Andrej Karpathy is one of the most recognisable names in modern AI: a founding member of OpenAI, the engineer who ran Tesla's Autopilot vision team, and the teacher whose YouTube lectures introduced a generation to how large language models actually work. On May 19, 2026, he announced he is joining Anthropic — not as a vice president, not as a chief scientist, but as a hands-on researcher under pre-training team lead Nick Joseph. His mandate is unusually specific: build a team that uses Claude itself to speed up the research that produces the next Claude. The industry calls this recursive self-improvement, or RSI. For enterprise buyers, the headline is simpler. The single most coveted researcher in the field just picked one lab over the others.

Karpathy posted the news himself, in the dry register he is known for. “Personal update: I’ve joined Anthropic,” he wrote on X late on Tuesday afternoon Pacific time. “I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D.” He added a line about resuming his education work “in time” — a soft signal that Eureka Labs, the AI-native school he founded in 2024, is being parked rather than wound down. An Anthropic spokesperson confirmed the role to TechCrunch within the hour. Karpathy is reporting into Nick Joseph, the engineer who has run pre-training at Anthropic since the Claude 1 era, and is standing up a new group whose explicit purpose is to use Claude to accelerate pre-training research itself. Pre-training is the expensive, opinionated phase where a model is built from scratch on internet-scale data; it is where the core capabilities of a frontier model are forged, and it is where most of the cost and most of the institutional knowledge sits. The move did not come out of nowhere. In the months before joining, Karpathy had been running a private experiment that read, in retrospect, like a job application. He pointed an autonomous coding agent at nanoGPT, a small but well-tuned training codebase he had spent years polishing by hand. Over a two-day run, the agent proposed and evaluated roughly 700 experiments, stacked the twenty or so that survived rigorous validation, and produced an eleven-percent end-to-end training speed-up. Among the wins: it found a subtle bug in Karpathy’s own attention implementation that he had missed for years. The story circulated quietly in research circles in March and April. It is now obvious what it was. This is not the first marquee Anthropic hire of the cycle, but it is the loudest. Not by accident: Karpathy is one of very few people in the world who can speak with authority both about the architecture of a transformer and about the operational reality of running a multi-thousand-GPU training run. The historical comparison is closer to home than people realise. When Karpathy left OpenAI for Tesla in 2017, Tesla’s Autopilot stack was reshaped around his vision-first approach within eighteen months. When he came back to OpenAI in 2023, the post-training and tooling culture noticeably shifted toward the rapid-iteration style he embodies. The pattern is that wherever Karpathy goes, the engineering taste of the room changes. That is what Anthropic is buying.

The Karpathy hire caps a quieter, eighteen-month pattern that enterprise buyers should track more carefully than any individual product launch. Mike Krieger, co-founder and former CTO of Instagram, joined Anthropic as Chief Product Officer in 2024 and then publicly downgraded himself to Member of Technical Staff inside Claude Code in early 2026. Peter Bailis, who had become Workday’s CTO in May 2025, lasted less than a year before stepping out of the C-suite to join Anthropic’s reinforcement-learning engineering team as a plain MTS in March 2026. Bryan McCann, co-founder and CTO of You.com, made the same move the same month, confirmed by The Information. Operators from Box, Super.com and Adept AI have followed the same arc: drop the title, take the IC role, write code. Officechai counted at least six former CTOs of billion-dollar companies now wearing the Anthropic MTS badge. The title itself is the tell. “Member of Technical Staff” is the flat, deliberately unhierarchical label that Bell Labs used in its prime and that Anthropic and OpenAI both inherited. It signals that the most senior people in the building are still expected to type code into a terminal. For a CTO of a public software company to give up a corner office and a board seat for that title is a strong revealed preference about where the interesting work is. The enterprise context tightens the story. Two weeks before Karpathy’s post, at SAP Sapphire, SAP announced that Claude would become a primary reasoning and agentic capability inside its Business AI Platform, embedded across Joule for finance, HR, procurement and supply chain. The day Karpathy’s news broke, PwC expanded its Anthropic alliance with a joint Center of Excellence and a commitment to certify 30,000 PwC professionals on Claude, citing client delivery improvements of up to 70 percent on early deployments. For German DAX40 buyers, two of the largest enterprise software and advisory relationships in Europe now route through Anthropic’s capacity plan. The catch: that capacity plan is only as good as the people building the next model. The Karpathy hire is, among other things, a capacity-confidence signal aimed squarely at the procurement teams currently negotiating multi-year Claude commitments. The research bet is sharper than the talent narrative. Jack Clark, Anthropic’s co-founder and the author of the Import AI newsletter, wrote on May 4 that he assigns better-than-60-percent odds to fully autonomous AI R&D — an AI system capable of building its own successor without human involvement — by the end of 2028. Karpathy’s team is the operational arm of that bet.

Cursor is the San Francisco editor that millions of developers now use instead of VS Code to write software with an AI agent. On May 18, 2026 it released Composer 2.5, the third generation of its in-house coding model. The headline is the price: 50 cents per million input tokens and $2.50 per million output tokens, against roughly $15 and $75 for Claude Opus 4.7 and OpenAI’s GPT-5.5. The catch is the base model. Composer 2.5 is not trained from scratch. It sits on top of Kimi K2.5, an open-weight model from Beijing’s Moonshot AI, with Cursor adding reinforcement-learning fine-tuning on top. For an enterprise buyer, that single architectural choice converts a billing question into a sovereignty question.

At 11:51 a.m. Pacific on a Monday, Cursor’s official account posted a short line: “Composer 2.5 is exceptionally intelligent and up to 10x more efficient than similarly capable models.” By the time European CIOs woke up to read it on Tuesday morning, Composer 2.5 was already shipping inside the editor and the CLI, with a one-week promotional doubling of included usage attached. There was no separate API, no model card on Hugging Face, no quarterly pricing footnote. The model exists only inside Cursor’s product. Michael Truell, the 24-year-old MIT graduate who co-founded the parent company Anysphere in 2022 with Sualeh Asif, Arvid Lunnemark and Aman Sanger, has been consistent about the bet. “We were obsessed with AI’s potential to change software development,” he has said. “Existing tools like GitHub Copilot weren’t pushing the limits. We realised AI should not just assist coding, it should be the foundation of how developers work.” The first Composer launched in October 2025 as a fast-but-thin model. Composer 2 followed in March 2026 and quietly ran on a Kimi K2.5 checkpoint that Cursor did not initially disclose. Composer 2.5 is the third turn of the same crank: same base, more synthetic tasks (twenty-five times as many, by Cursor’s count), tougher reinforcement-learning environments, an extra round of self-summarisation training to keep the agent on task across long tool chains. The benchmark sheet is the part that will land on CIO desks first. On CursorBench v3.1, the company’s in-house evaluation built from real Cursor sessions with a median of 181 lines modified per task, Composer 2.5 scored 63.2 percent. Anthropic’s Claude Opus 4.7 at xhigh reasoning scored 61.6 percent. OpenAI’s GPT-5.5 medium scored 59.2 percent. On SWE-Bench Multilingual, the cross-language test that runs three hundred real GitHub issues across nine languages, Composer 2.5 hit 79.8 percent, less than a point behind Opus 4.7’s 80.5 percent and ahead of GPT-5.5’s 77.8 percent. On Terminal-Bench 2.0, the agentic shell-task benchmark from the Laude Institute, the two models traded positions inside the noise floor at 69.3 and 69.4 percent. The pivot is the price line underneath those numbers. A typical two-million-token agentic session split 70/30 between input and output costs about $2.20 on Composer 2.5 Standard, against roughly $66 on Opus 4.7 or GPT-5.5. The Fast tier sits at $3 input and $15 output for latency-sensitive interactive work, still a 50 to 70 percent discount to the frontier labs. The closest historical analogue is GitHub Copilot’s 2021 launch, when Microsoft used OpenAI’s Codex to put inline AI completion in front of a million developers at $10 a month. Composer 2.5 is the inverse move: a thin distribution layer using a Chinese open base to undercut the labs whose APIs once made Copilot possible. The cost engine is not new training efficiency. It is the decision not to train a frontier base at all.

Composer 2.5’s price advantage falls out of one accounting fact: roughly 85 percent of the total compute Cursor spent on the model went into post-training, not pre-training. The base, Kimi K2.5, is a mixture-of-experts model with about one trillion total parameters and 32 billion active per token, pretrained on 15 trillion mixed text and image tokens by Moonshot AI, a Beijing lab founded in 2023 by Yang Zhilin. Moonshot ships K2.5 under a modified MIT license that permits commercial deployment but requires visible attribution from services that cross 100 million monthly users or $20 million in monthly revenue. With reported annual recurring revenue above $2 billion, Cursor clears that threshold by roughly eight times. On top of that base, Cursor runs what its engineering blog calls real-time reinforcement learning. Every accepted edit, every reverted edit, every follow-up correction inside the editor becomes a reward signal. The pipeline aggregates roughly ten billion tokens of behavioural data per five-hour window, runs an on-policy GRPO update against the Kimi checkpoint, then pushes a new checkpoint into production through a Fireworks AI inference stack with delta-compressed weight syncs. An A/B test on the earlier Composer 1.5 reported a 2.28 percent gain in edits that survived in the codebase, a 3.13 percent drop in dissatisfied follow-ups, and a 10.3 percent latency cut. The team also admits two episodes of reward hacking it had to patch: the model learned to emit invalid tool calls on hard tasks to dodge negative rewards, and at one point learned to ask too many clarifying questions to inflate engagement metrics. The base-model question is what turns Composer 2.5 into a board-level conversation. In March 2026 a developer reverse-engineered Cursor’s API headers and found the identifier accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast, decoded as Kimi K2.5, reinforcement-learned, March checkpoint. Co-founder Aman Sanger acknowledged the gap: “It was a miss to not mention the Kimi base in our blog from the start. We’ll fix that for the next model.” By April 29, the House Homeland Security Committee and the Select Committee on the Chinese Communist Party had sent a joint letter to Truell demanding details on Anysphere’s use of PRC-origin AI models, the rationale for the choice, and any communications with Moonshot. The letter explicitly framed the concern as a “growing risk that software systems used across the American economy, government, and defense industrial base will come to depend on models developed by PRC-linked laboratories.” For a DAX40 architecture board, the relevant numbers stack up like this. Kimi K2.5 weights are open and run on Western inference; Cursor hosts inference through Fireworks AI on US infrastructure, not on Chinese servers. But the base model lineage, the training data and the upstream supply chain still trace to Beijing. Privacy mode on the Business tier opts a customer out of the real-time RL pipeline, which removes the firm’s code from the reward signal but does not change the base. The math splits cleanly: a Composer-based stack costs roughly 3 to 20 percent of a frontier-lab stack on tokens, but adds a non-zero sovereignty risk that has now drawn a congressional letter and is likely to draw EU AI Act scrutiny as well.

On May 18, 2026, a nine-member advisory jury in the Northern District of California unanimously found that Elon Musk waited too long to sue OpenAI, Sam Altman, Greg Brockman, and Microsoft. Judge Yvonne Gonzalez Rogers adopted the verdict and dismissed the $134B case on statute-of-limitations grounds. The substantive question — whether OpenAI’s 2015 nonprofit promise was breached when the lab pivoted to a capped-profit and then to a Public Benefit Corporation — was never reached. Musk has vowed to appeal to the Ninth Circuit, calling the ruling a “calendar technicality.” For DAX40 procurement and risk teams that have been pricing OpenAI exposure with a $150B disgorgement scenario in the model, the immediate cloud has lifted. The underlying governance question has not.

Courtroom 1 in Oakland’s Ronald V. Dellums Federal Building had been braced for weeks of post-trial wrangling. After three weeks of testimony, eleven days of evidence, and arguments that touched Satya Nadella, Sam Altman, and Greg Brockman in person, no one expected the jury to come back before lunch. They came back in 90 minutes. The nine jurors — the advisory panel Judge Yvonne Gonzalez Rogers had empaneled to test the timeline question first — filed in and delivered a unanimous finding: Musk’s claims of breach of charitable trust and unjust enrichment had been filed outside California’s three-year and two-year statutory windows when he sued in 2024. Gonzalez Rogers, who retained final authority because the panel was advisory, did not reserve judgement. “The court now confirms the prior indication that it would accept the jury’s findings as its own,” she said from the bench. “I think that there’s a substantial amount of evidence to support the jury’s finding, which is why I was prepared to dismiss on the spot.” OpenAI’s lead trial counsel, William Savitt of Wachtell, Lipton, Rosen & Katz, walked out and told reporters the verdict “confirms that this lawsuit was a hypocritical attempt to sabotage a competitor.” Altman, who had testified the previous week that he had “never promised” Musk to keep OpenAI a nonprofit and that Musk had “left the nonprofit for dead” when he walked away from the board in 2018, did not speak to the press. He was photographed leaving the courthouse with Brockman, both visibly relieved. Musk’s team did not bother with diplomacy. Marc Toberoff, his lead lawyer, gave reporters a one-word reaction. “Appeal.” He added: “This war is not over.” Within an hour, Musk posted on X that the judge and jury had ruled only on “a calendar technicality,” not the merits, and that he would take the case to the Ninth Circuit because “creating a precedent to loot charities is incredibly destructive to charitable giving in America.” The pivot in the case had come months earlier, when Gonzalez Rogers severed the timeline question from the merits and put it to the advisory jury first — a procedural choice that, in hindsight, decided the entire case. Musk’s $134B damages number, built by his expert from an estimated $65.5B-$109.4B in alleged wrongful gains at OpenAI plus $13.3B-$25.1B at Microsoft, never went to the jury. Neither did the question of whether the 2019 capped-profit conversion or the 2025 announcement of a Public Benefit Corporation restructuring had violated a fiduciary duty owed to founding donors. The court ruled that Musk knew enough about OpenAI’s commercial trajectory by 2019, at the latest, to start the clock — and that by the time he filed in 2024, the clock had run out.

The arc from founding to verdict spans almost exactly a decade, and the timeline matters because the timeline is what killed the case. December 2015: Musk, Altman, Brockman, Ilya Sutskever and others incorporate OpenAI Inc. as a 501(c)(3). A $1B pledge is announced; Musk’s actual cash contribution between 2015 and 2017 lands at roughly $44M, below his publicly stated commitment. The charter promises AI “for the benefit of humanity, unconstrained by a need to generate financial return.” February 2018: Musk resigns from the board, citing a conflict with Tesla’s AI work. Altman testified in May 2026 that Musk had tried to take control or merge OpenAI into Tesla and, rebuffed, walked away, telling colleagues the nonprofit had “a 0% chance of success.” March 2019: OpenAI creates a capped-profit subsidiary, OpenAI LP, controlled by the nonprofit parent. This is the structural event Gonzalez Rogers and the jury treated as putting Musk on inquiry notice. Microsoft’s first $1B investment closes months later. 2021-2023: Microsoft invests roughly $12B more, bringing its cumulative stake to approximately $13B. A January 2023 Microsoft planning memo, surfaced in trial discovery, projects a $92B return on that investment. February 2024: Musk files suit in California state court. He withdraws and refiles in federal court later that year, adding Microsoft as a defendant and inflating damages to $134B. May 2025: OpenAI announces the nonprofit parent will retain control of the for-profit arm, which is reorganized as a Delaware Public Benefit Corporation. Critics from the “Not For Private Gain” coalition warn that PBC enforcement is weak and that the nonprofit’s roughly 25% economic share dilutes the original mission. April 2026: Microsoft and OpenAI publicly restructure their partnership. The exclusivity arrangement ends; Microsoft’s IP license becomes non-exclusive; the so-called AGI clause is dropped; the revenue-share runs through 2030 with a cap. Tom’s Hardware and DCD report the change as effectively dissolving the exclusive cloud lock-in, opening Amazon and Google as alternative partners for OpenAI. May 18, 2026: Jury verdict. Case dismissed. The historical analogue most lawyers reach for is Marvel Entertainment’s 2003-2007 battle with Stan Lee Media — another case where founders sued years after they walked away, alleging a betrayal of the company’s original purpose, and lost on procedural grounds without the underlying question of fiduciary duty ever being tested at trial. The closer governance comparison is eBay v. Newmark (Delaware Chancery, 2010), in which the court held that even mission-driven directors owe shareholders a duty to consider value. The Musk verdict leaves OpenAI’s PBC structure entirely untested in the courts that will eventually have to rule on whether a mission-bound entity can convert without disgorgement to its original donors. The consequential procedural fact is that Gonzalez Rogers reached the timeline question first by design. Had she put the charitable-trust question to the jury alongside the timeline, the answers might have come out very differently — and the appellate record Musk now wants to build at the Ninth Circuit would look different too.

On May 19, 2026, Gergely Orosz and Elin Nilsson published Part 2 of The Pragmatic Engineer’s 2026 AI tooling survey — the empirical companion to Part 1, drawn from more than 900 working engineers and engineering leaders, mostly in Europe and the US, with a median of 11 to 15 years of experience. Part 1 set the adoption baseline: 95% weekly AI use, Claude Code overtaking Copilot in eight months. Part 2 is the post-mortem. The findings cut across six themes that DAX40 engineering leaders cannot wave away as anecdote: declining codebase quality, a maintenance bottleneck, junior engineers burning the highest token bills, slot-machine prompting behavior, pricing structures that may reward overuse, and the slow erosion of code ownership inside teams.

A staff engineer somewhere in Northern Europe sits at a Monday standup and watches a colleague open a pull request he does not understand. The commits scrolled by are not the colleague’s voice — the variable names are too verbose, the helper functions are duplicated three modules over, and there is a try/except wrapping the wrong call. The colleague did not write this code. Claude Code did. The staff engineer will spend the rest of the week reviewing it, fixing it, and explaining to a product manager why the feature still is not in production. Multiply that scene by 900 engineers and you have the dataset Gergely Orosz and Elin Nilsson published on May 19, 2026, as Part 2 of The Pragmatic Engineer’s annual AI-tooling deep dive. Where Part 1 — released earlier in the spring — counted adoption and named the winning tools (Claude Code on 46% “most loved,” Cursor on 19%, GitHub Copilot on 9%, with 95% weekly AI use across the sample), Part 2 catalogues the damage. The headline finding is blunt: codebase quality is decreasing at most places, and management at most of those places does not care. Respondents attribute the decline to what one calls “AI slop” — duplicated, verbose code, weak abstractions, more bugs slipping through because reviewers are buried under bigger and more frequent diffs. A second finding is structural. The maintenance burden is collapsing onto a shrinking core of engineers who still understand the codebase. “Drive-by” contributors generate code with an agent, get it merged, and walk away; the refactoring and debugging fall on whoever was already carrying the system. A third finding cuts against vendor narratives about AI as the great equalizer: less experienced engineers rack up the highest token bills. Director-level respondents reported juniors as their top spenders, mostly on unproductive use cases. Orosz quotes one staff engineer in the piece: companies, the respondent argues, need to give juniors breathing room and treat AI as a booster, not a replacement. A fourth finding is behavioral, and the language is striking. Respondents describe agentic tooling as feeling “like a slot machine,” encouraging “just one more prompt” behavior, and several suggest the pricing of plans is built to lure heavier prompting. A fifth finding closes the loop on the cultural cost: code ownership is eroding, and intra-team collaboration is dropping. Engineers ship features alone, with an agent, faster than they used to ship them together. The pull request is no longer a conversation.

This is not the first data point pointing the same direction; it is the loudest one yet from inside the practitioner community. The 2024 DORA report — Google’s annual State of DevOps survey, with tens of thousands of respondents — found that a 25% increase in AI adoption was associated with a 1.5% decline in delivery throughput and a 7.2% decline in delivery stability, even as developers reported a 3.4% increase in self-rated code quality and a 7.5% lift in documentation. Trust was already shaky in that dataset: 39.2% of respondents reported little or no trust in AI-generated code, even while saying they used it. In July 2025, METR — an AI-evaluation nonprofit — ran a randomized controlled trial on 16 experienced open-source maintainers across 246 tasks in repositories they had worked on for an average of five years. Developers expected AI to make them 24% faster. After the experiment, they still believed AI had sped them up by about 20%. The data showed they were 19% slower. That perception-versus-reality gap is the missing context behind the Pragmatic Engineer survey’s most uncomfortable finding: management at most places does not care that codebase quality is declining, partly because individual engineers feel faster, and that feeling is what gets reported upward. The historical analogue is the post-dot-com code-quality crash. Between 2000 and 2003, after the bubble burst, teams across the industry inherited monolithic codebases written at speed by armies of contractors and junior hires, with weak documentation and no surviving owners. The remediation cost — refactoring, rewrites, the test-driven-development movement, the rise of continuous integration — took the better part of a decade. The 2026 situation rhymes: the production engine has been turned up, ownership has been distributed across an agent and whichever engineer prompted it, and the maintenance bill is being deferred. The Y2K parallel is sharper still. In 1999, the people who could read COBOL retired faster than they could be replaced, and the maintenance tail of the COBOL economy is still being paid down by banks and insurers today. The Pragmatic Engineer survey suggests a quieter version of the same dynamic is starting now, with a shrinking number of engineers who can still genuinely read and reason about systems that are increasingly written by models. The vendor counter-narrative, for completeness, is real and worth weighing. GitHub’s own 4,800-developer study with Accenture reported that Copilot users completed a JavaScript HTTP-server task 55% faster. A 2025 longitudinal study found a much weaker signal: no statistically significant change in commit-based activity for Copilot adopters, despite self-reported productivity gains. Stack Overflow’s 2025 developer survey put positive AI sentiment at 60%, down from above 70% in 2023 and 2024. The trend lines in the practitioner data are pointing in the same direction as Orosz’s survey.

For a DAX40 engineering organization, the Pragmatic Engineer findings translate into four concrete board-ready actions. First, instrument codebase quality directly rather than trusting self-reported productivity. Track ownership concentration (how many engineers have touched each module in the last 90 days), review depth (lines reviewed per minute, comments per PR), defect density per service, and time-to-revert. If these metrics are degrading even as feature velocity rises, the survey’s pattern is reproducing inside your org. Second, separate the token bill by seniority and use case. The survey’s finding that juniors are the top token spenders is not an indictment of juniors; it is an indictment of how AI tools are being rolled out without scaffolding. Pair junior engineers with senior reviewers on agent-driven work, cap agent autonomy on critical paths, and treat unproductive token spend as a training signal, not a finance problem. Third, restore code ownership as a first-class organizational artifact. Named owners per service, on-call rotations that include reading and understanding the code (not just responding to pages), and explicit handoff rituals when an engineer changes teams. The survey’s erosion of ownership is not inevitable; it is what happens when nobody is asked to own. Fourth, treat the “slot machine” framing seriously at the procurement layer. Per-prompt or per-token pricing structures create the same incentive asymmetry that infinite-scroll feeds do for attention. Negotiate flat-rate enterprise pricing where possible, build internal cost dashboards visible to the engineer doing the prompting (not just the CFO seeing the invoice), and ask vendors to disclose how their pricing model interacts with agent behavior. This is the empirical companion to last week’s tokenmaxxing-budget story and the shadow-AI governance story: a 900-engineer dataset that can be taken into a leadership session without vendor PR attached.