AI Daily · 2026-10-08

Today’s throughline is AI companies framing scientific discovery as national competitiveness: Anthropic pledged $150 million over three years to bring Claude to NASA, NIH, NSF and other Genesis Mission partners, while NVIDIA committed $1 billion over five years to US research in quantum, healthcare and energy. On models, Anthropic’s new Haiku 5.5 looks cheap at first but carries hidden costs from long-context pricing and tokenization; Liquid AI open-sourced d1-3B and d1-omni-600M edge decision models, and TII released Falcon-ASR, a multilingual speech model led by Arabic. In research, Google’s randomized trial with patent lawyers found an AI drafting assistant raised writing scores, but senior lawyers gained better judgment while junior lawyers showed no overall skill improvement. Agent tooling also advanced: LangChain showed a Restock agent that can actually pay via Stripe Link, added LangSmith fine-tuning, and shipped Managed Deep Agents 0.8/0.9; Claude Code fixed natural-language hooks that could wrongly allow blocked commands. On safety, OpenAI disrupted Russian and Iranian AI-enabled false-front influence operations, with the Russian case reaching category 5 on its breakout scale.

North America · First-hand

Anthropic

⭐⭐⭐ [Other] Building on our commitment to American scientific discovery

Anthropic has committed $150 million over three years to the Genesis Mission, a federal initiative to accelerate scientific and technological discovery through AI. The funding will make Claude available to more than 15 agencies participating in the mission, including NASA, the National Institutes of Health, and the National Science Foundation. Anthropic first announced its partnership with the US Department of Energy and the Genesis Mission last December, and has since worked with the DOE to bring Claude to scientists at national laboratories; the new commitment was announced at the White House Office of Science and Technology Policy's Science: A New Golden Age Summit in Washington, DC. Over the next three years, Anthropic will provide Claude, Claude Code, and API credits to several hundred Genesis Mission research projects, partner with agencies and national laboratories on priorities such as fusion energy and quantum computing, and offer training, onboarding, and technical support. The post also mentions earlier efforts including the Claude Science workbench, 10,000 free and discounted Claude seats for academic scientists, the AI for Science program, and a research preview of the Model Hardware Standard.

⭐⭐ [Product Update] 2.1.294

Claude Code released version 2.1.294, fixing an issue where prompt and agent hooks written as instructions (such as "Block commands that…") allowed the very commands they were meant to block. It also improves how prompt hooks on Stop and SubagentStop written as instructions (such as "Carry on if the build is broken") are judged, so Claude is less likely to stop early.

OpenAI

⭐⭐ [Other] Disrupting AI-enabled “false front” operations

OpenAI reported that it recently banned two covert influence operations, one from Russia and one from Iran, which combined its models with more traditional techniques to support sophisticated "false front" entities. The Iranian operation ran seven "journalist" personas that pitched long-form articles to small and medium online outlets worldwide, while the Russian operation appears to have co-opted unwitting people in Latin America to run an on-the-ground "think tank" and produced fake "leaked" documents and audio scripts. Both made heavy use of AI to draft internal reports and used questionable or deceitful methods to exaggerate their effectiveness. Scored on the IO Breakout Scale (1-6), the Russia-origin operation was assessed as Category 5—the first Category 5 operation OpenAI has disrupted since it began reporting—and the Iran-origin one as Category 4; both landed content in mainstream media. This fits a pattern across the 30 covert influence operations OpenAI has exposed in two and a half years: those aiming at real media outlets rather than fake social media distribution tend to have the highest potential reach and impact.

Google

⭐⭐⭐ [Research] Does better work always mean better workers?

In a three-month randomized controlled trial with practicing patent attorneys, researchers randomized access to a then-unreleased Google Labs AI patent-writing assistant (now part of Gemini Notebook) among 133 lawyers at 11 intellectual property law firms, giving two-thirds early access and leaving the rest as a control group. AI access raised drafting scores by 0.34 standard deviations after 10 days (roughly a 10-percentile-point climb) and by 0.38 SD after 90 days (about 11 percentile points); the gains came from fewer poor scores and more good scores, with no change in the number of excellent scores, and junior lawyers also saved 18 minutes. Effects on on-the-job learning depended on seniority: senior lawyers who used AI for 90 days demonstrated stronger judgment than those who did not, whereas junior lawyers showed no average skill gain, with their scores splitting into more strong and more low scores. The study was published by the National Bureau of Economic Research.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Multimodal open d1 decision models for the edge

Liquid AI released two open decision models in its d1 family: d1-3B and the experimental d1-omni-600M. d1-3B scores 48.57 on the Decision Index 0.2.1, the best result under 10B and ahead of Decider 35B-A3B (47.11); it accepts text and images, while d1-omni-600M accepts either text and images or text and audio. Unlike generative models, decision models do not produce tokens but answer in a single forward pass: d1-3B is trained from LFM2.5-VL-3B, and d1-omni-600M from LFM2.5-Encoder-350M with added vision and audio encoders. Across seven public datasets covering reading comprehension, toxicity detection, intent classification, medical QA and cross-lingual understanding, d1-3B averages 82.9 and d1-omni-600M 78.4. On speed, d1-3B answers a question in 16 ms on NVIDIA Jetson AGX Thor, 26 ms on AGX Orin and 50 ms on Orin Nano, and under 10 ms on RTX 4090 and AMD MI325X.

⭐⭐⭐ [Model Release] Claude Haiku 5.5

Anthropic released Claude Haiku 5.5, a new fast, low-cost model priced at $0.10/$0.50 per million tokens up to 100,000 tokens, rising 5x to $0.50/$2.50 beyond that. Simon Willison notes this matches OpenAI's GPT-6 Luna within 100k tokens and reports higher benchmark scores, but Luna becomes a better deal above 100k tokens. Haiku 5.5 uses a less generous tokenizer, so the same long prompt consumes roughly 1.25x more tokens than Haiku 4.5, amounting to a hidden price increase. The model does not allow reasoning to be disabled and defaults to medium; Willison tested it by generating pelican-on-a-bicycle SVGs across reasoning levels and reported the time and cost of each. Anthropic also halved cache-read prices for Sonnet 5.5 and introduced monthly API credits for Max and Team subscribers ($100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team), which do not roll over and can be used with auto-reload disabled.

⭐⭐ [Model Release] Introducing Falcon ASR

Abu Dhabi's Technology Innovation Institute (TII) introduced Falcon-ASR, a 1.6B-parameter speech recognition model focused on Arabic, especially the Emirati dialect, while also supporting English, French, Spanish and Portuguese. Across six Arabic test sets it achieved an average word error rate of 20.92%, 2.25 percentage points better than the best published result (23.17%) in the leaderboard snapshot used. In TII's internal Emirati evaluation it recorded 22.73% WER and 10.19% CER, the lowest among the systems compared and 4.07 percentage points below the next best, Qwen3-Omni. The model supports word-level timestamps and reached a mean WER of 5.74% on the seven public English test sets used by the Hugging Face Open ASR Leaderboard. All five languages share the same weights without a language flag, and it builds on the earlier Falcon3-Audio work.

Product Update

⭐⭐ [Product Update] Agents that can pay: building Restock with Stripe's Link and Managed Deep Agents

LangChain presents Restock, a sample agent that makes real purchases, running in Slack on Managed Deep Agents (MDA). It searches real products, builds a cart, and checks out through an API supporting the Machine Payments Protocol (MPP), using Zinc in the example for product search and retailer orders. Payment is handled by Link, Stripe's consumer wallet, so the agent never sees a card number; the user approves a spend request up to a set amount and the agent cannot spend above it. The architecture has four parts: Link holds payment methods and asks the user to approve payments; MPP defines the HTTP payment exchange, where the merchant responds with 402 Payment Required and payment instructions and the client retries with a payment credential; Zinc searches products and places orders via its MPP order API; and MDA hosts the agent and handles Slack, human review, credentials and saved state. The walkthrough follows one request for a 12-pack of blue ink pens under $25, ordered for $22.18 including fees (prices marked illustrative), and includes the restock-agent directory layout and a trimmed agent definition, with the user approving twice: the purchase in Slack and the payment in Link.

⭐⭐ [Product Update] Introducing LangSmith Fine-Tuning

LangChain launched LangSmith Fine-Tuning along with smithtune, a CLI that helps teams turn LangSmith trajectories into custom fine-tuned models for their agents. From a single CLI it handles the whole process: creating and preparing a dataset from LangSmith trajectories, training on Fireworks or Baseten, and evaluating the fine-tuned model with LangSmith while uploading eval results back to LangSmith for analysis; users can run it directly or have a coding agent execute commands and inspect results. smithtune is built for post-training and currently supports supervised fine-tuning (SFT), which trains a model on examples of good behavior by updating weights to imitate that behavior. LangChain says its trajectory format records what the model actually saw at every turn, including tool definitions and changing tool availability, so smithtune can pair each action with its true context, while trajectories held out of the training split are used for evaluation. LangChain partnered with Fireworks and Baseten to link trajectories to fine-tuning datasets, and the tool is available on GitHub with a smithtune skill that lets a coding agent drive fine-tuning end to end.

⭐⭐ [Product Update] Managed Deep Agents delivers a better user experience for agents in production

LangChain released version 0.8 of Managed Deep Agents, which it describes as the simplest way to build, deploy, and run agents in production. The release adds support for user-owned credentials, user-level memory, HTTP channels, file transfer in Slack, and a pre-built tool for web search.

⭐⭐ [Product Update] Revamping Skills in Deep Agents

LangChain describes a revamp of skills support in Deep Agents. A skill is a directory with a SKILL.md file—YAML frontmatter containing a name and description plus the instructions—optionally bundling files under scripts/, references/, and assets/. Skills use progressive disclosure: only each skill's name and description sit in the system prompt at startup, while the full instructions are read on demand via read_file and scripts or references are loaded only when needed. The update adds three capabilities: binding tools to skills so a bound tool's schema stays out of context until the skill is read (and, on models that accept new tools mid-conversation, adding it keeps the prompt cache intact); pinned skills so an app can load a user-requested skill before the next model call without read_file round trips; and skill reloading so a long-running thread can pick up added, edited, or deleted skills without starting over. The post uses a GTM agent with more than 50 skills as a running example.

⭐⭐ [Product Update] What's New in Managed Deep Agents: schedules, per-run configuration, and Slack reactions

LangChain released Managed Deep Agents v0.9 with three new capabilities. The Schedules SDK lets agents create reminders, follow-ups, and recurring tasks mid-conversation; schedules run as the requesting user with their permissions and connections, results post back to the originating channel, and one-time schedules (using at instead of cron) reply in the original thread. Per-run configuration defines the agent as a callable function that receives the runtime at the start of each run and returns a define_deep_agent definition, choosing that run's model, instructions, skills, MCP servers, and sandbox; because configuration is set before the model runs, the agent never sees skills or MCP tools outside its configuration. Slack reactions let agents acknowledge a message before replying. The post illustrates per-run configuration with a Slack agent loading finance or platform-team settings by channel and a coding agent loading different skills and models per repository.

Other

⭐⭐ [Other] NVIDIA Commits $1 Billion to Advance US Science Over the Next Five Years

NVIDIA announced commitments valued at $1 billion to build out the United States' capacity for superintelligence research and development in fields it describes as critical for U.S. leadership, including quantum computing, healthcare and energy security. The commitments cover the next five years.