compute overhang
In plain wordsA situation in which far more computing power is available than AI currently uses, so that a new method or a decision to scale up could suddenly produce a big jump in ability.
A situation in which the computing power available or affordable greatly exceeds what current AI systems use, so that once a sufficiently good algorithm arrives, or a decision is made to scale up, capabilities could jump abruptly rather than improve gradually. In the original hardware-overhang form, the worry is that if the software for human-level AI arrives late, abundant hardware would immediately allow running vast numbers of fast copies, favoring a fast takeofffast takeoff, n.: A scenario in which AI goes from about human level to far beyond it very quickly, in days, weeks or a few months, leaving people little time to react.. A related usage concerns training: if frontier training runs are held back, for instance by a pause, while chips keep getting cheaper, lifting the restriction could produce a sudden leap. Some laboratories have cited overhang as an argument for steady, incremental scaling rather than delay, while critics see this as a rationalization for racing. Inference overhang, the ability to run a newly trained model as many copies at once, and algorithmic overhang are variants.
A long pause followed by an unpause could leave a compute overhang bigger than the one it was meant to avoid.
Origin
Earliest use not established. Hardware overhang was discussed by 2010, on LessWrong and in Carl Shulman and Anders Sandberg's paper on software-limited singularities, and appears in Bostrom's Superintelligence (2014). 'Compute overhang' became common in the 2020s; OpenAI's February 2023 post 'Planning for AGI and beyond' invoked it in arguing for gradual progress.
Related
- fast takeofffast takeoff, n.: A scenario in which AI goes from about human level to far beyond it very quickly, in days, weeks or a few months, leaving people little time to react.
- AI pauseAI pause, n.: A proposed halt, temporary or indefinite, on building AI systems more powerful than today's, to give safety research and regulation time to catch up.
- computecompute, n.: The raw computing power used to train and run AI, usually measured by how many specialized chips are used and for how long. With data and algorithms, it is a main driver of AI progress.
- scaling lawsscaling laws, n.: Measured patterns showing that an AI model's error shrinks smoothly and predictably as the model is made bigger and given more data and computing power, so labs can forecast a giant model's performance from small, cheap experiments.
- takeofftakeoff, n.: The stretch of time between AI reaching roughly human level and AI becoming vastly superhuman. People argue over whether it would take decades, years, months or just days.
- intelligence explosionintelligence explosion, n.: The idea that once AI gets good enough at improving AI, each smarter version could build a smarter one faster still, so abilities would snowball far beyond human level in a short time.
Sources
- Shulman, C. & Sandberg, A. (2010). Implications of a Software-Limited Singularity. Proceedings of the European Conference on Computing and Philosophy (ECAP 2010).
Mentioned in
AI pauseAI pause, n.: A proposed halt, temporary or indefinite, on building AI systems more powerful than today's, to give safety research and regulation time to catch up., fast takeofffast takeoff, n.: A scenario in which AI goes from about human level to far beyond it very quickly, in days, weeks or a few months, leaving people little time to react., intelligence explosionintelligence explosion, n.: The idea that once AI gets good enough at improving AI, each smarter version could build a smarter one faster still, so abilities would snowball far beyond human level in a short time., scaling hypothesisscaling hypothesis, n.: The bet that today's AI methods, simply made bigger with more computing power and data, will keep getting smarter — possibly all the way to human-level intelligence and beyond — without needing fundamentally new ideas., effective computeeffective compute, n.: A way of measuring AI progress that counts better methods as if they were extra computing power: if a new technique lets a model match one trained with twice the computing, it counts as doubling effective compute.