Lumpy inventions are a crux between lines-on-graphs vs. scaling-pilled perspectives on AI timelines, even where they agree on chances of very near-term timelines. Different kinds of evidence become relevant, and different kinds of long timelines become plausible. METR Nov 2026 report on Anthropic isn't very relevant for the scaling-pilled perspective, future model scaling and its hardwareprecursors aren't very relevant for the lines-on-graphs perspective, recent mathbreakthroughs are relevant for both. On longer timelines, lumpy inventions ensure a permanent hazard rate of rapidly triggering strong ASI at any moment even when lines-on-graphs level off, and raw scaling has slowed down.
There is a tech tree, and human technological culture is climbing it at some pace. LLMs help climb it faster, and reach higher for the low-hanging tech fruit that's not too far from the current state of technological culture. Right now, if humanity stopped, but LLM infrastructure magically kept working, LLMs would fail to climb further. But in 1-2 years, LLMs would likely keep climbing even on their own (at some pace that's up for debate). Along the way, there are plentiful small incremental improvements to be found, which help with both the climbing speed and the distance that the LLMs can reach for the low-hanging fruit. The lines-on-graphs perspective on software-only singularity expects enough of it to sustain a high speed of climbing. The scaling-pilled perspective says that each new level of raw compute scale creates an overhang where the initial climbing speed is high (reaping incremental improvements), but there is a soft ceiling where that speed levels off, and it won't be sustained with incremental algorithmic fruit alone.
Lumpy inventions (paradigm-shifting algorithmic innovations that unhobble the AIs, such as continual learning in a strong sense) directly increase both the sustained pace of climbing the tech tree (as opposed to the ability to merely reach higher for the low-hanging fruit) and the current soft ceiling imposed by the raw compute scale. The lines-on-graphs perspective either doesn't need lumpy inventions, or expects that they are inherently part of the same statistical trends that rest on incremental improvements, and won't fundamentally disrupt those trends. The scaling-pilled perspective expects only a few lumpy inventions left before software-only singularity becomes possible, each one defining an entirely new lines-on-graphs regime, with high uncertainty about how difficult it is to encounter those inventions on the tech tree, how long it would take. Between the lumpy inventions, there are periods where the lines-on-graphs perspective fits, but it doesn't enter a runaway mode of software-only singularity before enough lumpy inventions arrive, which only happens if they happen to be sufficiently easy to be encountered soon (reaching high while climbing the tech tree only slowly), or enough raw compute scale has been exploited to reach them even if they are very difficult.