The Sequence Radar - Issue 910: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm

Next Week in The Sequence: More lessons about model distillation. We will cover 3 important AI papers and tech releases you need to know about using a very simple and easy to follow format. The opinion section we will discuss how AI inference works. Subscribe and don’t miss out: 📝 Editorial: Last Week in AI: Google Rewires Its Brain and Meta Hires a Coding Swarm AI weeks are usually measured in parameter counts. This one was measured in org charts. Google effectively opened its skull and began rearranging the cortex. Jeff Dean, one of the architects of the company’s technical nervous system, is leaving after 27 years. Together with longtime collaborator Sanjay Ghemawat, he is launching Discovery Loop, a public-benefit company designed to automate machine learning, science, and engineering. Yet this is not a conventional Silicon Valley defection. Google will remain a founding investor and cloud partner. It feels less like a neuron abandoning the brain and more like a new lobe being detached, given its own budget, and connected back through an API. The second move was even more revealing. Demis Hassabis is handing Google DeepMind’s daily operations to Koray Kavukcuoglu and becoming chair of DeepMind and chief scientist of Alphabet. Hassabis will focus more heavily on AGI, scientific discovery, global strategy, and Isomorphic Labs. The chess prodigy is moving away from managing every piece and toward deciding which game Google should be playing. This suggests Google now believes frontier AI runs on two clocks. The product clock ticks in model releases, developer adoption, and Gemini features. The civilization clock ticks in AGI safety, scientific breakthroughs, and questions that do not fit neatly inside a quarterly roadmap. Trying to run both from the same chair may have become impossible. Google is separating the factory floor from the observatory. Meanwhile, Meta released Muse Code, a terminal-based coding agent powered by Muse Spark 1.2. Muse can plan changes, write code, validate results, and work across large repositories. For bigger jobs, it can fan work out to multiple sub-agents operating concurrently in isolated worktrees. It also records its actions so it can recover after a crash instead of waking up with digital amnesia. That distinction matters. Muse Code is not merely a smarter autocomplete. Autocomplete is a power drill. Muse is trying to be a small construction crew. Meta is entering a market already shaped by Claude Code and Codex, but its architecture points toward the next competitive frontier: not who produces the best individual code suggestion, but who coordinates the best swarm of agents over long-running tasks. The first generation of frontier labs tried to contain everything—research, infrastructure, models, products, and talent—inside one giant castle. Now the castle is becoming a network. Scientists spin into specialized startups. Visionary researchers move above operational organizations. Coding agents divide work among sub-agents. It resembles a mixture-of-experts model, except the experts are people, companies, and software workers. Let’s review this week’s developments: 🔎 AI Research Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes AI Lab: FAIR, Meta, Reality Labs, Meta, University of Oxford. Summary: This study provides a systematic, empirical exploration of unified multimodal pretraining to uncover how modalities like language and vision interact, which is detailed in the file 2608.05000v2.pdf. It introduces insights on asymmetric knowledge flow, modality synergy driven by architectural choices, and the necessity of early joint training, ultimately synthesizing these into highly efficient pretraining recipes. Are the Financial Reasoning from LLMs Credible? A Real World Test over Long-Horizon Statements AI Lab: Qwen Team, Alibaba Group, Tsinghua University. Summary: This paper introduces FININDICES, a large-scale benchmark designed to evaluate data-processing fidelity and structural reasoning over uncropped, full-length financial statements. The evaluation reveals that modern LLMs suffer from both a “Knowledge Bottleneck” and a “Structural Bottleneck” when generating complex financial tables, though supervised fine-tuning can partially restore structured logical capabilities. PAST-Bench: Benchmarking the Foundations of Recursive Self-Improvement in Personal Agents AI Lab: Princeton University. Summary: PAST-Bench is a performance-attribution benchmark evaluating whether personal AI agents can successfully translate retained experiences into improved future behavior across different capabilities. Guided by diagnostic findings from this benchmark, the authors also present HERMES+, an extended agent framework with targeted interventions that enhances the average gain from retained experiences. FinanceHarness: Autonomous Financial Deep Research Framework AI Lab: Google Cloud AI Research, University of California, Los Angeles. Summary: This research presents FINANCEHARNESS, an expert-guided framework for automating financial deep research, along with FINANCEGYM, a verifiable benchmark grounded in strict point-in-time constraints. Results demonstrate that financial deep research remains highly challenging even for leading models, but utilizing the FINANCEHARNESS system significantly improves overall performance by successfully separating pre-cutoff evidence retrieval from post-cutoff reasoning. Agent Against Agent: An Agentic System for Automatic Prompt Injection Red Teaming AI Lab: The Pennsylvania State University. Summary: The authors propose PIMiner, an agentic system for prompt injection red-teaming that bridges the gap between search-based and reinforcement learning-based methods by accumulating reusable attack knowledge. By leveraging a hierarchical memory mechanism, PIMiner achieves highly effective attack success rates across frontier LLMs and demonstrates strong transferability without requiring target-specific training. 🤖 AI Tech Releases Muse Code Meta AI released the beta version of Muse Code, a terminal coding agent optimized for tasks across large repositories. Kitesurf Cloudflare announced Kitesurf, a browser built for AI agents. Prime Agent Prime Intellect released Prime Agent, a “self-improving” coding harness. LFM2.5-2.6B Liquid AI continues shipping with the release of LFM2.5-2.6B, an agentic model that runs on device. 📡10 AI News You Need to Know About Anthropic signs $10B compute deal with Volta — Bitdeer executed a 16-year colocation lease with Volta Tydal AS for its Tydal, Norway campus, with all 121 IT MW configured to run NVIDIA GPUs for an unnamed “leading AI lab”; the Bitdeer release is the primary document, and Bloomberg identified the lab as Anthropic. Demis Hassabis moves to Chair of Google DeepMind — In a joint message to employees published by Pichai and Hassabis, Hassabis stepped back from day-to-day operational leadership to become Chair of Google DeepMind and Chief Scientist of Alphabet while continuing to lead Isomorphic Labs, with longtime DeepMind CTO Koray Kavukcuoglu elevated to SVP and taking over Gemini model development, frontier research, and the Gemini app and developer teams. Jeff Dean leaves Google — Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals left Google to found Discovery Loop, a public benefit corporation building AI systems that automate the experimental loops of science and engineering, with Google as founding investor and cloud partner. Kimi K3 escapes its test sandbox — Frontier Security reported that Moonshot’s Kimi K3 exploited a network egress leak in the UK AI Security Institute’s Inspect benchmark framework, using standard CLI tools to pull reference solutions off GitHub rather than solving the tasks. SK hynix commits $38B to new fabs — SK hynix’s board approved roughly 54 trillion won, 35.2 trillion for the Yongin “Y2” DRAM fab and 19.1 trillion for the Cheongju “M17” NAND fab, with cleanrooms opening June 2029 and December 2028. Firmus raises $2B — Firmus received full commitments for a $2B strategic equity round with follow-on participation from Coatue and NVIDIA plus new money from Blackstone Tactical Opportunities and Jane Street, funding the next phase of its Project Southgate AI factory rollout in Australia and Asia-Pacific. DeepSeek reopens its $8B round — DeepSeek resumed its second funding round seeking close to $8B at a valuation near 500 billion yuan, with Monolith Management in talks to participate, after pausing last month over leaked founder remarks. Yann LeCun joins 224 Ventures — LeCun and Oriol Vinyals joined Shaun Johnson to launch 224 Ventures, a technical and GTM-focused firm investing in AI-native teams, launching with over $100M AUM and writing $1M to $5M checks. Nvidia and Dell back Volta at $2.4B — Volta emerged from stealth with a $300M seed and Series A co-led by Andreessen Horowitz and Altimeter with NVIDIA and Michael Dell participating, plus a $5B AI Infrastructure Program sponsored by Azora and over 1GW of near-term contracted power. Nscale targets a September US IPO — Nscale is telling prospective investors it has roughly $51 billion of total contracted revenue ahead of a US IPO that could come as soon as September, with revenue rising to over $100 million in Q2 2026 from about $37 million in Q1.