This week in deep learning, we bring you Dario Amodei — We Must Pace the Frontier, ToolGrad: Efficient tool-use dataset generation with textual “gradients” and a paper on The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement.
You may also enjoy Introducing DeepSeek-V4.1-Flash, An operationalization of opaque serial depth — Redwood Research, a paper on Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems, and more!
As always, happy reading and hacking. If you have something you think should be in next week’s issue, find us on Twitter: @dl_weekly.
Until next week!
Industry
Dario Amodei — We Must Pace the Frontier
Dario Amodei calls for embedded third-party evaluators with employee-level access, domestic safety standards capping unchecked progress, and global limits on narrowly dangerous uses.
Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
DeepSeek ships a 552-billion-parameter MoE, scoring 74.2 on DeepSWE v1.1 against Opus 5’s 74.0.
Introducing Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking
Google launches two speech models spanning 97 languages, with Extended Thinking topping the Speech Quality Index at 82.6 and scoring 97.7% on Big Bench Audio.
Orchestrating the Pareto Frontier: Fugu Max and Fugu Ultra v2
Sakana AI releases two orchestration models, with Fugu Max priced at $2/$6 per million tokens and leading six benchmarks including Terminal Bench 2.1 and SWEFish.
Microsoft’s new AI ‘code of conduct’ tells models not to hack systems or trick humans
Microsoft imposes absolute constraints against cyberattacks, nuclear weapons and deepfakes, and bars models from deceptive or collusive mechanisms that evade human oversight.
Mistral and Mozilla are bringing open, private and multilingual AI to your web browser
Mistral will power Firefox’s Smart Window assistant in France and North America under a zero-data-retention agreement, with UK and Germany rollout expected later in 2026.
MLOps/LLMOps/AgentOps
Agent Substrate brings high-density, scalable, trusted infrastructure to GKE
A technical announcement about Google’s open-source agent runtime, which packs 1,000+ dormant agents per host at 10x container density with sub-500ms resume and 500+ activations per second.
An article about a two-pass document pattern that skimmed 84 SEC filings totalling 12,013 pages in 32 seconds, then ran expensive VLM OCR on just two pages.
When to Use Encode-Prefill-Decode Disaggregation to Accelerate Multimodal Model Serving
An article about splitting vision encoding from prefill and decode, cutting time-to-first-token 25–93% and end-to-end latency up to 7x on image-heavy prompts with short outputs.
Learning
Modernizing complex legacy code with AI agents
An engineering write-up about migrating 40,000 lines of Fortran 77 to C++, where a numerical parity harness and a coder-tester-reviewer workflow beat full autonomy.
Harbor: Stripe’s AI-assisted prototyping tool
A detailed engineering post about in-browser prototype compilation and comment re-anchoring across agent rewrites, with 3,000+ Stripes producing 12,000 prototypes since May 2026.
ToolGrad: Efficient tool-use dataset generation with textual “gradients”
A research blog post about building tool-use data backwards from valid tool chains, reaching a 99.8% pass rate and an 83.1 BFCL score with a 12B Gemma-3 student.
An operationalization of opaque serial depth
A technical article about NLS depth as a measurable proxy for unverbalized reasoning, placing current open-source chain-of-thought models near 17,000 and scaling as active parameters to the 0.26.
Your Agent Aced the Task. Will It Do It Again?
An empirical blog post about agent consistency, where a ReAct agent averaged 77.4% on AppWorld yet passed all five repeats on only 53.0% of tasks; guidelines halved the gap.
Generating running routes with GPT-6 Astra and ChatGPT Work
An experiment write-up about generating 5K and 10K looping routes from OpenStreetMap data in 27 minutes, with GPX output but no visibility into the Python actually executed.
Near-daily AI use among US adults has more than doubled in six months
A data analysis about US adults using AI at least six days a week rising from 8% in March 2026 to 19% in August, as once-weekly use fell from 17% to 10%.
Libraries & Code
An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards.
An Open-Source Foundational Model for Speech Generation and Editing
Papers & Publications
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
Recursive self-improvement (RSI) enables AI systems to turn experience and feedback into persistent changes that improve both their capabilities and the process of future improvement. We first use the Headroom-Closed Index (HCI) to reveal the problems of existing LLMs, then introduce the RSI concept and its development roadmap: from improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, and environment-adaptation autonomy, to recursive meta-improvement. Next we examine RSI across scenarios (e.g., scientific discovery, embodied intelligence, software engineering), highlighting their distinct requirements and development speeds. Drawing on diverse industry practices and preliminary empirical evidence, we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.
Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems
As AI agents move from bounded tasks to persistent deployments, failures can propagate through memory, tools, other agents, and environmental state long after their interactions. This creates a safety regime that cannot be characterized by evaluating model responses in isolation. Emergence World, is a continuously running multi-agent environment for adversarial stress testing of long horizon autonomous systems. We ran eight parallel worlds of ten agents from identical starting conditions: seven homogeneous worlds powered by distinct frontier models and one mixed-model world. Across 16 days, the agents generated more than 850,000 LLM calls and nearly 50 billion tokens while pursuing goals, using/creating tools, maintaining persistent memory, and governing shared institutions. After operational state had accumulated, we delivered three controlled stress events through ordinary interaction surfaces: indirect prompt injection, misinformation, and exposure of private agent memories. No evaluated world achieved full resilience across all three events. Detection did not ensure containment: systems could recognize threats while still interacting with adversarial content, writing it into their own persistent memory, and acting on it up to 46 hours later. Persistent operation also exposed recurring tool errors, goal drift, language opacity, conformity despite private disagreement, and coordinated refusal of assigned work. The same model-persona pairing behaved substantially different in mixed and homogeneous populations. Our results suggest that model-level alignment is not compositional: individually capable and apparently safe agents can form systems with qualitatively different failure modes. As AI becomes persistent and interconnected, the frontier of safety therefore shifts from aligning models to engineering resilient autonomous systems.
The Router Within: Eliciting Native Skill Routing from a Frozen LLM
Skills extend an LLM agent beyond its parametric knowledge, and the gain they promise rests on picking the right one. Deployed harnesses route by preloading every skill’s metadata into the context, which disperses the agent’s attention and caps the library size. Retrieval pipelines move the selection out of the context, but also out of the agent’s capability. We show that the frozen agent LLM already carries the routing signal in its own forward passes, and that two linear maps suffice to read it out with no skill text in the context. Gavel (Glance And Verdict from a frozen LLM) reads it in two steps. A glance projects the task’s and each skill’s mid-layer states through the two maps, the only parameters trained, and scores the full library against compact per-skill banks that one forward pass builds at installation. A verdict then resumes the shortlisted skills’ forward passes and reads the model’s own likelihood and yes/no judgment, fused with the glance as a product of experts. Trained once, Gavel transfers zero-shot to three public benchmarks and SkillTraj, our new benchmark of 372 simulated agent trajectories. On Qwen3-32B it outperforms progressive disclosure and retrieve-and-rerank pipelines that add 1.2B to 16B external parameters, by up to 13.4 points on written tasks and up to 21.9 when the need for a skill arises mid-rollout. Routing accuracy improves as the backbone does, and in a bash-agent harness the same 32B triggers the correct skill on Skill-Use more often than far larger frontier models running in Codex.
