How AI labs view knowledge work.
Anthropic's recently published a case study in how they use Claude to develop internal research software. The focus of the article is on firstly, the astonishing volume of code produced; and secondly, how humans are used in the production of aligned code.
This is widely viewed as the future of programming.
AI companies are no stranger to the use of early modern production practices. After employment in the American entertaining arts sector collapsed, many creatives found work classifying training data. This work followed in the mold of Amazon's Mechanical Turk service; an intermediary employer of record operated a software platform which served as the only point of contact for the near-destitute creative.
These platforms host an interface in which “work packages” are presented, wherein you must, quickly, provide some simultaneously highly accurate and cookie-cutter evaluation of context-free snippets of content. Alongside, chat interfaces provide an irregular stream of work, encouragement, and threats. Your activity is monitored, there is not enough work to pay the bills, and only with the best evaluations (rapidly accepting work, perfect performance on the subjective tasks, etc) will your opportunity to bid for future work contracts be maintained.
Obviously, this is the future of programming.
Anthropic's recently published a case study in how a nurse's role may be transformed by AI. This appears to be by interdicting the relationship from the nurse to patient; with AI tools handling intake, diagnoses, assignment and monitoring on treatment pathways; and the nurse serving to explain the decisions (of which they had no way to develop culpability), and review the now much larger stream of decisions for accuracy. A review they cannot possibly make without the context denied by the structure of the work.
Obviously, this is the future of programming.
A executive suite can set strategic priorities; below this, the mill.
Task instructions are generated from any and all input desiderata. The owners of the system are the only ones permitted to directly input tasks, all others are generated by the system itself. This system is deployed and proven, at least to its ability to delight its operators, by OpenAI and Simon Willison’s “Gas Town”.
In Gas Town, AI's reduce tasks to step-by-step processes. AI's execute the processes, cross-checked by validation AI's. System maintenance and revision is performed by inserting further tasks into the queue. Everything is ordered, costed, and traceable.
All that's missing is taste. The human element. By inserting human reviewers into the review process, stochastically at the decisions most likely to be rolled back, culpability can be restored. Each worker can, over time, gain a accurate estimator of how well the changes they allowed produced profit (or otherwise), how much time was required to perform reviews; and how reliably the worker could provide the review. If a change is implicated in harm, every culpable worker can be automatically evaluated.
The programmer cannot be automated, but the programmer can be placed into the skin of an automaton and graded on their ability to perform.
Obviously, this is the future of programming.