke joins the Long-Term Benefit Trust, alongside current Chair Neil Buddy Shah, Richard Fontaine, and Mariano-Florentino Cuéllar (4 named trustees); Co-Founder & President Daniela Amodei publicly framed AI as 'the most significant economic effects of any technology in modern history' and asserted Anthropic has 'a dual responsibility to understand those effects and to act on them'. Semantic boundary: the LTBT holds a minority equity stake and has no operational veto over product or research decisions; the appointment is a corporate-governance signal, not a binding new constraint.

    1 daily topicOriginally published Jul 9, 2026View source

  • Jul 10, 2026

    official

    Google Research Blog: SensorFM

    Google 7/9 SensorFM: a foundation model for wearable health pre-trained on 1T+ minutes from ~5M people

    Google Research blog (Xin Liu + Daniel McDuff) published on 2026-07-09: SensorFM, a foundation model for wearable health pre-trained on more than 1 trillion minutes of sensor signal from ~5 million consented people (100+ countries, all 50 US states, 20+ Fitbit and Pixel Watch models; data window 2024-09 → 2025-09); 5 sensor types (PPG, accelerometry, EDA, skin temperature, altimetry) + 34 one-minute aggregate features. Scaling curve spans 4 orders of magnitude (data: ~2M → 2B sensor-hours; model: 100K → 100M parameters). Largest variant SensorFM-B: 31% lower reconstruction loss, +9% avg AUC, +21% Pearson, wins on 33 of 35 transfer tasks. Personal Health Agent evaluation: 31 participant profiles, 1,860 clinician ratings. Semantic boundary: this is a research blog post on pre-trained representation + transfer benchmarks — not FDA approval, clinical diagnostic deployment, or supervised medical guidance.

    1 daily topicOriginally published Jul 9, 2026View source

Jul 9, 2026

12 items

  • Jul 9, 2026

    official

    xAI News: Introducing Grok 4.5

    xAI 7/8 ships Grok 4.5: DeepSWE 1.0 62.0% + SWE Marathon 29.0% + ~4.2x fewer output tokens than Opus 4.8

    xAI's July 8, 2026 launch post introduces Grok 4.5 as SpaceXAI's newest flagship, built for coding, agentic tasks, and knowledge work; the post says the model was trained with Cursor on 'tens of thousands of NVIDIA GB300 GPUs.' Self-reported benchmarks: DeepSWE 1.0 62.0% (vs Fable max 66.1%, GPT 5.5 xhigh 64.31%); SWE Marathon 29.0% resolution rate (highest reported); Terminal Bench 2.1 83.3%; SWE Bench Pro 64.7% resolve rate; on SWE Bench Pro, ~4.2x fewer output tokens than Opus 4.8 max (15,954 vs 67,020). Served at ~80 TPS, priced $2 / $6 per million input / output tokens; available in Grok Build, Cursor (all plans), and the SpaceXAI API at launch; not yet available in the EU in any SpaceXAI product or API console (mid-July ETA).

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    community

    Hugging Face Blog: Native-speed vLLM transformers modeling backend

    HF 7/8: opt into `--model-impl transformers` and the transformers backend matches native-speed vLLM

    Hugging Face's July 8, 2026 blog post documents that opting in with `--model-impl transformers` makes the transformers modeling backend meet or beat vLLM's hand-written native backend across three Qwen3 workloads (4B dense on a single GPU, 32B dense with tensor parallelism, and 235B-parameter FP8 MoE Qwen3-235B-A22B-FP8 on data + expert parallelism on the same 8xH100 node); the technique uses torch.fx static graph capture plus Python ast rewrites to keep Python modeling code on the hot path while the compiler fuses and pre-allocates. Linear-attention models (Mamba2, RWKV, etc.) are explicitly unsupported in this pass.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    official

    Microsoft Research Blog: Flint — A Visualization Language for the AI Era

    Microsoft 7/8 surfaces Flint — an 'AI-era visualization language' built with Renmin's IDEAS Lab, with an MCP server

    Microsoft's July 8, 2026 blog surfaces Microsoft Flint — a visualization language project from Microsoft Research + IDEAS Lab @ Renmin University that pairs a chart-spec language with an MCP server so AI agents can produce Vega-Lite, ECharts, and Chart.js from short human-editable specs; the open-source repository microsoft/flint-chart first tagged v0.1.1 on 2026-06-28. The research paper is described as 'coming soon.'

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    repo

    Hugging Face: deepseek-ai/DeepSeek-V4-Pro-DSpark

    HF: DeepSeek-V4-Pro is live — 1.6T total / 49B activated MoE + Hybrid Attention + Muon optimizer

    Hugging Face's deepseek-ai/DeepSeek-V4-Pro-DSpark model card (uploaded 2026-06-27, last modified 2026-07-08): DeepSeek-V4-Pro is a 1.6T total / 49B activated MoE with 1M context, MIT license, using a Hybrid Attention design (CSA+HCA) plus the Muon optimizer and manifold-constrained hyper-connections; DSpark adds a confidence-scheduled speculative decoding module with semi-autoregressive generation.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    repo

    Hugging Face: zai-org/GLM-5.2

    HF: zai-org/GLM-5.2 — 753B MoE + Glm MoE DSA sparse attention + IndexShare, claims 2.9x FLOPs reduction at 1M context

    Hugging Face's zai-org/GLM-5.2 model card (uploaded 2026-06-16, last modified 2026-07-02): a 753B-parameter MoE (BF16 + F32, sharded safetensors) using Glm MoE DSA — a sparse attention mechanism that uses 'IndexShare' to share one indexer across every four sparse attention layers, claiming 2.9x per-token FLOPs reduction at 1M context; MIT license, no regional restrictions; self-reported scores include GPQA-Diamond 91.2, SWE-bench Pro 62.1, DeepSWE 46.2, HLE default 40.5 / HLE with tools 54.7, AIME 2026 99.2, Terminal Bench 2.1 (Best Harness) 82.7.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    repo

    Hugging Face: tencent/Hy3

    HF: tencent/Hy3 is live — Tencent Hunyuan 3rd-gen MoE, 295B total / 21B activated, 256K context + MTP layer

    Hugging Face's tencent/Hy3 model card (uploaded 2026-07-02, last modified 2026-07-06): Tencent Hunyuan third-generation MoE, 295B total / 21B activated, Apache-2.0, 256K context, 80 layers, with a multi-token prediction (MTP) layer; self-reported scores include GPQA Diamond 90.4, SWE-bench Verified 78, SWE-bench Pro 57.9.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    repo

    Hugging Face: meituan-longcat/LongCat-2.0

    HF: meituan-longcat/LongCat-2.0 is live — Meituan 2nd gen, 1.6T total / ~48B activated + LongCat Sparse Attention

    Hugging Face's meituan-longcat/LongCat-2.0 model card (uploaded 2026-07-05, last modified 2026-07-08): Meituan LongCat second generation is a 1.6T total / ~48B activated MoE, MIT license, introducing LongCat Sparse Attention (Streaming / Cross-Layer / Hierarchical Indexing) and a 135B N-gram embedding, trained on 35T+ tokens.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    repo

    Hugging Face trending — Chinese MoE wave cluster thread

    First week of July: four Chinese open-weight MoEs land on HF in one week — DeepSeek-V4-Pro + GLM-5.2 + Tencent Hy3 + Meituan LongCat-2.0

    In the first week of July, four Chinese open-weight MoE models landed on Hugging Face within one week: DeepSeek-V4-Pro (1.6T total / 49B activated / Hybrid Attention / Muon); GLM-5.2 (753B MoE / Glm MoE DSA / IndexShare 2.9x FLOPs at 1M context); Tencent Hy3 (295B / 21B / 256K + MTP); Meituan LongCat-2.0 (1.6T / ~48B / LongCat Sparse Attention + 135B N-gram embedding). Shared direction: sparse attention at long context.

    1 daily topicOriginally published Jul 8, 2026View source

  • Jul 9, 2026

    community

    Hugging Face Blog: LeRobot v0.6.0 — Imagine, Evaluate, Improve

    HF 7/7: LeRobot v0.6.0 — world-model policies + 5 new VLAs + 2 reward models + 6 new simulation benchmarks

    Hugging Face's July 7, 2026 blog announces LeRobot v0.6.0 'Imagine, Evaluate, Improve': new world-model policies (VLA-JEPA on Qwen3-VL-2B with JEPA world-model supervision that disappears at inference; LingBot-VA autoregressive video-action model on a single 24-32 GB GPU; FastWAM pairs a ~5B video-generation expert with a compact action expert, fine-tunes from lerobot/fastwam_base); 5 new VLAs (GR00T N1.7 replacing N1.5, MolmoAct2, EO-1, Multitask DiT ~450M, EVO1 0.77B); 2 reward models (Robometer on Qwen3-VL-4B with >1M trajectories, RSS 2026 paper arxiv:2603.02115; TOPReward zero-shot Qwen3-VL log-prob wrapper); 6 new simulation benchmarks (LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench) runnable via `lerobot-eval`; `lerobot-rollout` deployment CLI with DAgger strategy; HF Jobs cloud training `--job.target=a10g-small`; FSDP training; ~2x faster data loading; PyTorch 2.7-2.11 support; ~40% fewer base dependencies.

    1 daily topicOriginally published Jul 7, 2026View source

  • Jul 9, 2026

    official

    Anthropic: Fable 5 cyber safeguards + CJS scale thread

    Anthropic 7/2 Fable 5 cyber safeguards + CJS scale: five bands CJS-0 through CJS-4, four-axis scoring + three Glasswing partners

    Anthropic's July 2, 2026 post introduces a draft Cyber Jailbreak Severity (CJS) scale co-developed with 'Glasswing partners' (Amazon, Microsoft, Google): five bands CJS-0 Informational (score 0), CJS-1 Low (1-3.5), CJS-2 Medium (4-6.5), CJS-3 High (7-8.5), CJS-4 Critical (9-10), scored on four axes (capability gain, breadth of capability gain, ease of weaponization, discoverability); the final level may be raised but not lowered relative to the initial calculation.

    1 daily topicOriginally published Jul 2, 2026View source

  • Jul 9, 2026

    official

    Anthropic: Fable 5 cyber safeguards (HackerOne bounty thread)

    Anthropic 7/2 Fable 5: HackerOne bounty live + four-bucket cyber-activity classifier, feedback at cyber-safeguards@anthropic.com

    Anthropic's July 2, 2026 post on Fable 5 cyber safeguards launches a HackerOne bug-bounty program at hackerone.com/anthropic-cyber-jailbreak accepting Fable 5 cyber jailbreak submissions; feedback address cyber-safeguards@anthropic.com. Classifiers sort cyber activity into four buckets (Prohibited, High-risk dual use, Low-risk dual use, Benign) with a larger safety margin than prior models.

    1 daily topicOriginally published Jul 2, 2026View source

Jul 8, 2026

10 items

  • Jul 8, 2026

    community

    Hugging Face Blog: From Hugging Face to Amazon SageMaker Studio in one click

    HF 7/7: Hugging Face models enter Amazon SageMaker Studio in one click, with inline GPU quota

    Hugging Face's July 7 blog announces a deep-link integration with Amazon SageMaker AI — model pages expose 'Customize on SageMaker AI' and 'Deploy on SageMaker AI' buttons that open the corresponding Studio workflow with model context preserved; new Studio environments are auto-provisioned with the AmazonSageMakerModelCustomizationCoreAccess managed policy (SFT/DPO/RLVR/RLAIF fine-tuning + SageMaker AI/Bedrock endpoint deployment); instance selection UI surfaces G5/G6 GPU quota inline and links to Service Quotas.

    2 daily topicsOriginally published Jul 7, 2026View source

  • Jul 8, 2026

    community

    Hugging Face Blog: Hugging Face Models on Foundry Managed Compute

    HF 7/7: Microsoft Build 2026 puts HF open-weight models in Foundry Model Catalog, refreshed weekly, on A100/H100/MI300X preview

    Hugging Face's July 7 blog records Microsoft Build 2026's Foundry Managed Compute announcement: a curated catalog of HF open-weight models in the Foundry Model Catalog, refreshed weekly, deployable in one click, spanning text/vision/audio/multimodal (LLMs, VLMs, ASR, embeddings, segmentation, image generation), screened for license compliance and security, shipping only in SafeTensors. Microsoft pre-stages weights in Azure storage, builds and CVE-scans runtime images (vLLM, SGLang, TensorRT-LLM, NIM, TEI, llama.cpp, hf-serve), validates each model + runtime + accelerator combination before publication. Preview on NVIDIA A100, H100, and AMD MI300X in Global and Data Zone scopes.

    2 daily topicsOriginally published Jul 7, 2026View source

  • Jul 8, 2026

    community

    Hugging Face Blog: zero-egress storage with SkyPilot

    HF 7/7: Hub Storage becomes a first-class SkyPilot backend, zero-egress across 20+ clouds

    Hugging Face's July 7 blog: Hugging Face Storage becomes a first-class SkyPilot backend via a new `store: hf` option and `hf://` URLs — Buckets (read-write) or any model/dataset/Space repo (read-only) mount into a SkyPilot task using the `hf-mount` FUSE driver with MOUNT and COPY modes and lazy reads; data can be read onto GPUs across 20+ clouds, Kubernetes, Slurm, and on-prem without per-cloud copies or transfer taxes (HF charges no egress or CDN fees). A Qwen3.5-4B SFT benchmark loaded free in about 30 seconds at up to 500 MB/s, with checkpoints streaming to Buckets at 112-168 MB/s across AWS, GCP, and Lambda. Buckets are Xet-backed with content-defined chunking for incremental checkpoint deduplication.

    2 daily topicsOriginally published Jul 7, 2026View source

  • Jul 8, 2026

    official

    Anthropic Features: The Making of Claude Code

    Anthropic's 7/6 features post: Claude Code started as an internal CLI at Anthropic

    Anthropic's July 6 features post 'The Making of Claude Code' reveals Claude Code's internal origin — engineers first used it in the terminal; the team deliberately chose terminal-native ergonomics (minimal surface area, transparent tool use, tight integration with existing developer environments) over a separate IDE; only later was it released more broadly.

    2 daily topicsOriginally published Jul 6, 2026View source

  • Jul 8, 2026

    official

    Anthropic Features: The Making of Claude Code (positioning thread)

    Anthropic's 7/6 features post: Claude Code is positioned as an agent that works alongside engineers, not one that displaces them

    Anthropic's July 6 features post positions Claude Code as a product that respects the developer's existing workflow (users can interrupt, redirect, or partially accept changes), supports both cloud and local configurations; Anthropic publicly states the next leap in developer tooling is agents that collaborate with engineers, not tools that try to replace them.

    2 daily topicsOriginally published Jul 6, 2026View source

  • Jul 8, 2026

    official

    Anthropic Features: The Making of Claude Code (origin thread)

    Anthropic features post: the 'internal CLI to terminal-first' origin thread, isolated as its own signal

    Isolated signal for the internal-CLI origin thread of Claude Code: engineers first used it in the terminal, deliberately chose terminal-native ergonomics, explicitly avoided an IDE-shaped product; early users reported 'less like autocomplete, more like a collaborator holding context across long sessions.'

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 8, 2026

    repo

    huggingface/transformers v5.13.0 release

    HF transformers 7/3 v5.13.0: 8 new models + HfExporters unified export subsystem

    huggingface/transformers v5.13.0 released July 3, adding 8 new models: Kimi K2.5/2.6/2.7 (Moonshot's multimodal agent for long-horizon coding, front-end design, and swarm-based task orchestration across Rust, Go, and Python); Xiaomi MiMo-V2-Flash (27T-token MoE, native 32k context, extendable 256K window); Zyphra ZAYA1 (760M active / 8.4B total MoE, Compressed Convolutional Attention); OpenBMB MiniCPM3-4B (MLA from DeepSeek-V2 + SwiGLU); vision: Google DeepMind VideoPrism (general-purpose video encoder pretrained on 36M video-caption pairs and 582M noisy clips) + NVIDIA RADIO (distills CLIP, DINOv2, SAM into a single variable-resolution ViT); ASR: NVIDIA Nemotron 3.5 ASR + Nemotron ASR Streaming (80/160/560/1120 ms chunk configurable) + Qwen3 ASR (Whisper-style audio encoder + Qwen3 decoder + forced-aligner head); plus the HfExporters subsystem (DynamoExporter / OnnxExporter / ExecutorchExporter subclasses + prefill/decode auto-split).

    2 daily topicsOriginally published Jul 3, 2026View source

  • Jul 8, 2026

    repo

    huggingface/transformers v5.13.0 (ASR cluster thread)

    HF transformers v5.13.0 ASR cluster: Nemotron 3.5 / Nemotron Streaming / Qwen3 ASR

    The ASR subset of transformers v5.13.0: NVIDIA Nemotron 3.5 ASR (multilingual, FastConformer-RNNT backend); Nemotron ASR Streaming (English streaming, cache-aware FastConformer-RNNT, 80/160/560/1120 ms chunk size for tunable latency); Qwen3 ASR (Whisper-style audio encoder + Qwen3 decoder with a forced-aligner head for timestamp prediction).

    1 daily topicOriginally published Jul 3, 2026View source

  • Jul 8, 2026

    repo

    huggingface/transformers v5.13.0 (HfExporters subsystem)

    HF transformers v5.13.0 HfExporters: unified ONNX/ExecuTorch export with auto prefill/decode split

    transformers v5.13.0 introduces the HfExporters subsystem — a unified base class with DynamoExporter, OnnxExporter, and ExecutorchExporter subclasses giving a single API for PyTorch/ONNX/ExecuTorch export with automatic prefill/decode splitting for generative models; standardizes layer declarations, mask/cache construction, and hybrid-attention handling so models cleanly export to ONNX, torch.export, and ExecuTorch.

    2 daily topicsOriginally published Jul 3, 2026View source

Jul 7, 2026

10 items

  • Jul 7, 2026

    repo

    Hugging Face Hub: kernels repository directory

    HF Hub launches the kernels directory: 127 repositories, kernels-community leads by downloads

    Alongside the July 6 Kernels revamp, huggingface.co/kernels goes live as the [kernel] repository directory — 127 repos across 5 pages. Top-downloaded: flash-attn3 (6.79k), flash-attn2 (6.63k). Most-recently-updated: vllm-moe (~18 hours ago), flashrt-fp8-swiglu-ffn (~15 hours), aiter-kernels (4 days). Publishers include kernels-community, kernels-staging, flashrt, MiniMaxAI, and many individual contributors.

    1 daily topicView source

  • Jul 7, 2026

    community

    Hugging Face Blog: LeRobot v0.6.0 — Imagine, Evaluate, Improve

    LeRobot v0.6.0: imagine, evaluate, improve — world-model policies, Robometer, and 6 new simulation benchmarks

    Hugging Face's July 7 LeRobot v0.6.0 release closes the robot learning loop: three new world-model policies VLA-JEPA (arXiv 2602.10098), LingBot-VA (arXiv 2601.21998), and FastWAM; new vision-language-action models GR00T N1.7, MolmoAct2, EO-1 (arXiv 2508.21112), EVO1 (arXiv 2511.04555), and Multitask DiT; a unified reward-models API debuting Robometer (lerobot/Robometer-4B, Qwen3-VL-4B based) and zero-shot TOPReward producing per-frame progress curves for reward-aware behavior cloning. Datasets gain custom video encoding (NVENC, VideoToolbox, VAAPI, QSV), end-to-end depth capture via Intel RealSense, a VLM-powered lerobot-annotate CLI, and up to 2x faster loading. Six new simulation benchmarks — LIBERO-plus, RoboTwin 2.0, RoboCasa365, RoboCerebra, RoboMME, VLABench — unify under lerobot-eval alongside LIBERO, Meta-World, and IsaacLab-Arena. Deployment streamlined via lerobot-rollout with DAgger-style human-in-the-loop corrections; training scales through FSDP + one-flag cloud training on HF Jobs. ~40% fewer base dependencies, PyTorch 2.7-2.11 + CUDA 12.8, Foxglove streaming, pip-installable plugin packages.

    1 daily topicView source

  • Jul 7, 2026

    official

    Anthropic News: Alberta Government + Claude cybersecurity case study

    Alberta government used Claude Code to scan 466M lines of code in 20 hours — equivalent to 6.5 years of manual work

    Anthropic's July 6 case study: Canada's Alberta Ministry of Technology and Innovation used Claude Code (Opus + Sonnet) running ~50 parallel agents to scan 466 million lines of code across ~1,280 applications and 3,400 repositories in 20 hours — a task the team estimates would have taken 6.5 years by traditional methods. The pipeline is a two-stage scan (rules engine then verifier citing exact file/line) plus auto-fix, test writing, and full rewrites (e.g., a 25-year-old Java subsidy portal rebuilt in 4-5 days). Specialized red-team and blue-team agents built on the Claude Agent SDK continuously check ~95 security controls per application.

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 7, 2026

    official

    Anthropic Research: A global workspace in language models

    Anthropic interpretability research: Claude has an internal 'J-space' workspace that causally drives multi-step reasoning

    Anthropic's July 6 interpretability post 'A global workspace in language models' uses a Jacobian lens (J-lens) to read an emergent internal 'J-space' in Claude — a small set of neural patterns where Claude holds concepts that causally drive multi-step reasoning. Swap experiments ('Soccer'->'Rugby', 'spider'->'ant', 'France'->'China') change downstream behavior; J-space accounts for less than a tenth of Claude's internal activity, and ablation eliminates higher-order reasoning while preserving fluency and simple recall. Neel Nanda (DeepMind) independently replicated some findings.

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 7, 2026

    community

    Hugging Face Blog: 🤗 Kernels — Major Updates

    HF revamps 🤗 Kernels: trusted publishers + Sigstore cosign + Torch Stable ABI + first non-Torch framework (Apache TVM FFI)

    Hugging Face's July 6 revamp of custom compute kernel packaging, distribution, and consumption: a new [kernel] repository type live on the Hub; two new security layers — trusted kernel publishers (requiring explicit opt-in via trust_remote_code) and kernel code signing via Sigstore cosign (ephemeral keys + GitHub workflow verification); the kernels and kernel-builder CLIs cleanly separated; new Torch Stable ABI support (targeting ~2 years of Torch versions) and Apache TVM FFI as the first non-Torch framework for cross-framework kernels; an agent-optimized CLI, backend-specific skills, and HF Jobs integration for cross-hardware benchmarking; one-click environment install script; Terraform guide for ephemeral instances; auto system cards per kernel; has_kernel()/get_kernel_variants() compatibility-check APIs; a switch to dynamic libstdc++ linking via the official manylinux_2_28 toolchain to fix corruption/segfault issues from static linking.

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 7, 2026

    community

    Hugging Face Blog: Photoroom PRX Part 4 — Our Data Strategy

    Photoroom PRX Part 4: re-caption every image with Qwen3-VL-8B, stream data via Lance + Mosaic Shards

    Photoroom's July 6 'PRX Part 4 — Our Data Strategy' HF blog post: re-caption every PRX training image with Qwen3-VL-8B; argue long accurate captions beat short ones; validate JPEG quality 92 as visually indistinguishable from PNG for training; show that the captioner choice lowered FID, CMMD, and DINO-MMD on PRX pre-training benchmarks. Data is stored in Lance for exploration and rewritten into Mosaic Data Shards with resolution and aspect-ratio bucketing for streaming; light caption-based filtering and perceptual-hash deduplication are applied via skip lists rather than rewriting shards.

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 7, 2026

    official

    OpenAI API Changelog: GPT-Realtime-2.1 + GPT-Realtime-2.1 mini

    OpenAI ships GPT-Realtime-2.1 + GPT-Realtime-2.1 mini: alphanumeric recognition, noise handling, interruption behavior

    OpenAI's July 6 API changelog — the only entry in the 7/1-7/7 window — ships GPT-Realtime-2.1, an updated realtime reasoning model with improved alphanumeric recognition, silence and noise handling, and interruption behavior, alongside GPT-Realtime-2.1 mini, a faster lower-cost distilled reasoning model for realtime voice. Both run on the v1/realtime endpoint.

    1 daily topicOriginally published Jul 6, 2026View source

  • Jul 7, 2026

    official

    Anthropic News: Fable 5 cyber safeguards + jailbreak framework

    Anthropic details Fable 5 cyber safeguards + proposes a CJS jailbreak scoring framework

    Anthropic's July 2 post details the cyber safeguards baked into Claude Fable 5 and — with Glasswing partners — proposes a Cyber Jailbreak Severity (CJS) framework that scores jailbreaks 0-4 across four axes (capability gain, breadth of capability gain, ease of weaponization, discoverability), alongside a HackerOne submission program.

    1 daily topicOriginally published Jul 2, 2026View source

  • Jul 7, 2026

    paper

    arXiv 2607.02087: SUNTA — Hierarchical Video Prediction with Surprise-based Chunking

    arXiv SUNTA: surprise-driven chunking lets a hierarchical SSM predict video 250 timesteps without breakdown

    arXiv 2607.02087 'SUNTA: Hierarchical Video Prediction with Surprise-based Chunking' (Iiyama / Suzuki / Matsuo, submitted 2026-07-02): chunk boundaries in a hierarchical state-space model (HSSM) should be driven by prediction errors ('surprise') rather than fixed length or similarity. The paper addresses hierarchical collapse via decoupled training and uses top-down internal inconsistency as an open-loop surprise signal to identify chunk boundaries in imagined rollouts. On 2D and 3D video prediction, SUNTA maintains accurate predictions over 250 timesteps while all baselines degrade within 10 timesteps.

    1 daily topicOriginally published Jul 2, 2026View source

Jul 6, 2026

10 items

Jul 3, 2026

10 items

Jul 2, 2026

10 items

Jul 1, 2026

11 items

Jun 30, 2026

15 items

Jun 29, 2026

3 items

Jun 23, 2026

8 items

Jun 22, 2026

3 items

Jun 21, 2026

2 items

Jun 17, 2026

5 items

Jun 16, 2026

3 items