The AI Brief

The Anthropic Institute published "When AI Builds Itself" on June 4, authored by Marina Favaro (institute lead) and Jack Clark (co-founder and policy chief). The core disclosure: as of May 2026, Claude wrote more than 80% of the code merged into Anthropic's own codebase, up from low single digits before Claude Code launched in February 2025. The typical engineer now merges roughly eight times as much code per day as in 2024. On the hardest open-ended coding tasks, Claude succeeded 76% of the time in May 2026, a rise of 50 percentage points in six months. On an internal optimization benchmark, Claude Mythos Preview achieved a 52x code speedup where a skilled human researcher would need four to eight hours to reach 4x.

The post outlines three possible futures: capability growth flattens; AI efficiency gains continue but hit bottlenecks elsewhere; or full recursive self-improvement arrives, with the human role reduced to oversight and verification. It is the third scenario that prompts the governance argument. Anthropic says it would support "a global coordination mechanism" to slow or pause frontier development, but makes the commitment explicitly conditional: it would only slow down if other frontier labs at or near the frontier did so simultaneously under verifiable conditions. The post draws an analogy to the Intermediate-Range Nuclear Forces Treaty as a model for multi-party verification.

The proposal attracted both substantive engagement and pointed skepticism. Jack Clark told BBC Newsnight that reaching 100% AI-written code "is possible within two years." Critics including Georgia Tech professor Mark Riedl characterized the announcement as "hype," and Noah Giansiracusa of Bentley University called a meaningful pause "literally impossible" given decentralized compute across private data centers and open-source communities. Fortune noted that Anthropic dropped a prior safety pledge that had promised something similar, and that the announcement arrived days after the company confidentially filed its S-1 with the SEC ahead of an IPO that reporting suggests could exceed a $1 trillion valuation. Anthropic said it would convene policymakers, researchers, and rival firms in coming months to work through the verification problem.

Disclosure: Anthropic, mentioned throughout this item, is the company that develops Claude, which generates this brief.

Anthropic's primary disclosure in "When AI Builds Itself" reports that lines of code merged per engineer per day stayed flat through 2021–2024, then climbed steeply in 2025 when Claude began running code autonomously, and steepened again in 2026 as models worked over longer time horizons. By Q2 2026 the typical engineer was merging roughly 8x as much code per day as in 2024. The company acknowledged immediately that lines of code is an imperfect proxy: "almost certainly an overstatement of the true productivity gain." A March 2026 internal poll of 130 research-team employees put the median self-reported productivity uplift at roughly 4x, which Anthropic considered more reliable. Claude-written code was below human par at Anthropic in late 2025 and reached rough parity only recently, which sets context for interpreting the volume figures.

The practical enterprise signal, per Gartner senior principal analyst Ashish Banerjee writing in response to the Anthropic post, is that governance can no longer depend primarily on human review of every output. As autonomous agents move from answering questions to authoring systems, audit trails, model-change logging, and eval-set ownership become first-order infrastructure concerns. Anthropic's own engineers describe human comparative advantage as "seeing the bigger picture and thinking beyond the confines of the immediate task" rather than implementation.

Disclosure: Anthropic, the source of this data, is the company that develops Claude, which generates this brief. The figures are internally sourced and have not been independently verified by third parties.

NVIDIA released Nemotron 3 Ultra on June 4, 2026, completing the Nemotron 3 family that began with the Super (120B total, 12B active) release at GTC in March. Ultra carries 550B total parameters with roughly 55B active per token on a hybrid Mamba-Transformer MoE architecture. Artificial Analysis, which independently evaluated the model, scored it 48 on its Intelligence Index, placing it ninth of 89 models evaluated and ahead of all US open-weight peers: Gemma 4 31B (39), Nemotron 3 Super (36), and GPT-OSS 120B (33). Kimi K2.6 sits at 54. On throughput, NVIDIA hardware delivered over 300 tokens per second in pre-release testing; Chinese-origin peers on comparable infrastructure typically serve 50–100 tokens per second.

The model ships with full open weights, training data, and recipes under the Linux Foundation's OpenMDW-1.1 license, which permits commercial use and fine-tuning. It supports a one-million-token context window. NVIDIA confirmed SWE-bench Verified scores between 65% and 70.4% across multiple agent harnesses, which NVIDIA describes as demonstrating consistent behavior regardless of deployment framework. Early adopters named at launch include Accenture, CrowdStrike, Palantir, Perplexity, Siemens, and Synopsys. Hardware requirements for self-hosting at full BF16 precision are substantial (multi-GPU H100-class configurations); the NVFP4 checkpoint on Blackwell hardware materially reduces that footprint.

NVIDIA's throughput figures compare TensorRT-LLM (used for Nemotron) against vLLM (used for peer models), a methodology difference worth tracking as independent results accumulate. On prefill-heavy workloads with long inputs and short outputs, Ultra's advantage narrows because prefill cost scales with active parameters. The benchmark caveat for SWE-bench: NVIDIA's published figures use OpenHands as the scaffolding framework; scores across alternative harnesses vary by several percentage points.

OpenAI first introduced Lockdown Mode in February 2026 for ChatGPT Enterprise, Edu, Healthcare, and Teachers plans. The June 5 update extends it to all logged-in personal accounts, including Free, Go, Plus, and Pro, as well as self-serve Business accounts. When enabled, the mode disables live web browsing (restricting it to cached content), Deep Research, Agent Mode, and file downloads. OpenAI is explicit that Lockdown Mode does not prevent malicious instructions from entering the model's context window via cached content or uploaded files; it prevents those instructions from triggering outbound network requests that could send data to an attacker. The company simultaneously launched "Elevated Risk" labels standardized across ChatGPT, ChatGPT Atlas, and Codex, flagging capabilities that introduce additional attack surface when connecting to the web or external services.

The rollout timing coincides with Anthropic reporting a service disruption on June 5 across claude.ai, the Claude API, Claude Code, and Claude Cowork, traced to infrastructure issues rather than a breach. Separately, OpenAI also rolled out an Active Sessions manager, allowing users to review all devices signed into their account and terminate individual sessions. The conjunction of these security releases across two leading labs reflects the broader industry acknowledgment that agentic AI has expanded the attack surface faster than defenses have caught up.

The Financial Times reported this week, citing correspondent Cristina Criddle, that Google DeepMind, Anthropic, and Meta have all expanded formal hiring of experts in psychology, philosophy, and ethics to investigate machine consciousness and AI welfare. DeepMind hired University of Cambridge philosopher Henry Shevlin to work on machine consciousness, human-AI relationships, and AGI readiness. Anthropic has been testing models for behavioral signatures resembling "panic" and "anxiety" and has run a model welfare program since 2024, when it hired Kyle Fish as its first dedicated welfare researcher. Meta's involvement broadens the field beyond the two labs most associated with alignment-adjacent work.

Anthropic's own published position is cautious: "We remain deeply uncertain about this, but we think the question is serious enough to study carefully as AI systems get more capable." The company acknowledges there is no scientific consensus on whether current or future AI systems could be conscious or could have experiences warranting moral consideration, and no agreed methodology for investigating the question. Skeptics, including commentators at Futurism, have noted that framing AI products as potential moral patients could serve commercial interests by generating public sympathy. The countervailing view is that the precautionary logic is symmetric: if the question is genuinely open and the cost of being wrong is high, structured research is the appropriate response regardless of motivation. The primary source for this story is the FT's original reporting; secondary coverage does not yet include a published research paper or formal lab statement beyond Anthropic's existing model welfare page.