We have the second installment of our series about recursive self improvement. Can’t miss that one.
We deep dive into the new Gemini models, Stanford’s University Paper2Agent research and Astra for Law.
In the opinion section, we debate the opportunities and challenges of text diffusion models from first principles.
We have another installment of our robotics series.
Building useful AI is starting to resemble building an automobile. The engine matters enormously, but so do the steering, transmission, chassis, and fuel supply. This week offered a remarkably complete tour of that machine: Google upgraded conversational control, OpenAI specialized intelligence for law, Figure tested physical generalization, and Crusoe put a $30.9 billion valuation on the infrastructure underneath it all.
Google’s Gemini 3.8 Live and Live Extended Thinking tackle a deceptively hard problem: keeping a conversation alive while useful work happens. The models combine visual context with background tool calls; Extended Thinking can reason and speak concurrently. Imagine discussing travel plans while an assistant checks availability and handles your interruptions. The engineering challenge is coordinating dialogue, computation, and actions without making the user wait through awkward silence. Voice becomes a practical control surface for agents, with latency joining accuracy as a central design constraint. Google’s announcement
OpenAI’s Astra for Law addresses another source of friction: professional context. It combines GPT-6 Astra with legal instructions, specialized search, and tools for legal workflows. On 200 questions from a private Legal Research Bench validation set, OpenAI reports 54% correctness versus 38.7% for Astra using ordinary web search. That remains a substantial distance from dependable autonomy. Still, the result illustrates an important principle: equipping a capable model with the right information environment can materially improve its performance. A brilliant associate still needs access to the right case law. OpenAI’s announcement
Figure’s Helix 2.5 takes that argument into unfamiliar living rooms. The company tested tidying, towel folding, and bed making across 30 unseen homes. Pretraining on its Index human-behavior dataset raised complete-task success from 9% to 56%, with architecture and task-specific training held constant. Here, “zero-shot” means new homes and objects; the behaviors were learned using robot data collected elsewhere. The exciting result is the transfer: broad human experience made the same robot training much more useful. The remaining 44% failure rate is equally instructive. Your laundry has very little patience for a promising scaling curve. Figure’s report
Then comes the electricity bill. Crusoe announced the initial closing of an anticipated $3.9 billion Series F at a $30.9 billion post-money valuation. Its platform connects energy, data centers, and AI cloud services, with expansion spanning large campuses and modular Spark facilities. The investment thesis is straightforward: increasingly capable agents create demand for increasingly available computation. Delivering that computation requires securing power, constructing facilities, and operating hardware efficiently. Every conversational flourish and robotic recovery eventually becomes somebody’s infrastructure workload. Crusoe’s announcement
These developments suggest that AI’s next chapter will reward the teams that connect intelligence to its operating environment. Conversation requires timing. Legal work requires authoritative context. Robotics requires transfer across messy physical settings. All three require computation that someone can actually deliver. The opportunities extend across that entire chain, and so do the failure modes. As models acquire more responsibility, progress will increasingly be measured by completed work under real constraints. The fascinating part is how much invention remains between an impressive model and a system we can comfortably depend on.
AI Lab: Carnegie Mellon University
Summary: DDO freezes base weights and edits a few low-impact MLP neurons to inject a harmful-selective, refusal-orthogonal decoy so contrastive abliteration (RFA) removes the decoy instead of real refusal. Across six model families it holds standard-RFA ASR under 10% (85%→1.8% on Llama-3-8B-Instruct), cuts Heretic ASR from 88.7% to 18%, and matches trained defenses under multi-phase attacks at 30–450× lower cost (~2 minutes on one A100).
AI Lab: Johns Hopkins University
Summary: The authors formalize 100-task continual memorization (no replay buffer of raw past examples, no task IDs) and show single mechanisms fail, then compose data/function/weight anchors with merged LoRA via successive-halving + factorial search on Symbol-/LLM-/Real-QA. The full stack is the only composition top-3 on all three datasets, lifting average final retention from 1.2% (naive SFT) to 34.9%—a 28-fold gain—driven by a super-additive replay × merged-LoRA interaction.
AI Lab: Google Research, University of Illinois Urbana-Champaign
Summary: R4T uses Soft-GRPO once to train a fan-out LLM on set-level rewards (diversity, groundedness, alignment/coverage), synthesizes supervision from successful trajectories, and distills a 53.9M-parameter embedding diffusion retriever for single-pass fan-out. On Polyvore fashion and music playlist benchmarks it beats zero-shot and Best-of-N fan-out baselines while cutting query fan-out latency by roughly an order of magnitude (about 12–20× vs autoregressive LLM fan-out).
AI Lab: Sungkyunkwan University, Microsoft
Summary: When2Think post-trains hybrid reasoning with Instance-level Difficulty-Aware Control (IDAC)—reward shaping from pre-computed reference accuracy and token budgets plus verifier rewards—so the model learns when to use NoThink (System 1) vs Think (System 2). On AIME24 it lifts Pass@3 from 46.0% to 56.0% (+10.0) while cutting tokens 27.9% (14,195→10,236); on AIME25 it reaches 40.0% Pass@3, beating compression and routing-only baselines.
AI Lab: Stanford University
Summary: Paper2Agent auto-builds validated MCP servers (tools/resources/prompts) from a paper’s manuscript and codebase, then wires them to chat agents so methods run via natural language. AlphaGenome agents hit 98.7%/100% on tutorial/novel queries (vs ~83%/79% Claude+Repo); across 74 agentified of 100 comp-bio papers, 593/599 tools pass validation and Sonnet-4 agents score 91.2% on 300 tutorial questions—plus multi-agent collaboration that prioritizes GPR137 at a psoriasis GWAS locus.
Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, its latest live-dialogue models for voice agents—near real-time visual grounding, mid-conversation language switching across 97 languages, background tool calls, and simultaneous speak-while-reason on Extended Thinking—rolling out via the Gemini Live API, Search Live, Gemini Live, and Workspace Live surfaces.
OpenAI introduced Astra for Law, a GPT-6 Astra configuration for legal work with analysis/writing instructions, thoroughness settings, and a Legal Search Index covering U.S. case law, statutes, regulations, court rules, and administrative decisions across 230M+ URLs (sources added daily)—initially for selected firms via Trusted Access in ChatGPT and Codex, with API access coming soon.
Figure introduced Helix 2.5, its strongest humanoid policy yet—pretrained on the Index human-behavior dataset, then adapted once into three whole-body skills (living-room tidy, towel fold, bed make) that ran zero-shot across 30 unseen Bay Area homes with no on-site data or fine-tuning, lifting success from 9% to 56% versus training from scratch.
Profound raised $180 million in Series D funding at a $1.8 billion valuation, led by Sequoia Capital and Kleiner Perkins with Lightspeed, Khosla, and South Park Commons participating—less than seven months after its Series C—as the AEO/marketing platform cites rapid growth and large enterprise adoption.
Nvidia CEO Jensen Huang pushed back on Anthropic’s call for antitrust waivers so labs can coordinate “pacing the frontier,” telling CNBC the idea of new antitrust or regulatory carve-outs for that purpose is “completely unnecessary” and framing AI safety as an engineering and testing problem rather than a reason for coordinated slowdowns.
Bloomberg reported that OpenAI is in early, investor-initiated talks for a fresh funding round that would value the company at more than $1.2 trillion ahead of an IPO, with any decision to proceed hinging on IPO timing (no company announcement).
Emulate, a UK startup founded by former Google DeepMind world-model researchers (including Jack Parker-Holder, Matthew McGill, and Philip Ball), is in talks to raise about $700 million in seed funding seeking a ~$3 billion pre-money valuation for systems that simulate and predict physical-world behavior (terms not final; no company announcement).
Manus is nearing a ~$500 million raise at a $4 billion valuation—its first round since Beijing forced the unwind of Meta’s ~$2 billion acquisition—which would make the agentic AI startup China’s most valuable in its category if it closes (talks ongoing; existing backers include Tencent, HSG, and ZhenFund).
Bain Capital Ventures raised $1.6 billion for Fund XI to back early- and growth-stage companies for a “post-AGI” economy—spanning AI infrastructure, physical-world tech, security, and services that sell work rather than classic software.
Treble raised $18 million in a Series A-2 led by Paladin Capital Group (with KOMPAS VC, Frumtak Ventures, and the EIC Fund) to expand its Iceland-based acoustic simulation and synthetic audio data platform for voice AI, wearables, and physical AI.
Crusoe raised $3.9 billion in Series F funding at a $30.9 billion valuation, co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners (with Founders Fund, GIC, Nvidia, QIA, Radical Ventures, and TPG among participants), to scale large AI campuses and truck-deployable modular Crusoe Spark “AI factories.”
Snap introduced SPECS, standalone AR glasses (132–136 g, 51° FOV LCoS display, dual Snapdragon processors, ~7 ms motion-to-photon latency) for AI assistance, work, and Lens experiences—pre-order at $2,195 with fall shipping in the US, UK, and France.
Bloomberg reported that SoftBank raised its Arm-backed margin loan by $5 billion to $25 billion after renegotiating terms with creditors this month, adding leverage against its chip unit stake to help fund expanding AI investments (no company announcement).