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 maximalist case, argued at book length: that there is no version of building superintelligence under current conditions that humanity survives.
Why read it. Read the strongest form of the argument you disagree with, from the people who hold it most firmly.
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.
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.
Called for a six-month moratorium on training systems more powerful than GPT-4. It did not happen.
Why read it. A resolved prediction about coordination, and a useful calibration on what open letters achieve.
One sentence: mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war.
Why read it. The signature list is the argument. Read who signed and who pointedly did not.
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 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.
Traces alignment from present-day bias and reward hacking through to the long-horizon version of the problem.
Why read it. The best bridge between documented near-term harms and speculative long-term ones.
Puts explicit probabilities on existential risks by source and defends the practice of doing so.
Why read it. It is the reason people now answer the P(doom) question with a number instead of a mood.
Argues the failure is in the standard model itself: systems built to maximise a fixed objective will pursue it past the point anyone intended.
Why read it. The clearest technical statement of why the problem is structural rather than a bug.
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.
The book that moved the control problem from science fiction into philosophy departments and, eventually, into boardrooms.
Why read it. Almost every later risk argument is either a development of this or a reply to it.