AI Coding Stack Shift | AI Field Notes #22

AnalysisA repository called andrej-karpathy-skills on GitHub distils Karpathy's January 26, 2026 X post on LLM coding failures into a single CLAUDE.md file (the project-level configuration Claude Code reads at session start) that drops into any project. Built by Forrest Chang, the file organises four principles: Think Before Coding, Simplicity First, Surgical Changes, and Goal-Driven Execution. Each addresses a failure mode Karpathy named: silent wrong assumptions, unnecessary complexity, sprawling diffs, and unverifiable success criteria. The repo has crossed 120,000 stars, an outlier number for a single-file project, suggesting these failure modes are widespread enough to spook working engineers.

AnalysisOpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as the ChatGPT default and as 'chat-latest' in the API, starting May 5. On the AIME 2025 math benchmark, the new model scores 81.2 versus 65.4 for its predecessor; multimodal reasoning on MMMU-Pro improved from 69.2 to 76. The model also gains context memory, pulling from past conversations, linked files, and Gmail to personalize responses. GPT-5.3 stays accessible to paid API users for three months, then goes offline. Developers who pinned their integrations to 'chat-latest' already absorbed the change; those on a hard model ID have until roughly August before OpenAI forces the cutover.

AnalysisOn May 6, Anthropic signed an agreement with SpaceX to access the full compute capacity of Colossus 1, a supercluster in Memphis, adding 300 megawatts and over 220,000 NVIDIA GPUs to Anthropic's available resources within a month. The company used the same announcement to double Claude Code's five-hour rate limits across Pro, Max, Team, and Enterprise plans and remove peak-hour throttling for Pro and Max users. API rate limits for Opus-class models also increased. The sequence was deliberate: more hardware unlocked immediately higher throughput for paying customers, with no lag between the infrastructure deal and the limit change.

AnalysisAmazon Web Services launched Bedrock AgentCore Payments in preview on May 7, letting AI agents execute stablecoin micropayments independently, without a human approving each transaction. The service uses the x402 protocol, a standard for automated machine-to-machine payments, and integrates with Coinbase's CDP wallet infrastructure and Stripe. Coinbase's companion x402 Bazaar MCP server (Model Context Protocol, a standard for connecting agents to external tools) lists over 10,000 API endpoints agents can pay to access. Session-level spending limits are enforced at the AWS infrastructure layer and logged through standard observability tools, giving compliance teams an audit trail without blocking autonomous agent operation.

AnalysisCursor shipped version 3.3 on May 7 with three changes targeted at multi-step AI-assisted development. PR review now runs inside the agents window, with inline comments and full commit history visible without leaving the IDE. Build in Parallel runs independent plan segments simultaneously rather than sequentially, cutting wall-clock time on large refactors. Split PRs divides a large change into smaller logical chunks with backup snapshots before each split, reducing the diff size reviewers see and making rollback cleaner. Together, the features reduce the overhead of handing AI-generated changes to a human review team.