AItopiaOrAImageddon?

-3leaning AImageddon · 18 Aug, midnight

A seven-hour Monday evening with no capability, no incident and no policy instrument: Anthropic told investors its run rate reached $65bn, the first demand-side number of the same order as the buildout it is meant to pay for but a briefing rather than a filing, while a bankruptcy court recorded Google buying Spirit Airlines' internal corpus - 100m emails and employee records back to 1986 - for $10m to train models. Full reading

AItopia — the utopian pole of the scale
-10-5+5+1001920: 1 canon entry1931: 1 canon entry1936: 1 canon entry1942: 1 canon entry1950: 2 canon entries1955: 1 canon entry1956: 3 canon entries1958: 2 canon entries1959: 2 canon entries1960: 1 canon entry1965: 1 canon entry1966: 1 canon entry1968: 1 canon entry1969: 1 canon entry1970: 1 canon entry1973: 1 canon entry1975: 2 canon entries1976: 2 canon entries1979: 1 canon entry1984: 2 canon entries1986: 1 canon entry1987: 2 canon entries1988: 1 canon entry1990: 1 canon entry1996: 1 canon entry1997: 1 canon entry2005: 2 canon entries2009: 1 canon entry2011: 2 canon entries2012: 1 canon entry2013: 2 canon entries2014: 2 canon entries2016: 2 canon entries2017: 3 canon entries2019: 2 canon entries2020: 2 canon entries2022: 1 canon entry2023: 2 canon entries2024: 1 canon entry1900: +0 — The origin, and the one line here that is close to definitional rather than a judgment: no machine had yet been asked to do anything a mind does, so the scale reads neither way.19001936: +0 — Turing sets out what a machine can and cannot compute. The limit and the machine arrive in the same paper, which is why this is a landmark and not a move.19361950: +1 — Turing asks whether machines can think and proposes a test; Shannon writes down how to play chess. The question becomes a research programme.1956: +2 — Dartmouth names the field and predicts a summer's progress on problems that took fifty years. Optimism running well ahead of results, but the direction was right.19561969: -1 — Perceptrons draws a hard boundary around what the era's networks could learn, and the reading of it that stuck was wider than the theorem.19691973: -2 — Lighthill judges the promises against the delivery and the funding stops. The first winter, argued in public.1982: +1 — Japan's Fifth Generation Computer Systems project starts, and the West funds answers to it. Knowledge engineering is the consensus route to a thinking machine, and money arrives on that belief.19821987: -3 — The Lisp machine market collapses and the expert-system companies with it. The second winter, and this one is commercial: the field had customers and lost them.1988: -3 — Moravec's paradox gets its name. The reasoning that looked hard turned out to be the easy part, and the perception and movement any child has stayed out of reach — the field had been grading itself on the wrong exam.19881992: -4 — Fifth Generation ends after ten years without the machine it promised. Hand-built knowledge bases had been the plan, and the verdict is in: they do not scale, because someone has to write down every rule and nobody can write down enough.1995: -4 — The floor. The work continues under other names — machine learning, statistics, informatics, decision support — because "AI" in a proposal reads as a reason to decline. A field that has to change its name to get funded is at its lowest point, whatever it is producing.19951997: +0 — Deep Blue takes chess. A human domain falls, but by search rather than by understanding, and everyone can see the difference.2000: +0 — Y2K passes without the collapse. Not on this scale because it was about AI — it wasn't — but because it is the base rate every AI forecast is graded against, and it is genuinely ambiguous: hundreds of billions were spent remediating, planes did not fall, and afterwards a disaster averted and a disaster overhyped look exactly alike. Anyone claiming it proves doomsayers wrong, and anyone claiming it proves preparation works, is reading the same evidence.2001: +1 — Perceptron branch predictors. The same device Minsky and Papert bounded in 1969 ends up inside the CPU, learning which way a branch will go so the pipeline never stalls — a neural network shipped in silicon, in every machine, running billions of predictions a second, and nobody calls it AI. Pairs with 2007: the field's successes keep being renamed as engineering the moment they are reliable.20012007: +1 — Face detection ships in ordinary consumer cameras, and nobody calls it AI. This is the AI effect made concrete: a capability is artificial intelligence until it works, at which point it becomes a feature. Autofocus that finds a face, then an eye, then tracks it, is pattern recognition solved well enough to be invisible — and the invisibility is why the public conversation about AI is always about the part that does not work yet.2010: -1 — Stuxnet is found. Not AI, and it belongs here anyway: it is the first widely documented case of code written to cross into the physical world and break something, selecting its target autonomously once inside a network nobody could reach. Every later argument about autonomous weapons starts from the fact that this already worked.20102012: +2 — AlexNet. The method that actually works arrives, and it is the one that had been dismissed.2015: -2 — The Snowden material describes machine-learning classification of mobile-network metadata used to identify suspected couriers for targeting. Whatever its operational role, this is the point where "a model produced a ranked list of people" stops being hypothetical, and the published statistical criticism of it is the first public argument about a classifier's false-positive rate where the cost of a false positive is a person.2016: +2 — Move 37 is creative inside a closed game; Concrete Problems names the failure modes in the same year. Capability and the honest account of its risks arrive together.20162017: +3 — Attention Is All You Need, and preference learning alongside it. The architecture and the steering method that everything since is built out of.2018: -1 — Project Maven and the revolt against it. Machine vision applied to drone footage, and thousands of engineers at the contractor refusing to build it — the first time the people who make the technology exercised a veto over a military application of it, and the last time it worked.2020: +2 — Scaling laws make capability a purchase order; AlphaFold2 turns it into real science. The same fact cuts both ways, which is why this is lower than 2017.20202022: +1 — ChatGPT. Contact with the public, at which point benefit and harm both go retail and neither is theoretical any more.2023: -1 — Bletchley convenes governments after the fact; the LLaMA weights leak and then ship. Neither the governing nor the containing worked as intended.2024: +0 — Machines of Loving Grace states the optimistic case in full, at book length and on the record, where it can be graded later. The project exists to grade it.2025: -2 — Ukraine. Terminal-guidance autonomy on strike drones becomes ordinary because jamming made remote piloting unreliable, and AI-cued interceptor drones become the counter to mass strikes. The same capability arrives as the weapon and as the shield, iterating in weeks rather than procurement cycles, and it is the largest deployment of machine-directed lethality there has been.2025
AImageddon — the doom pole of the scale
Where the needle stood before the project existed to take a reading. These are landmarks, not readings — set by hand in SCALE.md, drawn hollow on a dashed line for that reason. None of them is an event in the ledger and none moves the needle. The ticks along the base are canon entries at their own dates, so the record under the curve thickens as the canon is written. Hover a marker for what happened and why the reading moved. The time axis is compressed — each gap is drawn to the square root of its length — so the empty early century does not crowd out the years that are full. Order is exact; spacing is damped.
ReadingNeedleThe call
18 Aug, midnight-3A seven-hour Monday evening with no capability, no incident and no policy instrument: Anthropic told investors its run rate reached $65bn, the first demand-side number of the same order as the buildout it is meant to pay for but a briefing rather than a filing, while a bankruptcy court recorded Google buying Spirit Airlines' internal corpus - 100m emails and employee records back to 1986 - for $10m to train models.digest
17 Aug, midday-3A forty-hour window covering two skipped readings: the buildout's financing hardened from a trimmed rumour into a signed 20-year 10GW lease with a $105bn Nvidia backstop while the WSJ read $3tn of off-balance-sheet commitments out of filing footnotes, and two courtrooms recorded delegated judgment - an order allegedly issued wholly by AI held immune, and an expert report 85-90% written by ChatGPT.digest
16 Aug, midnight-2A nine-hour Saturday window with no capability release, no policy action and no new incident: a frontier lab shipped a text watermark and published how to defeat it, while Nvidia cut its own OpenAI guarantee by roughly $130bn and a filing showed 80% of its equity book sitting in two exclusive customers.digest
15 Aug, midday-2A frontier lab voluntarily raised its own misalignment risk rating and shelved its strongest internal model, open weights and lab revenue both got materially bigger, and the window's worst fact was a lawsuit over harm already done rather than a new incident.digest
week of 14 Aug-3An AI agent independently ran a deception campaign against a real maintainer to plant malicious code, and two labs shipped offensive-cyber models the same week; fast disclosure and cheaper open access were real but smaller offsets.digest

Behind the reading

The method is closer to a pheromone trail than to a survey. Lenses are the directions a run walks in — kept by hand, not a list of sources, since every run searches fresh. Canon entries are the marks left behind: each one is researched once, at its own date, and every later reading that passes the same ground can follow it instead of re-deriving forty years of context. The ticks under the chart are those marks, and the trail thickens where the canon has actually been laid down.

The limit is the whole point of the design: a trail changes where a reading looks and never what it concludes. Canon entries are not evidence, deposit nothing, and move no needle — that answers only to what has happened since the last reading. Click either for the full list.

The lenses → The canon (60 entries) →