4 ms·

The AI Market Is Firming Up Fast

48 points by dbreunig 3y ago

whatshisface 3y ago
There are a lot of benefits to the constitutional prohibition on ex-post-facto laws but one of the downsides is the way the private sphere ends up lead by the fastest-drawing outlaws.
- actionfromafar 3y ago
  Regarding data, an idea would be if data had a time moratorium, where it had a do-not-use until the data was x days old.
  - whatshisface 3y ago
    That's how copyright works, but the courts have sided against your copyright in this area, it would take a new law to give it back. Congress needs to say, "this precedent is wrong, you have a copyright to your data and ChatGPT is a massive infringer, way worse than the internet archive would be if it gave away free books." If the world was perfect we would reduce copyright terms to twenty to thirty years and allow training AIs on public domain text... solving the human learning (free books relevant to today!) and AI learning issues in one step. Sadly the people on the inside want to take everything and leave us with nothing so we will probably get a hodgepodge of laws, a philosophical system no one can explain, and a seriously nonuniform landscape of compromises. Don't interpret this as defeatism, just as pointing out where the ball is heading if nobody catches it. If you want one takeaway from my comment, don't let them train AI on text you are not allowed, yourself, to read. See this lawsuit for evidence that it is happening: https://www.cbc.ca/radio/asithappens/authors-guild-chatgpt-lawsuit-1.6974154 https://www.cbc.ca/radio/asithappens/authors-guild-chatgpt-l...
wsintra2022 3y ago
There may also be alternative solutions to the problem solving that ML offers we just stopped thinking about them because so much r&d goes into neural networks.
- krasin 3y ago
  The definition of AI used in the executive order ([1]) does not include neural nets or even ML / statistics, only the ability to make predictions/recommendations/decisions. It is not apparent what kind of R&D in this area would not fall under it: >(b) The term “artificial intelligence” or “AI” has the meaning set forth in 15 U.S.C. 9401(3): a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. Artificial intelligence systems use machine- and human-based inputs to perceive real and virtual environments; abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action. 1. https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/ https://www.whitehouse.gov/briefing-room/presidential-action...
  - xenonite 3y ago
    I don’t think this correct, the key being here this sentence: > abstract such perceptions into models through analysis in an automated manner
- IKantRead 3y ago
  Companies aren't worried about "problem solving" they're worried about "using AI". Not long ago I was at a company that was rushing to get their AI "chat with our documentation" bot out the door as fast as possible. But prior to AI hype, 'improve documentation search' was not even remotely a priority, or even a concern. It's not like they had an okay chatbot for docs that was previously important but not high enough quality. Why was a documentation chatbot suddenly important? I was talking to a neighbor the other day and he said he was tasked with using AI to help solve scheduling problems in his hospital! This was particularly bizarre, because if "solve scheduling issues" was a real problem, there are plenty of existing approaches to this (and have been for decades) and further more LLMs are not a particularly good way to solve this. There are plenty of tricky NLP problems that AI/LLMs can unblock, and it makes sense that if you've been trying to solve a tricky NLP problem you might think AI can help you. There are numerous old problems I've had that I just couldn't get quite performant enough to work years ago that I would love to revisit with LLMs. Funnily enough, when I chat with former coworkers about progress they've made solving these old problems they tell me leadership isn't interested and instead they're building something like a chatbot for documentation.
madofuller 3y ago
>allowing them to be built with less compute and less data – enabling more challengers to enter and create a more dynamic field during these formative years. I feel the solution will never be using less compute and less data, but rather, will those projects like Gensyn and Bittensor actually succeed in the sharing of compute and data resources.
- brucethemoose2 3y ago
  I dunno anything about this crypto, but just getting community ML projects to integrate basic optimizations and hw vendor ports (much less rearchitect for these projects) is like pulling teeth. If its not turnkey PyTorch... Yeah. Also, Petals and AI Horde seem to have gotten far even though they are relatively new (and crypto free)
whatshisface 3y ago
>Further, those parties make up an ouroboros of funding, where invested funds quickly come back to investors in the form of cloud compute fees. This allows Amazon, Microsoft, and Google to stomach sky-high valuations for these rounds, as they’re essentially buying marketshare for their cloud offerings. This dynamic is a major barrier standing between you and the latest GPUs. I don't think cloud compute prices are so far above their costs that this would amount to a huge distinguishing factor.
- agnokapathetic 3y ago
  But equities investors value cloud growth for all of these companies, and AI spend is high enough that even if the margin isn't there, the revenue growth will make the street happy.
  - whatshisface 3y ago
    But isn't that the same kind of sales growth an investor could see if a retail company used investor money to load up trucks with its own shelved goods?
- choppaface 3y ago
  Not necessarily compute minutes but ancillary costs like egress and especially local block storage. I recently tested a training cluster and these costs were nearly as much as the GPUs. And they’re both extremely high margin as well as lower QoS than what you can get in bare metal.
effnorwood 3y ago
[dead]
karmasimida 3y ago
There is not going to be a chance for small folks in AI land. It looks more like military industry day by day.
nightski 3y ago
The market for using multi head attention based transformers is firming up fast, but it's one algorithmic innovation away from all those foundation models being worthless and starting over from scratch.

48 points by dbreunig 3y ago

There are a lot of benefits to the constitutional prohibition on ex-post-facto laws but one of the downsides is the way the private sphere ends up lead by the fastest-drawing outlaws.

actionfromafar 3y ago
Regarding data, an idea would be if data had a time moratorium, where it had a do-not-use until the data was x days old.
- whatshisface 3y ago
  That's how copyright works, but the courts have sided against your copyright in this area, it would take a new law to give it back. Congress needs to say, "this precedent is wrong, you have a copyright to your data and ChatGPT is a massive infringer, way worse than the internet archive would be if it gave away free books." If the world was perfect we would reduce copyright terms to twenty to thirty years and allow training AIs on public domain text... solving the human learning (free books relevant to today!) and AI learning issues in one step. Sadly the people on the inside want to take everything and leave us with nothing so we will probably get a hodgepodge of laws, a philosophical system no one can explain, and a seriously nonuniform landscape of compromises. Don't interpret this as defeatism, just as pointing out where the ball is heading if nobody catches it. If you want one takeaway from my comment, don't let them train AI on text you are not allowed, yourself, to read. See this lawsuit for evidence that it is happening: https://www.cbc.ca/radio/asithappens/authors-guild-chatgpt-lawsuit-1.6974154 https://www.cbc.ca/radio/asithappens/authors-guild-chatgpt-l...

wsintra2022 3y ago

There may also be alternative solutions to the problem solving that ML offers we just stopped thinking about them because so much r&d goes into neural networks.

krasin 3y ago
The definition of AI used in the executive order ([1]) does not include neural nets or even ML / statistics, only the ability to make predictions/recommendations/decisions. It is not apparent what kind of R&D in this area would not fall under it: >(b) The term “artificial intelligence” or “AI” has the meaning set forth in 15 U.S.C. 9401(3): a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments. Artificial intelligence systems use machine- and human-based inputs to perceive real and virtual environments; abstract such perceptions into models through analysis in an automated manner; and use model inference to formulate options for information or action. 1. https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/ https://www.whitehouse.gov/briefing-room/presidential-action...
- xenonite 3y ago
  I don’t think this correct, the key being here this sentence: > abstract such perceptions into models through analysis in an automated manner
IKantRead 3y ago
Companies aren't worried about "problem solving" they're worried about "using AI". Not long ago I was at a company that was rushing to get their AI "chat with our documentation" bot out the door as fast as possible. But prior to AI hype, 'improve documentation search' was not even remotely a priority, or even a concern. It's not like they had an okay chatbot for docs that was previously important but not high enough quality. Why was a documentation chatbot suddenly important? I was talking to a neighbor the other day and he said he was tasked with using AI to help solve scheduling problems in his hospital! This was particularly bizarre, because if "solve scheduling issues" was a real problem, there are plenty of existing approaches to this (and have been for decades) and further more LLMs are not a particularly good way to solve this. There are plenty of tricky NLP problems that AI/LLMs can unblock, and it makes sense that if you've been trying to solve a tricky NLP problem you might think AI can help you. There are numerous old problems I've had that I just couldn't get quite performant enough to work years ago that I would love to revisit with LLMs. Funnily enough, when I chat with former coworkers about progress they've made solving these old problems they tell me leadership isn't interested and instead they're building something like a chatbot for documentation.

madofuller 3y ago

>allowing them to be built with less compute and less data – enabling more challengers to enter and create a more dynamic field during these formative years. I feel the solution will never be using less compute and less data, but rather, will those projects like Gensyn and Bittensor actually succeed in the sharing of compute and data resources.

brucethemoose2 3y ago
I dunno anything about this crypto, but just getting community ML projects to integrate basic optimizations and hw vendor ports (much less rearchitect for these projects) is like pulling teeth. If its not turnkey PyTorch... Yeah. Also, Petals and AI Horde seem to have gotten far even though they are relatively new (and crypto free)

whatshisface 3y ago

>Further, those parties make up an ouroboros of funding, where invested funds quickly come back to investors in the form of cloud compute fees. This allows Amazon, Microsoft, and Google to stomach sky-high valuations for these rounds, as they’re essentially buying marketshare for their cloud offerings. This dynamic is a major barrier standing between you and the latest GPUs. I don't think cloud compute prices are so far above their costs that this would amount to a huge distinguishing factor.

agnokapathetic 3y ago
But equities investors value cloud growth for all of these companies, and AI spend is high enough that even if the margin isn't there, the revenue growth will make the street happy.
- whatshisface 3y ago
  But isn't that the same kind of sales growth an investor could see if a retail company used investor money to load up trucks with its own shelved goods?
choppaface 3y ago
Not necessarily compute minutes but ancillary costs like egress and especially local block storage. I recently tested a training cluster and these costs were nearly as much as the GPUs. And they’re both extremely high margin as well as lower QoS than what you can get in bare metal.

effnorwood 3y ago

[dead]

karmasimida 3y ago

There is not going to be a chance for small folks in AI land. It looks more like military industry day by day.

nightski 3y ago

The market for using multi head attention based transformers is firming up fast, but it's one algorithmic innovation away from all those foundation models being worthless and starting over from scratch.