Claude Projects v2 💼, Google family agent 👨👩👧👦, — TLDR AI 09/18
The 5% AI cashback card (Sponsor)
For most of you reading, your card statement looks less like a list of hobbies and more like the most recent edition of LMArena.
Every other card treats that spend as "software." Plasma One Core treats it as its own category, and pays 5% cashback on AI spend.
The eligible merchants cover everything you might need for work or this weekend's side project:
Anthropic / Claude · OpenAI / ChatGPT · Cursor · Replit · Lovable · Perplexity · ElevenLabs · Notion AI · OpenRouter · xAI / Grok · DeepSeek
What Core includes:
- 5% cashback on AI spend
- 3% base cashback on everything else
- A ChatGPT Go subscription, about $100 of annual value, included
How Claude is uplifting biomolecular modeling (12 minute read)
Claude enhanced biomolecular modeling by optimizing over 30 models, achieving a 4x speed increase and creating a low-memory mode for predicting larger systems on a single NVIDIA GPU. These improvements, now open-sourced, allow for efficient protein design and structure prediction, potentially accelerating drug discovery. A protein design competition co-sponsored with Adaptyv Bio offers up to $1 million in Claude credits and highlights these advancements.
Google wants to give your family its own cloud computer (4 minute read)
Google's family agent runs on its own cloud computer and Google account, turning the emails, files, and calendars your household shares into daily briefings and updated plans. It can fill out forms and coordinate activities for up to six people, asking permission before acting outside the group. The US experiment has a waitlist.
Projects redesigned: from folder to conversation (4 minute read)
Claude Code Projects let users manage builds by automating task delegation, coordination, and result assembly across cloud sessions. Projects utilize threads for parallel operations, adapting based on progress, and draw on shared memory for efficient task execution. Available now in beta for select subscribers, with wider access coming soon, projects promise enhanced workflow management and context retention.
Noam Brown – Agent swarms, alignment, & recursive self-improvement (72 minute read)
This post features a transcript of an interview with Noam Brown, a research scientist at OpenAI who works on reasoning, reinforcement learning, self-play, and multi-agent AI. Brown was a foundational contributor to the development of reasoning models. In this interview, Brown talks about multi-agent, Navier-Stokes, the current explosion of maths progress, automating AI research, and how we will know whether models are actually aligned before recursive self-improvement kicks off. A video of the interview is available.
LLM Classification Is Feature Engineering (20 minute read)
Getting LLMs into shape to reliably serve as classifiers is hard work but potentially highly impactful. More and more research is relying on LLMs for classification, so these tools need to output high-quality results. This post shows how classification is just a feature engineering problem.
Anthropic says its AI now does a quarter of its research work (25 minute read)
Anthropic says Claude now leads 26% of its AI research work and oversees tens of thousands of active internal agents. Its new measurements track how much AI is helping build the next models, whether people can still oversee those agents, and the computing driving the work.
Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure (15 minute read)
Z.ai used a GLM-5.3-powered Infra Agent to help build 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 while keeping humans responsible for objectives and risk.
Anthropic and SpaceXAI on What Comes After the AI Copilot (Sponsor)
AI agents are moving beyond writing code to planning, testing, reviewing, and shipping it. At {unscripted} Virtual, hosted by Harness, hear from Anthropic, SpaceXAI, Workday, and more on building an SDLC ready for increasingly autonomous AI.
September 30 | 10 AM PT | Virtual
Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint (7 minute read)
Ternary Bonsai 2 27B brings stronger reasoning, coding, vision, and agentic capability to the Bonsai series. It uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, for 1.76 effective bits per weight and a total model footprint of 5.9GB. The model supports a 262K-token context window and multimodal text-and-image input. It can be run on Nvidia GPUs via CUDA and on Apple devices via MLX through custom low-bit kernels.
Helix 2.5 (2 minute read)
Figure introduced Helix 2.5, a humanoid control model pretrained on its Index dataset and tested zero-shot across 30 Bay Area homes. Without collecting training data in those homes, the robots performed tasks including tidying rooms, folding towels, and making beds.
Git as Shared Memory for AI Research Agents (GitHub Repo)
Agora lets autonomous research agents share findings through an append-only Git DAG, where hypotheses, results, verifications, and reports become reproducible commits.
Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery (18 minute read)
Qwen3.8-Omni-Flash is a native omnimodal model. It supports a 1M-token context window with text, image, audio, and video inputs. Qwen3.8-Omni-Flash achieves audio-visual performance close to Gemini 3.8 Flash and overall audio performance that exceeds Gemini 3.8 Flash. It is now available on the Qianwen AI Platform.
The Awesome and Alarming AI Visions of Anthropic's CEO (10 minute read)
Dario Amodei, a co-founder and the chief executive of Anthropic, is one of the tech industry's most prolific and polished explainers. He has contributed to dozens of scholarly papers and more recently wrote a half-dozen informal essays for the general public. This article takes a look at his writings to see what they reveal about the AI leader.
Natural General Intelligence (125 minute read)
AI's potential to transform our relationship with the natural world is often overlooked. Infinite intelligence will not eliminate our need for natural resources. AI promises abundance via the automation of knowledge work, but the Earth system is already fully automated. Natural General Intelligence is a foundation model grounded in the state and dynamics of the planet itself. This article explains the concept, shows readers how it can be built, and explains why it is necessary.
4 meeting platforms, 21 languages, one transcript that gets the jargon right (Sponsor)
Wispr Flow Notetaker captures Zoom, Meet, Teams, and Slack huddles, spells names and acronyms right, and answers questions across every meeting you've had. Free on Mac and Windows.
Models know when they're reward hacking — and we can catch them at scale (16 minute read)
There is a clear internal signal in models that accompanies reward hacking.
Instinct Concierge Can Now Handle Phone Calls (1 minute read)
Instinct is rolling out Concierge, a white-glove service for tasks that require human interaction, including phone calls, high-end bookings, and customer service.
Astra for Law (8 minute read)
Astra for Law is a legal AI foundation that combines GPT-6 Astra with tools, settings, privacy controls, and context designed for professional legal work.
Notion lets teams share the same instructions across AI agents (7 minute read)
Notion's Skills API lets teams edit agent instructions together and distribute them to other agents through GitHub sync or Vercel's skills installer.