Intel
Page 10
AssemblyAI ships Vet Mode, a veterinary correction pass for speech-to-text
AssemblyAI released Vet Mode in beta, a post-transcription correction pass enabled with domain=veterinary-v1.
Why it mattersSpeech-to-text builders in veterinary workflows can fix drug-name swaps like carprofen to ibuprofen by adding one request parameter, at $0.15/hr on top of the base model. Accuracy claims are vendor examples, not benchmarks.

Stop Shipping LLM Settings That Only Win on Paper - Rohan Pandey & Balaji Varadarajan, DigitalOcean
DigitalOcean engineers walk through an agent that tunes vLLM and Kubernetes settings on AMD MI355 servers, measuring throughput under real latency targets and restoring the old setup when a change loses.
Why it mattersServing-config gains can be noise, so changes should be validated against latency targets with repeat runs before shipping. The talk shows an agent loop that automates this with safe rollback on AMD MI355 servers.

Inkling 2 Small now available on AI Gateway
Vercel adds Thinking Machines' Inkling 2 Small, a multimodal reasoning model with text and image input, to AI Gateway, usable from the AI SDK and from coding agents such as Claude Code, Codex and Cursor.
Why it mattersEngineers on Vercel AI Gateway can call thinkingmachines/inkling-2-small through the AI SDK or route it into Claude Code, Codex and Cursor via the Vercel CLI setup, with usage and cost tracking.

Cyber Chaos Monkey
Shopify Engineering describes an agentic vulnerability hunter that attacks its own production infrastructure.
Why it mattersShows how a large engineering team points AI agents at its own production systems with live credentials and full system context to find bypasses. Useful if you are designing agentic security testing.

Gemini 4 Hydrogen now available on AI Gateway
Vercel announces Gemini 4 Hydrogen on AI Gateway.
Why it mattersA new Flash-class Google model aimed at agentic coding and tool use is available now via AI Gateway as google/gemini-4-hydrogen, usable from the AI SDK and coding agents.
Artificial Analysis reran VIALS-AA without tools: scores fall from about 47% to 25% on average. Haiku 5.5 with tools (49%, $0.07 per task) beats Fable 5.2 without them (44%, $0.73). On cell counting, Fable drops from 65% to 9%.
Why it mattersShows that a cheap model in an agent harness can beat a frontier model used one-shot on visual reasoning at about a tenth of the cost. Tools matter most for counting tasks, so harness design deserves as much attention as model choice.

Your SKILL.md Is 70% Stuff the Model Already Knows - Nikolay Petrov, Intel
Nikolay Petrov of Intel presents a method for auditing agent skills.
Why it mattersYou can measure which lines of a skill change model behavior by ablating each claim, instead of guessing. This lets you prune skills safely and keep the details that matter.

codex rust-v0.167.2 — 0.167.2
Codex CLI 0.167.2 makes incremental tool updates the default on Google Cloud Vertex AI.
Why it mattersIf you run Codex against Vertex AI, tool updates now stream incrementally by default. Set incremental_tools_legacy_gcp = false under [features] in config.toml to restore the old behavior.

crewAI 1.15.40
CrewAI 1.15.40 drops chromadb from the embedding function layer, adds exact historical state lookup by row id, scaffolds standalone agents as JSON by default.
Why it mattersCrewAI users get a fix so after_llm_call hooks run on async calls. The embedding layer no longer depends on chromadb, which can change dependency and storage setups.

Lock Your GitHub Actions
Hands-on guide to GitHub Actions workflow lockfiles.
Why it mattersGitHub Actions now supports lockfiles, so you can use readable version tags while commits stay pinned. This hardens CI against tag-repointing supply-chain attacks without manual SHA bookkeeping.
Gemini Decisions API
Simon Willison tests Google's Gemini Decisions API, a decision-model endpoint supporting yes/no, choice and rubric scoring across text, images, audio, video and PDFs.
Why it mattersGemini Decision 1 costs 10 cents per million input tokens with free output, and unlike Jev it accepts video, audio, images and PDFs. It returns graded probabilities instead of hard 0 or 1 answers.

Your Spec Has Bugs. Catch Them Before Your Agent Codes. - Rajath Mahendrappa Kotyal, AWS
AWS's Rajath Mahendrappa Kotyal explains how Kiro checks specs before design or code is generated, combining LLMs with SMT solvers and treating model disagreement as a sign of ambiguity.
Why it mattersSpec-driven agent workflows build whatever the spec says, bugs included. Using SMT solvers plus LLM disagreement to flag contradictory or ambiguous requirements catches errors before code generation.

Listeners blind-ranked 32 ElevenLabs voices; the pro tier lost
Woodstitch ran a blinded pairwise study of 32 ElevenLabs Voice Library voices reading one nonfiction passage with 125 listeners.
Why it mattersIf you pick TTS voices for audio products, this study found Voice Library recording tier and popularity did not predict listener preference, while louder clips won 54% of pairs. The per-judgment data is published.

United Airlines bans AI agents from booking flights on behalf of travelers
United Airlines says it is restricting unauthorized third-party AI agents from booking, check-in and account actions, both on its platforms and through travel agencies and aggregators.
Why it mattersAgents that book United flights or act in MileagePlus accounts will be blocked, including through travel aggregators. Builders of travel agents should plan for first-party APIs or partner channels instead of credential sharing.
aws-emulator
Simon Willison highlights AWS's official service emulator, written in Rust by a swarm of agents.
Why it mattersAWS now offers an official local emulator for testing against its services without paid API access. It was reportedly built by an agent swarm in three weeks, a notable scale for agentic development.

225,000 Lines of Tests, Written and Merged by Agents - Kush Agrawal, Duolingo
Duolingo's iOS team built a scheduled pipeline where coding agents write unit tests, a lifecycle manager shepherds CI fixes, AI review and merges.
Why it mattersShows an agent loop that picks untested code, opens PRs, fixes CI and merges under human review, with measured results (700+ PRs, coverage 11% to 24%). It is a reusable pattern for large migrations.
Déjà View: a looped 117M model rivals billion-parameter 3D reconstruction
NVIDIA's Déjà View loops a single transformer block to recover camera poses and dense geometry from multi-view photos.
Why it mattersReusing one transformer block for K refinement steps beat an identical variant with independent per-step parameters, so depth can be bought with compute at inference instead of parameters.

Llama 3.2 1B Qwen3.8 27B Draft released
NVIDIA released a Llama-3.2-1B draft model for Qwen3.8-27B.
Why it mattersIt shows a Llama-family 1B model rebuilt to speculate for a Qwen target, a practical route when no same-family small model exists for the model you serve.

Qwen3.5 4B Qwen3.8 27B Draft released
NVIDIA released a 4B draft model for speculative decoding with Qwen3.8-27B, trained with Model Optimizer.
Why it mattersIf you serve Qwen3.8-27B, this off-the-shelf 4B draft model can cut decode latency through speculative decoding without changing the target's output distribution.
ElevenLabs announced Interaction Engine, which routes each conversation turn to cheaper or frontier LLMs, plus memory, routines and the Architect assistant in ElevenAgents, the ElevenAI work assistant, and planned fine-tuning of TTS, STT and open LLMs.
Why it mattersVoice-agent builders get per-turn LLM routing to cut token spend, cross-conversation memory, and pre/post-call routines inside one platform. Enterprise fine-tuning on own data is promised but not yet dated.

Agents Made Code Free. Maintaining It Didn't Get Cheaper. - Dmitry Buykin, Maersk
Maersk's Dmitry Buykin describes what breaks months into a codebase largely written by coding agents.
Why it mattersAgent-written code is cheap to produce but costly to maintain. The talk gives concrete practices (task replay, expert-owned evals, runtime guardrails, weekly pruning) and an 18-question audit for AI-coded repos.
Google DeepMind released Gemini Decision 1, scoring 66 on the Artificial Analysis Decision Index, 7 behind leaders at 73. It leads knowledge and reasoning evals but costs $345 per 1M tasks, about 3x Drex Max v1.1, at $0.10 per 1M input tokens.
Why it mattersGives pricing and benchmark positioning for Google's first decision model: it leads on knowledge and reasoning and is well calibrated, but costs about 3x more per task than the leader.

Gemini Decisions API: Fast, calibrated multimodal decisions
Google DeepMind introduced the Gemini Decisions API and gemini-decision-1.
Why it mattersRouting, policy checks and grading can run as one call that returns confidence scores instead of generated text. That cuts cost and latency versus prompting a general model per question.

Fable 5.2 or Opus? When to use each model in Claude Code
Anthropic's guide to choosing between Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5 in Claude Code.
Why it mattersWith four Claude models available, picking the right one per task controls cost and quality in Claude Code. The video covers when to start on Opus, escalate to Fable 5.2, drop to cheaper models, or split work.

python-genai v3.2.0
The python-genai 3.2.0 release adds the Decisions API (client.decisions.create) and introduces self-hosted BYOS mode for environments.
Why it mattersGoogle's SDK now exposes client.decisions.create and a self-hosted environments mode, so Gemini users can call the new decision model and bring their own sandbox hosting from Python.
Praxis 1 Is Now Available
Runway open-sourced Praxis-1, a 27B video-action world model for robot arms, with weights, inference and LoRA fine-tuning code under Apache 2.0.
Why it mattersRobotics developers can fine-tune a 27B world action model with full or LoRA training on their own hardware under a commercial license. The reported 46.2% on RoboLab-120 leads the open VLAs listed.

llm 0.37.2
LLM 0.37.2 fixes two advisories where OpenAI-compatible custom endpoints could receive the user's OpenAI API key.
Why it mattersIf you point LLM at non-OpenAI endpoints, versions before 0.37.2 could send your OpenAI key to them. Upgrade to stop the credential leak.

The Epoch Brief - November 11, 2026
Epoch AI's brief covers an open-weight model autonomously exploiting six Firefox versions, a new record on its Capabilities Index, compute capacity doubling every 9 months, country-level AI chip imports.
Why it mattersEpoch measured an open-weight model producing working browser exploits at roughly $1,000 each, and reports AI compute capacity doubling about every 9 months. Both bear on security planning and capability trends.

Ripple-65B Preview, an open agentic coding model, lands on Ollama
Ricursive Intelligence's Ripple-65B Preview, a 65B dense model trained from scratch on 12T tokens for agentic coding, is on Ollama under Apache 2.0 with up to 1M tokens of context.
Why it mattersYou can run an Apache 2.0, agent-trained coding model locally and point Claude Code, OpenCode or other harnesses at it with one command. Treat the Terminal-Bench result as a vendor claim until independently reproduced.

Validating GPU Platforms with Runtime AI Inventory via the NVIDIA DSX AI Factory
NVIDIA's AI Cluster Runtime now bundles k8s-aibom, which emits CycloneDX AI bills of materials for workloads on Kubernetes.
Why it mattersYou can inventory which models and serving runtimes run in a GPU Kubernetes cluster at roughly 2.2 millicores of CPU overhead, with records in CycloneDX 1.6 format.

Faster Retrieval for Better, More Efficient RAG
Apple researchers train small neural networks offline against a fixed key database to route queries toward their best match faster, exploiting the gap between short user questions and long document keys in RAG.
Why it mattersRAG retrieval cost grows with millions of keys. Training small networks on your expected query distribution can speed up and sharpen retrieval compared with query-agnostic clustering, which matters when your traffic is stable.
Nyc Summit
ElevenLabs summarizes its NYC Summit launches.
Why it mattersVoice-agent builders get domain-tuned speech-to-text and fine-tuning options for open and voice models. Verify availability and pricing in the full post, since the excerpt cuts off.

unsloth v0.1.917-beta — Dynamic v3.1 GGUFs + Performance Fixes
Unsloth 0.1.917-beta ships Dynamic v3.1 GGUFs for Qwen3.8-27B, lower VRAM for FLUX.2 and FLUX.3 decoding, an image queue, HTML and PDF chat export.
Why it mattersLocal-model users get Qwen3.8-27B quantizations and can run Claude Code or Codex against MLX models via unsloth start. FLUX decoding also needs 36-43% less VRAM.

langchain langchain==1.5.0
LangChain 1.5.0 requires Python 3.11 and langchain-core 1.8.0, and adds model router and auto mode middleware, which is marked beta. Python 3.10 users continue to resolve 1.4.x.
Why it mattersPython 3.10 projects stay on langchain 1.4.x, while 3.11+ users get model router middleware (beta). Check your Python and langchain-core versions before upgrading.

OpenRouter launches Extractions API for typed field extraction
OpenRouter launches the Extractions API in alpha.
Why it mattersInstead of prompting and parsing JSON, you declare typed fields and get values back with confidence, spans and null probabilities. Low-confidence fields can go to human review, and one API covers three providers.
Dan Shipper: generation is cheap, ranking is the human edge
Every CEO Dan Shipper contrasts aggregate benchmarks like GDPval with close-up team use of the same model on one survey.
Why it mattersArgues that as default model capability spreads, value shifts to the human perspective that selects among outputs.

langchain langchain-core==1.8.0
langchain-core 1.8.0 removes end-of-life Python 3.10 support, so the minimum is now 3.11, and adds new decision models. Python 3.10 users keep resolving 1.7.x.
Why it mattersIf you build on LangChain, langchain-core 1.8.0 requires Python 3.11 or newer. Environments on 3.10 stay on 1.7.x, so check your pins and CI images before upgrading.
Anthropic's IPO roadshow video
Simon Willison pulls compute details from Anthropic's IPO roadshow video.
Why it mattersAnthropic says it serves Claude on Trainium, TPUs, NVIDIA and AMD, and that a model ported most AMD kernels with little human help. That bears on future model availability and on agents doing low-level systems work.
Runway releases Praxis-1, a 27B mixture-of-experts diffusion transformer that predicts video and actions together, with open weights under Apache 2.0. It reports first place on NVIDIA's RoboLab-120 board and 80.0% on SimplerEnv after fine-tuning on RT-1.
Why it mattersAn Apache-licensed 27B robot control model trained from video pretraining can be fine-tuned on your own hardware. Teams exploring embodied agents get an open baseline to test against closed LLM-driven control.

AI Code Rating – a 3-character disclosure of how AI was used in a repo
A proposed disclosure spec in which a three-character code in a repo's ACR.md states who maintains the project, what share of its code AI wrote.
Why it mattersGives maintainers and readers a compact, machine-readable way to state how much AI wrote a codebase and how closely a human reviewed it. That helps when judging dependencies written with coding agents.
AlphaGenome 2: Deciphering how cells read DNA to better understand disease
DeepMind announces AlphaGenome 2, a sequence model that predicts how DNA variants affect expression across hundreds of human cell types, linking single-cell biology with human genetics.
Why it mattersA new model predicts how DNA variants alter gene expression per cell type, extending last year's AlphaGenome and its all-variants atlas. Biology tooling builders gain finer-grained variant prioritization.
Praxis 1 Video Pretraining To Robot Control
Runway's Praxis-1 report describes an open-weight world action model that turns video pretraining into robot control.
Why it mattersAn open-weight model pretrained on video reports 46.2% on RoboLab-120, ahead of listed VLA baselines such as pi0.5. That points to video pretraining as a data source for robot policies.
Atuin Threads: All your agent sessions on every machine
Atuin 18.24 adds Threads, which syncs coding-agent sessions across machines with end-to-end encryption.
Why it mattersAgent sessions are scattered across machines and tools and are hard to search. Atuin Threads gives them one encrypted, searchable history and lets you start a task in Claude Code and resume it in Codex.
Can a Linear Program Do Weight-Space Programming?
Nathan Barry reinterprets an MIT CSAIL weight-space programming experiment on Qwen3.5-0.8B.

Google SWE points 48 AI agents at NASA Lidar, finds unmapped Maya pyramid in 72h
A Google engineer split NASA lidar imagery of about 1,200 km² of Maya jungle across 48 general-purpose agents guided by a written checklist.
Why it mattersShows a general-purpose multimodal agent fan-out, with a detailed guide and no trained model, can screen a large image dataset and surface a candidate discovery overnight. The pattern applies to any large-scale visual screening task.
DeepMind says its AlphaProof Code framework verified the security of Signal's messaging protocol at the code level. The claim comes from the announcement alone; the proof scope, assumptions and methods were not available.
Why it mattersIf it holds, AI-driven formal verification of production cryptographic code is becoming practical for real systems like Signal. Check the full write-up for what was and was not proven before relying on it.
[AINews] Mercor $750m Series D at $20B valuation, $3B run rate
Latent Space's AINews issue covers Mercor's $750M Series D at a $20B valuation and $3B run rate, including its claim of 3x inference-to-salary spend.
Why it mattersMercor reports spending 3x more on inference than on salaries while reaching $3B annualized revenue, a datapoint on inference economics. The issue also tracks a Gemini 4 Carbon launch with pricing details.

servers v2.4.0
Release v2.4.0 of the official MCP reference servers fixes six vulnerabilities.
Why it mattersSix vulnerabilities are fixed in the official filesystem, git and fetch MCP servers, including symlink race conditions that could expose files outside allowed directories.

ai ai@6.0.330
Vercel AI SDK 6.0.330 fixes excessive CPU usage when streaming long arrays with Output.array or streamObject by parsing and validating completed elements incrementally instead of reprocessing the full accumulated array per text chunk.
Why it mattersIf you stream long arrays of structured output with the AI SDK, parsing used to reprocess the whole accumulated array on every chunk and burned CPU. This release parses completed elements incrementally.

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model
Tether AI Research's qvac team introduces BotPsy, a 0.8B model fine-tuned from Qwen3.5-0.8B for function calling and agentic workflows on-device, positioned as a private, low-cost alternative to frontier cloud models for routine tasks.
Why it mattersA sub-billion-parameter tool-calling model targets on-device agents, avoiding cloud cost swings and provider access risk for routine workloads. Check its evaluation results before relying on it.
An index of the vibe-coding frontier. Corrections welcome.