[00:00.000 --> 00:00.600]  Awesome.
[00:00.860 --> 00:01.580]  My name is Roll Time.
[00:01.840 --> 00:05.980]  This talk is called AI, Solar Punk, and an Uncertain Future in Computing.
[00:06.480 --> 00:13.600]  Before I get started with it, I want to thank everybody at NYC Resistor, Free Radical, the Fediverse.
[00:14.040 --> 00:20.960]  Thank you all for being here, and thank you to everyone in my life who helped me out in the frantic two weeks leading up to this.
[00:21.500 --> 00:24.540]  So, now that you're all, like, placated, let's talk about me.
[00:25.080 --> 00:26.200]  My name is Roll Time.
[00:26.600 --> 00:30.300]  I'm a hacker, a software engineer, and an inveterate optimist.
[00:30.660 --> 00:40.660]  About a year ago, I finished my bachelor's degree, I wrote my thesis on Solar Punk and AI, and I am now the lead engineer at a startup where I do neither of those things.
[00:41.840 --> 00:45.580]  Despite that fact, I want to bring my findings from my BA back here.
[00:45.840 --> 00:53.360]  It is my hope that all of you leave this talk with newfound understanding and passion, and hopefully you learn something new.
[00:54.080 --> 00:56.360]  This actually isn't my first hope talk.
[00:56.640 --> 01:07.020]  Two years ago, at A New HOPE, I spoke about ActivityPub, the Federated Social Networking Protocol, which was thrust into the spotlight following Elon Musk's acquisition of Twitter.
[01:07.780 --> 01:12.620]  I spoke about my vision for a better kind of social media and how I felt we could get there.
[01:12.800 --> 01:14.600]  And that talk was fairly popular.
[01:14.600 --> 01:17.440]  It got a lot of feedback in the form of YouTube comments.
[01:17.960 --> 01:22.440]  I really value feedback, so I want to take the time to review some of that feedback now.
[01:23.320 --> 01:29.260]  So, at RedNafi said, thank you for explaining the core architecture of a Federated System in a single slide.
[01:29.400 --> 01:30.560]  That was incredibly helpful.
[01:31.120 --> 01:32.300]  I'm glad to hear someone say that.
[01:32.360 --> 01:35.340]  It took a lot of time to make that diagram, so I'm glad it helped somebody out.
[01:35.940 --> 01:37.540]  Forrest Wilkins says, love this.
[01:37.680 --> 01:38.080]  Great talk.
[01:38.220 --> 01:38.840]  Thank you, Forrest.
[01:39.760 --> 01:47.540]  At SegFault01 says, it's unfortunate that the speaker sprayed his political slash ideological view on the audience.
[01:47.760 --> 01:49.600]  But apart from that, it was a good talk.
[01:52.100 --> 01:54.080]  So, I value feedback.
[01:54.500 --> 02:06.720]  And I want to let everybody know, it is solely for the reason of this comment, that this talk is about radical social movements, ending capitalism, and curbing climate change.
[02:07.260 --> 02:09.280]  SegFault01, this one is on you.
[02:10.760 --> 02:13.060]  You have only yourself to thank.
[02:16.520 --> 02:17.040]  Yeah.
[02:18.480 --> 02:21.200]  Okay, but what are we going to blame SegFault01 for?
[02:21.980 --> 02:27.860]  I think all of you might agree with me when I say that our relationship with technology is broken.
[02:28.160 --> 02:28.240]  Right?
[02:28.360 --> 02:36.380]  And not just technology as it exists, but how we develop it, how we let it influence our lives, the role we let it play in our day-to-day experiences.
[02:37.540 --> 02:38.880]  So, we're going to dive into that.
[02:39.080 --> 02:46.140]  We're going to talk about a key example of a problem in terms of modern AI landscape, such as it is.
[02:46.380 --> 02:51.320]  We're going to talk about a beautiful future, as imagined by solar punk and related aesthetics.
[02:51.920 --> 02:56.660]  We're going to talk about real-world applications of solar punk and other wonderful things.
[02:57.200 --> 03:03.340]  And we're going to talk about challenging the AI mainstream with a really cool new technology called Approximate Multipliers.
[03:04.160 --> 03:07.860]  But before we can get into the good, let's talk about the bad.
[03:08.120 --> 03:12.080]  The never-ending bubble, the last four years of hell.
[03:12.320 --> 03:21.020]  So, in case you have been living under the largest rock since the crypto bubble, let's get into a very brief history, very brief, of how all this worked out.
[03:21.420 --> 03:22.760]  The year is 2016.
[03:23.180 --> 03:33.660]  The state-of-the-art is recurrent neural networks for sequential modeling tasks like text, CNNs for image classification, and GANs, Generative Artificial Networks for image generation.
[03:33.920 --> 03:41.500]  This was right about the time when GANs started generating the first photorealistic images of human faces, which at the time was considered pretty scary.
[03:42.560 --> 03:53.200]  In 2017, Ashish Faswani and a team at Google DeepMind released the Transformer model, and it swiftly became the state-of-the-art for all of the tasks that I just talked about.
[03:53.420 --> 04:03.440]  And it's not that important that you understand exactly how Transformers work, but all you need to know is that they achieved unprecedented performance and also unprecedented scaling.
[04:03.440 --> 04:13.380]  So by increasing the amount of data you could feed to a model, by increasing the amount of hardware you ran the Transformer on, you could also scale up the performance for your task.
[04:14.740 --> 04:24.740]  OpenAI was quick to realize exactly what Transformers could do, and through GPT-1, 2, and eventually 3, they created the first models that you might call Cogent.
[04:25.060 --> 04:28.720]  And the second they had something that could produce a real sentence, they went for-profit.
[04:31.220 --> 04:34.920]  2021 saw a spate of succession models.
[04:35.020 --> 04:38.460]  AlphaFold 2 made some really interesting strides in microbiology.
[04:38.720 --> 04:42.700]  Stable Diffusion showed us how to generate images from a single line of text.
[04:43.260 --> 04:47.300]  And things really took off in 2023 with models like GPT-4.
[04:47.580 --> 04:52.380]  GPT-4, of course, the first robot to pass the bar exam, unless you count Alan Dershowitz.
[04:53.880 --> 05:01.220]  Swiftly followed by Anthropic, Anthropic's Cloud 2, Facebook's Llama, which is the best-known open model.
[05:01.380 --> 05:07.960]  So if you have a cheap source of electricity and, like, four GPUs, you can run Llama on your own machine.
[05:08.680 --> 05:11.140]  In 2024, we saw some interesting stuff.
[05:11.500 --> 05:15.900]  GPT-4-0 was released a few months ago, a smaller, lighter version of GPT-4.
[05:16.220 --> 05:20.500]  Cloud 3.5 made some really interesting strides in programming skills.
[05:20.500 --> 05:29.560]  And AlphaFold 3 continued to be pretty much the only useful model, once again setting the state of the art for protein folding and other microbiology tasks.
[05:30.040 --> 05:35.640]  So where we are now in 2024 is that the entire field is soaked in cash.
[05:36.560 --> 05:40.720]  AI Ventures attracted $96 billion of investment in 2023.
[05:40.720 --> 05:52.160]  The favorite fact that I'd like to share about this is that in 2021, the peak of the crypto bubble, when all that anyone could talk about was NFTs, Crypto Ventures received $30 billion in private investment.
[05:52.540 --> 05:56.500]  The same year, AI companies received $42 billion in private investment.
[05:56.820 --> 06:01.660]  This isn't a new phenomenon, it's a bubble that's been building and building for four years.
[06:01.660 --> 06:10.300]  Mostly to the benefit of NVIDIA, who briefly became the world's most valuable company earlier this year, with a market cap of over $3 trillion.
[06:10.980 --> 06:11.940]  An obscene number.
[06:12.860 --> 06:16.460]  For all of that, we have an answer about, is it useful?
[06:17.140 --> 06:17.760]  Eh, kind of.
[06:17.880 --> 06:18.160]  I don't know.
[06:18.840 --> 06:22.020]  About 30% of people use ChatGPT on a weekly basis.
[06:22.580 --> 06:24.700]  That is a lot of people, I want to be clear.
[06:25.140 --> 06:31.380]  ChatGPT is actually the product with the second fastest user growth in history behind threads.
[06:32.020 --> 06:36.200]  That number, by the way, the 30%, is probably higher in the U.S.
[06:37.280 --> 06:39.220]  So, 30% is a global estimate.
[06:39.220 --> 06:45.000]  In the U.S., where we live increasingly networked lives, it's probably more than 30% of people.
[06:45.200 --> 06:51.640]  Most of this information, by the way, comes from the Stanford AI Index Report, which is released yearly.
[06:51.800 --> 06:54.100]  The 2024 version came out a couple months ago.
[06:54.420 --> 06:56.220]  A lot of interesting information in there.
[06:56.220 --> 07:04.820]  But the thing I find most interesting is not anything inside the index report itself, but the ever-growing page count of the report.
[07:05.580 --> 07:15.140]  You can see if we plot it against this exponential regression, we can project that the AI Index Report may yet reach a length of 1,500 pages by 2030.
[07:16.780 --> 07:20.080]  Unfortunately, past performance is not a predictor of future results.
[07:20.860 --> 07:26.040]  Many people are now saying what artists have been saying for months.
[07:26.220 --> 07:28.620]  Hey, this doesn't work that well.
[07:29.300 --> 07:32.180]  Um, and it's common to see people complaining.
[07:32.180 --> 07:38.580]  There's a really good article, uh, called, I Will Fucking Piledrive You If You Mention AI Again by someone called Ludicity.
[07:38.980 --> 07:39.280]  Yeah!
[07:40.340 --> 07:41.700]  Uh, just to share a quote.
[07:41.700 --> 07:56.540]  Quote, I'm going to ask ChatGPT how to prepare a garrote, and then I am going to strangle you with it, and you will simply have to pray that I rolled the 10% chance that it freaks out, and tells me that a garrote should consist entirely of papier-mâché and malice.
[07:57.660 --> 08:02.660]  Um, so, AI companies may be not winning the war of public discussion right now.
[08:02.780 --> 08:14.320]  In fact, even Sequoia Capital, uh, well-known for perpetrating or monetarily enabling FTX, the largest fraud of the 21st century so far, uh, are claiming that AI is unsustainable.
[08:14.320 --> 08:22.060]  They've released an article saying that it'll need to generate an additional $600 billion of revenue in the next few years to become profitable.
[08:22.420 --> 08:33.820]  When you consider that OpenAI, by far the largest player in the AI in space with a lion's share of revenue, only pulled down less than $4 billion last year, you start to understand the scale of the problem.
[08:34.620 --> 08:36.580]  So, AI's not making anybody any money.
[08:37.180 --> 08:38.780]  How's it doing for the environment?
[08:39.900 --> 08:41.320]  It's pretty bad, guys.
[08:41.320 --> 08:46.020]  Um, this is a mad scramble to convert billions of dollars into waste heat.
[08:46.260 --> 08:55.780]  And just to quantify that, one person flying from New York to Los Angeles results in a staggering, like, absolutely staggering two metric tons of CO2 emissions.
[08:56.380 --> 08:57.200]  That's a lie.
[08:57.540 --> 09:06.480]  Trading GPT-4 cost OpenAI $100 million in hardware and electricity, and it produced the equivalent of 2,500 of these flights.
[09:06.480 --> 09:15.340]  That's 156 person years of CO2 for the average American, for one model over the course of just a few months.
[09:16.280 --> 09:22.880]  It's also the lifetime carbon emissions of 100 cars, from production to 20 years of driving to scrap.
[09:23.720 --> 09:29.300]  Training is energy intense, but depending on model lifecycle, inference can be an even bigger problem.
[09:29.300 --> 09:38.460]  In 2023, it's estimated that ChatGPT was consuming about half a gigawatt hour of power every single day to serve 200 million requests.
[09:38.720 --> 09:49.440]  If we account for user growth and the larger models that are now in use at OpenAI, we can make a conservative estimate that ChatGPT is producing more than 100 flights per day of CO2.
[09:51.360 --> 10:01.120]  Since 2019, Google's carbon emissions are up 48%, mostly due to new data center demand, and they've directly stated that AI is at least partially responsible.
[10:01.460 --> 10:01.700]  I quote,
[10:09.180 --> 10:14.760]  They've also claimed that their AI-powered search is 10 times more expensive than the standard version.
[10:15.020 --> 10:24.220]  Based on that higher cost, we can estimate that if Google were to use AI for every search, the system's power usage might approach 30 terawatt hours per year.
[10:24.640 --> 10:30.480]  To place that number in context, I've often heard that AI is using as much power as a small country.
[10:30.480 --> 10:32.000]  And that's actually false.
[10:32.300 --> 10:33.320]  It's as much as a large one.
[10:34.240 --> 10:52.860]  30 terawatt hours per year has Google alone tied with Bulgaria as the 66th largest consumer of electricity in the world, just below Ireland, producing 1.8 megatons of CO2 per year, or almost 2,500 flights per day, and all so that it can tell you to eat glue.
[10:55.360 --> 10:59.180]  These numbers show a real, meaningful impact on the climate.
[10:59.180 --> 11:05.220]  That's scary, especially at a time when reducing carbon emissions should be humanity's number one priority.
[11:05.640 --> 11:08.040]  All of which raises the question...
[11:08.960 --> 11:11.660]  All of which raises the question, what do we do about it?
[11:12.280 --> 11:17.240]  I feel like there's this urge, especially amongst younger people, to say, Why don't we just stop?
[11:17.600 --> 11:19.280]  Why don't we just ban AI?
[11:20.300 --> 11:22.360]  Well, unfortunately, the cat is out of the bag.
[11:22.840 --> 11:29.640]  In our modern, globally interwoven, technologically advanced society, there's no way to effectively ban AI.
[11:29.980 --> 11:34.640]  I want to be clear, I'm a proponent of regulation, but that's really the most we can hope for.
[11:35.200 --> 11:43.180]  In a similar vein, a lot of people seem to think that the solution to climate change is radical de-urbanization.
[11:43.180 --> 11:46.100]  Why don't we all just go live in the country?
[11:46.340 --> 11:47.860]  Why don't we all just go somewhere else?
[11:49.000 --> 11:51.500]  I think this is a dangerous set of ideas.
[11:51.500 --> 11:58.440]  It's part of a wider body of thought, which I'm gonna call neopastoral anarcho-primitivist cottagecore bullshit.
[12:00.560 --> 12:01.000]  Okay.
[12:01.540 --> 12:05.880]  And I understand there's this human urge to just fuck off into the woods.
[12:06.920 --> 12:09.960]  Grow your own food in a garden and be completely self-sufficient.
[12:10.120 --> 12:12.380]  We're all gonna live in isolated rural communities.
[12:12.560 --> 12:13.840]  We're gonna be self-sufficient.
[12:14.000 --> 12:14.880]  We're gonna be off-grid.
[12:15.900 --> 12:20.860]  That's unrealistic, to say the least, because if we all move to the country, it won't be the country anymore.
[12:21.300 --> 12:25.220]  I also want to point out that density isn't inherently bad for the environment.
[12:25.220 --> 12:33.400]  The way that capitalism has used density to increase profit margins is, but density can make it easier to produce enough stuff for everyone.
[12:34.400 --> 12:44.340]  This human urge for a simpler, greener time or a cottage with your lesbian lover with a garden is, I think, very common, but it gets pretty dangerous.
[12:44.960 --> 12:51.560]  It's impossible to disentangle these urges from histories of racism and exploitation in the United States.
[12:51.560 --> 13:07.080]  And just to give you one example, in the 1950s, it's that same urge for greener pastures that drove white, rich Americans to leave cities for newly created suburbia in the hopes that it would solve the fact that they were unhappy.
[13:10.140 --> 13:20.320]  White flight drove cities into debt spirals and prevented the black Americans who couldn't amass the money to leave from amassing meaningful wealth post-segregation.
[13:21.020 --> 13:24.200]  It's hollowed out the cores of our cities and destroyed our infrastructure.
[13:25.240 --> 13:27.860]  This isn't the first time we've done something like this.
[13:28.480 --> 13:31.500]  You can look back at the 1920s... uh, 1820s, pardon me.
[13:31.840 --> 13:40.400]  Andrew Jackson believed that white yeoman farmers would settle the empty Midwest, forming green, pastoral, self-sufficient communities.
[13:40.400 --> 13:43.840]  That was Andrew Jackson's dream, and it required genocide.
[13:44.680 --> 13:46.220]  We need a different dream.
[13:46.840 --> 13:51.040]  We need to believe in a world where the practical merges with the beautiful.
[13:51.540 --> 13:56.880]  Where dense urban environments thrive with lush greenery and sustainable food systems.
[13:57.180 --> 14:01.400]  Where optimism isn't a fallacy, but the path to a brighter future.
[14:01.400 --> 14:02.980]  That's why we need solar punk.
[14:04.160 --> 14:08.660]  Solar punk is what happens when we dare to ask the question, what if we could do better?
[14:09.380 --> 14:11.040]  It's a movement that dares to hope.
[14:12.200 --> 14:15.820]  It happens as a response to cyber punk in the late 2010s.
[14:15.980 --> 14:18.640]  If cyber punk is dystopian, solar punk is utopian.
[14:18.920 --> 14:20.520]  Or at least attempts to be.
[14:21.460 --> 14:25.080]  Solar punk holds that science fiction is a form of activism.
[14:25.460 --> 14:32.220]  That beautiful solutions are better because they inspire out-of-the-box thinking, which brings better lives to us all.
[14:32.940 --> 14:34.860]  Solar punk is post-capitalist.
[14:35.140 --> 14:36.460]  It's counterculture, right?
[14:36.720 --> 14:41.860]  In solar punk, to progress is to move past competition and towards collaboration.
[14:42.900 --> 14:50.720]  So I guess my question to all of you is, what do you get when you combine radical optimism, out-of-the-box thinking, and counterculture?
[14:51.080 --> 14:56.540]  What do you get when you form communities centered around bold new ideas and inspired problem-solving?
[14:57.040 --> 14:58.140]  You get hackers.
[14:59.120 --> 15:02.760]  Solar punk is how we hack our way towards a better world.
[15:03.100 --> 15:04.020]  So let's do it.
[15:04.360 --> 15:10.740]  I'm going to share with you guys some absolutely beautiful art made by solar punk, and then we're going to get into some cool technology.
[15:11.000 --> 15:12.220]  This is Heliostat No.
[15:12.300 --> 15:14.080]  1 by Stella Jaku.
[15:14.640 --> 15:18.740]  It is a DIY heliostat that Stella built at NYC Resistor.
[15:18.740 --> 15:28.460]  Over the course of a few months, it acts to redirect sunlight into her bedroom, which calls attention to how sunlight can be a privileged resource in dense urban New York City.
[15:28.940 --> 15:30.260]  This is something I made.
[15:30.260 --> 15:31.880]  It's called the Farming Web of Trust.
[15:31.880 --> 15:36.200]  It's a soil moisture sensor attached to a LoRa-enabled microchip.
[15:36.600 --> 15:45.260]  It gives farmers the opportunity to share information about weather conditions with each other without relying on the Internet or centralized services.
[15:45.980 --> 15:48.980]  Solar punk is also a genre of music called so-fi.
[15:50.120 --> 15:59.360]  This video was published by Acorn Land Labs, and they make some really wonderful videos about their permaculture setup and how they farm off the grid.
[15:59.520 --> 16:00.340]  It's a wonderful channel.
[16:00.720 --> 16:02.460]  But let's look at some tech.
[16:03.040 --> 16:04.120]  Some bits and bytes.
[16:04.440 --> 16:05.940]  More specifically, let's hack the web.
[16:07.300 --> 16:08.760]  First, we're going to play a little bit of a game.
[16:09.720 --> 16:11.300]  It's an interview question, I guess.
[16:11.640 --> 16:12.640]  I pull out my phone.
[16:12.800 --> 16:13.820]  I go to facebook.com.
[16:14.120 --> 16:15.000]  I hit enter.
[16:15.720 --> 16:18.300]  There's a long series of events, right?
[16:18.300 --> 16:39.920]  From DNS to BGP to rattling around Facebook's internal network, meeting a server that receives my request, and then aggregates data from anywhere between 50 and 200 other servers, collecting, sorting, de-anonymizing, and algorithmically optimizing data in order to get it to me as fast as possible.
[16:40.200 --> 16:42.280]  For Facebook, speed is profit.
[16:42.540 --> 16:49.260]  The shorter the time between I hit enter and when I see some stupid political bullshit, the more money they make from me.
[16:50.980 --> 16:52.580]  What if we thought about this a different way?
[16:53.220 --> 17:00.260]  Instead of centering our designs around latency, around speed, what if we designed in an energy-centric way?
[17:01.040 --> 17:02.320]  We would get the Solar Protocol.
[17:02.960 --> 17:12.300]  And I don't want you guys to do this right now, but at some point I want you guys to go on your phone or your computer and type in solarprotocol.net, that's solarprotocol.net.
[17:12.800 --> 17:15.060]  When you hit enter, you get DNS.
[17:15.920 --> 17:27.420]  Then rather than getting bounced around a data center, you get redirected to one of 20 to 30 small microcomputers, globally distributed and all solar-powered.
[17:27.880 --> 17:37.480]  Every couple of minutes, these microcomputers talk amongst each other, discuss who has the most available solar energy, and nominate a leader.
[17:37.740 --> 17:41.360]  And that leader receives all of the requests to solarprotocol.net.
[17:41.600 --> 17:44.330]  What that means is, the Solar Protocol...
[17:44.780 --> 17:52.880]  If I'm accessing the Solar Protocol website at midnight in the United States, I might be served from Kenya or somewhere else where the sun is shining.
[17:52.880 --> 18:02.980]  The Solar Protocol is energy-centered, because it places the energy that is available to the servers as the primary method of optimizing traffic.
[18:03.340 --> 18:04.320]  It's planet scale.
[18:04.540 --> 18:12.780]  And what that means is, that while any one server at some point might have enough shade that it goes offline, the sun is always shining somewhere.
[18:12.980 --> 18:14.900]  And the Internet is a global network.
[18:15.100 --> 18:20.940]  The Solar Protocol can take advantage of geographical distribution in order to increase resiliency.
[18:20.940 --> 18:23.220]  It's naturally intelligent.
[18:23.620 --> 18:36.260]  Because this network communicates in a way that's based around collaboration, we can give ourselves the best opportunity for efficient, energy-centered design.
[18:36.480 --> 18:45.800]  The Solar Protocol is an exercise in permacomputing, which is computing that seeks to keep itself sustainable, and even carbon-neutral or carbon-negative if possible.
[18:46.980 --> 18:48.460]  Here's a really cool diagram.
[18:48.460 --> 18:57.460]  I recognize that it's a little hard to see, but the Solar Protocol has servers in Canada, the U.S., China, Chile, Kenya, other countries around the world.
[18:57.660 --> 18:59.700]  And this is a graph of four of them.
[18:59.860 --> 19:01.720]  What you can see is... Can I do this?
[19:01.820 --> 19:02.360]  I can't do that.
[19:03.200 --> 19:09.540]  Over the past 72 hours, solar energy has been distributed across the globe in this specific bar chart.
[19:09.540 --> 19:16.500]  You can see when I created this... or when I downloaded this graph later last night than I'd really like to admit to any of you.
[19:17.200 --> 19:22.300]  When I downloaded this graph, we were just coming off of sunset in the East Coast.
[19:22.500 --> 19:23.320]  So you can see that here.
[19:23.700 --> 19:25.600]  The sun is rising in Kenya.
[19:25.780 --> 19:34.620]  And so if I had accessed the site about two hours later, I might have been redirected to the Kenyan server because it was the best one poised to serve me.
[19:37.220 --> 19:39.800]  How can we apply energy-centered design to AI?
[19:40.900 --> 19:42.660]  That brings us to approximate multipliers.
[19:42.920 --> 19:45.080]  Essentially, we're taking the logic of the Solar Protocol.
[19:45.300 --> 19:48.660]  We're disincentivizing traditional motivation for profit.
[19:48.660 --> 19:53.060]  And we're asking, how do we make technology revolve around energy?
[19:53.620 --> 19:55.940]  This is a summary of my BA thesis.
[19:56.220 --> 19:59.880]  And I think to start, we really have to ask, what do we mean by efficient?
[20:00.360 --> 20:00.480]  Right?
[20:00.800 --> 20:02.320]  There's a lot of ways to define efficient.
[20:02.320 --> 20:09.340]  We're concerned about power usage, but somebody else might be concerned about reducing the amount of memory that your AI model uses.
[20:09.340 --> 20:11.700]  Because that reduces the amount of money you have to spend on hardware.
[20:12.380 --> 20:17.120]  It increases the number of customers you can serve with a given amount of memory.
[20:17.340 --> 20:19.420]  We might be concerned about reducing compute.
[20:19.700 --> 20:25.820]  So reducing the number of floating-point multiplication operations necessary to train and infer on the network.
[20:26.140 --> 20:29.180]  That increases speed and it reduces power requirements.
[20:29.820 --> 20:34.840]  So some of those techniques from the EI mainstream, we're going to talk about quantization and pruning.
[20:35.460 --> 20:51.280]  Quantization is a way of reducing the memory required by a network by reducing the amount of... reducing the bit width of floating-point numbers we use to store weights, which isn't a thing you need to understand beyond the fact that it reduces the amount of memory a model requires to operate.
[20:52.040 --> 20:59.360]  Pruning, on the other hand, doesn't affect the weights themselves, it just sets some of the weights to zero, so that we don't have to worry about them.
[20:59.580 --> 21:04.260]  Similarly to how the brain operates, we can trim unneeded neurons while maintaining performance.
[21:06.440 --> 21:07.760]  There's some problems here.
[21:08.560 --> 21:16.680]  What they both have in common is that they attempt to reduce the amount of multiplication done, or reduce the computational cost of multiplication in software.
[21:17.100 --> 21:18.760]  They also require retraining.
[21:18.920 --> 21:31.280]  What that means is, if you're going to make a model quantized or pruned, you might have to execute some additional training steps and basically plan ahead in order to give yourself the best chance of a successful result.
[21:32.700 --> 21:35.160]  Let's talk about multiplication in a little more detail.
[21:35.740 --> 21:39.580]  And consider, if you will, from elementary school, long multiplication.
[21:39.920 --> 21:44.000]  To multiply these two three-digit numbers requires nine operations.
[21:44.340 --> 21:48.220]  Multiplication is actually one of the most inefficient things computers do every day.
[21:48.220 --> 21:49.200]  It's n squared.
[21:49.560 --> 21:50.860]  It's quadratic complexity.
[21:52.220 --> 21:57.040]  Now, let me tell you that 99% of the compute in a deep neural network is multiplication.
[21:58.220 --> 22:06.960]  So if we could find a way to make multiplication a little more efficient, we could probably vastly reduce the power requirements of neural networks.
[22:08.680 --> 22:11.360]  Recall, if you will, from high school math, we're bumping you up a grade.
[22:11.360 --> 22:12.800]  Log multiplication.
[22:13.260 --> 22:14.960]  This is the logarithm product rule.
[22:15.720 --> 22:18.560]  And this idea gives rise to approximate multipliers, right?
[22:18.660 --> 22:24.780]  Because it turns out that multiplying log of a times b is equal to the sum of log a and log b.
[22:26.200 --> 22:29.340]  We can calculate binary log very quickly on a computer.
[22:29.540 --> 22:31.100]  It's basically a bit shift operation.
[22:31.380 --> 22:38.620]  It has high error, and what that means is that we might not get an exact result, but we can do multiplication really fast.
[22:38.620 --> 22:42.340]  Approximate multipliers can be used in training.
[22:46.520 --> 22:48.380]  Approximate multipliers can be used in training.
[22:48.720 --> 22:54.920]  They can vastly reduce the power of neural networks because they affect the underlying framework of how the model operates.
[22:55.160 --> 23:05.920]  If we build approximate multiplication hardware, we can run networks trained on precise hardware without any retraining and a vast reduction in power.
[23:07.640 --> 23:24.240]  The crux here is that while precise multiplication might be necessary for a traditional computing setup, where it's very important that we get the right answer out of a multiplication operation, AI are fuzzy by design.
[23:25.440 --> 23:27.400]  They can handle small errors.
[23:27.400 --> 23:39.300]  So if we introduce an error-prone design specifically to evaluate this kind of network, basically an ASIC for DNNs, we can vastly reduce their power requirements.
[23:39.520 --> 23:41.040]  That's energy-centered design.
[23:42.060 --> 23:44.180]  Again, we need no change in architecture.
[23:44.200 --> 23:45.860]  We need no retraining.
[23:48.060 --> 23:49.760]  There's minimal performance degradation.
[23:50.180 --> 23:59.760]  In fact, in some cases, if you're using approximate multipliers during training, you can see an increase in performance because the error acts as a form of stochastic regularization and prevents overfitting.
[24:00.100 --> 24:01.360]  They're also modular.
[24:01.680 --> 24:11.460]  You can use approximate multipliers alongside pruning, alongside quantization, alongside sparsification, and any other form of software optimization.
[24:12.340 --> 24:16.560]  They also have massive energy savings, on the order of 90%.
[24:18.000 --> 24:25.380]  Just by switching over to an approximate hardware model, we can reduce the amount of power that Google is using by 90%.
[24:26.360 --> 24:33.240]  So for my BA thesis, I did a case study on AlexNet, which is a CNN from like eight years ago now.
[24:34.080 --> 24:42.660]  It's large enough to be a useful test model in this case, but it was small enough for me to test easily in the month I have left to finish my thesis.
[24:43.380 --> 24:49.560]  So I implemented a precise version of AlexNet, and a simulated approximate version of AlexNet on my computer.
[24:49.800 --> 24:53.900]  I trained them on a large dataset for 50 epochs, and this was the results.
[24:54.240 --> 24:57.020]  I'm not going to dive into this because it's a little too technical.
[24:57.360 --> 25:00.580]  The long and the short of it is that they arrive at the same result.
[25:01.260 --> 25:10.000]  Approximate multipliers get off to a slower start learning, but both of them end up at about an 81% accuracy and validation, and they also had equal loss.
[25:11.520 --> 25:14.280]  So the question must be entering your mind at this point.
[25:14.560 --> 25:16.300]  Like, what's the catch, right?
[25:17.920 --> 25:21.440]  These companies are not ignorant to the fact that power costs money.
[25:21.500 --> 25:24.160]  If this was this easy, they would already be doing it.
[25:24.440 --> 25:27.220]  There's a couple things that make up what I'm calling the catch.
[25:27.220 --> 25:33.300]  And the first one is that NVIDIA has the largest monopoly in the history of the last 20 years of technology.
[25:33.380 --> 25:38.140]  If you're not buying GPUs from them, you're really just not going to get anywhere.
[25:38.400 --> 25:40.260]  You're not going to be able to train large models.
[25:40.420 --> 25:41.640]  You're not going to be able to do it quickly.
[25:41.800 --> 25:48.600]  You're going to fall behind OpenAI, Anthropic, Facebook, and Google, right?
[25:48.760 --> 25:54.280]  To build efficient approximate multipliers to catch up to NVIDIA's precise ones would take a lot of time.
[25:55.040 --> 25:59.720]  And that's a problem for these companies because they're concerned with grabbing up all the market share.
[26:00.280 --> 26:06.880]  For them, they don't care about how much it costs to build AI because they're certain that it's going to be profitable in the long run.
[26:07.700 --> 26:14.140]  Unfortunately for us, what that means is that they're dumping CO2 into the atmosphere without any regard for those consequences.
[26:14.420 --> 26:16.840]  Those consequences shouldn't be an externality.
[26:17.060 --> 26:21.520]  And yet for Google, for Anthropic, for OpenAI, they are.
[26:25.560 --> 26:27.080]  Like, why are we here?
[26:30.100 --> 26:37.200]  What possessed me to write this talk and practice it at least once and come and bring it up in front of you guys?
[26:38.440 --> 26:40.120]  And there's a few reasons for that.
[26:40.120 --> 26:52.760]  I think because energy-centered design is a way of ensuring that advances in technology can propagate, can change the world because AI can have positive impacts.
[26:53.040 --> 26:59.020]  Energy-centered design gives it the opportunity to do that without destroying the environment.
[27:01.940 --> 27:05.220]  On a wider level though, I'm here because of the name of a conference.
[27:06.020 --> 27:06.540]  HOPE.
[27:08.420 --> 27:19.620]  I'm here because of this strain of climate doomerism that has established itself among young people, especially, and it feels like most people in America these days.
[27:19.620 --> 27:23.260]  They don't think that we can fix the world.
[27:23.500 --> 27:24.520]  Voting doesn't matter.
[27:24.720 --> 27:26.060]  Corporations control everything.
[27:26.320 --> 27:26.840]  There's no point.
[27:27.180 --> 27:31.300]  I think worse than not believing you can change the world is believing that you shouldn't try.
[27:33.060 --> 27:35.160]  Solarpunk isn't about saving the world.
[27:35.420 --> 27:37.360]  But it is about trying our hardest to.
[27:38.000 --> 27:45.160]  Through beauty and experimentation and collaboration and community, Solarpunk is about trying to save the world.
[27:46.100 --> 27:56.340]  And so if the rest of this talk means nothing to you, if you've been sitting on your laptop the whole time and this is all gobbledygook, then please look at me now and remember this.
[27:57.680 --> 27:58.720]  You are the future.
[27:59.560 --> 28:05.480]  Everyone in the room with me, everyone watching live, everyone doing the recording, you are the future.
[28:06.380 --> 28:13.020]  The freaks, the hackers, the weirdos, all of us have a role to play in saving the world.
[28:13.580 --> 28:15.940]  And you guys being here is the first step.
[28:16.100 --> 28:18.020]  You guys listening is the first step.
[28:18.280 --> 28:19.640]  It's up to you to take the next one.
[28:20.880 --> 28:22.420]  So take this forward, please.
[28:22.780 --> 28:29.580]  The more people believe that it is possible to change the world, the more likely it is to happen.
[28:30.240 --> 28:32.720]  So get out there and make it happen.
[28:33.320 --> 28:34.000]  Thank you.
[28:48.120 --> 28:50.160]  I have some links for you all.
[28:51.120 --> 28:56.740]  Up there is my Fediverse handle and below it is a QR code that takes you to my website.
[28:57.200 --> 28:58.960]  Someone please tell me if the website is broken.
[28:59.160 --> 29:00.100]  I fixed it an hour ago.
[29:01.120 --> 29:09.920]  On the website should be a massive collection of links, documents, and resources related to solarpunk, efficient AI, and other miscellanea.
[29:10.000 --> 29:14.320]  I hope you find them interesting and helpful in delving further into any of these topics.
[29:14.940 --> 29:19.740]  If anyone has questions, I have a bunch of stuff about the math here that we can take a look at.
[29:20.100 --> 29:26.840]  But if anyone has questions or would like to know more about any of what I've said, I rushed through this like hell so we have a lot of time.
[29:28.100 --> 29:28.780]  Come on up.
[29:29.580 --> 29:32.320]  Yeah, you're going to have to come to the microphone, so I'm...
[29:34.100 --> 29:34.500]  Yeah.
[29:35.500 --> 29:48.060]  What about the next, say, six to eight years with the development of solar power in America?
[29:48.940 --> 29:49.860]  Is that...
[29:49.860 --> 29:55.060]  Are we considering, like, how much our power generation is going to change?
[29:55.560 --> 29:56.980]  That's a really good question.
[29:58.100 --> 30:02.740]  And the short answer is that, yeah, our power grid is becoming more carbon efficient.
[30:02.740 --> 30:08.060]  The carbon cost of electricity is dropping, but it's not dropping nearly fast enough.
[30:08.320 --> 30:25.760]  One of the big problems is that not only are we unable to produce enough renewable energy to supplant existing carbon-based methods, AI and other data center demand is also preventing the closure of older carbon-based generation facilities.
[30:25.960 --> 30:34.940]  So there's a Kansas City coal plant that was slated to be closed, and that closure has been canceled because of the opening of a new data center in the area.
[30:35.660 --> 30:36.660]  So it's not necessarily...
[30:37.480 --> 30:38.680]  Like, what you say is true.
[30:39.120 --> 30:45.460]  The increasing amount of solar and wind and geothermal power is going to decrease the cost of electricity for the planet.
[30:45.600 --> 30:50.540]  But it's not happening fast enough, and it's actually being impeded by increasing AI compute.
[30:51.780 --> 30:53.100]  All right, let's go to the right first.
[30:54.000 --> 30:55.560]  You actually front-run my question.
[30:55.700 --> 31:07.100]  It was a really similar question, which was, you know, looking at the trade-off in making AI more efficient versus, say, looking at accelerating the facilitation of silicon production.
[31:07.400 --> 31:07.880]  Right?
[31:08.020 --> 31:09.840]  And so it's kind of lead in on the same question.
[31:09.980 --> 31:11.340]  So I think you covered it.
[31:11.340 --> 31:12.780]  I just wanted to say your website works.
[31:13.060 --> 31:13.880]  Oh, thank God.
[31:14.160 --> 31:14.720]  Yes.
[31:15.980 --> 31:31.380]  I'm curious if you've been following attempts at new architecture to significantly reduce the number of nodes in neural nets, like liquid neural nets out of CSAIL, and what you think about those different architectures.
[31:31.380 --> 31:32.800]  I think they're really cool.
[31:33.140 --> 31:43.000]  I also saw BitNet, which is an attempt to remove multiplication from nets by using just like three, like zero, one, negative one, which means you can just use addition.
[31:43.360 --> 31:44.160]  That's really cool.
[31:44.320 --> 31:45.100]  It's a similar thing.
[31:45.360 --> 31:51.740]  I think approximate multipliers haven't been pursued enough, particularly because it's so difficult to build this new hardware.
[31:52.660 --> 31:55.880]  But I'm all for software solutions alongside hardware ones.
[31:56.140 --> 31:56.220]  Yeah.
[31:57.480 --> 32:00.560]  Thanks for imagining AI that doesn't suck.
[32:00.560 --> 32:02.060]  At least for me.
[32:02.280 --> 32:03.300]  I rather...
[32:03.300 --> 32:05.420]  I can't imagine a solar punk AI.
[32:05.700 --> 32:07.920]  So this is starting on a road for me.
[32:08.900 --> 32:09.820]  My question.
[32:10.200 --> 32:12.160]  So solar punk is not just an aesthetic.
[32:12.460 --> 32:15.160]  It's also like a political...
[32:15.160 --> 32:16.720]  It's very politically charged.
[32:17.760 --> 32:23.380]  And AI to me strikes me as a hammer looking for nails.
[32:23.900 --> 32:27.140]  So a solution looking for a problem in a lot of ways.
[32:27.140 --> 32:34.840]  So I wonder what your thoughts would be in application in terms of like, say, community oriented AI.
[32:35.160 --> 32:40.300]  What problems can AI solve in a solar punk world?
[32:41.360 --> 32:42.040]  Yeah.
[32:42.320 --> 32:45.400]  I think we've already seen some really interesting AI applications.
[32:45.400 --> 32:47.960]  I think one I highlighted was AlphaFold 3.
[32:48.360 --> 32:49.360]  That has done...
[32:49.360 --> 32:53.160]  Like, we've already seen four or five revolutionary papers because of AlphaFold.
[32:53.280 --> 32:54.980]  It's a really impressive piece of technology.
[32:55.420 --> 33:01.280]  Similarly, AlphaDev reduced the amount of assembly instructions needed to sort lists, which is really cool and important.
[33:02.140 --> 33:09.160]  CNNs are hitting equal accuracy to doctors to detecting tumors on CT scans.
[33:09.560 --> 33:12.980]  So I think AI is a useful tool.
[33:13.200 --> 33:16.120]  I think it absolutely has a place in a solar punk world.
[33:16.340 --> 33:17.960]  I think it'll certainly be scaled down.
[33:18.240 --> 33:22.480]  I think that, like, the impact of LLMs has been vastly overstated by profit-seeking companies.
[33:23.620 --> 33:29.500]  But there are clear implications of this technology and they could be really good for the world.
[33:29.680 --> 33:36.720]  It's just a question of how we want to spend our time, our resources, our intellectual capital in developing these technologies.
[33:36.980 --> 33:38.380]  That's, I think, what really matters.
[33:39.000 --> 33:39.080]  Yeah.
[33:42.200 --> 33:52.220]  Just to put that in the microphone, we've heard that apparently a new AI system targeted some specific cancer treatments in patients who are in hospice and managed to provide a medical breakthrough.
[33:52.220 --> 33:53.100]  Am I saying that right?
[33:54.020 --> 33:54.580]  That's what we've heard.
[33:54.780 --> 33:55.020]  Yes.
[33:56.680 --> 34:00.200]  This goes back a little bit to when you were talking about the Solar Protocol.
[34:00.660 --> 34:00.760]  Yeah.
[34:00.760 --> 34:18.340]  But do you know of any research or any projects into making that work for, like, smaller individuals or smaller organizations that don't have the ability to host servers worldwide?
[34:18.340 --> 34:19.180]  Like, is there...
[34:19.180 --> 34:19.560]  Yeah.
[34:20.200 --> 34:23.280]  Is it with, like, hosting VMs or containers or is it something new?
[34:24.560 --> 34:29.600]  Well, I think one of the requirements of their particular architecture is geographical distribution.
[34:29.780 --> 34:32.860]  That's how they achieve resiliency in the network.
[34:34.100 --> 34:38.160]  If you want to go and look, there's an online zine called Low Tech Magazine.
[34:38.540 --> 34:39.740]  Their website is solar powered.
[34:39.920 --> 34:42.620]  And they solved the problem with a really big battery.
[34:43.880 --> 34:45.220]  It's a solar powered website.
[34:45.600 --> 34:47.560]  It's located on a central server.
[34:47.680 --> 34:48.240]  There's a big battery.
[34:48.420 --> 34:55.960]  One of the cooler things about it is that as the amount of available battery energy decreases, it'll actually start scaling down the website.
[34:55.960 --> 34:58.000]  It'll send you lower resolution images.
[34:58.000 --> 34:59.340]  It'll stop sending videos.
[35:00.160 --> 35:04.980]  It'll, like, do some dithering to make everything more energy efficient.
[35:05.260 --> 35:16.800]  So there's definitely ways to place energy at the center of your design and then basically say, Okay, well, when I don't have power, I'm going to give up user experience rather than giving up...
[35:17.460 --> 35:21.820]  We're giving up user experience for power rather than power for user experience.
[35:22.000 --> 35:26.500]  So there's a lot of interesting stuff there and all that's in the list of stuff.
[35:27.380 --> 35:27.620]  Yeah.
[35:28.600 --> 35:38.480]  That was just what was really novel to me about what you talked about is that the system avoided the dependency on a battery by switching where it was hosted.
[35:38.480 --> 35:44.540]  And I think, like, that would be very... that would make it a lot more even.
[35:45.080 --> 35:45.480]  Yeah.
[35:45.640 --> 35:49.380]  I mean, like, batteries are definitely one of the more carbon intensive things to produce.
[35:49.660 --> 35:53.500]  I will point out the Solar Protocol web servers, they do have batteries, they're just not that large.
[35:53.700 --> 36:01.840]  And so it's like, you can get, I think, four or five hours of server time off of the battery, which is more than enough to cover until sunrise in the next location.
[36:02.380 --> 36:04.480]  So you could use a much smaller battery.
[36:07.900 --> 36:10.320]  Hey, I really enjoyed the presentation.
[36:10.560 --> 36:11.040]  Great work.
[36:11.540 --> 36:21.820]  I was wondering, as someone who lives in a tiny New York apartment with poor light, but would love to experiment with solar powered, like, hosting, like we've seen here.
[36:22.120 --> 36:24.600]  Do you have any, like, tips on how to do that?
[36:26.100 --> 36:29.660]  That's a really good question that I don't have an immediate answer to.
[36:29.660 --> 36:34.320]  I think finding friends with roofs is a pretty good way to start.
[36:34.560 --> 36:36.300]  You have raised your hand.
[36:36.780 --> 36:39.380]  Yeah, actually, I have a direct answer to that.
[36:39.620 --> 36:45.900]  Low Tech Magazine actually just released an article on how they're using windowsill solar panels to...
[36:46.280 --> 36:49.120]  and a DC, like a DC to DC.
[36:49.420 --> 36:51.400]  Like, they don't do the DC to AC translation.
[36:51.540 --> 36:53.500]  They just go straight DC for their lights.
[36:54.180 --> 36:57.400]  That article was, yeah, it was just released, like, I think, two weeks ago.
[36:57.640 --> 36:58.380]  Excellent stuff.
[36:58.780 --> 36:59.760]  I'll have to check that out.
[36:59.860 --> 37:00.540]  Yeah, that's really cool.
[37:01.120 --> 37:01.360]  Awesome.
[37:01.980 --> 37:05.020]  Yeah, there's a link to the Low Tech Magazine in the...
[37:05.020 --> 37:06.020]  Yeah, so, thank you.
[37:07.440 --> 37:11.340]  I loved your talk and your thoughts on climate humorism and Solarpunk.
[37:11.480 --> 37:21.040]  I'm wondering if you have any resources or, like, favorite media in Solarpunk that you recommend just to explore better the topic, like books or movies or...
[37:21.040 --> 37:22.340]  Yeah, good question.
[37:23.220 --> 37:26.260]  There's some really wonderful stuff, I think, from...
[37:27.040 --> 37:32.440]  So, Solar Protocol got kicked off by Tega Brain, who's an NYU person.
[37:32.840 --> 37:40.180]  And if you look into their wider body of research, they integrated Solarpunk into, like, I think, five or six other technological and aesthetic projects.
[37:41.160 --> 37:43.200]  And then published that pretty recently.
[37:43.440 --> 37:45.860]  So, if you dive into that, their whole body of work is really cool.
[37:46.060 --> 37:47.740]  There's a lot of hopeful stuff in there.
[37:47.820 --> 37:51.520]  There's a lot of wonderful, beautiful pieces of art and cool tech demos.
[37:51.520 --> 37:52.940]  So, worth looking into for sure.
[37:53.760 --> 37:54.100]  Thank you.
[37:54.260 --> 37:54.480]  Yeah.
[37:57.660 --> 37:58.800]  I loved your talk.
[37:58.920 --> 37:59.620]  It was wonderful.
[37:59.840 --> 38:00.540]  Thank you so much.
[38:00.540 --> 38:05.320]  Um, there's a side of Solarpunk, um, called Lunarpunk.
[38:05.440 --> 38:06.480]  I don't know if you're familiar.
[38:06.840 --> 38:08.380]  Uh, you're a little away from the microphone.
[38:08.560 --> 38:09.500]  Oh, Lunarpunk?
[38:09.660 --> 38:10.120]  I don't know if...
[38:10.120 --> 38:10.480]  Lunarpunk.
[38:10.580 --> 38:10.720]  Okay.
[38:11.040 --> 38:12.020]  I've not heard of this.
[38:12.260 --> 38:16.540]  Some people perceive it as, like, a gothic aesthetic to the Solarpunk.
[38:16.700 --> 38:24.340]  Um, but on a more philosophical level, it can be seen as the shadow work that needs to be done to make Solarpunk possible.
[38:24.340 --> 38:40.360]  And one thing that we're talking about, like, with the Solar Protocol and using Internet that's run off of solar power, it kind of puts the fear in me that, you know, you showed the example where, if the sun is in Kenya, we're gonna pull resources from Kenya.
[38:40.960 --> 38:46.700]  Um, that corporations would capitalize on that and put solar panels everywhere.
[38:46.700 --> 38:46.900]  Right.
[38:46.900 --> 38:54.440]  We wouldn't be able to own land, and it would be taken a lot of, I mean, the target would be other marginalized communities.
[38:54.780 --> 38:58.360]  And I guess my question is, you made a comment that Solarpunk is political.
[38:58.880 --> 39:12.460]  Did you ever see any research or any conversations happening by politicians around the world or in Congress to kind of get ahead of this and, um, just kind of start that conversation?
[39:13.320 --> 39:16.940]  I mean, I've seen a lot of discussion on further exploiting resources.
[39:17.340 --> 39:17.660]  Yeah.
[39:17.900 --> 39:32.040]  Um, I think, uh, not off the top of my head, I think that's like, because Solarpunk is such a community-oriented movement, it is really up to us to figure out how to architect systems in a way that we're all helping each other.
[39:32.440 --> 39:52.480]  That's what kind of powered my farming club of trust project was the idea that you would be broadcasting information about, uh, weather and, and climate conditions and all of these things to people around you, and as, as part of a system where, of mutual aid whereby you would receive information in
[39:52.480 --> 39:52.740]  return.
[39:53.020 --> 39:59.960]  So, I would love to see, like, systems and communities built up to answer that question, but I think it hasn't been discussed enough.
[39:59.960 --> 40:01.180]  Thank you.
[40:01.440 --> 40:01.560]  Yeah.
[40:03.480 --> 40:07.640]  Yeah, so, I have a bit of a practical question, and I love the talk as well, by the way.
[40:07.900 --> 40:23.440]  Um, so, in a world where, for most people, the website, or the Internet is about seven websites run by monopolies, how do we bring forth this new idea of design and the changes on usage and consumption that it would bring to everyday users?
[40:24.860 --> 40:25.880]  Can you repeat that?
[40:26.000 --> 40:26.740]  A little bit slower.
[40:27.580 --> 40:28.140]  Sounds good.
[40:28.140 --> 40:29.000]  I talk fast.
[40:29.000 --> 40:34.940]  Um, so, like, the Internet for, like, my mom is, like, six websites.
[40:35.200 --> 40:35.360]  Okay.
[40:35.480 --> 40:40.960]  And it's all Facebook and Google and these big companies that would never consider this type of thing.
[40:41.280 --> 40:41.480]  Yeah.
[40:41.800 --> 40:49.500]  So, as regular people who maybe run their own blog or whatever, how do we make the Internet we want to see in the world, basically?
[40:50.000 --> 40:51.900]  That is a really good question.
[40:52.120 --> 40:57.280]  I actually spent an hour talking about it last time, and I can't say that I got a great answer.
[40:57.280 --> 41:01.500]  Um, and so, like, one of the answers to that question is something like ActivityPub, right?
[41:01.740 --> 41:09.700]  So, let's build an indie web that duplicates or extends the features of these larger conglomerative social media platforms.
[41:09.700 --> 41:14.240]  Um, but we can do it ourselves and we can make it solar powered and all these wonderful things.
[41:14.480 --> 41:26.480]  And I think the unfortunate reality of the situation is that building duplicates of Facebook is never going to work because Facebook is Facebook.
[41:26.760 --> 41:29.560]  Why would you go on a Facebook that doesn't have your friends on it?
[41:29.560 --> 41:30.020]  Yes.
[41:30.540 --> 41:42.420]  Um, and so, the, the, the strategy I've kind of formalized in the couple years since I gave that talk is that what we should try and do is build something that Facebook couldn't.
[41:42.800 --> 41:45.840]  Facebook, it's not just that Facebook wouldn't build the Solar Protocol.
[41:46.200 --> 41:48.140]  They couldn't build the Solar Protocol.
[41:48.140 --> 41:53.960]  They couldn't muster the level of community organization and collaboration required to maintain this network.
[41:54.380 --> 42:04.780]  So, building, proselytizing, and thinking about how we create beautiful systems that are wholly unprofitable is, I think, the best way to take the steps that you're talking about.
[42:05.440 --> 42:05.800]  Beautiful.
[42:05.860 --> 42:05.860]  Beautiful.
[42:06.040 --> 42:06.300]  Thank you.
[42:06.540 --> 42:06.700]  Yeah.
[42:08.480 --> 42:09.020]  Thanks, Greg.
[42:10.220 --> 42:10.580]  Okay.
[42:10.940 --> 42:16.580]  On the five minute warning, is there anyone in the front with the cool hair or are you just using your pencil?
[42:16.920 --> 42:17.100]  Okay.
[42:18.780 --> 42:20.780]  And, um, if I can add a...
[42:21.360 --> 42:21.860]  So, I...
[42:21.860 --> 42:22.820]  You are more than welcome to.
[42:23.140 --> 42:23.840]  Thanks, thanks.
[42:24.100 --> 42:27.020]  Uh, yeah, cause I'm, I'm probably gonna do my masters in solar punk.
[42:27.220 --> 42:27.880]  So, I'm...
[42:27.880 --> 42:28.320]  Incredible.
[42:28.680 --> 42:29.620]  Going, going nuts here.
[42:29.860 --> 42:33.620]  But, um, to, to, uh, to, to go back on that, that question.
[42:33.620 --> 42:37.540]  How can we change the, make the Internet that we wanna, we wanna see?
[42:37.820 --> 42:47.680]  Um, I think that is really crucial not to compete, not to copy with the mainstream web, but to offer alternatives that are actually attractive.
[42:48.040 --> 42:51.080]  Uh, mainstream web destroys context.
[42:51.580 --> 42:51.660]  Yeah.
[42:52.040 --> 42:52.620]  It's globalized.
[42:53.400 --> 42:59.000]  It's, it's like you have no idea who's at the other side of the computer screen.
[42:59.240 --> 43:02.440]  Um, and we live in a world where community's dead.
[43:04.040 --> 43:06.080]  Um, and, uh, horror is dying.
[43:06.460 --> 43:18.440]  So, my answer to that question is, okay, how can we create an Internet with context that brings you closer to your community, uh, in your, in your, in your city, in your town, what have you?
[43:18.440 --> 43:32.100]  How can we use technology to extend our physical reality rather than, than, than, than, than have us be more connected with people at the other side of the world than the people that are right in front of us that we see from day to day?
[43:32.100 --> 43:43.400]  Uh, to me, solar punk, there's a localism aspect to it, like bringing things back home, decentralizing, um, and, and, and, and, and, and adding.
[43:43.400 --> 43:46.340]  So that just has nothing to do with AI, just cause we have five minutes.
[43:46.520 --> 43:54.100]  But, um, uh, but I think, I think technology, solar punk tech can really help us bring us back home in a way.
[43:54.440 --> 44:01.300]  Now there's this, uh, there's a lot of fiction that's being, that's being written, uh, using solar punk as, as basis.
[44:01.300 --> 44:08.000]  Uh, half, half built garden is one that tried to imagine network that's localized, but global at the same time.
[44:08.040 --> 44:09.900]  And I think activity pub is one answer to that.
[44:10.140 --> 44:10.440]  Yeah.
[44:10.700 --> 44:11.060]  Uh, yeah.
[44:11.280 --> 44:11.540]  So.
[44:12.120 --> 44:12.480]  Awesome.
[44:12.720 --> 44:12.880]  Yeah.
[44:12.960 --> 44:21.160]  Actually the conclusion I came to at the end of my talk was that we should just all host Minecraft servers, which I think is like, yeah, is, is a similar thing to what you're talking about.
[44:21.260 --> 44:21.320]  Yeah.
[44:21.420 --> 44:21.540]  What's up?
[44:21.620 --> 44:21.800]  Yes.
[44:22.040 --> 44:28.040]  Uh, I always thought this was a great talk and also agree with the Minecraft servers, um, thing, um, because yeah.
[44:28.040 --> 44:44.440]  And, but, um, this was my first time hearing about solar punk and it's phenomenal just because, um, a few years ago I was in Puerto Rico and built, uh, like a storm station, which is to prevent against, um, further hurricanes, like Hurricane Maria in Puerto Rico.
[44:44.660 --> 44:46.340]  But that was solar paneled.
[44:46.420 --> 44:52.480]  And like you said, they, it worked, um, the solar panel worked to power battery supply or to provide Wi-Fi.
[44:52.860 --> 44:53.100]  Yeah.
[44:53.100 --> 45:08.120]  And that worked because of the, of the battery, like the battery and then how you can can connect that to potentially like a solar punk movement, um, is that it can connect through Laura to like an owl network thing to satellites.
[45:08.440 --> 45:08.580]  Yeah.
[45:08.760 --> 45:10.040]  If that can be hacked.
[45:10.140 --> 45:14.020]  And I, I know that the amateur radio people know how to hack satellite stuff.
[45:14.020 --> 45:23.240]  Um, it's, this is like a cool, like, um, movement and I like the optimism, but I think it's also like, I, I didn't even know about lunar punk either.
[45:23.420 --> 45:24.660]  So this is like really cool.
[45:24.780 --> 45:25.800]  I just want to say it's really cool.
[45:25.980 --> 45:31.460]  And do you see any caveats happening, um, in, in the whole movement, um, in, in general?
[45:32.240 --> 45:33.720]  Will I send them some more time, please?
[45:34.000 --> 45:37.540]  Oh, the question was like, I guess, do you see any caveats happening?
[45:38.260 --> 45:46.920]  Well, my question was really like, how would I connect like a wood structure that's like a storm station that's solar paneled to like a solar punk protocol?
[45:47.320 --> 45:47.320]  Yeah.
[45:47.720 --> 45:53.520]  I mean, I think you already have, like, like you, you, building a solar power to anything is solar punk, right?
[45:53.720 --> 45:57.020]  And so are you, are you asking how do I connect it in a technical sense?
[45:57.020 --> 46:00.500]  I, yeah, like maybe it's, it's a protocol, right?
[46:00.680 --> 46:01.660]  So maybe it would...
[46:01.660 --> 46:01.880]  Uh, oh, sorry.
[46:01.940 --> 46:02.520]  The solar protocol.
[46:02.660 --> 46:03.060]  I see what you mean.
[46:03.160 --> 46:04.420]  It's, it's a website or...
[46:04.900 --> 46:05.340]  Interesting.
[46:05.540 --> 46:12.560]  So in this case, I think the solar protocol is designed to serve a single website, which it's mostly advertising the existence of the solar protocol.
[46:12.840 --> 46:24.260]  I think if you were going to apply a similar practical technique, you would want multiple servers hosting the same content kind of as globally, geographically distributed as possible.
[46:24.260 --> 46:33.400]  Um, something that I'd love to see but haven't is a mutual aid web server network, where if your server hosts my stuff, my server will host your stuff.
[46:33.660 --> 46:36.740]  If anyone knows of something like that, I'd love to hear about it.
[46:36.880 --> 46:41.200]  And it's the kind of thing that needs to be technologically developed, I guess.
[46:41.380 --> 46:44.960]  The solar protocol is really cool, but it is in many ways a tech demo.
[46:45.240 --> 46:45.980]  Mm-hmm.
[46:46.480 --> 46:46.780]  Thank you.
[46:47.000 --> 46:47.600]  Thank you, Greg.
[46:47.800 --> 46:48.400]  And thank you guys.
[46:48.560 --> 46:49.280]  It means a lot that you were here.
[46:49.360 --> 46:50.140]  I really appreciate it.