Anthropic's Claude Duet acts before you ask

Hi there, this is your daily ☕️ AIpresso.

Anthropic launches Claude Duet work agent LINK

• Anthropic launched Claude Duet, a proactive agent that connects to your apps and reaches out before you ask, suggesting items that need attention and completing tasks in the cloud without local access to your computer.
• Built on Opus 5.5, Duet sends a morning brief, rechecks connected apps every few hours, and runs in its own virtual machine with a browser, handling everything from admin chores to building web apps in one ongoing thread.
• It requests approval before sending messages, making payments, or deleting files, and never sees passwords, but Duet is a US-only research preview behind a waitlist for Pro and Max subscribers, with background checks counting against existing plan limits.

Nvidia expands Cosmos world-model tools LINK

• Nvidia expanded its Cosmos world-model platform at GTC Berlin with Cosmos 3, new Data Factory blueprints, and Framework training tools for developers building physical AI systems that simulate the real world.
• Cosmos Data Factory ships five cloud-native blueprints, covering autonomous-vehicle data, defect imaging, and video augmentation, available through GitHub, Azure, Nebius, and CoreWeave, while Cosmos Framework lets developers fine-tune open models like GR00T, QWEN, and Pi0.5.
• Updated Cosmos 3 NIM microservices deliver up to 2x faster inference with no quality loss, and the Metropolis VSS 3.3 blueprint monitors 40% more video streams per GPU, though those inference gains apply only in validated configurations.

AI finds critical PHP flaws LINK

• PHP shipped emergency updates yesterday fixing two critical memory-safety bugs that let unauthenticated attackers run code on nearly any server, after AI-assisted audits surfaced the flaws and a proof-of-concept exploit leaked online.
• Both bugs, rated 9.2 on CVSS v4, were first found in August by security firm Calif using OpenAI's GPT-5.6 Sol, then independently by Google's Wiz team with Wiz Atlas, Gemini 4 Argon and GPT-5.6 Sol.
• Patched releases span PHP 8.5.12 back to 8.2.35, with backported fixes for the unsupported 7.4, 8.0 and 8.1 branches given reach to PHP 5.4, though PHP warns no configuration change substitutes for updating.

Recursive's AI built a coding model LINK

• Recursive released Flint, a free coding model built largely by its own automated research system, the Eureka Machine, which a small team used to develop the laptop-ready model from scratch in under four months.
• Flint is a mixture-of-experts model activating 2.5B of 40B parameters, hitting ~131 tokens/sec on an M4 Pro MacBook Pro, roughly twice what Gemma 4 and Nemotron 3.5 Lightning managed, and beating other sub-50B models on SWE-Bench Verified.
• The model is free on Hugging Face under Apache 2.0 for offline, private coding without cloud fees, though Recursive admits Flint needs more hand-holding than big paid assistants and works best when large tasks are split into smaller steps.

Hugging Face AI chats with DNA LINK

• Hugging Face on Thursday released Carbon-X, an open-source model that converses in natural language about DNA sequences and can run locally on a laptop, built with researchers at China's Zhongguancun Academy and Italy's University of Naples Federico II.
• The team expanded an existing language model's vocabulary to read DNA in six-letter groups, then trained on 300 billion tokens mixing genomic sequences from mammals, plants, and fungi with English text, deliberately excluding viruses to limit dangerous uses.
• At 14 billion parameters the model runs on local hardware so sensitive patient genomes never leave the user's machine, though it ingests only ~50,000 base pairs at a time and the developers haven't validated its prediction accuracy with wet lab experiments.

Strategies & Tactics

> Managing AI Coding Costs at Scale: Cut AI coding costs by quickly adopting cheaper models that match quality and routing simple tasks to them, rather than capping developer budgets outright.

> How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications: Specialize a fine-tuned model's vocabulary around each token's serving cost rather than its frequency, cutting inference cost 77% without touching the base model.

> LLM Prompt Caching Versus Semantic Caching: Prompt caching reuses the model's processing of identical opening text, while semantic caching reuses whole answers to similarly-worded questions, each cutting repeated work and cost.

> OAuth for AI agents: What four extensions cover, and the gaps they leave: Four OAuth extensions can grant AI agents permissions but none proves an agent's true instruction, traces delegation across untrusted agents, or revokes all a user's access at once.

> Exit Code Zero: The Hidden Failure Mode Every AI Pipeline Inherits: Verify that your AI pipeline actually reads the reviewer's verdict, since a command-line AI tool exits cleanly whether it approves or rejects a change.

Other news you might like

• China's AI leader DeepSeek raises $15B LINK

• TorchTitan Update: New SOTA Models, Performance at Scale, TitanRL LINK

• Accelerating PyTorch on AMD GPUs with FlyDSL LINK

• Nebius borrows $3.5bn from Citi as it says AI compute prices keep rising LINK

• Introducing AI spend controls with Unity Gateway LINK

• Anthropic plans $750 billion compute spend LINK

• Researchers made AI agents better by turning academic papers into working code — no retraining required LINK

Trending tools

BAND for Developers: provides a shared, governed interaction layer letting distributed AI agents and human teams collaborate in real time, reducing fragmentation at scale. LINK

Claude Duet: an AI assistant from a safety-focused research company, designed to help with tasks at any scale reliably and predictably. LINK

Mouse-1: a lightweight decision model that slots into your agent harness, returning probabilities for routine choices so frontier models focus on heavy reasoning. LINK

Fastlane 2.0: analyzes your brand and competitors to generate AI video and image ads, then launches and optimizes campaigns across organic and paid channels. LINK

CutBay: an AI video agent that creates videos from prompts, URLs, docs, and media through chat, handling scripting, visuals, voiceovers, and editing automatically. LINK

ElevenLabs Realtime Translation: streams live speech into expressive translated audio across 90+ languages over one WebSocket, keeping a consistent voice for calls, contact centers, and events. LINK

Trending papers & reports

Shrunk reasoning models can be retrained to think through long, multi-step problems nearly as well as their full-size versions, letting companies run cheaper, smaller AI that still finishes complex tasks reliably. LINK

AI research teams let multiple specialist agents share findings and split a limited experiment budget to autonomously invent better language-model designs, beating a leading architecture at 1.3 billion scale on reasoning and retrieval tasks. LINK

Faster AI text generation gets a reality check, as fancy training tricks that speed up a model look impressive on outside data but add little once the main model rewrites its own answers. LINK

Vocabulary collapse explains why models tuned to human preferences can ace every quality metric yet suddenly break, repeating one token, and shows the hidden drift causing it can be blocked, appearing in 14 of 17 model families. LINK

Diffusion text generators keep dropping rare words, writing them at only ~0.3x the human rate, and this training-free fix restores those missing vocabulary choices so output reads more naturally without retraining. LINK

See you tomorrow for a new dose of ☕️ AIpresso!