Intel
Page 44
Terminally online Mistral Vibe.
Mistral's terminal coding agent gets configurable.
Why it mattersVibe 2.0 adds custom subagents, slash-command skills, and a clarify-before-execute step to Mistral's terminal coding agent. Devstral 2 also moves behind paid API access, which matters if you were relying on the free tier.
articlemistral.ai
Tencent Hy@TencentHunyuan
HPC-Ops is a production LLM inference operator library written in CUDA and CuTe, reporting 30% QPM gains on Tencent's own models and 17% on DeepSeek, with kernel speedups up to 2.22x over FlashInfer/FlashAttention, 1.88x over DeepGEMM and 1.49x over TensorRT-LLM on FusedMoE.
Why it mattersServing throughput is mostly kernel quality; these are the operators already running Tencent's production inference, with measured margins over FlashAttention, DeepGEMM and TensorRT-LLM.
articleLinkedIn Engineering
Contextual Agent Playbooks And Tools How Linkedin Gave Ai Coding Agents Organizational Context
Coding agents hit a ceiling without knowing your services, frameworks and data systems.
Why it mattersThe gap between a capable coding agent and a useful one is organizational context. This describes how internal knowledge, safe tool access, and executable workflows get packaged so agents can act in a two-decade-old microservice estate.
Configuring Stripe using Terraform and AI agents
Stripe on why agents authoring Terraform beats direct API calls for Stripe configuration.
Why it mattersRouting agent actions through declarative config turns an irreversible API call into a reviewable diff.
articleMichael Selander
Tencent Hy@TencentHunyuan
An 80B MoE (13B active) that reasons over an input image before editing it, using a native chain-of-thought schema trained with Tencent's MixGRPO. Targets add/remove/modify edits that leave untouched regions intact, plus multi-image fusion. The web demo is PC-only.
Why it mattersInstruction-following edits that leave non-target regions intact are where most generative editors still fail; this one puts an explicit reasoning pass in front of synthesis.
articleLinkedIn Engineering
Reimagining Linkedins Search Stack
Query understanding produces an embedding, exhaustive CUDA vector search assembles candidates, and LLM ranking refines them.
Why it mattersShows what it takes to serve LLM semantic retrieval and ranking under real-time QPS budgets — including exhaustive GPU vector search instead of ANN, and where model quality was traded for latency.
Skypilot
Shopify runs training jobs through SkyPilot as a launcher over its own persistent Kubernetes clusters on several clouds.
Why it mattersA concrete pattern for multi-cloud GPU training: use SkyPilot as a scheduler over clusters you already run rather than as a provisioner, keeping data resident per cloud while a plugin layer handles fair scheduling and cost attribution.
articleShopify Engineering
Categories of Inference-Time Scaling for Improved LLM Reasoning
An in-depth educational article that categorizes and explains inference-time scaling techniques for improving LLM reasoning.
Why it mattersIt gives a clear, practical taxonomy of inference-time scaling methods with code and experimental results.
articleSebastian Raschka, PhD
Andrew Ng@AndrewYNg
Ng's argument from Davos: scattered bottom-up AI experiments have failed to pay off. Using a loan pipeline as the worked example, he shows why replacing one manual review with an agent is a minor efficiency gain, and why compressing the whole flow to ten minutes changes the product itself.
Why it mattersSwapping a human step for an agent buys minutes; redesigning the workflow around the new latency changes the product. A concrete frame for deciding where an agentic system actually pays off.
Signals: Toward a Self-Improving Agent
Factory's account of moving past descriptive agent analytics.
Why it mattersSession duration and completion rates tell you an agent failed but not why; this lays out a loop that turns detected failures into automatic fixes, and what instrumentation that requires.
articlefactory.ai
Theturingreel
Runway showed 1,043 people twenty clips each, half real and half generated from the same first frame, none regenerated or edited.
Why it mattersWith detection accuracy at 57.1% and under 10% of viewers beating chance, human judgment is no longer a workable check on video authenticity, leaving provenance metadata as the practical lever.
articleRunway editorial sitemap
Designing AI-resistant technical evaluations
Anthropic's performance engineering team describes iterating through three versions of a take-home coding test after successive Claude models (Opus 4, then Opus 4.5) matched or beat top human candidates under time constraints.
Why it mattersOffers a concrete, tested case study in designing technical evaluations that resist LLM saturation, useful for anyone building hiring processes or benchmarks in a world where frontier models increasingly match human experts under time constraints.
articleAnthropic Engineering
Claude New Constitution
Anthropic replaced its list-of-principles constitution with a reasoned document written primarily for Claude, released CC0.
Why it mattersThe document that governs Claude's behavior is public and CC0, along with an account of how it feeds training and synthetic data, making it the clearest available map of which model behaviors are intended versus accidental.
articleAnthropic News
Heaps do lie: debugging a memory leak in vLLM.
No crash, no error, just system memory climbing 400 MB a minute under production-like traffic until OOM.
Why it mattersA step-by-step hunt for a 400 MB/minute leak in vLLM that only appeared with disaggregated serving and graph compilation on, showing how heap profilers mislead and where dependency layers hide allocations.
articlemistral.ai
Flux 2
FLUX.2 spans managed endpoints to open-weight checkpoints under one release.
Why it mattersTen-image multi-reference conditioning and 4MP editing are available in every variant, plus a flex tier exposing step count and guidance. Those controls determine whether image generation holds up inside a production brand pipeline.
articleBlack Forest Labs editorial sitemap
Flux2 Klein Towards Interactive Visual Intelligence
FLUX.2 [klein] packs text-to-image, editing and multi-reference into one compact model at under half a second.
Why it mattersSub-second unified generation and editing on a consumer GPU, with the 4B variant under Apache 2.0 and undistilled base checkpoints released for LoRA and finetuning, make this a genuinely permissive option at this speed tier.
articleBlack Forest Labs editorial sitemap
24 11 06 Ultra
FLUX1.1 [pro] gains two modes.
Why it mattersFour megapixel generation at roughly ten seconds and $0.06 an image, plus a raw mode that trades the polished default aesthetic for candid-photography realism and more diverse human subjects.
articleBlack Forest Labs editorial sitemap
Flux 1 Krea Dev
Built with Krea AI, FLUX.1 Krea [dev] is the open-weights version of Krea 1, an intentionally opinionated model trained to avoid the oversaturated textures typical of text-to-image output.
Why it mattersAn open-weight text-to-image checkpoint tuned for photorealism over the oversaturated default aesthetic, architecturally compatible with the FLUX.1 [dev] ecosystem so existing finetunes and tooling still apply.
articleBlack Forest Labs editorial sitemap
Lenny Rachitsky
The non-technical PM’s guide to building with Cursor | Zevi Arnovitz (Meta)
podcast
Listen
Why I Stopped Using nbdev
Hamel Husain on why he stopped using nbdev, and the more general point that a tool feels like a secret weapon only while it matches how you happen to think.
Why it mattersA concrete case study of how AI coding tools invert old tooling tradeoffs — worth reading before committing to any unconventional development workflow.
articleHamel Husain
The Data Center Backlash, Answered With Numbers
A contrarian case that carbon, electricity-price, and water objections to data centers are overstated.
Why it mattersData-center siting decides how much compute exists and what it costs.
articleAndrew Ng
Tencent Hy@TencentHunyuan
Tencent's 3D creation pipeline opens to everyone with no application required. PartGen 1.5 raises component partitioning to 1536³ and adds brush-based manual editing, while HY 3D 3.1 accepts up to eight input views and improves texture fidelity.
Why it mattersPart decomposition and eight-view input are what make generated 3D assets editable inside a production pipeline instead of one-shot meshes you cannot revise.
Andrew Ng@AndrewYNg
A course on document extraction that starts from OCR's failure modes — lost reading order, orphaned table headers, uncaptioned charts — then builds agentic pipelines that parse pages visually, map fields to a schema with bounding boxes for validation, and feed event-driven RAG.
Why it mattersExplains why plain OCR breaks on forms, tables and charts, and how page-as-image parsing with schema mapping and bounding-box grounding produces validated, LLM-ready structured output.
The Design & Implementation of Sprites
How Sprites are actually built.
Why it mattersExplains the mechanics behind creating a root Linux VM in a second or two with a durable filesystem that sleeps when idle — the design constraints that matter if you are building or choosing an agent execution environment.
articlefly.io
Lenny Rachitsky
Why most AI products fail: Lessons from 50+ AI deployments at OpenAI, Google, and Amazon
podcast
Listen
Ucp
The design behind the Universal Commerce Protocol.
Why it mattersAgentic commerce needs a capability-negotiation layer, not a fixed API: UCP has merchants declare what they support and agents negotiate down to what they can handle, a pattern worth studying for any agent-to-service protocol.
articleShopify Engineering
Code And Let Live
Sprites are persistent, second-to-boot Linux VMs with durable filesystems and instant checkpoints, pitched as a replacement for the throwaway read-only sandbox that agent harnesses currently default to.
Why it mattersArgues ephemeral agent sandboxes throw away expensive setup, and shows an alternative: root VMs that boot in ~1s, keep a durable filesystem, snapshot to named checkpoints, and sleep when idle.
articlefly.io
Tencent Hy@TencentHunyuan
The interactive world model gets the pieces that make it usable outside a demo: training code you can point at your own data, a lite 5B checkpoint that fits small-VRAM GPUs, inference and memory optimizations for real-time interaction, and an open hosted app with no waitlist.
Why it mattersOpen training code plus a 5B checkpoint sized for small-VRAM GPUs moves interactive world models from a hosted demo to something you can post-train and run yourself.
Let Me Droid That For You: What 780K Agent Searches Reveal
Factory mined 780,000 web searches its Droid agent ran in production to see what coding agents seek outside the codebase, from API docs onward.
Why it mattersReveals what a coding agent actually looks up beyond the repo — mostly API and library documentation — which informs what you should pre-load, cache, or index instead of paying for repeated web lookups.
articlefactory.ai
Andrew Ng@AndrewYNg
Ng proposes the Turing-AGI Test: a subject, human or machine, gets a computer with internet, browser and Zoom, then a judge designs a multi-day work experience — training, task execution, feedback — and the system passes if it performs like a skilled hire.
Why it mattersA testable definition of AGI framed as multi-day remote work — training period, real task, ongoing feedback — which is a far sharper target for evaluating agents than benchmark scores on isolated tasks.
Prompting Long Context
A measured study of two long-context recall techniques.
Why it mattersTwo concrete long-context recall techniques, quote extraction before answering and cross-section few-shot examples, plus a repeatable recipe for generating your own eval set from a document published after the training cutoff.
articleAnthropic News
Tencent Hy@TencentHunyuan
Tencent open-weights a translation pair sized for two very different deployments: a 1.8B model claiming a 1GB footprint and 0.18s latency per 50 tokens on consumer hardware, and a 7B model positioned against far larger general systems. Weights and code are public.
Why it mattersA 1.8B translation model with a 1GB footprint and 0.18s latency for 50 tokens makes on-device translation practical without a hosted API in the loop.
Dwarkesh Patel
Dwarkesh Podcast: Adam Marblestone on the Brain
podcast
Listen
The State Of LLMs 2025
An in-depth year-in-review article covering the major developments in large language models throughout 2025, including DeepSeek R1, RLVR, GRPO, inference-time scaling.
Why it mattersIt distills a chaotic year of LLM research — DeepSeek's cost disruption, the rise of RLVR/GRPO for reasoning, and inference-time scaling.
articleSebastian Raschka, PhD
Tencent Hy@TencentHunyuan
A billion-parameter flow-matching DiT turning text into 3D character animation, released with weights, code and a technical report. Tencent claims the first motion model trained through a complete pretrain→SFT→RL loop, spanning 200+ motion categories.
Why it mattersOpen weights for text-to-motion at 1B scale with outputs that drop into standard 3D animation pipelines, plus a documented pretrain→SFT→RL recipe for applying RL post-training outside the language-model setting.
Tencent Hy@TencentHunyuan
Two open translation models on a shared lineage: a 1.8B aligned by on-policy distillation for consumer hardware at a 1GB footprint and 0.18s latency for 50 tokens, and a 7B upgrade of Tencent's WMT25 winner. Custom terminology, long-dialogue context and format preservation supported.
Why it mattersThe on-device and cloud models share a lineage, so you can fall back between local and hosted translation without output drift. The 1.8B claims 0.18s for 50 tokens inside a 1GB footprint.
Andrew Ng@AndrewYNg
Ng pushes back on the advice to skip learning and just ship, citing candidates who rebuilt standard RAG chunking and agent evaluation techniques badly. His prescription: courses plus deliberate building, with research papers as an optional third leg.
Why it mattersNames the recurring cost of skipping foundations — badly reinvented RAG chunking, duplicate agent eval schemes, messy context management — and argues for pairing courses with building rather than choosing one.
Dwarkesh Patel
Thoughts on AI progress (Dec 2025)
podcast
Listen
Lenny Rachitsky
The coming AI security crisis (and what to do about it) | Sander Schulhoff
podcast
Listen
The Shape of AI: Jaggedness, Bottlenecks and Salients
An essay by Ethan Mollick analyzing why AI capabilities improve unevenly (the 'jagged frontier') and how single bottlenecks hold back entire systems until they suddenly break through.
Why it mattersGives builders a practical lens — watch which single capability is bottlenecking a workflow rather than chasing benchmark deltas.
articleEthan Mollick
Andrew Ng@AndrewYNg
Ng argues LLMs are general but far less general than the hype implies: after pretraining consumed the public web, progress has become piecemeal, with labs targeting individual tasks. The practical consequence is that breadth of competence is now assembled, not emergent.
Why it mattersCapability gains are now largely per-task work by labs rather than broad generalization, so assume a model is strong only where someone deliberately made it strong — and plan evals accordingly.
Introducing Mistral OCR 3
A document-understanding upgrade aimed at the cases that break pipelines, including handwriting, bad scans, nested tables, and forms.
Why it mattersMistral OCR 3 parses PDFs and images to markdown with HTML-reconstructed tables at $2 per 1,000 pages ($1 batched), claiming a 74% win rate over OCR 2 on forms, handwriting, low-quality scans and complex tables.
articlemistral.ai
Tencent Hy@TencentHunyuan
Tencent open-sources a streaming video diffusion model that generates explorable 3D worlds from text or an image at 24 FPS, holding geometric consistency over long horizons by rebuilding context from past frames rather than letting memory decay. Keyboard and mouse control use a dual action represent
Why it mattersLong-horizon consistency in streaming world models is usually traded against speed; the Reconstituted Context Memory mechanism is a concrete published attempt at holding both, with weights and a report to check it against.
Evaluating Context Compression for AI Agents
Factory.ai built an evaluation framework to measure how well different context-compression strategies preserve useful information during long-running agent sessions (debugging, code review, feature implementation).
Why it mattersIf you run agents over long sessions, compaction strategy silently determines what your agent forgets.
articleFactory
Lenny Rachitsky
Why humans are AI’s biggest bottleneck (and what’s coming in 2026) | Alexander Embiricos (OpenAI Codex Product Lead)
podcast
Listen
How People Use Claude For Support Advice And Companionship
Why it mattersPuts real numbers on how often people bring emotional needs to a general assistant and how rarely the model resists, which is the behavior anyone building a consumer-facing assistant has to design around.
articleAnthropic News
Donating The Model Context Protocol And Establishing Of The Agentic Ai Foundation
Anthropic hands MCP to the Agentic AI Foundation, a Linux Foundation fund co-founded with OpenAI and Block, where it joins goose and AGENTS.md. Governance and maintainers stay put.
Why it mattersMCP's stewardship now sits with a neutral foundation alongside OpenAI and Block, and the Nov 25 spec adds async operations, statelessness and server identity. Both matter if you depend on the protocol's direction.
articleAnthropic News
Introducing Runway Gwm 1
Runway's first general world model family generates frame by frame in real time and takes actions, including camera pose, speech, and robot commands, as input.
Why it mattersA real-time, action-conditioned world model with a fine-tuning path means policy evaluation, counterfactual rollouts and synthetic training data can run in simulation instead of on hardware.
articleRunway editorial sitemap
Political Even Handedness
Anthropic defines political even-handedness as its bias criterion, spells out the target behaviors.
Why it mattersA reproducible, open-sourced eval for measuring political bias across models, with published cross-model results. It is usable directly if you need to evidence neutrality claims about a model you ship on.
articleAnthropic News
Introducing: Devstral 2 and Mistral Vibe CLI.
Mistral's coding model family is now open at two scales: a 123B dense model with 256K context scoring 72.2% on SWE-bench Verified, and a 24B Apache 2.0 variant deployable locally.
Why it mattersDevstral 2 reaches 72.2% on SWE-bench Verified at 123B parameters with a 256K context window under a modified MIT license.
articlemistral.ai
An index of the vibe-coding frontier. Corrections welcome.