Bb5a31f81ClaimComputer scientist Rich Sutton consolidated 'The Bitter Lesson', which posits that general computational methods utilizing compute and data consistently outperform human-designed domain heuristicsAgreeDisagree
Bb5a31f81ClaimInception reported that Mercury 3 scaled compute roughly 10 times relative to Mercury 2AgreeDisagree
Bb5a31f81ClaimIncreasing parameters rather than recursive depth can expand AI capabilities while maintaining greater model interpretabilityAgreeDisagree
Bb5a31f81ClaimComputer scientist Richard Sutton published the essay titled The Bitter Lesson in 2019AgreeDisagree
Bb5a31f81ClaimPre-training frontier artificial intelligence models continues to rely on scaling laws requiring exponentially greater compute resourcesAgreeDisagree
Bb5a31f81ClaimAnthropic and METR reported that artificial intelligence automation roughly doubled the internal pace of AI research progress over a six-to-twelve-month periodAgreeDisagree
Bb5a31f81ClaimThe GPT-4 technical report stated that OpenAI predicted GPT-4's final training loss using smaller models trained with 10,000 times less computeAgreeDisagree
Bb5a31f81ClaimThe GPT-4 technical report stated that final model loss was predictable using models trained with 10,000 times less computeAgreeDisagree
Bb5a31f81ClaimThe compute used to train the final model in the 2017 Transformer paper was roughly 1/10,000th of the training compute required for GPT-3AgreeDisagree
Bb5a31f81ClaimIncreasing raw computing power is the primary factor in accelerating neural network training performanceAgreeDisagree
Bb5a31f81ClaimMETR found that the pace of capability improvement for Anthropic AI models roughly doubled over the course of a yearAgreeDisagree
Bb5a31f81ClaimMETR determined that the acceleration of Anthropic model capabilities was driven primarily by AI assistance in research and development rather than fully autonomous contributionsAgreeDisagree
Bb5a31f81ClaimAnthropic defines its recursive self-improvement danger threshold as achieving as much AI capability progress in one year as occurred between 2018 and 2024AgreeDisagree
Bb5a31f81ClaimMistral Large 4 was trained using an estimated compute budget of 10 to the 25th floating-point operationsAgreeDisagree
Bb5a31f81ClaimMeta's Llama 3 was released in July 2024 and was trained using 3.8 times 10 to the 25th floating-point operationsAgreeDisagree
Bb5a31f81ClaimMaintaining a decisive strategic advantage achieved via a rapid takeoff is vulnerable to erosion over time through model theft, model distillation, and diffusionAgreeDisagree
Bb5a31f81ClaimCurrent AI scaling laws show that an exponential increase in compute is required to achieve a linear increase in performance, which argues against a sudden, fast takeoffAgreeDisagree
Bb5a31f81ClaimAlan Chan argues for collecting more data on AI progress as a function of compute and labor costs to estimate the "returns to research effort"AgreeDisagree
Bb5a31f81ClaimA 'bitter lesson' for robotics is that progress will be driven by scaling computation and data over specialized engineeringAgreeDisagree
Bb5a31f81ClaimA challenge in applying the Bitter Lesson to robotics is that identifying the 'right' kind of data is difficultAgreeDisagree