I’m so impressed thoroughly researched. you’ve been in some of the best approaches to this evolving and still unfinished area that’s often full of tricky nuances. .
This is such a great introduction to important issues. People need to be thinking about beyond the Clickbait trigger headlines about AI doom or saviour. How do we consider and make the best of AI reality rather than the fantasy?
They were also a couple of points I did want to slightly challenge or perhaps offer alternative viewpoint in a friendly way of course 😊
The first is on AI bias which is of course a really important topic there is not well understood by the public and you really explained the fundamentals of issue well .
And you also acknowledge the difficulties in defining what bias is so well.
I guess my thinking currently on this goes a bit like this it goes a bit further to say I wonder if the whole concept of an unbiased AI is a bit of a fallacy or unachievable ?
I think you mentioned in the interview that an unbiased AI should aim to hold all your points which I totally agree with you but thinking about this this is would seem practically impossible or unachievable .
A truly unbiased AI would not only need to hold and consider all views that we know of currently but all future views as well that might be possible in the future because otherwise you know it would just be different types of bias.
What Really worries me much more about the idea of an unbiased AI or in belief that a more an unbiased position is possible which is something certain media organises organisations claim, eg UK bbc news! is its potentially a way of avoiding scrutiny or indeed way of hiding bias behind a mask of unbiasedness which I would propose is practically in Impossible to achieve.
So for me, I would propose that even aiming for an unbiased model or claiming to have an unbiased model potentially very unhelpful and damaging for confidence and trust and even truth because I don’t think these are achievable things and claims to be alluding to this could be highly misleading .
I was also really glad you quoted the Google Nazi images diversity example I’ve used this myself to illustrate how potentially hypocritical illusions to unbiasedness can be ending up potentially doing the exact opposite and being incredibly biased
Anyway, I suppose what I’m saying is so I have a different suggestion for how we think about bias.
rather than believe that we can be unbiased or even that we can move in a less biased position which I would challenge, an alternative and perhaps more radical approach would be to Enforce transparency of all models of there bias, and not only that but to actually assume by default that all AI models and indeed all human beings are fundamentally biased, because each of these are limited systems so by definition cannot hold all points of view so by definition cannot be unbiased?
I think, if we actually drop the illusion to having none or little bias, and instead focus on regulation to enforce transparency of bias and just assume that there is bias and that we have an obligation or companies have an obligation to therefore reveal the biases of all those models then that also gives people choice about what models they use, which is a point you also made in a different context and I think you’re absolutely right about this and I actually think giving consumers choice by enforcing AI model transparency on bias is not only realistic but also inform peoples agency in regard to bias rather than believing we can have less bias AI models which I think is a pipe dream even worse potentially away hiding bias behind a mask of claiming unbias.
Second I wanted to challenge slightly is the argument and case about worrying about the increasing energy use of Ai.
For clarity I’d definitely agree that energy use of AI is increasing dramatically and it would make sense for us to reduce this so I completely agree with this .
What I would challenge is that the threat of this to the environment is overstated .
As some would say context is everything, and I think what can be misleading about the way AI energy use is framed that can be unhelpful is looking at increasing energy use of AI in isolation and of course in isolation the increases in energy use are quite dramatic but is that a fair analysis? if we are concerned about the impact on the environment?
I’m not sure that it is to me. It would seem a fair analysis is to put AI energy use in the context of other types of energy use and I have actually looked into several studies on this and that show actually if we look at global energy use and carbon emissions , that the AI use that contributes to carbon emissions from total carbon emissions is around 3%.
Yes of course that is kind of a new thing and you know it’s something we want to reduce but that kind of means that 97% of the global carbon emissions comes from any other sources such as industry and domestic use which I think roughly 30 and 40% of carbon emissions respectively roughly I’ve Get you the stats. I’ve quoted them in one of my articles if need to be but that’s roughly right..
I think this argument can be a bit misleading because it’s often used to frame AI as a bit of a bogeyman in terms of climate change and I think this is the bit that’s actually highly misleading because if we’re really serious about reducing the major sources of carbon emissions then surely it makes sense that we should focus most on the major causes of carbon emissions in the world and that is certainly not AI at 3% of global emissions right ?
I wonder if we should be focusing on reducing the major causes of carbon emissions which include the top two I believe is industry and domestic energy use modest reductions carbon emissions in these two sectors would have a huge effect on reducing carbon emissions compared to focusing all our worry and concern on the 3% of carbon emissions coming from AI.
This is what seems to be highly misleading about climate impact of AI. It’s affect is taken completely out of any context at all and in isolation and the numbers and the increase of energy use of AI is a course big and dramatic but only when considered an isolation not in the context of global carbon emissions of which it’s actually quite a small amount. .
The other aspects that I think is maybe a bit misleading is the sources of energy usage so for example if we are using renewable any sources for AI model training and usage such as solar wind or hydro then as much of a problem because of course these are contributing to carbon emissions and actually increasing Not all but some of these uses are turning to renewable sources simply because of cost actually if nothing else so I ironically even corporate greed and profit maybe helping to move AI training a bit more towards renewable sources.
So again, I think it could be a bit misleading to paint AI energy usage as a bit of a bogeyman. I think number one it’s a tiny fraction of carbon emissions so it doesn’t seem to be that it’s the major threat to climate change or should be necessarily our major worry compared to the other other major causes of carbon emissions.
And second not all energy use by AI is bad even if it’s a lot it really depends on the sources of energy obviously if it’s using fossil fuels that’s terrible and we want to reduce that obviously but actually increasingly not only AI training but you know all of our energy uses generally is going towards renewable in many countries across the west and the east so it would naturally also make sense that AI use just like all other energy Uses are generally moving towards renewable.
So I’m not sure it really makes sense that using more energy by default is a bad thing. It really depends on if the energy generation is generating carbon emissions which of course renewable energy which is increasing used every doesn’t so I can’t see why increasing energy usage by AI using renewable energy is such a bad thing unless of course it takes it away from domestic useful availability but that’s separate issue.
I’m just challenging the idea that AI increasing energy usage is by default a bad thing or should be a major source of concern for climate emissions.
Would love to hear your thoughts and responses on this 🙂🙏🏾
- Yes, I completely agree that transparency is a necessary and complementary proponent to the bias conversation. Achieving a perfect equilibrium within AI models is similar to striving to perfection in anything - resource intensive and an ever-evolving target.
- On your arguments around AI and climate usage, the context you provide is helpful though I haven't researched/substantiated the stats you use. What I will share is this Atlantic piece just published that speaks to AI's nuanced impacts on climate - because, to your point, if AI's impact specifically on carbon emissions is only 3%, then there should be other more pressing areas for us to focus on to reduce emissions specifically. And AI could, if implemented into certain applications, improve our intelligence around effective sustainable solutions. BUT - AI (and the frontier companies at its helm) can also be misused in this instance as demonstrated in the piece, if used by certain actors (i.e. Chevron) to advance unsunstainable practices: https://www.theatlantic.com/technology/archive/2024/09/microsoft-ai-oil-contracts/679804/?gift=lhL3dXSYCcu9vqTqEbg0ONK94WYTD2WAw8xP-bSGBLY&utm_source=copy-link&utm_medium=social&utm_campaign=share
I thank you for your comments and for the discussion! As we know - there aren't many definitive answers in this space, and the dialogue we have is essential to ensure people have diverse perspectives and research so that they can form their own views.
Yes good points e.g. about the Microsoft/Chevron case.
Here are the energy usage stats...
So the IEA (International energy agency):
'The IEA report said data centers, cryptocurrencies and AI combined used 460 terawatt-hours (TWh) of electricity worldwide in 2022, almost 2% of total global electricity demand.'
So yes, 2 years out of date unfort. Also, there are only stats for datacenters, which as we can see is not just AI but other use such as for crypto, but it doesnt say how much of each.
So I think 3% is probably a conservative estimate for energy use by AI now in 2024, its probably less given the stats for data centers include crypto etc.
What a fantastic podcast episode!
I’m so impressed thoroughly researched. you’ve been in some of the best approaches to this evolving and still unfinished area that’s often full of tricky nuances. .
This is such a great introduction to important issues. People need to be thinking about beyond the Clickbait trigger headlines about AI doom or saviour. How do we consider and make the best of AI reality rather than the fantasy?
They were also a couple of points I did want to slightly challenge or perhaps offer alternative viewpoint in a friendly way of course 😊
The first is on AI bias which is of course a really important topic there is not well understood by the public and you really explained the fundamentals of issue well .
And you also acknowledge the difficulties in defining what bias is so well.
I guess my thinking currently on this goes a bit like this it goes a bit further to say I wonder if the whole concept of an unbiased AI is a bit of a fallacy or unachievable ?
I think you mentioned in the interview that an unbiased AI should aim to hold all your points which I totally agree with you but thinking about this this is would seem practically impossible or unachievable .
A truly unbiased AI would not only need to hold and consider all views that we know of currently but all future views as well that might be possible in the future because otherwise you know it would just be different types of bias.
What Really worries me much more about the idea of an unbiased AI or in belief that a more an unbiased position is possible which is something certain media organises organisations claim, eg UK bbc news! is its potentially a way of avoiding scrutiny or indeed way of hiding bias behind a mask of unbiasedness which I would propose is practically in Impossible to achieve.
So for me, I would propose that even aiming for an unbiased model or claiming to have an unbiased model potentially very unhelpful and damaging for confidence and trust and even truth because I don’t think these are achievable things and claims to be alluding to this could be highly misleading .
I was also really glad you quoted the Google Nazi images diversity example I’ve used this myself to illustrate how potentially hypocritical illusions to unbiasedness can be ending up potentially doing the exact opposite and being incredibly biased
Anyway, I suppose what I’m saying is so I have a different suggestion for how we think about bias.
rather than believe that we can be unbiased or even that we can move in a less biased position which I would challenge, an alternative and perhaps more radical approach would be to Enforce transparency of all models of there bias, and not only that but to actually assume by default that all AI models and indeed all human beings are fundamentally biased, because each of these are limited systems so by definition cannot hold all points of view so by definition cannot be unbiased?
I think, if we actually drop the illusion to having none or little bias, and instead focus on regulation to enforce transparency of bias and just assume that there is bias and that we have an obligation or companies have an obligation to therefore reveal the biases of all those models then that also gives people choice about what models they use, which is a point you also made in a different context and I think you’re absolutely right about this and I actually think giving consumers choice by enforcing AI model transparency on bias is not only realistic but also inform peoples agency in regard to bias rather than believing we can have less bias AI models which I think is a pipe dream even worse potentially away hiding bias behind a mask of claiming unbias.
Second I wanted to challenge slightly is the argument and case about worrying about the increasing energy use of Ai.
For clarity I’d definitely agree that energy use of AI is increasing dramatically and it would make sense for us to reduce this so I completely agree with this .
What I would challenge is that the threat of this to the environment is overstated .
As some would say context is everything, and I think what can be misleading about the way AI energy use is framed that can be unhelpful is looking at increasing energy use of AI in isolation and of course in isolation the increases in energy use are quite dramatic but is that a fair analysis? if we are concerned about the impact on the environment?
I’m not sure that it is to me. It would seem a fair analysis is to put AI energy use in the context of other types of energy use and I have actually looked into several studies on this and that show actually if we look at global energy use and carbon emissions , that the AI use that contributes to carbon emissions from total carbon emissions is around 3%.
Yes of course that is kind of a new thing and you know it’s something we want to reduce but that kind of means that 97% of the global carbon emissions comes from any other sources such as industry and domestic use which I think roughly 30 and 40% of carbon emissions respectively roughly I’ve Get you the stats. I’ve quoted them in one of my articles if need to be but that’s roughly right..
I think this argument can be a bit misleading because it’s often used to frame AI as a bit of a bogeyman in terms of climate change and I think this is the bit that’s actually highly misleading because if we’re really serious about reducing the major sources of carbon emissions then surely it makes sense that we should focus most on the major causes of carbon emissions in the world and that is certainly not AI at 3% of global emissions right ?
I wonder if we should be focusing on reducing the major causes of carbon emissions which include the top two I believe is industry and domestic energy use modest reductions carbon emissions in these two sectors would have a huge effect on reducing carbon emissions compared to focusing all our worry and concern on the 3% of carbon emissions coming from AI.
This is what seems to be highly misleading about climate impact of AI. It’s affect is taken completely out of any context at all and in isolation and the numbers and the increase of energy use of AI is a course big and dramatic but only when considered an isolation not in the context of global carbon emissions of which it’s actually quite a small amount. .
The other aspects that I think is maybe a bit misleading is the sources of energy usage so for example if we are using renewable any sources for AI model training and usage such as solar wind or hydro then as much of a problem because of course these are contributing to carbon emissions and actually increasing Not all but some of these uses are turning to renewable sources simply because of cost actually if nothing else so I ironically even corporate greed and profit maybe helping to move AI training a bit more towards renewable sources.
So again, I think it could be a bit misleading to paint AI energy usage as a bit of a bogeyman. I think number one it’s a tiny fraction of carbon emissions so it doesn’t seem to be that it’s the major threat to climate change or should be necessarily our major worry compared to the other other major causes of carbon emissions.
And second not all energy use by AI is bad even if it’s a lot it really depends on the sources of energy obviously if it’s using fossil fuels that’s terrible and we want to reduce that obviously but actually increasingly not only AI training but you know all of our energy uses generally is going towards renewable in many countries across the west and the east so it would naturally also make sense that AI use just like all other energy Uses are generally moving towards renewable.
So I’m not sure it really makes sense that using more energy by default is a bad thing. It really depends on if the energy generation is generating carbon emissions which of course renewable energy which is increasing used every doesn’t so I can’t see why increasing energy usage by AI using renewable energy is such a bad thing unless of course it takes it away from domestic useful availability but that’s separate issue.
I’m just challenging the idea that AI increasing energy usage is by default a bad thing or should be a major source of concern for climate emissions.
Would love to hear your thoughts and responses on this 🙂🙏🏾
What a comment! A few responses to your notes -
- Yes, I completely agree that transparency is a necessary and complementary proponent to the bias conversation. Achieving a perfect equilibrium within AI models is similar to striving to perfection in anything - resource intensive and an ever-evolving target.
- On your arguments around AI and climate usage, the context you provide is helpful though I haven't researched/substantiated the stats you use. What I will share is this Atlantic piece just published that speaks to AI's nuanced impacts on climate - because, to your point, if AI's impact specifically on carbon emissions is only 3%, then there should be other more pressing areas for us to focus on to reduce emissions specifically. And AI could, if implemented into certain applications, improve our intelligence around effective sustainable solutions. BUT - AI (and the frontier companies at its helm) can also be misused in this instance as demonstrated in the piece, if used by certain actors (i.e. Chevron) to advance unsunstainable practices: https://www.theatlantic.com/technology/archive/2024/09/microsoft-ai-oil-contracts/679804/?gift=lhL3dXSYCcu9vqTqEbg0ONK94WYTD2WAw8xP-bSGBLY&utm_source=copy-link&utm_medium=social&utm_campaign=share
I thank you for your comments and for the discussion! As we know - there aren't many definitive answers in this space, and the dialogue we have is essential to ensure people have diverse perspectives and research so that they can form their own views.
Yes good points e.g. about the Microsoft/Chevron case.
Here are the energy usage stats...
So the IEA (International energy agency):
'The IEA report said data centers, cryptocurrencies and AI combined used 460 terawatt-hours (TWh) of electricity worldwide in 2022, almost 2% of total global electricity demand.'
https://www.dailysabah.com/business/tech/is-ai-major-drain-on-worlds-energy-supply-consumption
So yes, 2 years out of date unfort. Also, there are only stats for datacenters, which as we can see is not just AI but other use such as for crypto, but it doesnt say how much of each.
So I think 3% is probably a conservative estimate for energy use by AI now in 2024, its probably less given the stats for data centers include crypto etc.