cuppa

brewed 03:12 AM
← Jul 06

Tuesday

jul07

2026

Jul 08 →

the brief

Concrete launches anchored the day: Tencent’s Apache‑2 Hy3 MoE landed, Cloudflare put a tiered cache in front of Workers, and Hugging Face pushed LeRobot 0.6. On the research side, Anthropic’s J‑space interpretability work and fresh TMLR papers stood out, while Next.js, Claude Code, OpenSSH, and a 7MB in‑browser embed kept builders busy.

the poursit · sip · 12 items

pulse

(07)
  • simonw/blog· feedJul 6, 11:57 PM

    Tencent releases Apache‑2 Hy3 MoE

    New 295B‑parameter Mixture‑of‑Experts with 21B active parameters and a multi‑token prediction layer, released under Apache 2.0, signaling a stronger open model push from China.

    tencent/Hy3 — <p><strong><a href="https://huggingface.co/tencent/Hy3">tencent/Hy3</a></strong></p> New Apache 2.0 licensed model from Tencent in China:</p> <blockquote> <p>Hy3 is a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters and 3.8B MTP layer parameters, developed by the Tencent Hy Team. Following the Hy3 Preview launch in late April, we gathered feedback from 50+ products and scaled up post-training with higher quality data. Today, we introduce Hy3, which outper...

    signal 9hype 1model_releaseopen_source_licensemoesource ↗
  • cloudflare/blog· feedJul 6, 01:00 PM

    Cloudflare launches regionally tiered Workers Cache

    Regionally tiered cache sits directly before Worker entrypoints, configured via standard HTTP headers, cutting latency and origin load while staying fully composable inside Workers.

    Your Worker can now have its own cache in front of it — We are launching Workers Cache, a regionally tiered cache that sits directly in front of your Worker entrypoints. Infinitely composable, configured via standard HTTP headers

    signal 8hype 3cloudflareworkersedge_computesource ↗
  • huggingface/blog· feedJul 7, 12:00 AM

    LeRobot 0.6 adds imagine–evaluate–improve

    The HF robotics stack introduces new data generation, offline evaluation, and iterative improvement workflows, lowering friction to build, benchmark, and refine real robotic policies.

    LeRobot v0.6.0: Imagine, Evaluate, Improve

    signal 4hype 1library_releaseroboticshuggingfacesource ↗
  • anthropics/claude-code· feedJul 6, 10:51 PM

    Claude Code adds workflow sizing controls

    Release adds a Dynamic workflow size setting and OTel attributes for workflow runs, plus stability fixes, improving observability and control for multi‑agent coding workflows.

    v2.1.202 — What's changed Added a "Dynamic workflow size" setting in /config for controlling how large Claude generally makes dynamic workflows (small/medium/large agent counts) — an advisory guideline, not an enforced cap Added workflow.run_id and workflow.name OpenTelemetry attributes to telemetry emitted by workflow-spawned agents, so a workflow run's activity can be reconstructed from OTel data Fixed a crash in the inline Ctrl+R history search when accepting or cancelling while the search...

    signal 9hype 1release_notesclaude_codeworkflowsource ↗
  • vercel/next.js· feedJul 7, 12:01 AM

    Next.js canary refines RSC HMR and scroll

    Latest canary enables the new fragment scroll handler by default, aborts superseded RSC HMR requests, and tunes Turbopack and persistence ahead of 16.3.

    v16.3.0-canary.79 — Misc Changes docs: remove incorrect statement that force-cache is the default for …: #95235 Turbopack: add all keys to dynamic exports before sealing the object: #93334 Disable supportsImmutableAssets with config.output: #95521 turbo-persistence: skip directory fsync on Windows: #95497 Abort superseded Server Components HMR requests on the client: #95463 [ci] Add stats comment with partial stats: #95401 [fragment-scroll] Enable new scroll handler by default: #95378 Upgrade...

    signal 8hype 1release_notesnextjsturbopacksource ↗
  • hn/frontpage· feedJul 6, 11:06 PM

    7MB embedding model runs in browser

    Ternlight ships a tiny WASM inference stack delivering on‑device embeddings without servers, enabling fully client‑side semantic search and privacy‑preserving personalization.

    Ternlight – 7 MB embedding model that runs in browser (WASM) — Article URL: https://ternlight-demo.vercel.app/ Comments URL: https://news.ycombinator.com/item?id=48811644 Points: 96 # Comments: 30

findings

(04)
  • huggingface/blog· feedJul 6, 03:30 PM

    PRX details scalable data strategy

    Hugging Face and partners outline sourcing, filtering, and quality controls behind PRX, offering concrete, reproducible practices for curating high‑signal training data.

    PRX Part 4: Our Data Strategy

    signal 6hype 2data_strategydatasetopen_modelsource ↗
  • techmeme· feedJul 6, 10:15 PM

    Anthropic maps Claude’s hidden J‑space

    Researchers identify a small set of internal neural patterns correlating with latent “thoughts” not in outputs, hinting at new levers for steering and interpretability of LLMs.

    Anthropic researchers detail J-space, a small set of neural patterns in Claude that reveals internal thoughts that don't appear in the model's output (Anthropic) — Anthropic: Anthropic researchers detail J-space, a small set of neural patterns in Claude that reveals internal thoughts that don't appear in the model's output — As you read this sentence, circuits in your brain are adjusting your posture, controlling your breathing, and transforming lines and curves on the screen into recognizabl...

    signal 9hype 2interpretabilityresearchclaudesource ↗
  • tmlr-pub.bsky.socialJul 7, 12:19 AM

    Variance‑gated ensembles quantify epistemic uncertainty

    TMLR paper gates ensemble members by predictive variance to better capture epistemic uncertainty, offering a practical route to more trustworthy model confidence estimates.

    Variance-Gated Ensembles: An Epistemic-Aware Framework for Uncertainty Estimation Martin Gillis, Isaac Xu, Thomas Trappenberg Action editor: Michele Caprio https://openreview.net/forum?id=fNMZjV1gje #ensembles #ensemble #uncertainty

    signal 4hype 1paperuncertainty_estimationensemblessource ↗
  • tmlr-pub.bsky.socialJul 6, 12:20 PM

    Rule‑based RL stabilizes multimodal reasoning

    MM‑Eureka combines rule‑based reinforcement learning with policy distillation to stabilize multimodal reasoning traces, pointing to scalable training recipes beyond pure SFT.

    MM-Eureka: Toward Stable Multimodal Reasoning via Rule-based Reinforcement Learning with Policy D... Fanqing Meng, Lingxiao Du, Zongkai Liu et al. Action editor: Pablo Castro https://openreview.net/forum?id=8y1ch6y24H #reinforcement #multimodal #reward

    signal 5hype 1paperreinforcement_learningmultimodalsource ↗

voices

(01)
  • jackclark/importai· feedJul 6, 12:31 PM

    Import AI on AI R&D automation

    Jack Clark spotlights AI writing competitive GPU megakernels and the feedback loops emerging in automated R&D, a sharp read on where agentic tooling may be headed.

    Import AI 464: Fables writes GPU kernels; AI automation; and analog computation — Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe now Fable writes a decent GPU kernel, hinting at broader AI R&D automation:…The start of an RSI loop…Fable has written “the first genuine (and fastest) megakernel ever […]

    signal 6hype 2newsletteranalysisai_researchsource ↗