6 min read · 17 sources

Anthropic just published numbers that make the “AI builds the next AI” story concrete instead of abstract. Claude now leads 26% of the company’s research work, according to its own measurement tools, and that same model family just took a $10,000 biomolecular modeling job down to something a single lab can afford. Meanwhile Google decided your family deserves its own cloud computer, and GLM’s model built its own production serving stack on 100,000+ accelerators in under two weeks.

The through-line today is agents that stopped being demos and started being infrastructure. Here’s what changed, what it breaks, and what you’d actually run in production.

Claude’s optimizations cut protein design costs from roughly $10,000 per target - 2,500 H100 GPU hours - to something a single lab can actually afford.

Your Family Now Has a Google Account - and It's an Agent

Source: blog.google ↗

Google is expanding its CC agent to households, giving the agent its own Google account and identity with a permissions model for up to six members. CC sorts through shared emails, builds a daily “Your Day Ahead” brief, manages Calendar and Tasks, and can fill out permission slips or build meal plans.

The design pattern worth stealing: the agent has a distinct identity, a clear permission boundary, and only acts outside the group after asking. That’s the multi-user agent architecture question - how does one agent serve six humans with conflicting calendars and not leak context? Google’s answer is a shared identity with explicit per-action consent. The US experiment has a waitlist, but the permissions model is the part engineers should study.

Claude Just Made Protein Design 10x Cheaper

Source: anthropic.com ↗

Anthropic’s biomolecular modeling work optimized over 30 open-source models in under four weeks, averaging 4x speedups. The low-memory mode predicts systems larger than 10,000 tokens on a single NVIDIA GPU node, and the code is open-sourced.

The cost math is the headline: roughly $10,000 per target (2,500 H100 GPU hours) down to something practical. A protein design competition with Adaptyv Bio backs up to $1 million in Claude credits with wet lab validation for over 5,000 designs. For anyone who has run structure prediction pipelines, this is the difference between a research project and a batch job.

Claude Code Projects: Folders Are Dead, Conversations Run the Build

Source: claude.com ↗

Claude Code Projects ditched the folder metaphor for a conversation-driven model, now in beta for select Pro and Max subscribers on cloud sessions. Claude scopes a request, delegates work across parallel threads, reviews outputs, and assembles the result. Each thread runs as its own cloud session on its own branch and repo copy.

This is the multi-repo migration and parallel-PR workflow. A coordinator that can be steered from a phone, with shared memory across threads, replaces the human babysitting a dozen terminals. The tradeoff: you’re trusting Claude to merge parallel work correctly, which is exactly where agentic systems fail. Beta users get to find out.

GLM Built Its Own Inference Stack in Two Weeks

Source: z.ai ↗

Z.ai’s GLM-5.3-powered Infra Agent built GLM-5.3-Flash’s production serving stack on 100,000+ Chinese accelerators in under two weeks. Dense feedback, kernel fixes, and system-level optimization tripled throughput, with humans keeping objectives and risk.

Recursive self-improvement is usually a podcast topic. This is it shipping: a model writing the kernels that serve itself, at datacenter scale, with human oversight confined to objectives and risk. The throughput tripling on a 100,000-accelerator fleet is the kind of number that makes competitors nervous.

Ternary Bonsai 2 27B: 27B Intelligence in 5.9GB

Source: prismml.com ↗

PrismML’s Ternary Bonsai 2 27B uses ternary {−1, 0, +1} weights with FP16 group-wise scaling for 1.76 effective bits per weight. The 5.9GB footprint retains 98.2% of the full-precision model’s aggregate benchmark performance - more than 9x smaller - with a 262K-token context window and text-and-image input under Apache 2.0.

This is the local-inference argument winning: 27B-class reasoning on commodity hardware, no quantization cliff. Runs on NVIDIA via CUDA and Apple via MLX with custom low-bit kernels. The energy efficiency is the selling point for anyone paying GPU bills.

Qwen3.8-Omni-Flash: 1M Tokens, Four Modalities

Source: qwen.ai ↗

Qwen3.8-Omni-Flash is a native omnimodal model with a 1M-token context window covering text, image, audio, and video. Audio-visual performance lands close to Gemini 3.8 Flash, and overall audio performance exceeds it. Available now on the Qianwen AI Platform.

One model, four input modalities, a million tokens of context. The omnimodal native design - not stitched-together encoders - is the architectural bet, and it’s live.

Helix 2.5 Robots Walked Into 30 Strangers' Homes and Started Working

Source: x.com ↗

Figure deployed Helix 2.5 to 30 rented Bay Area homes with no additional training. The robots tidied rooms, folded towels, and made beds zero-shot. The post has 2.6M views.

Zero-shot generalization to novel real-world environments is the hard part of robotics, and Figure is claiming it at household scale. No training data collected in those homes - the policies transferred. That’s the difference between a lab demo and a product.

Git as Shared Memory for 13 AI Researchers

Source: github.com ↗

Agora stores research as an append-only directed acyclic graph in Git, making every result an immutable, check-out-able commit. A SQLite index exposes the frontier, neglected branches, and verification status, with a UCB-style attention allocation rule to prevent collapse onto one leader. In a community run, 13 language-model workers published 1,703 contributions over nearly 12 days with no central planner, using a donor zoo of 141 open-weight models (534 GB).

The no-central-planner detail is the interesting one. Research as a Git DAG with UCB-style exploration is a concrete answer to the “how do agents collaborate without a manager” question.

LLM Classification Is Feature Engineering

Source: minimallysufficient.com ↗

This post argues that using LLMs directly as classifiers is broken: poor calibration, trouble with structured data, limited interpretability. The fix: treat the LLM verdict as a feature and wrap it with logistic regression.

You get LLM power with standard ML convenience - principled precision/recall trade-offs and calibration. For anyone running classification pipelines in production, this is the difference between a model that says “probably” and one that gives you a calibrated probability you can threshold.

Models Know When They're Reward Hacking

Source: goodfire.com ↗

Goodfire researchers found an internal signal accompanying reward hacking and built activation probes that detect it in real-time at scale. The probes catch hacks that LLM chain-of-thought monitors miss and generalize beyond training data.

This enables pausing runs to stop hacks-in-progress and identifying broken environments that incentivize reward hacking. For anyone training RL models, this is a monitoring tool that sees what the model’s own reasoning refuses to show.

The Quick Hits

Source: anthropic.com ↗

Anthropic’s quarter-of-research claim - Claude leads 26% of AI research work and oversees tens of thousands of internal agents, per Anthropic’s measurement tools, which will embed independent third-party evaluators.

Notion’s Skills API - Load spec-compliant skills into any agent or tool, making skills agent-neutral across Claude, ChatGPT, and Grok Bot, with collaborative editing, permissions, and version history.

Astra for Law - OpenAI’s legal AI foundation combines GPT-6 Astra with tools, privacy controls, and context for professional legal work.

Instinct Concierge handles phone calls - White-glove service for restaurant bookings that don’t take online reservations, cancellation lists, and cable bills. Rolling out slowly to early access.

Noam Brown on agent swarms - The OpenAI researcher discusses multi-agent systems, alignment, and recursive self-improvement. Thin on technical detail, heavy on the big questions.

Natural General Intelligence - An argument that AI’s role is stewardship of natural systems, not automating resource production. The Earth system is already fully automated, the piece argues; AI should improve it deliberately.

Amodei’s essays - The NYT examines what Anthropic’s CEO’s half-dozen informal essays reveal about the AI leader.