• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI

The bottlenecks that could make AI timelines longer than they look

A commentator quoted in a post argues that experimental data and the process of adopting improvements could limit how quickly even highly capable AI delivers results.

Dimitris PapailiopoulosDP
Pedro DomingosPD
Bits&Bytes.BI
9 Sources, 20d ago, first seen 20d ago

TLDR

A post quotes a commentator arguing that AI timelines can look short while remaining long, depending on the bottleneck. They point to cancer research’s need for experimental data and question whether even superintelligence could instantly improve New York or San Francisco transit, given the process of getting agencies to adopt changes. Progress across different domains, they suggest, could keep the singularity feeling close while very important domains take much longer to unlock.

Combined views

22.2K

9 Sources, first seen 20d ago

207 likes26 comments34 saves24 reposts

Combined views

22.2K

9 Sources, first seen 20d ago

207 likes26 comments34 saves24 reposts

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

Sentiment

Positive——Negative

Summary

Not enough discussion yet.

No sentiment analysis available yet.

9 Sources

Bits&Bytes.@A_DecenTech_Guy“As excited as I am about AI and its potential impact on technology, economics, wealth, health, society & culture - I do think the most bullish ‘AI hard takeoff’ scenarios underestimate how slowly the physical world can move. Intelligence may scale rapidly, but compute does not exist in a vacuum. There are hardware, energy, infrastructure and supply-chain constraints. You can’t instantly manufacture and deploy millions of GPUs or build the infrastructure required to power them. AI may advance exponentially in the digital world, but its impact still has to propagate through the physical economy and that takes time.” Thoughts From The Latent Space: The AI Engineer Podcast, with guest @RichardSocher20d
Timothy B. Lee@binarybitsI really appreciated this reality check from @natolambert. A lot of AI people underestimate how much the physical world matters, and how little impact AI has had on it so far. https://www.interconnects.ai/p/when-will-average-people-feel-ais20d
MTS@MTSlive.@fleetingbits on Richard Ngo’s fear that the next decade of AI could feel like a Shepard tone, with the singularity forever seeming just around the corner: "It's reasonable for timelines to appear short while being longer, and a lot of it depends on what the bottleneck actually is." "You just have to run some number of experiments in order to figure out how to cure cancer. It doesn't matter how good your machine learning model is. If you don't have the data, you can't solve it." "It could feel like we're constantly approaching the singularity because different domains are falling all the time, but the very important domains take a long time to fully unlock." "Even if we had artificial superintelligence today, it's hard to know if we could instantly get the transit systems in New York or San Francisco to run better. There's gonna be some process to get the MTA or the BART to adopt whatever the improvements are."20d
Bharath Ramsundar@rbhar90I am starting to think of AI progress like Moore's law. A truly astounding amount of capital poured in make for good probability models find new axes to continue improving. Like Moore's law, progress may continue for several years (decades?) But like Moore's law itself, progress must come to an end in time. Moore's law has had a tremendous impact on society, and yet we have not yet seen the singularity. I am starting to think of AI the same. The tremendous number of math advances alone will lead to many downstream advances in time. But shifts in everyday life will take a time to come, especially outside tech circles.20d
Ash Jogalekar@curiouswavefnExcellent set of thoughts from @rbhar90. "Here is where I perhaps fall more skeptical than the average frontier lab researcher and argue that I don’t think so. AI training is increasingly a physical act. Vast amounts of power, tremendous data centers, and extreme infrastructure are used for the latest AI models. These are physical artifacts. It may be possible to use the “mathematical telescope” to optimize parts of this (perhaps designing chips more quickly, or creating more efficient rack layouts, or implementing smarter protocols to make use of available energy) but the hard constraints of the physical world remain. We have seen powerful new optimization tools in the past. The advent of linear programming, semidefinite programming, convex optimization, and even deep learning have each heralded large shifts in scientific disciplines but without rocking the foundations of society wholesale. Each new optimization shift just sheds light on the next bottleneck. AI advances are not fundamentally dissimilar." https://deepforest.substack.com/p/how-do-we-process-ai-math-advances?utm_source=post-email-title&publication_id=254613&post_id=216990247&utm_campaign=email-post-title&isFreemail=true&r=4do01&triedRedirect=true&utm_medium=email18d
Dimitris Papailiopoulos@DimitrisPapailIt seems to me that "extracted value" as a function of AI (sorry i meant SI) capabilities is not only very sublinear but also has a ceiling. I think we are likely seeing to see a saturation not of capabilities, but how much value the enterprise/industry can extract, since the biggest source of friction is like dealing with people (and the physical layer) and bespoke enterprise infra/data. One way to improve this "value vs capability" scaling is to optimize the infra and increase margins (aka pace the frontier). I am personally happy with any model from this list: Opus 4.5/4.6/5.5, GPT-5.6, and Astra. I have only seen mind-blowing gaps in math, not coding, experimentation, or agency. Perhaps some improvements in creativity/writing too but not mindblowing. If you told me today that capabilities will stop increasing, I would not care. A higher DeepSWE/HLE/TB4 score by itslef won't resolve this and the idea that full enterprise/SWE automation is coming soon seems doubtful. Ergo, my speculation is many labs will focus on industries beyond SWE where capability, not integration, is still the bottleneck, e.g., bio, chem, or robotics to increase multiplicative factor of margin.18d
Pedro Domingos@pmddomingosThe impact of AI is accelerating even as the research progress slows down. That's what happens when one S-curve triggers another.18d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    9 Sources

    Bits&Bytes.@A_DecenTech_Guy“As excited as I am about AI and its potential impact on technology, economics, wealth, health, society & culture - I do think the most bullish ‘AI hard takeoff’ scenarios underestimate how slowly the physical world can move. Intelligence may scale rapidly, but compute does not exist in a vacuum. There are hardware, energy, infrastructure and supply-chain constraints. You can’t instantly manufacture and deploy millions of GPUs or build the infrastructure required to power them. AI may advance exponentially in the digital world, but its impact still has to propagate through the physical economy and that takes time.” Thoughts From The Latent Space: The AI Engineer Podcast, with guest @RichardSocher20d
    Timothy B. Lee@binarybitsI really appreciated this reality check from @natolambert. A lot of AI people underestimate how much the physical world matters, and how little impact AI has had on it so far. https://www.interconnects.ai/p/when-will-average-people-feel-ais20d
    MTS@MTSlive.@fleetingbits on Richard Ngo’s fear that the next decade of AI could feel like a Shepard tone, with the singularity forever seeming just around the corner: "It's reasonable for timelines to appear short while being longer, and a lot of it depends on what the bottleneck actually is." "You just have to run some number of experiments in order to figure out how to cure cancer. It doesn't matter how good your machine learning model is. If you don't have the data, you can't solve it." "It could feel like we're constantly approaching the singularity because different domains are falling all the time, but the very important domains take a long time to fully unlock." "Even if we had artificial superintelligence today, it's hard to know if we could instantly get the transit systems in New York or San Francisco to run better. There's gonna be some process to get the MTA or the BART to adopt whatever the improvements are."20d
    Bharath Ramsundar@rbhar90I am starting to think of AI progress like Moore's law. A truly astounding amount of capital poured in make for good probability models find new axes to continue improving. Like Moore's law, progress may continue for several years (decades?) But like Moore's law itself, progress must come to an end in time. Moore's law has had a tremendous impact on society, and yet we have not yet seen the singularity. I am starting to think of AI the same. The tremendous number of math advances alone will lead to many downstream advances in time. But shifts in everyday life will take a time to come, especially outside tech circles.20d
    Ash Jogalekar@curiouswavefnExcellent set of thoughts from @rbhar90. "Here is where I perhaps fall more skeptical than the average frontier lab researcher and argue that I don’t think so. AI training is increasingly a physical act. Vast amounts of power, tremendous data centers, and extreme infrastructure are used for the latest AI models. These are physical artifacts. It may be possible to use the “mathematical telescope” to optimize parts of this (perhaps designing chips more quickly, or creating more efficient rack layouts, or implementing smarter protocols to make use of available energy) but the hard constraints of the physical world remain. We have seen powerful new optimization tools in the past. The advent of linear programming, semidefinite programming, convex optimization, and even deep learning have each heralded large shifts in scientific disciplines but without rocking the foundations of society wholesale. Each new optimization shift just sheds light on the next bottleneck. AI advances are not fundamentally dissimilar." https://deepforest.substack.com/p/how-do-we-process-ai-math-advances?utm_source=post-email-title&publication_id=254613&post_id=216990247&utm_campaign=email-post-title&isFreemail=true&r=4do01&triedRedirect=true&utm_medium=email18d
    Dimitris Papailiopoulos@DimitrisPapailIt seems to me that "extracted value" as a function of AI (sorry i meant SI) capabilities is not only very sublinear but also has a ceiling. I think we are likely seeing to see a saturation not of capabilities, but how much value the enterprise/industry can extract, since the biggest source of friction is like dealing with people (and the physical layer) and bespoke enterprise infra/data. One way to improve this "value vs capability" scaling is to optimize the infra and increase margins (aka pace the frontier). I am personally happy with any model from this list: Opus 4.5/4.6/5.5, GPT-5.6, and Astra. I have only seen mind-blowing gaps in math, not coding, experimentation, or agency. Perhaps some improvements in creativity/writing too but not mindblowing. If you told me today that capabilities will stop increasing, I would not care. A higher DeepSWE/HLE/TB4 score by itslef won't resolve this and the idea that full enterprise/SWE automation is coming soon seems doubtful. Ergo, my speculation is many labs will focus on industries beyond SWE where capability, not integration, is still the bottleneck, e.g., bio, chem, or robotics to increase multiplicative factor of margin.18d
    Pedro Domingos@pmddomingosThe impact of AI is accelerating even as the research progress slows down. That's what happens when one S-curve triggers another.18d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet