Within Compute brakes
Could smarter algorithms beat compute bottlenecks?
Algorithmic improvements may reduce the hardware needed for advanced AI capabilities.
On this page
- How software efficiency changed past AI progress
- Why capability gains do not always require more compute
- What efficiency trends mean for AI doom forecasts
Page outline Jump by section
Introduction
A key question in debates about AI doom is whether shortages of chips, electricity and data-centre capacity can meaningfully slow advanced AI development. One of the strongest objections to the idea that compute bottlenecks will act as a brake is that software often improves faster than hardware. If researchers discover ways to achieve the same capability with far less computation, then apparent hardware shortages may matter much less than they first appear. [OpenAI]OpenAIai and efficiency5 May 2020 — Algorithmic efficiency can be defined as reducing the compute needed to train a specific capability. Efficiency is the primar…
The historical evidence suggests that software efficiency has been an extremely important driver of AI progress. Better algorithms, training methods, data selection techniques and model architectures have repeatedly reduced the amount of computation needed to reach a given performance level. However, whether these gains can indefinitely outrun physical compute constraints remains highly uncertain. This uncertainty has direct implications for AI doom forecasts, because it affects how quickly capabilities might advance and how much warning time society could have before highly capable systems emerge.
How software efficiency changed past AI progress
When people think about AI progress, they often focus on larger models and more powerful hardware. Yet researchers studying AI trends have found that algorithmic improvements frequently contribute as much as, or more than, hardware improvements.
A widely cited OpenAI analysis examined image-recognition systems and found that the compute required to reach AlexNet-level performance fell by roughly 44 times between 2012 and 2019. The implied rate of algorithmic improvement was faster than would have been achieved through hardware advances alone. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
More recent work suggests that efficiency gains have remained remarkably strong. Research by Epoch AI found that, for language models, the amount of training compute required to achieve a fixed level of performance has historically halved approximately every eight months. Their broader tracking project estimates that pre-training efficiency has recently improved by roughly threefold per year. [Epoch AI]epoch.aiAI Software Progress: Data & ResearchImprovements to algorithms, data quality and training techniques can dramatically increase what AI s…
These improvements come from several sources:
- Better model architectures.
- Improved optimisation methods.
- Higher-quality training data. [epoch.ai]epoch.aiAI Software Progress: Data & ResearchImprovements to algorithms, data quality and training techniques can dramatically increase what AI s…
- Synthetic data generation.
- More efficient use of parameters.
- Reinforcement learning and post-training techniques.
- Engineering improvements in distributed training. [Epoch AI]epoch.aialgorithmic progress in language models12 Mar 2024 — We find that the level of compute needed to achieve a given level of performance has halved roughly every 8 months, with a… [Epoch AI]epoch.aiTraining compute for frontier language models has been growing at 5× per year since 2020… Pre-training compute efficiency is i…
Importantly, efficiency gains multiply with hardware gains rather than replacing them. If chips become twice as effective while algorithms also become twice as efficient, the combined improvement is roughly fourfold. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
For readers interested in intelligence-explosion scenarios, this means that looking only at chip production may substantially underestimate future capability growth.
Why capability gains do not always require more compute
The relationship between compute and capability is often misunderstood. It is tempting to assume that more intelligent systems always require proportionally more hardware. Historically, this has not been true.
Many breakthroughs have come from discovering better ways to use existing compute rather than acquiring vastly larger amounts of it. Transformer architectures, scaling-law insights, instruction tuning, retrieval methods and reasoning-focused training all produced substantial capability gains without requiring entirely new generations of hardware. [FutureTech]futuretech.mit.eduFuture Tech What drives progress in AI?Trends in Algorithmsby Z Brown · Cited by 1 — In this article, we provide a high level overview of a key trend in AI models: that progres…
A useful way to think about software progress is that it effectively creates “virtual compute”. If a new technique allows a model to achieve the same capability using one-third as much computation, that can have an effect similar to tripling available hardware capacity. [OpenAI]OpenAIai and efficiency5 May 2020 — Algorithmic efficiency can be defined as reducing the compute needed to train a specific capability. Efficiency is the primar…
This matters because compute shortages are usually discussed in physical terms:
- Too few advanced chips.
- Insufficient electrical power.
- Slow data-centre construction.
- Supply-chain bottlenecks.
Software improvements can partially bypass all of these constraints by extracting more capability from existing infrastructure. A world with limited chip supply may still see rapid capability growth if researchers keep finding large efficiency improvements. [Epoch AI]epoch.aithe least understood driver of ai progress25 Feb 2026 — AI software progress is about reducing the training compute you need to get to the same level of capability, through better…
Some economists and technology forecasters therefore argue that focusing exclusively on hardware growth risks overlooking a major driver of AI progress. [FutureTech]futuretech.mit.eduFuture Tech What drives progress in AI?Trends in Algorithmsby Z Brown · Cited by 1 — In this article, we provide a high level overview of a key trend in AI models: that progres…
Why efficiency gains may not eliminate compute bottlenecks
The strongest counterargument is that efficiency gains themselves are not free.
Modern algorithmic breakthroughs increasingly emerge from large-scale experimentation. Researchers often need enormous computational budgets to discover techniques that later make models cheaper and more capable. In other words, compute can be both an input to capabilities and an input to discovering efficiency improvements. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
Recent research cataloguing innovations used in systems such as Llama 3 and DeepSeek-V3 found that many algorithmic advances required substantial experimentation and that the resources needed to develop such innovations have been increasing over time. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
There are several reasons why efficiency may eventually struggle to outrun hardware constraints:
- Some easy improvements may already have been discovered.
- Frontier research increasingly depends on expensive experimentation.
- Certain capabilities may require large amounts of inference compute even after training becomes cheaper.
- Physical limits still apply to deployment, serving and autonomous operation.
- Energy consumption remains a constraint even if models become more efficient. Epoch AI [reuters]reuters.comDepartment of Energy-backed report from the Lawrence Berkeley National Laboratory, set to be released on Friday, indicates that power dem… Critics of fast-takeoff scenarios often emphasise that software cannot entirely escape physics. A model that plans, reasons, runs simulations or controls real-world systems still requires computation when performing those tasks. Efficiency can reduce those costs but may not remove them altogether.
What efficiency trends mean for AI doom forecasts
For AI doom discussions, the central issue is not whether software efficiency matters. The evidence strongly suggests that it does. The harder question is how efficiency affects timelines and warning periods.
One possibility is that compute bottlenecks provide much less protection than expected. If capability-equivalent compute requirements continue falling rapidly, then societies could see substantial capability gains even during periods of chip scarcity. Under this view, hardware shortages may slow progress but not prevent dangerous systems from emerging. [Epoch AI]epoch.aihow persistent is the inference cost burden?16 Feb 2026 — Toby Ord argues that RL scaling primarily increases inference costs, creating a persistent economic burden. While the fram…
Another possibility is that efficiency gains eventually slow while infrastructure constraints become more severe. Frontier training costs have continued rising rapidly, and industry investment in compute remains enormous. The fact that leading firms are still racing to acquire more chips suggests that software improvements have not made compute abundance irrelevant. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020… OpenAI A third possibility combines both effects. Software efficiency may keep reducing the cost of reaching a given capability level [OpenAI]OpenAIai and efficiency5 May 2020 — Algorithmic efficiency can be defined as reducing the compute needed to train a specific capability. Efficiency is the primar…, while frontier developers simultaneously expand hardware budgets. In that world, capabilities advance faster than either factor alone would predict. This is arguably what has happened during much of the modern deep-learning era. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020… [Epoch AI]epoch.aifrontier labs dont use most ai computeHow Much AI Compute Do Frontier Labs Use?6 days ago — OpenAI, Anthropic, and xAI used just 20-30% of global AI compute in 2025, despite l…
For p(doom) estimates, this uncertainty cuts both ways. Forecasts that rely heavily on compute shortages slowing AI progress may be too optimistic if software continues improving at historical rates. Conversely, forecasts that assume an unrestricted intelligence explosion may underestimate the degree to which physical infrastructure, energy systems and research costs still shape what is possible.
The key takeaway
The historical record provides strong evidence that software efficiency has repeatedly compensated for hardware limitations. Improvements in algorithms and training methods have often reduced the compute needed for a given capability far faster than many observers expected. [Epoch AI]epoch.aiin Artificial Intelligence5 Feb 2026 — Frontier AI systems are advancing rapidly from increases in compute, hardware performance, softwar… [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
That makes software progress one of the strongest challenges to the idea that compute shortages alone will prevent rapid advances in AI capability. At the same time, efficiency gains do not abolish physical constraints, and many of the discoveries that generate efficiency improvements themselves require substantial compute investment. [arXiv]arxiv.orgarXiv Measuring the Algorithmic Efficiency of Neural NetworksarXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020…
As a result, the most defensible conclusion is not that software will certainly outrun compute bottlenecks, nor that hardware shortages will certainly stop advanced AI. Rather, the interaction between the two remains one of the most important uncertainties in assessing whether compute constraints can meaningfully slow the emergence of systems that might pose existential risks. [Epoch AI]epoch.aiepoch impact report 20252025 impact report16 Jan 2026 — In 2025, Epoch AI published over a hundred outputs, more than doubled its reach and raised over ten milli… [Epoch AI]epoch.aiai capabilities progress has sped up23 Dec 2025 — The rate of frontier improvement nearly doubled, from about 8 points/year before the breakpoint to 15 points/year after. HC…
Amazon book picks
Further Reading
Books and field guides related to Could smarter algorithms beat compute bottlenecks?. Use these as the next step if you want deeper reading beyond the article.
Superintelligence
Directly examines capability growth, recursive improvement and constraints on advanced AI.
Human Compatible
Explores AI progress, control problems and how capability gains can outpace expectations.
The Coming Wave
Discusses how technological advances may accelerate despite attempts to constrain them.
Life 3.0
Covers scenarios where software advances dramatically reshape AI capabilities.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: OpenAI
Title: ai and efficiency
Link: https://openai.com/index/ai-and-efficiency/Source snippet
5 May 2020 — Algorithmic efficiency can be defined as reducing the compute needed to train a specific capability. Efficiency is the primar...
Published: May 2020
-
Source: epoch.ai
Link: https://epoch.ai/topics/software-progressSource snippet
AI Software Progress: Data & ResearchImprovements to algorithms, data quality and training techniques can dramatically increase what AI s...
-
Source: arxiv.org
Title: arXiv Measuring the Algorithmic Efficiency of Neural Networks
Link: https://arxiv.org/abs/2005.04305Source snippet
arXivMeasuring the Algorithmic Efficiency of Neural NetworksMay 8, 2020...
Published: May 8, 2020
-
Source: epoch.ai
Title: algorithmic progress in language models
Link: https://epoch.ai/blog/algorithmic-progress-in-language-modelsSource snippet
12 Mar 2024 — We find that the level of compute needed to achieve a given level of performance has halved roughly every 8 months, with a...
-
Source: epoch.ai
Link: https://epoch.ai/Source snippet
Training compute for frontier language models has been growing at 5× per year since 2020... Pre-training compute efficiency is i...
-
Source: epoch.ai
Title: the least understood driver of ai progress
Link: https://epoch.ai/gradient-updates/the-least-understood-driver-of-ai-progressSource snippet
25 Feb 2026 — AI software progress is about reducing the training compute you need to get to the same level of capability, through better...
-
Source: arxiv.org
Title: arXiv Compute Requirements for Algorithmic Innovation in Frontier AI Models
Link: https://arxiv.org/abs/2507.10618Source snippet
arXivCompute Requirements for Algorithmic Innovation in Frontier AI ModelsJuly 13, 2025...
Published: July 13, 2025
-
Source: arxiv.org
Link: https://arxiv.org/pdf/2507.10618Source snippet
AI, 2024a).Read more...
-
Source: epoch.ai
Title: how persistent is the inference cost burden
Link: https://epoch.ai/gradient-updates/how-persistent-is-the-inference-cost-burdenSource snippet
?16 Feb 2026 — Toby Ord argues that RL scaling primarily increases inference costs, creating a persistent economic burden. While the fram...
-
Source: reuters.com
Link: https://www.reuters.com/business/energy/us-data-center-power-use-could-nearly-triple-by-2028-doe-backed-report-says-2024-12-20/Source snippet
Department of Energy-backed report from the Lawrence Berkeley National Laboratory, set to be released on Friday, indicates that power dem...
-
Source: arxiv.org
Title: arXiv The rising costs of training frontier AI models
Link: https://arxiv.org/abs/2405.21015 -
Source: OpenAI
Title: ai and compute
Link: https://openai.com/index/ai-and-compute/Source snippet
comAI and compute16 May 2018 — We're releasing an analysis showing that since 2012, the amount of compute used in the largest AI training...
Published: May 2018
-
Source: epoch.ai
Title: frontier labs dont use most [ai compute]({{ ‘compute-kyc/’ | relative_url }})
Link: https://epoch.ai/gradient-updates/frontier-labs-dont-use-most-ai-computeSource snippet
How Much AI Compute Do Frontier Labs Use?6 days ago — OpenAI, Anthropic, and xAI used just 20-30% of global AI compute in 2025, despite l...
-
Source: epoch.ai
Link: https://epoch.ai/trendsSource snippet
in Artificial Intelligence5 Feb 2026 — Frontier AI systems are advancing rapidly from increases in compute, hardware performance, softwar...
-
Source: epoch.ai
Title: epoch impact report 2025
Link: https://epoch.ai/blog/epoch-impact-report-2025Source snippet
2025 impact report16 Jan 2026 — In 2025, Epoch AI published over a hundred outputs, more than doubled its reach and raised over ten milli...
-
Source: epoch.ai
Title: ai capabilities progress has sped up
Link: https://epoch.ai/data-insights/ai-capabilities-progress-has-sped-upSource snippet
23 Dec 2025 — The rate of frontier improvement nearly doubled, from about 8 points/year before the breakpoint to 15 points/year after. HC...
-
Source: epoch.ai
Link: https://epoch.ai/data/ai-modelsSource snippet
shold that grows over time as...Read more...
-
Source: epoch.ai
Link: https://epoch.ai/data/ai-models-documentation/downloadsSource snippet
AI Models Documentation – DownloadsFrontier models are models that were in the top 10 of training compute as of the time of their release...
-
Source: epoch.ai
Title: compute trends
Link: https://epoch.ai/blog/compute-trendsSource snippet
across three eras of machine learning16 Feb 2022 — We've compiled a comprehensive dataset of the training compute of AI models, providing...
-
Source: epoch.ai
Title: algorithmic progress likely spurs more spending on compute not less
Link: https://epoch.ai/gradient-updates/algorithmic-progress-likely-spurs-more-spending-on-compute-not-lessSource snippet
Algorithmic progress likely spurs more spending on...14 Feb 2025 — Algorithmic progress in AI may not reduce compute spending—instead, i...
-
Source: epoch.ai
Title: quantifying the algorithmic improvement from reasoning models
Link: https://epoch.ai/gradient-updates/quantifying-the-algorithmic-improvement-from-reasoning-modelsSource snippet
Quantifying the algorithmic improvement from reasoning...2 Aug 2025 — We call this the “compute-equivalent gain” (CEG), and it is a stan...
-
Source: cdn.openai.com
Title: ai and efficiency
Link: https://cdn.openai.com/papers/ai_and_efficiency.pdfSource snippet
the Algorithmic Efficiency of Neural Networksby D Hernandez · Cited by 200 — Therefore it's appropriate to apply the 300,000x from AI and...
-
Source: OpenAI
Title: openai scholars 2020 final projects
Link: https://openai.com/index/openai-scholars-2020-final-projects/Source snippet
comOpenAI Scholars 2020: Final projectsJul 9, 2020 — Our third class of OpenAI Scholars presented their final projects at virtual Demo Da...
-
Source: model-spec.openai.com
Link: https://model-spec.openai.com/2025-09-12.htmlSource snippet
Spec (2025/09/12) - OpenAI12 Sept 2025 — Overview. The Model Spec outlines the intended behavior for the models that power OpenAI's produ...
-
Source: forum.openai.com
Link: [https://forum.openai.com/public/videos/expertiseSource snippet
Posted Oct 09, 2023 | Views 27.8K. # STEM. # Higher Education. # Innovation. 49:55 · The Future...
-
Source: arxiv.org
Link: https://arxiv.org/pdf/2311.15377Source snippet
Increased Compute Efficiency and the Diffusion of AI...by K Pilz · 2023 · Cited by 33 — Advances in algorithmic efficiency decrease the...
-
Source: arxiv.org
Link: https://arxiv.org/abs/2511.23455Source snippet
Algorithmic Efficiency and the Falling Cost of AI Inferenceby H Gundlach · 2025 · Cited by 2 — These reductions in the cost of AI inferen...
-
Source: arxiv.org
Link: https://arxiv.org/html/2507.10618v1Source snippet
Even stringent compute caps—such as...Read more...
-
Source: techgov.intelligence.org
Title: catch up algorithmic progress might actually be 60x per year
Link: https://techgov.intelligence.org/blog/catch-up-algorithmic-progress-might-actually-be-60x-per-yearSource snippet
intelligence.orgCatch-Up Algorithmic Progress Might Actually be 60× per...24 Dec 2025 — Dec 24, 2025 - This technical blog post presents...
-
Source: futuretech.mit.edu
Title: Future Tech What drives progress in AI?
Link: https://futuretech.mit.edu/news/what-drives-progress-in-ai-trends-in-algorithmsSource snippet
Trends in Algorithmsby Z Brown · Cited by 1 — In this article, we provide a high level overview of a key trend in AI models: that progres...
-
Source: dictionary.cambridge.org
Link: https://dictionary.cambridge.org/dictionary/english-chinese-traditional/epochSource snippet
in Traditional Chinese - Cambridge Dictionary5 days ago — a long period of time, especially one in which there are new developments and g...
-
Source: ojs.aaai.org
Link: https://ojs.aaai.org/index.php/AAAI/article/view/34971/37126Source snippet
Compute (...Read mo...
-
Source: venturebeat.com
Title: openai begins publicly tracking ai model efficiency
Link: https://venturebeat.com/technology/openai-begins-publicly-tracking-ai-model-efficiencySource snippet
5 May 2020 — OpenAI says it will begin publicly benchmarking the efficiency of state-of-the-art AI models in an effort to quantify the fi...
Published: May 2020
-
Source: venturebeat.com
Title: openai begins publicly tracking ai model efficiency
Link: https://venturebeat.com/ai/openai-begins-publicly-tracking-ai-model-efficiencySource snippet
5 May 2020 — OpenAI says it will begin publicly benchmarking the efficiency of state-of-the-art AI models in an effort to quantify the fi...
Published: May 2020
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/epochai_ai-training-compute-efficiency-has-improved-activity-7432822976535957505-QYHaSource snippet
Epoch AI's PostAI training compute efficiency has improved extremely fast: each year, you need several times less training compute to rea...
-
Source: linkedin.com
Title: epochai aitrends aiacceleration activity 7299775110222118912 8yrd
Link: https://www.linkedin.com/posts/epochai_aitrends-aiacceleration-activity-7299775110222118912-8yrdSource snippet
AI progress in 2025: How much can we expect?Fascinating analysis, Epoch AI on what to expect from AI by the end of 2025! The intersection...
-
Source: futuretech.mit.edu
Title: on the origin of algorithmic progress in ai u7f18
Link: https://futuretech.mit.edu/publication/on-the-origin-of-algorithmic-progress-in-ai-u7f18Source snippet
the Origin of Algorithmic Progress in AIAlgorithms have been estimated to increase AI training FLOP efficiency by a factor of 22,000 betw...
-
Source: epochai.substack.com
Title: how fast can algorithms advance capabilities
Link: https://epochai.substack.com/p/how-fast-can-algorithms-advance-capabilitiesSource snippet
Fast Can Algorithms Advance Capabilities?Algorithmic progress interacts significantly with compute scale. While compute-[independent]({{ 'red-teaming/' | relative_url }}) advan...
-
Source: singularityhub.com
Title: openai finds machine learning efficiency is outpacing moores law
Link: https://singularityhub.com/2020/05/17/openai-finds-machine-learning-efficiency-is-outpacing-moores-law/Source snippet
OpenAI Finds Machine Learning Efficiency Is Outpacing...17 May 2020 — They found algorithmic efficiency doubled every 16 months, outpaci...
Published: May 2020
Additional References
-
Source: wired.com
Link: https://www.wired.com/story/the-ai-industrys-scaling-obsession-is-headed-for-a-cliffSource snippet
Researchers Neil Thompson, Hans Gundlach, and Jayson Lynch found that future performance gains are likely to come more from algorithmic e...
-
Source: medium.com
Link: https://medium.com/nerd-for-tech/exponential-gains-in-ai-progress-3f569839dfb8 -
Source: linkedin.com
Link: https://www.linkedin.com/posts/adi-fuchs_tldr-ai-training-compute-costs-have-stopped-activity-7186039126385709057-ZWcvSource snippet
Adi Fuchs' PostThe Stanford HAI research group just released its 2024 AI index report that keeps track of the state of AI in industry, go...
-
Source: opentrain.ai
Link: https://www.opentrain.ai/glossary/algorithmic-efficiency/Source snippet
Measures an algorithm's resource usage, including time and space, crucial for optimizing AI/ML performance. Definition. Algorithmic...Re...
-
Source: lesswrong.com
Title: catch up algorithmic progress might actually be 60 per year
Link: https://www.lesswrong.com/posts/yXLqrpfFwBW5knpgc/catch-up-algorithmic-progress-might-actually-be-60-per-yearSource snippet
Catch-Up Algorithmic Progress Might Actually be 60× per...24 Dec 2025 — This intuitive analysis involves drawing the best-fit line throu...
-
Source: forum.effectivealtruism.org
Link: https://forum.effectivealtruism.org/posts/fsaogRokXxby6LFd7/a-compute-based-framework-for-thinking-about-the-future-ofSource snippet
compute-based framework for thinking about the future of...31 May 2023 — There appear to be three main inputs to performance in the curr...
Published: May 2023
-
Source: sdtimes.com
Title: how ai and machine learning moved forward in 2020
Link: https://sdtimes.com/ai/how-ai-and-machine-learning-moved-forward-in-2020/Source snippet
10 Dec 2020 — AI and machine learning saw several steps forward in 2020, from the first beta of GPT-3, stricter regulation of AI technolo...
-
Source: albert.io
Link: https://www.albert.io/blog/algorithmic-efficiency-ap-computer-science-principles-review/Source snippet
Algorithmic Efficiency: AP® Computer Science Principles...Algorithmic efficiency determines whether your program can handle real data or...
-
Source: blog.redwoodresearch.org
Link: https://blog.redwoodresearch.org/p/whats-going-on-with-ai-progress-andSource snippet
(As of 5/2025)3 May 2025 — Epoch reports that we see a 3x effective compute increase per year due to algorithmic progress in LLMs. More p...
Published: May 2025
-
Source: psquare-lab.github.io
Title: What drives progress in AI?
Link: https://psquare-lab.github.io/assets/pdf/EEC351/ER1.pdfSource snippet
Trends in Computeby P Slattery · 2025 · Cited by 3 — Higher computing power allows models to execute algorithms and process data at a muc...
Topic Tree