3 ms·

As if you could ever stop research in what is essentially just maths. The concept of neural networks has been around since before many of those people were bor

As if you could ever stop research in what is essentially just maths. The concept of neural networks has been around since before many of those people were born. The statistical and mathematical underpinnings of AI predate anyone alive today. It might seem to an outsider that ChatGPT and the other household name AI came out of nowhere, but they didn’t. The research has been ongoing for decades and will continue with or without the big name AI companies.

truculent 3y ago
The math has been around for decades, but the hardware and capital requirements are huge.
- truculent 3y ago
  Likewise, given the ideas have been around for decades, the question of why it took so long for the fire to light is instructive: https://gwern.net/scaling-hypothesis#prospects https://gwern.net/scaling-hypothesis#prospects
_heimdall 3y ago
Theoretical math is one thing, capital and hardware requirements are very different. You need people willing to do the research, money to pay them, and a mountain of hardware. If you are concerned with at least putting on a show of risk management you need expensive lawyers, a large PR budget, and teams working on safety and ethics to at least act like it matters. If you actually are concerned with risks you need time to first consider what you do upon successfully developing an AGI. How would you recognize it? Does it have rights? Would turning it off be murder? More importantly, how do you ensure that it doesn't escape controls you put in place, if that possible at all?
- mr_toad 3y ago
  For now, but in a decade or two that hardware will become cheaper and more widely available. What happens when people can run the equivalent of GPT-4 on a desktop GPU?
  - _heimdall 3y ago
    The hardware market would very likely be much different if those years don't include a heavily funded AI research industry driving up innovation and demand for that hardware, but even if nothing changed we would at least be buying a decade or two for us to be better prepared and more thoughtful about any technological advancements we want to pursue with AI. Today the industry is driven entirely by the promise of future profits and the love of innovation with no real regard for actual risks that could be in play.
- tavavex 3y ago
  Given the ever-growing computing power, computation itself becoming cheaper and the extremely wide reach of information on the internet, the outcome is inevitable either way. Even if you time-traveled 10 years back and somehow erased all modern ML research from existence, there'd still be a point in time where, eventually, some guy in a garage would do the work. Eventually, the work we use massive server clusters for today would be run on consumer-grade computers, and someone would try their hand at it. It has always been the case that new technology doesn't ask for permission to exist. Even if we lived in this society that was extremely risk-averse and suspicious of any new developments, at some point the bar would get so low that, if companies refused to invest and do the research, individuals would.
  - _heimdall 3y ago
    Are you effectively using something akin to a multiverse scenario to guarantee that one outcome is inevitable? By multiverse here I'm mainly pointing to the idea of branching of all possible outcomes, not strictly trying to say that the idea I'd about jumping to an actual different universe. If you could go back 10 years and erase all ML research, for example, there would be fundamental differences. The number of people that understand how those systems would be much smaller. Hardware would be different, for example we wouldn't have tensor chips and GPUs would still be primarily for playing Call of Duty. Cell phones, search, even more benign systems like traffic control on city streets would look different. There's simply way too much changed even with that one jump to possibly guarantee that we'd find ourselves in the same boat inventing AI or AGI on roughly the same timeline.

truculent 3y ago

The math has been around for decades, but the hardware and capital requirements are huge.

truculent 3y ago
Likewise, given the ideas have been around for decades, the question of why it took so long for the fire to light is instructive: https://gwern.net/scaling-hypothesis#prospects https://gwern.net/scaling-hypothesis#prospects

_heimdall 3y ago

Theoretical math is one thing, capital and hardware requirements are very different. You need people willing to do the research, money to pay them, and a mountain of hardware. If you are concerned with at least putting on a show of risk management you need expensive lawyers, a large PR budget, and teams working on safety and ethics to at least act like it matters. If you actually are concerned with risks you need time to first consider what you do upon successfully developing an AGI. How would you recognize it? Does it have rights? Would turning it off be murder? More importantly, how do you ensure that it doesn't escape controls you put in place, if that possible at all?

mr_toad 3y ago
For now, but in a decade or two that hardware will become cheaper and more widely available. What happens when people can run the equivalent of GPT-4 on a desktop GPU?
- _heimdall 3y ago
  The hardware market would very likely be much different if those years don't include a heavily funded AI research industry driving up innovation and demand for that hardware, but even if nothing changed we would at least be buying a decade or two for us to be better prepared and more thoughtful about any technological advancements we want to pursue with AI. Today the industry is driven entirely by the promise of future profits and the love of innovation with no real regard for actual risks that could be in play.
tavavex 3y ago
Given the ever-growing computing power, computation itself becoming cheaper and the extremely wide reach of information on the internet, the outcome is inevitable either way. Even if you time-traveled 10 years back and somehow erased all modern ML research from existence, there'd still be a point in time where, eventually, some guy in a garage would do the work. Eventually, the work we use massive server clusters for today would be run on consumer-grade computers, and someone would try their hand at it. It has always been the case that new technology doesn't ask for permission to exist. Even if we lived in this society that was extremely risk-averse and suspicious of any new developments, at some point the bar would get so low that, if companies refused to invest and do the research, individuals would.
- _heimdall 3y ago
  Are you effectively using something akin to a multiverse scenario to guarantee that one outcome is inevitable? By multiverse here I'm mainly pointing to the idea of branching of all possible outcomes, not strictly trying to say that the idea I'd about jumping to an actual different universe. If you could go back 10 years and erase all ML research, for example, there would be fundamental differences. The number of people that understand how those systems would be much smaller. Hardware would be different, for example we wouldn't have tensor chips and GPUs would still be primarily for playing Call of Duty. Cell phones, search, even more benign systems like traffic control on city streets would look different. There's simply way too much changed even with that one jump to possibly guarantee that we'd find ourselves in the same boat inventing AI or AGI on roughly the same timeline.