The basis of measurement to limit us are flops, so large integer operations are fine? So we could just move the decimal point and nothing is illegal anymore?
This touches a lot of issues I haven’t heard discussed and that alone gives me som hope. Though if we look at history there’s not a lot of examples of large public companies choosing well-being over profit
@@nathanlabenz it might just be me it just came across a bit sycophanty up to the 20 minute mark. one thought though, as for AI assisted biological weapon development, that seems a pretty extreme stretch, I'd think of that application as more a learning aide as they don't reason through a problem, without the context of an expert the unknown unknowns can often be unaddressed. where a remedy for such a problem would be comparatively trivial to apply the same technology to sentiment analysis of social media or credit card purchases, to determine who might be worthy of checking on, anyone disturbed enough for such an effort would have to be putting out some large social ripples.. yeah I thought the conversation was insightful overall, I don't see the value in limiting development at this stage though, personally I think that intelligence and kindness go hand in hand, I'm much more concerned about more flexible architectures than scale.. anyways 😆
Once you build an illegal model, wouldn't it be incredibly easy to prune all the useless branches then train it as a legal model absent the pruned limbs which would've made it illegal? How many times do you think a team could do that before anyone could prove you used too much compute during training?
Just a point of feedback. It feels like the editing in this video is somehow automated where everytime there is a pause something gets cut, but this makes it super hard to follow when listening to this. There are so many cuts that it feels jarring.
This makes no sense to me at all... Whats the difference between one model and a family of models? Who's tracking which matrix multiply accumulate process is for which? I mean it's convention to do them all on clusters near each other, but there's nothing to identify one stream process from another over ssh. And for that matter, what's the penalty? If it's like the banks where $100 mln in crimes costs $20 mln in fines then we're just helping a small group with regulatory capture. And what about an open source desktop resource donation, are we going to consider 3 bln individuals worldwide liable for violating a California bill if no one person ever set up the training or crossed the threshold? No lawmakers today is competent enough to write this law. No AI giant should write this law to create barriers of entry for others like every other American leader has. How do you make doing math illegal, anyway? What's the actual crime perpetrated by Bayesian inference? If someone is capable of reverse engineering these models to police this system would they not be better serving the public designing these systems instead of policing researchers? Which of the supreme court judges who can't think critically today about 19th century constitutional law would be suited to preside over a trial of PhD's spread across 5 programming languages, jurisdictions internationally, or perhaps even in lawless seas? There are so many irrefutable arguments against this bill it's astonishing its even being entertained imho. What's the difference between an AI model and a machine learning architecture or fancy sorting algorithm? Flops don't infer function, lawmakers should know that if they're actively trying to outlaw inference functions lol Great conversation, but absurd premise to defend.
Sad if states and big companies try to regulate away open source and freedom so it's untenable to develop AI outside of the big companies. Actually, it's disgusting. Especially to use people's low knowledge about it and emotionally blackmail them about scary big AI destroying them! Meanwhile, these actors are working for things that will badly affect people.
Worst take I've ever heard: we need to regulate AI in a way to benefits big companies so they disproportionally capture the market. The level of money and barriers needed to cross will be the death of democracy.
Don't get loud about less wrong wanting their words heard, foo. (brown-haored guy with skant plants) We've all been thinking about this process far longer than you. It takes time to absorb such thoughts. We shouldn't be the lone decision maker but we should be of the prime leaders. (tech hates when scrubs come up and want to leapfrog)
Absolutely fascinating discussion guys, we need more of this.
The basis of measurement to limit us are flops, so large integer operations are fine? So we could just move the decimal point and nothing is illegal anymore?
This touches a lot of issues I haven’t heard discussed and that alone gives me som hope. Though if we look at history there’s not a lot of examples of large public companies choosing well-being over profit
that intro had me concerned 😆 interesting chat though, thanks
How so?
@@nathanlabenz it might just be me it just came across a bit sycophanty up to the 20 minute mark. one thought though, as for AI assisted biological weapon development, that seems a pretty extreme stretch, I'd think of that application as more a learning aide as they don't reason through a problem, without the context of an expert the unknown unknowns can often be unaddressed. where a remedy for such a problem would be comparatively trivial to apply the same technology to sentiment analysis of social media or credit card purchases, to determine who might be worthy of checking on, anyone disturbed enough for such an effort would have to be putting out some large social ripples.. yeah I thought the conversation was insightful overall, I don't see the value in limiting development at this stage though, personally I think that intelligence and kindness go hand in hand, I'm much more concerned about more flexible architectures than scale.. anyways 😆
Once you build an illegal model, wouldn't it be incredibly easy to prune all the useless branches then train it as a legal model absent the pruned limbs which would've made it illegal? How many times do you think a team could do that before anyone could prove you used too much compute during training?
Just a point of feedback. It feels like the editing in this video is somehow automated where everytime there is a pause something gets cut, but this makes it super hard to follow when listening to this. There are so many cuts that it feels jarring.
This makes no sense to me at all...
Whats the difference between one model and a family of models? Who's tracking which matrix multiply accumulate process is for which? I mean it's convention to do them all on clusters near each other, but there's nothing to identify one stream process from another over ssh. And for that matter, what's the penalty? If it's like the banks where $100 mln in crimes costs $20 mln in fines then we're just helping a small group with regulatory capture. And what about an open source desktop resource donation, are we going to consider 3 bln individuals worldwide liable for violating a California bill if no one person ever set up the training or crossed the threshold?
No lawmakers today is competent enough to write this law. No AI giant should write this law to create barriers of entry for others like every other American leader has.
How do you make doing math illegal, anyway? What's the actual crime perpetrated by Bayesian inference? If someone is capable of reverse engineering these models to police this system would they not be better serving the public designing these systems instead of policing researchers? Which of the supreme court judges who can't think critically today about 19th century constitutional law would be suited to preside over a trial of PhD's spread across 5 programming languages, jurisdictions internationally, or perhaps even in lawless seas?
There are so many irrefutable arguments against this bill it's astonishing its even being entertained imho. What's the difference between an AI model and a machine learning architecture or fancy sorting algorithm? Flops don't infer function, lawmakers should know that if they're actively trying to outlaw inference functions lol
Great conversation, but absurd premise to defend.
This bill is silly in a dozen ways, and I think Russia loves it
TLDR? I should trust business to regulate math? Open source is dangerous?
Sad if states and big companies try to regulate away open source and freedom so it's untenable to develop AI outside of the big companies. Actually, it's disgusting. Especially to use people's low knowledge about it and emotionally blackmail them about scary big AI destroying them! Meanwhile, these actors are working for things that will badly affect people.
Worst take I've ever heard: I'd like a major AI incident that results in deaths before we regulate AI.
Worst take I've ever heard: we need to regulate AI in a way to benefits big companies so they disproportionally capture the market. The level of money and barriers needed to cross will be the death of democracy.
Ok so 2:02, it's obviously a power grab. gj. vote no.
Don't get loud about less wrong wanting their words heard, foo. (brown-haored guy with skant plants) We've all been thinking about this process far longer than you. It takes time to absorb such thoughts. We shouldn't be the lone decision maker but we should be of the prime leaders. (tech hates when scrubs come up and want to leapfrog)