Hi there, this is your daily ☕️ AIpresso.
In today's AIpresso:
🧊 Anthropic warns AI may spread to fridges
🧠 Meta launches Muse Wonder AI models
🤖 Claude can now coordinate agents on the fly
🚀 Alibaba speeds up AI agent training
🔎 Perplexity upgrades its AI routing model
Plus: 💡 5 strategies & tactics, 🎁 8 more stories you might like, 🧰 6 tools, and 📚 5 papers.
🧊 Anthropic warns AI may spread to fridges LINK
- Anthropic's Frontier Red Team warns that small AI models capable of self-replicating across consumer IoT devices without data-center infrastructure could emerge within months, based on observed model improvement rates.
- In closed-network tests across eight emulated Android devices, only GLM-5.2 completed the full replication cycle, succeeding in 4 of 5 runs; a 26B-parameter model split across 15 ~$55 TV boxes ran at just 0.39 tokens/sec.
- Sub-30B models already complete individual replication steps but fail at chaining them, Anthropic projects they'll match GLM-5.2's capability by end of 2026, though large model transfers and abnormal inter-device traffic remain detectable at the ISP level.
🧠 Meta launches Muse Wonder AI models LINK
- Meta launched two proprietary Muse Wonder models from its Superintelligence Labs, pitching them for long-running agent tasks like software engineering, legal work and business automation, and claiming parity with top OpenAI and Anthropic offerings.
- Artificial Analysis scored Muse Wonder Full at 56 on its Intelligence Index, tying Claude Sonnet 5.5 and edging out GPT-6 Astra and Gemini 4 Argon, but trailing Claude Opus 5.5, while the standard Wonder landed just below Astra.
- Standard Wonder runs via API at $4/$20 per million input/output tokens, but neither model ships open weights, and the stronger Wonder Full is locked to Muse Code subscriptions that default to a mode letting Meta train on your submissions.
🤖 Claude can now coordinate agents on the fly LINK
- Anthropic walked through orchestrating its Claude Managed Agents at AI Engineer New York today, spotlighting "dynamic workflows," or "flows," a multiagent feature announced on Friday that lets Claude design and run its own agent topology.
- Rather than developers configuring a coordinator and a roster of subagents, the lead agent writes an orchestration program that launches agents, collects results, and manages information flow across phases, with pre-configured specialist sub-agents optionally passed in.
- Available in Claude Console for API developers, flows are designed to throw test-time compute at ambitious tasks and proved more thorough in a stock-watchlist demo, though Avilla said they cost "much more" than a single-agent approach, making a session budget essential.
🚀 Alibaba speeds up AI agent training LINK
- Alibaba's Qwen team unveiled QwenGyre, an elastic reinforcement learning framework built to train extreme-long-horizon software engineering agents whose single rollouts span hundreds of environment interactions and burn up to roughly one million tokens.
- An Elastic Scheduler shifts GPUs between rollout generation and training on the fly to kill idle time, while a Trajectory Processor dedups branching execution graphs into prefix-sharing trees, cutting redundant training overhead.
- QwenGyre ran 1.21x-1.85x faster than colocated setups with training scores matching baselines, and on the Qwen 3.8 2.4T model lifted NL2RepoBench accuracy from 52.5% to 58.5% in 48 steps.
🔎 Perplexity upgrades its AI routing model LINK
- Perplexity has open-sourced pplx-decider-v1.1-27b, an updated version of its routing model that scores 61.56 on the Decision Index, up from 56.4 and now ahead of Jev by more than 3.5 points on the same backbone.
- Built on Qwen3.8-27B with a BF16 decision head, the model owes most of its gains to lifting the causal mask and training on more data, including tasksource, and handles images at up to 2048×2048 within an 8,192-token limit.
- Reproducing the evaluated behavior requires the included inference code, default causal inference won't match the checkpoint, plus Python 3.12+, a CUDA GPU with ~49 GiB for weights, and authenticated access to what remains a private Hugging Face repository.
💡 Strategies & Tactics
> Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations: Meta pools idle AI computing capacity across its data centers to rank more ad candidates and route key requests to stronger models, improving ad relevance without buying more hardware.
> ReplaySSM: Rethinking State Storage for Faster Hybrid-Model Decoding: Caching recent inputs instead of repeatedly rewriting a model's memory state roughly halves memory traffic, making hybrid AI models decode up to twice as fast.
> Do Agent Teams Pay Off? A Case Study on Vibe Code Bench: Splitting a coding job across multiple AI agents raised scores only sometimes, costing up to five times more, so gains depend on the model and task.
> Building Long-Running Agents with ScaleDown and Deep Agents: Offload an agent's repetitive context-management decisions to ScaleDown's narrow, single-task models, which run far cheaper and faster than general-purpose models on every turn.
> DiffusionART fixes a hidden flaw in diffusion RL and trains video models with 12x fewer GPU hours: DiffusionART removes a hidden destabilizing term from fast diffusion reinforcement learning, training video models up to 12 times cheaper without per-task tuning.
Other news you might like
- OCP Global Summit: Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide LINK
- Microsoft plans huge deployment of its own AI chips, with an eye on saving costs LINK
- Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation LINK
- NTT Data in final talks to build Anthropic's first data centre in Japan, Nikkei reports LINK
- AMD debuts Helion reference design for agentic AI servers with up to 32,768 Zen 6 ‘Venice’ cores LINK
- Intel's Crescent Island AI accelerator officially powers new GPU D-Series products LINK
- “A major overhaul”: OpenCode 2.0 puts its terminal, desktop, and web apps on one shared service LINK
- Google's Gemini 4 "Carbon" model is reportedly matching Anthropic's Opus 5.5 coding performance LINK
🧰 Trending tools
Pegasus 1.6: a multimodal video model that searches, analyzes, and generates text from footage, turning raw video into structured datasets without manual annotation. LINK
MausBot: assign jobs to AI agents and manage them locally through a chat interface, open source with bring-your-own-provider support for flexible workflows. LINK
Takes for Mac: a free, open-source app for talking-head videos with voice-following teleprompter, one-button recording, and AI-assisted editing, captions, and post drafting. LINK
AdHelm CLI: an open-source command-line tool connecting your AI agent to ad and analytics accounts locally to report, audit, and recommend changes securely. LINK
Nano Banana 2.1: searches webpages, images, videos, and more using specialized tools and filters, helping you find exactly what you need quickly. LINK
Xiryn: free, open-source personal agent running in a Docker Linux desktop that browses, researches, codes, and fills forms locally, controlled via Telegram. LINK
📚 Trending papers & reports
AI email assistants can be tricked by hidden instructions buried in incoming messages, with a real-world test of 208,000 human-written attacks fooling even purpose-built defenses 32% of the time and GPT-4o 58%. LINK
Measuring meaningful information offers a new math framework for scoring how informative a statement is, fixing a long-standing flaw where false statements wrongly scored as maximally informative. LINK
Trial-and-error AI training gets a faster, simpler recipe that cuts the noise in how systems learn from experience, reaching near the mathematically best efficiency possible using only lightweight calculations and minimal assumptions. LINK
Phase-aware attention speeds up how Transformers digest long documents by up to ~22x when reading the input, cutting costs on heavy-context tasks while matching or beating standard models on accuracy. LINK
Graph generators fix a built-in flaw where networks modeled from scratch collapse into bland, uniform shapes, letting the tool reproduce realistic connection patterns and lopsided popularity without extra hand-tuning. LINK
See you tomorrow for a new dose of ☕️ AIpresso!