Deconstructing The Grift
Google's real-time AI-generated 1993 DOOM remake: every frame produced by a neural network on the fly. Either the most impressive tech demo of the decade, or proof that superintelligence will spend its first years clearing demons in a hallway.
Records: 15 | Sort: Date Desc | Status: Active
AI doomerism: a philosophy of measured risk that got gentrified into an identity by people who desperately needed a new religion. Here's what it actually is, where it came from, and why the loudest version is almost entirely wrong.
How AI regulation can become a moat, why ‘safety’ can turn into an incumbent advantage, and what to watch for if you don’t want policy to be a product feature.
A practical anti-doomscrolling protocol for staying informed about AI without letting the discourse hijack your nervous system. Being well-informed and being terminally online about AI doom are not the same thing, and your cortisol knows the difference.
A guide to reading doom probabilities as the summary statistic they are, not the identity they're trying to become. A number typed with confidence is not the same as evidence — here's how to tell the difference without catching the anxiety.
Benchmarks are useful. They’re also easy to game, easy to over-interpret, and terrible at predicting ‘works in production.’ Here’s how to think in evals instead of vibes.
A timeline reality check: what’s actually being predicted, what the data says, and why ‘2027’ is usually a vibes-based calendar invite. Every few months someone rediscovers the year like a prophecy — here's how to read the claim without catching the certainty.
Alignment isn’t a single algorithm. It’s a set of failure modes, measurement problems, and governance constraints pretending to be a philosophy debate.
RLHF makes models nicer and more useful. It does not magically turn ‘optimize the loss’ into ‘share your values.’ Here’s what it fixes, and what it doesn’t.
Recursive self-improvement is a hypothesis with prerequisites, not a magic word you say to skip the hard parts of forecasting. Let's talk about what it actually requires — and why those prerequisites are doing a lot of heavy lifting that nobody wants to measure.
Instrumental convergence, goal misgeneralization, and why 'it'll turn us into paperclips' is a metaphor — not a screenplay you should bet your nervous system on. The thought experiment is useful; treating it like a mechanism map is where the discourse goes off the rails.
A field guide to the incentives that turn ‘maybe risky’ into ‘extinction on par with nukes,’ and why fear sells—even when the mechanism is missing.
A mechanism-first way to sort AI-doom claims from ‘maybe’ to ‘touch grass,’ and why p(doom) often behaves like astrology with better graphs.
Zador and LeCun wrote the paper your doomer friend will never read: AI doesn't have survival instincts because evolution didn't build it that way. The Terminator scenario requires a biological premise that silicon structurally lacks.
The definitive, exhaustive field guide to the ten prophets, patrons, and panic-mongers convinced that linear algebra is going to eat the human race.
Alignment 'risk' is a research field in search of a real problem. The counterfactual exists: uncensored models and anti-aligned models are all over Hugging Face, and Earth is still here. Here's what the discourse gets wrong.
This site is dedicated to one impolite premise: the world is not ending because another founder discovered caps lock. Humanity can build useful systems and still laugh at recycled apocalypse sales decks.
We track how fear gets productized into books, consulting, and speaking tours, then we publish receipts. Licensed under Creative Commons Attribution-ShareAlike 4.0 International.