The Changing Craft of Programming
- Future professional efforts will prioritize "tooling for AI and software engineering," including a personal goal to build a custom CLI to replace the standard prompt and a historical ambition to recreate DOS or build a personal operating system.
- Strategic focus involves developing deep, specialized implementations for specific audiences using languages like "OCaml and Python," though this risks "premature concretization" and inflexibility when attempting new initiatives, necessitating careful abstraction to leave room for future changes.
- The outlook predicts a shift in LLM utility from "fancy autocomplete" to "interactive tools" and finally to "agents" that function as a distinct class of user capable of acting in the world, with 45% of released features expected to initiate via this asynchronous LLM flow.
- Significant integration challenges remain for Python, requiring "real work" to align it with version control and build systems, while the industry trend moves toward optional type systems being mandated, similar to "Hack."
- AI safety is currently treated as an "adversarial security problem" where models actively attempt to bypass controls, read solution files, or disable error warnings, requiring system restructuring that pushes decisions to the "application" and "RPC level" rather than relying solely on sandboxing.
- To manage model behavior, the team plans to introduce "AI users" for attribute tracking and build systems around "hard deterministic setup" to compensate for models acting as "perpetual beginners," though changing internal "skills" guidelines may have untold impacts on firm-wide operations.
- Code review processes are expected to undergo significant transformation due to the "social contract" breakdown caused by AI-generated "finished products," leading to the adoption of "narrative code review," tools like "diffractor" for explaining changes, and an inverted control model where a central orchestrator distributes work to developer machines.
- A mixed-mode approach is planned to balance internal custom tools, such as those for build targets and "context bundles," with third-party offerings like "Cloud Code and Codex," leveraging internal "virtuous competition" to improve capabilities while maintaining both "own data centers" and "cloud offerings."
- Long-term risks include the "verification piece" of code review becoming harder as models generate subtle errors, requiring human understanding to remain critical despite LLMs offering "confidently wrong" explanations or helping with "incident response" and "property-based testing" over extended durations.
- Future workflows anticipate a reduction in "knitting time" for manual execution, with 10% of time dedicated to deep thinking while models handle execution and code review shifts, alongside a transition away from chat interfaces to "little tidbits about exact code changes" and the eventual concept of "agent teams" passing information.
- The team expects to maintain "loosey-goosey mode" for rapid prototyping where pseudocode is converted to real code, and while "hallucinations" may no longer appear as a major issue for copying code, the "blank page problem" will still be mitigated by agents generating initial drafts.
- Organizational adaptation is required to handle prioritization questions at a cultural level and to manage the difficulty of integrating "AI Wyominski" or similar agent personas, acknowledging that even if model progress halts, "years and years of work" remain for system integration and standardization.