Skip to content
Last Dictionary
Plane
Training & Post-trainingGovernance, Policy & LawAGI & Superintelligence

effective compute

nounAlso: effective FLOP, effective training compute, compute multiplier, compute-equivalent gain

In plain wordsA 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.

  1. Training computetraining compute, n.: The total number of arithmetic operations performed to train a model. It is the standard yardstick for the size of an AI training effort, and some laws use it to decide which models face extra rules. adjusted for algorithmic efficiency: the amount of physical compute a reference method — for example, a fixed older architecture and recipe — would need to match the performance of the model in question. A technique that reaches a baseline's loss with half the FLOP is a '2× compute multiplier', and multipliers from architecture, data, optimizer and other improvements compound with growth in physical compute, so effective compute grows faster than raw FLOP. The concept underlies forecasts of AI progress that count orders of magnitude of effective compute from spending, hardware and algorithmic progressalgorithmic progress, n.: Improvements in AI methods — not faster chips — that let a model reach the same ability with less computing power. Researchers estimate this has cut the computing needed by half every eight to sixteen months.. It also entered frontier-safety policy: the first version of Anthropic's responsible scaling policyresponsible scaling policy, n.: A public rulebook in which an AI company promises to test each new model for dangerous abilities and to add stronger safeguards, or slow down, as its models become more capable. (September 2023) called for capability evaluations after every 4× increase in effective compute, and its October 2024 revision used 4× effective compute, or six months of accumulated post-training enhancements, to trigger comprehensive assessments. Estimates depend on the model, benchmark and baseline; gains measured at small scale may not transfer to large scale; and no standard measurement procedure exists.

    The new architecture is worth about a 3× compute multiplier, so the run is a 12× jump in effective compute.

Origin

Measuring algorithmic progress in compute-equivalent terms goes back at least to Danny Hernandez and Tom Brown (2020) and AI-forecasting work such as Ajeya Cotra's biological-anchors report (2020). 'Effective compute' was in use by at least September 2023, when Anthropic's Responsible Scaling Policy defined it as roughly the compute needed to train a model if no improvements in pretraining or fine-tuning techniques were counted, and spread through Leopold Aschenbrenner's 'Situational Awareness' (2024), which called algorithmic gains 'compute multipliers'.

Usage note

Not a measured physical quantity but a modeled one; comparisons across labs or eras inherit all the assumptions of the chosen baseline. Distinct from unhobblingunhobbling, n.: Fixing the simple ways a capable AI model is held back — such as by giving it tools, memory or time to think — so that abilities it already has can actually be used., which Aschenbrenner counted separately.

Related

See also
  • training computetraining compute, n.: The total number of arithmetic operations performed to train a model. It is the standard yardstick for the size of an AI training effort, and some laws use it to decide which models face extra rules.
  • algorithmic progressalgorithmic progress, n.: Improvements in AI methods — not faster chips — that let a model reach the same ability with less computing power. Researchers estimate this has cut the computing needed by half every eight to sixteen months.
  • compute overhangcompute overhang, n.: A 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.
  • responsible scaling policyresponsible scaling policy, n.: A public rulebook in which an AI company promises to test each new model for dangerous abilities and to add stronger safeguards, or slow down, as its models become more capable.
  • AI timelinesAI timelines, n.: Forecasts of when AI will reach major milestones, such as matching humans at most work. People with short timelines expect this within years; others expect decades, or never.
  • unhobblingunhobbling, n.: Fixing the simple ways a capable AI model is held back — such as by giving it tools, memory or time to think — so that abilities it already has can actually be used.
  • 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.

Sources

  1. Hernandez, D. & Brown, T. B. (2020). Measuring the Algorithmic Efficiency of Neural Networks. arXiv:2005.04305. arxiv.org/abs/2005.04305
  2. Aschenbrenner, L. (2024). Situational Awareness: The Decade Ahead — I. From GPT-4 to AGI: Counting the OOMs. situational-awareness.ai/from-gpt-4-to-agi

Mentioned in

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., algorithmic progressalgorithmic progress, n.: Improvements in AI methods — not faster chips — that let a model reach the same ability with less computing power. Researchers estimate this has cut the computing needed by half every eight to sixteen months., training computetraining compute, n.: The total number of arithmetic operations performed to train a model. It is the standard yardstick for the size of an AI training effort, and some laws use it to decide which models face extra rules., biological anchorsbiological anchors, n.: A way of forecasting when AI could transform the world, by estimating how much computing it would take to match the human brain and when that much computing would be affordable., Leopold AschenbrennerLeopold Aschenbrenner, prop. n.: A German-born former OpenAI researcher whose 2024 essay series 'Situational Awareness' predicted human-level AI around 2027 and a government-led race to build superintelligence., Situational Awareness (essay)Situational Awareness (essay), prop. n.: A long 2024 essay by a former OpenAI researcher arguing that AGI could arrive around 2027, that superintelligence could follow soon after, and that the US government should run a national AGI project., unhobblingunhobbling, n.: Fixing the simple ways a capable AI model is held back — such as by giving it tools, memory or time to think — so that abilities it already has can actually be used.