Thursday 4 June 2026

Uber has capped per-developer AI tool spend at $1,500/month to control costs, surfacing a broader debate about unpredictable and unsustainable pricing for agentic coding tools like Claude Code. The discussion highlights that enterprise AI tooling costs are opaque and difficult to budget for at scale.

Discussion exploring the nature of LLMs as "made out of weights" — probing conceptual gaps around what AI models actually are and what that means for how developers reason about, trust, and build on top of them.

Ted Chiang argues AI is not conscious, prompting a 636-comment debate about the nature of AI systems. The underlying friction is the widespread conceptual confusion about AI capabilities that shapes how developers and organizations over- or under-rely on these tools.

Codeium's Cascade AI assistant in Windsurf IDE suffers persistent RPC connection resets to the backend server, causing code edits to be abandoned mid-operation and leaving the codebase in a broken intermediate state. Multiple failures occur within a single hour with no resume capability.

Codeium's chat feature on WebStorm with the pro model repeatedly fails with "insufficient space for system prompt and completion tokens," silently clearing chat history rather than giving the user an actionable error or graceful degradation. The token budget management is opaque and causes total chat failure.

LangGraph's PostgreSQL checkpointer `setup()` method throws an `InterfaceError: row must be included between 0 and 0` on first-time migration, making the documented getting-started flow broken out of the box for new users.

LangGraph's `AsyncPostgresSaver` crashes with `operator does not exist: text = bytea` when querying an empty `checkpoint_ns` under psycopg 3 binary mode, because the library fails to register the correct string type dumper. The reporter provides a workaround of manually registering `StrBinaryDumper` on the connection after construction.

Opinion piece on guiding principles for software development, touching on the tension between building for maintainability/correctness versus shipping fast — a friction point increasingly salient as AI-generated code trades long-term quality for short-term velocity.