Not a catalogue of scary AI stories, and not a reading list for optimists. Both cases are given the same seriousness, because a person deciding what to think needs the strongest version of each.
The compressed-21st-century argument: that powerful AI could deliver 50 to 100 years of biological progress in five to ten.
Why read it. The most specific, most falsifiable version of the abundance case, written by someone who also puts double-digit odds on catastrophe.
An extrapolation-driven argument that AGI by 2027 follows from straight lines on graphs, with national-security consequences.
Why read it. Whether or not the trendlines hold, this document shaped how a lot of capital and policy attention got allocated.
Estimates that around 80 percent of US workers have at least 10 percent of tasks exposed to LLMs.
Why read it. The most-quoted employment number in the debate. Read what exposure actually means before quoting it.
A field study of 5,000 customer-support agents: productivity up about 14 percent, with gains concentrated among novices.
Why read it. Measured effects in a real workplace, not a benchmark.
Proposes evaluating models for dangerous capabilities and for alignment, before deployment rather than after.
Why read it. The technical foundation under most frontier-safety policies now in force.
An early, contested claim that GPT-4 showed general-intelligence characteristics.
Why read it. A case study in the evaluation problem: the disagreement was about what counts, not about what the model did.
A method for training model behaviour against a written set of principles rather than case-by-case human labels.
Why read it. One of the few concrete answers to 'so what would you actually do about it?'
The clearest existing case of AI producing a genuine scientific result rather than a demonstration.
Why read it. When someone asks for evidence that AI accelerates science, this is the strongest single answer.
The foundational statement of the present-harms critique: environmental cost, data provenance, and the illusion of understanding.
Why read it. The most-cited counterweight to existential-risk framing, and it predates the current wave.
Showed that loss falls predictably with compute, data and parameters, which turned capability into a budgeting exercise.
Why read it. The empirical basis for every timeline argument in either direction.
Seven hundred words arguing that general methods leveraging computation beat human-designed structure, every time, eventually.
Why read it. The shortest item in this library and among the most consequential.
The transformer architecture. Everything in the current debate runs on top of this paper.
Why read it. Provenance: the whole argument has a technical origin, and this is it.
Turned abstract worry into five engineering problems: side effects, reward hacking, scalable oversight, safe exploration, distributional shift.
Why read it. Where safety stopped being philosophy and became a research agenda.