Intel
Page 43
Tencent Hy@TencentHunyuan
An RL post-training framework aimed squarely at interactive world models, released with training code you can point at your own data, rewards or base model, plus a checkpoint. The reported gain is following compositional action instructions and staying coherent over long interactions.
Why it mattersAn open RL post-training stack for interactive world models means you can tune action-following on your own rewards and data rather than accepting a vendor checkpoint's behavior.
Unfortunately, Sprites Now Speak MCP
Sprites gain an MCP interface so agents can spin up their own disposable VMs.
Why it mattersMCP access to disposable agent VMs ships, alongside a concrete critique: every MCP server burns context on tool descriptions, where a CLI an agent can discover and read costs nothing until used.
articlefly.io
Andrew Ng@AndrewYNg
Context Hub (npm i -g @aisuite/chub) is an open CLI that feeds coding agents curated, current API documentation on demand, aimed squarely at stale-endpoint and hallucinated-parameter failures. Agents can annotate docs with workarounds, building a shared record over time.
Why it mattersCoding agents routinely generate against outdated APIs and hallucinate parameters. Context Hub is an open CLI that hands the agent curated current docs and lets it save workarounds it discovers so they persist between sessions.
Where Stand Department War
The supply-chain-risk designation is confirmed in writing and headed to court.
Why it mattersThe supply-chain-risk designation is now formally issued and being litigated; its reach is limited to Claude used directly under Department of War contracts, and access continues at nominal cost during any transition.
articleAnthropic News
Tencent Hy@TencentHunyuan
HY-WU synthesizes instance-conditioned LoRA updates from hybrid image-instruction representations and injects them into a frozen backbone during the forward pass, producing a per-input adapter with no test-time optimization. The premise is that shared LoRAs compromise across objectives while single-domain ones fail to generalize.
Why it mattersA concrete alternative to fine-tuning for personalization: the adapter is generated per request rather than trained per user, so there is nothing to store per user and no optimization step at inference.
FlashAttention-4: Algorithm and Kernel Pipelining Co-Design for Asymmetric Hardware Scaling
As GPU throughput outpaces memory bandwidth, kernels must evolve.
Why it mattersFlashAttention is the backbone of efficient transformer inference and training.
articleTogether AI
Cache-aware prefill–decode disaggregation (CPD) for up to 40% faster long-context LLM serving
Serving long prompts doesn't have to mean slow responses.
Why it mattersIf you're serving long-context LLMs and fighting slow time-to-first-token, CPD shows how separating cache-warm and cache-cold workloads across prefill and decode stages can lift throughput ~40%.
articleTogether AI
Enabling Claude Code To Work More Autonomously
Claude Code gains a native VS Code extension with inline diffs, a refreshed terminal with searchable prompt history, and automatic checkpointing you can rewind with /rewind.
Why it mattersCheckpoints let you rewind Claude's edits (code, conversation, or both) after a bad autonomous run, and the renamed Agent SDK now exposes subagents and hooks for building custom agents.
articleAnthropic News
Evals Skills for Coding Agents
A set of skills for AI product evals, built to guard against the mistakes seen across 50-plus companies and 4,000-plus students, now that coding agents instrument applications and run the experiments themselves.
Why it mattersIf your coding agent is instrumenting and evaluating your AI product, these skills encode the error-analysis discipline that keeps it from lumping distinct failure modes into one useless score.
articleHamel Husain
Lenny Rachitsky
The design process is dead. Here’s what’s replacing it. | Jenny Wen (head of design at Claude)
podcast
Listen
Statement Comments Secretary War
A US government supply-chain-risk designation lands on a frontier lab after negotiations stalled over two refused use cases.
Why it mattersIf you hold Department of War contract work, Claude use on that contract is in scope; commercial API, claude.ai and non-DoW work are explicitly not.
articleAnthropic News
Accelerating Robot Policy Evaluation
Runway benchmarks its robotics world model as an evaluation substitute.
Why it mattersA world model that predicts real robot policy outcomes at 0.95 correlation from starting images alone removes scanning and scene reconstruction from the evaluation loop, and replaces hardware runs during iteration.
articleRunway editorial sitemap
Simgym
SimGym runs up to 2,000 concurrent cloud browsers, each driven by an LLM with a persona, budget and shopping intent, against two storefront variants.
Why it mattersShows LLM browser agents used as a measurement instrument, not a product: thousands of concurrent cloud-browser sessions producing a simulated A/B result in minutes for stores with too little traffic to reach significance.
articleShopify Engineering
Speed Structure And Smarts The Notion Ai Way
Notion describes routing each task class to a different model on quality/latency/cost, fine-tuning cheap models for high-volume database autofill.
Why it mattersRouting by task category and running LLM-as-a-judge evals continuously rather than as a one-time gate lets you swap in new models without regressions.
articleNotion Blog
Using LLMs to amplify human labeling and improve Dash search relevance
Dash retrieval is only as good as the relevance labels its ranker trains on, and human judgments do not scale to billions of enterprise documents.
Why it mattersIn RAG, the ranker decides which handful of documents the model ever sees. This is a worked approach to scaling relevance labels with LLMs when human judgments are too scarce to train on.
articleIlya Yakovlev,Andrew Cheung,Binoy Dash,Simran Jumani,Dmitriy Meyerzon,Mark Breitenbach,Ishan Mishra,Kazuaki Okumura,Mike White,Kevin Altschuler,Facundo Agriel,Ishan Mishra,Eric Wang,Dmitriy Meyerzon,Dmitriy Meyerzon
Responsible Scaling Policy V3
Version 3 of the conditional-commitment framework governing frontier model release, with an unusually candid retrospective.
Why it mattersThe RSP sets the capability thresholds and safeguards that determine when a frontier model ships and under what deployment restrictions; v3 revises those thresholds and adds transparency and accountability measures.
articleAnthropic News
A Dream of Spring for Open-Weight LLMs: 10 Architectures from Jan-Feb 2026
A technical roundup comparing ten open-weight LLM architectures released in early 2026, including Arcee Trinity, Kimi K2.5, Step 3.5 Flash, Qwen3-Coder-Next, and GLM-5.
Why it mattersA technical walkthrough of ten open-weight LLM architectures — what actually changed in MoE routing, attention and multi-token prediction.
articleSebastian Raschka, PhD
Generative Recommendations
Treating a shopper's history as a raw event sequence instead of a feature summary.
Why it mattersSequence models replacing hand-summarized features is the live shift in recommendation systems; this covers the training choices that actually moved metrics and what it costs to serve one inside production latency budgets.
articleShopify Engineering
Can you reverse engineer our neural network?
A hand-built network with integer weights that outputs zero for nearly every input, and the step-by-step reverse engineering that cracked it.
Why it mattersShows the practical moves of mechanistic interpretability — reading weight matrices directly, working backward from the final layers — on a network deliberately built so that backpropagating from the output tells you nothing.
blogblog.janestreet.com
Detecting And Preventing Distillation Attacks
An attribution write-up of industrial-scale capability extraction.
Why it mattersLarge-scale distillation via fraudulent API accounts is detectable through traffic structure and infrastructure fingerprints, and models trained this way inherit capabilities without the safety training that constrained them.
articleAnthropic News
How speech models fail where it matters the most and what to do about it
State-of-the-art speech models like Whisper and Deepgram score near-human on benchmarks — then fail 39% of the time on street names.
Why it mattersIf you're building on Whisper or Deepgram, near-human aggregate benchmark scores can hide catastrophic failures on critical entities like street names and proper nouns.
articleTogether AI
The 2028 Global Intelligence Crisis
A speculative macro-economic thought experiment framed as a 2028 research memo, arguing that continued AI capability gains could trigger a 'human intelligence displacement spiral'.
Why it mattersIt offers agentic engineers and AI builders a concrete, mechanistic framework for thinking about AI's downstream economic tail risks.
articlewww.citriniresearch.com
Claude Code Security
A research-preview security scanner inside Claude Code on the web.
Why it mattersClaude Code Security reasons about data flow and access control instead of matching known patterns, and every finding is self-verified and severity-rated before an analyst sees it, with patches gated behind human approval.
articleAnthropic News
articleLinkedIn Engineering
Scaling Llm Based Ranking Systems With Sglang At Linkedin
Ranking with an LLM means one forward pass over a shared prefix and the final token's logits — no sampling, no beam search.
Why it mattersPrefill-only scoring is a different serving workload from chat generation; this names which parts of an open-source LLM server — batching, prefix reuse, GPU execution, Python runtime behavior — have to change to meet strict latency SLAs.
Lenny Rachitsky
Head of Claude Code: What happens after coding is solved | Boris Cherny
podcast
Listen
Minions: Stripe’s one-shot, end-to-end coding agents—Part 2
Part two on Minions, Stripe's homegrown coding agents behind more than a thousand merged pull requests a week: humans review the code, but minions write it end to end.
Why it mattersThe follow-up carries the implementation detail behind a coding agent fleet running at production scale.
articleAlistair Gray
Consistency diffusion language models: Up to 14x faster inference without sacrificing quality
Standard diffusion language models can't use KV caching and need too many refinement steps to be practical.
Why it mattersIf you're deploying diffusion language models, CDLM removes their two biggest practical blockers — no KV caching and excessive refinement steps.
articleTogether AI
A Guide to Which AI to Use in the Agentic Era
An explanatory guide that breaks down how to choose AI tools in 2025 by understanding three layers.
Why it mattersIt gives you a reusable three-layer framework — models, apps, and harnesses — for reasoning about which AI tool fits a given task, cutting through the confusion of overlapping ChatGPT/Claude/Gemini coding agents and desktop products.
articleEthan Mollick
Aidan Gomez@aidangomez
Cohere Labs released Tiny Aya, a family of small multilingual models at 3.35B parameters targeting 70+ languages and on-device execution. The pitch is coverage of languages that frontier models underserve, at a size that fits local inference rather than a hosted endpoint.
Why it mattersA 3.35B model claiming usable quality across 70+ languages moves multilingual inference onto phones and edge hardware, where per-language hosted API calls were previously the only option.
Tencent Hy@TencentHunyuan
GradLoc uses distributed binary search to pin a gradient spike to the exact culprit token in O(log N) with negligible overhead. It surfaced layerwise gradient heterogeneity, specifically tokens with safe importance-sampling ratios that still blow up individual layers, a problem addressed by LayerClip, a per-layer clipping method.
Why it mattersTurns RLVR training instability from guesswork about dirty data into a locatable defect, and shows global gradient clipping can mask per-layer explosions that only local clipping controls.
Dwarkesh Patel
Dario Amodei — "We are near the end of the exponential"
podcast
Listen
Andrew Ng@AndrewYNg
Google Cloud and IBM Research teach the Agent2Agent protocol: wrapping ADK and LangGraph agents as A2A servers so they discover and call each other, chaining them sequentially and hierarchically, pulling external data through MCP, and deploying on IBM's open-source Agent Stack.
Why it mattersIf you run agents across different frameworks, A2A is the emerging interop standard, and ACP folding into it removes a competing protocol from the decision.
Lenny Rachitsky
“Engineers are becoming sorcerers” | The future of software development with OpenAI’s Sherwin Wu
podcast
Listen
lexfridman.com
OpenClaw
podcast
Listen
Tencent Hy@TencentHunyuan
Quantization-aware training takes a 1.8B model to a 2-bit representation, roughly a 0.3B bit-equivalent footprint at 600MB, while keeping dual chain-of-thought reasoning. Reported prefill speeds are 3-8x faster on Apple M4 and Dimensity 9500, with 2-3x faster generation. The model ships as GGUF.
Why it mattersQAT to 2 bits with chain-of-thought reasoning intact means a 600MB model that fits inside a mobile app bundle, and the GGUF release drops straight into existing edge inference engines.
Making Diffusion LLMs Faster: A Survey of Speedup Techniques
A video walkthrough of methods researchers use to speed up diffusion-based language models.
Why it mattersIf you're tracking diffusion LLMs as a lower-latency alternative to autoregressive decoding, this maps the whole speedup landscape — caching tricks, unmasking schedules, block diffusion.
videoJulia Turc
Minions: Stripe’s one-shot, end-to-end coding agents
Stripe introduces Minions, its homegrown coding agents behind more than a thousand merged pull requests a week: humans review the code, but minions write it from start to finish.
Why it mattersA thousand agent-authored PRs a week is a working existence proof, and the post explains the machinery behind it.
articleAlistair Gray
Lenny Rachitsky
The rise of the professional vibe coder (a new AI-era job) | Lazar Jovanovic (Professional Vibe Coder)
podcast
Listen
Tencent Hy@TencentHunyuan
A training-ready 3D corpus: 252k+ filtered high-fidelity objects, 240k+ part-level decompositions for fine-grained control, and 125k synthesized assets for class balance, shipped with Hunyuan3D-2.1-Small as a lightweight reproducible baseline.
Why it mattersPart-level decomposed data at this scale is the missing ingredient for training controllable 3D generation, and the small baseline makes results reproducible without a large cluster.
What do LLMs think when you don't tell them what to think about?
What do language models generate when you don't tell them what to generate?
Why it mattersUnderstanding that GPT defaults to code/math, Llama to narratives, DeepSeek to religious content, and Qwen to exam questions helps engineers anticipate model biases when prompts are underspecified.
articleTogether AI
Building a C compiler with a team of parallel Claudes
Anthropic researcher Nicholas Carlini describes 'agent teams,' a harness that runs multiple Claude Code instances in parallel, unsupervised, coordinating via git-based task locks to build a Rust-based C compiler from scratch.
Why it mattersProvides concrete, battle-tested harness design patterns — git-based task locks, decentralized coordination without an orchestrator agent — for anyone building or scaling multi-agent coding systems beyond toy demos.
articleAnthropic Engineering
I design with Claude more than Figma now
A designer at Jane Street describes abandoning Figma for most work.
Why it mattersPrototyping directly in the production codebase collapses the mockup-to-implementation handoff: the artifact you iterate on is the one that ships, and unlimited cheap iteration makes small UX refinements worth doing at all.
blogblog.janestreet.com
Voxtral transcribes at the speed of sound.
Two speech-to-text models split by job.
Why it mattersVoxtral Realtime ships Apache 2.0 open weights with a genuine streaming architecture and latency configurable below 200ms, while Mini Transcribe V2 adds diarization, word-level timestamps and context biasing in 13 languages.
articlemistral.ai
articleAnthropic News
Protecting Well Being Of Users
Why it mattersDocuments a working pattern for sensitive-conversation handling: a lightweight classifier over live sessions driving a product-layer intervention, backed by published system prompts, rather than relying on model training alone.
Apple Xcode Claude Agent Sdk
Xcode 26.3 ships a native Claude Agent SDK integration.
Why it mattersApple developers get full Claude Code behavior inside Xcode, including agents that iterate against captured SwiftUI Previews, and Xcode now speaks MCP so a CLI agent can pull those Previews itself.
articleAnthropic News
Vision Transformer, Diffusion Transformer, and Multimodal Diffusion Transformer Explained
An educational video tracing the architectural evolution from the Vision Transformer (ViT) to the Diffusion Transformer (DiT) to the Multimodal Diffusion Transformer (MMDiT), explaining how Transformers displaced CNNs as the standard architecture for vision tasks.
Why it mattersTraces how the Transformer displaced CNNs in vision and became the backbone of modern image generation, explaining the conditioning mechanisms that most practitioners use without understanding.
videoJulia Turc
Tencent Hy@TencentHunyuan
CL-bench measures whether models can apply knowledge that exists only in their context: 1,899 expert-curated tasks over 500 complex contexts, gated by 31,000+ validation rules. Ten frontier models complete 17.2% on average, which the authors call a bottleneck for real context-dependent work.
Why it mattersThis quantifies the gap that appears when an agent must work inside a private repo or fresh document set rather than recall pretraining. A 17.2% average is a hard ceiling to design around.
Lex Fridman
#490 – State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI
podcast
Listen
Aidan Gomez@aidangomez
Cohere's Model Vault gives its models a dedicated, fully managed home: isolated VPC, no noisy neighbours, no shared rate limits. It targets the gap between calling a public API and standing up your own inference stack for regulated or throughput-sensitive workloads.
Why it mattersTeams blocked from hosted LLM APIs by tenancy or rate-limit constraints get a self-hosting-like isolation boundary for Cohere models without running the infrastructure themselves.
Tencent Hy@TencentHunyuan
Tencent released HunyuanImage 3.0-Instruct under open weights, publishing full and distilled checkpoints on Hugging Face plus a GitHub repo. The lab claims top-tier standing on Arena's Image Edit leaderboard for open-source image-to-image editing.
Why it mattersAn open-weight image-to-image editing model with published weights and a distilled variant gives builders a self-hostable alternative to closed editing APIs, with a leaderboard result to size it against.
An index of the vibe-coding frontier. Corrections welcome.