Could Local AIs be Powered by Food Waste?
And would it be synthetic biology that makes this possible?
Could Future AIs Be Powered by Food? | Synbio Meets AI
As a bit of a computer dork, something I think about on a regular basis is compute efficiency. I know the meme is a bit old, but frankly it’s my Roman Empire in that I probably think about it every day. What particularly boggles my mind is how even though I keep on hearing “Moore’s law is dead!” (particularly from you, Jensen Huang), this actually feels a bit like the boy who cried wolf – My MacBook for instance with its M1 Max processor can easily run my editing software on a paltry 5-10 watts without breaking a sweat. Whereas my old Windows laptop was glugging down more like 80 for the same task [here’s some more numbers for you wattage nuts like me].
Being plagued by thoughts of compute efficiency daily is already distracting. Yet alas, I am cursed with dorkiness in multiple fields – I’m also a huge nut for sustainability and synthetic biology (hence, of course, why I am a part of ValleyDAO). If computers are like my Roman Empire, then those two are like my Greek and Persian empires; with the three constantly in conflict in my head – and even intermingling at the borders.
And so it was; plagued with these thoughts whilst eating my morning toast and perusing Twitter for comments to respond to, that an idea came to me.
“What if my MacBook could be powered by food like me?”
“What if it ate food to charge its battery?”
And then I looked over to the kitchen, where the leftovers I forgot to eat last week called out stinkily to me; reminding me of the waste I caused in not eating them. And then I thought, “What if my MacBook could eat those leftovers and make them go away?”
I was just wanting the voices to stop, of course... but from this serendipitous idea, I realised I’d stumbled onto something quite profound. If we had a means of converting food waste into energy in our own homes, we could not only prevent said waste from being burned in municipal waste facilities, we could also power our own electronics. In the same way that one might with solar panels and a Tesla Powerwall. Not only that – but if you were to run an LLM like ChatGPT off said electronics, you would have designed an AI… that is powered entirely by consuming food.
Alright, so that’s a fun idea. But it’s totally impractical, right? OpenAI employs 1000s of GPUs to run its chat service; their services use up the same electricity as a small city - equivalent to 33,000 households. No way that food waste could ever power a computer, yet alone an AI like ChatGPT. And the generators in municipal waste generators are enormous, and would need to be fed huge amounts of waste to generate sufficient power. And even then, you’re still contributing to greenhouse gas emissions, seeing as they give off ~400 kilogrammes of CO2 per tonne of waste - and who even knows what else!
Well, funny thing there, reader. Because from the major momentum in synbio, to the rate at which locally-run AI is advancing, the tech is actually converging in such a way that all those things might well be solvable. Let’s break it down:
Energy usage.
Running LLMs takes energy. They need to become optimised enough to run on small amounts of it. (We’re taking single digit watts, or even ideally milliwatts).
Energy production.
Our food needs to be efficiently converted into energy, and hopefully lots of it. It wouldn’t have to match the usage necessarily; charging up a battery overnight could help in that regard.
Size.
Our rig for converting food waste into energy needs to be small enough to fit in a house, and hopefully discreet enough that it’s not a major eyesore.
Step 1, energy usage.
Just a year ago, the smartest commercially available language model, Llama-2 70B-chat, required enormous amounts of compute to use. We’re talking datacentre sizes of memory. And it didn’t even match GPT-3.5 in many regards, yet alone GPT-4.
Fast forward to today, and major advances in compression algorithms mean I’m able to run updated versions of Llama-2 70B-chat, including ones that trade blows with GPT-4 like the latest Senku-70B, completely locally on my MacBook – using anywhere from 3 to 20 watts depending on the prompt size. Models that are smaller than this can run even faster, thereby consuming less electricity. It’s not infeasible to imagine that we’ll be able to run a pseudo-sentient Spot the Dog like this one which was powered by GPT-4 in just a few years, so long as this rate of progress keeps up. And if Senku-70B can run off my MacBook, then this future Spot the Dog could potentially use a measly 3 to 20 watts to power this LLM-based brain. Or even less.
(Considering the human brain runs on an estimated 10-20 watts, I think that would be extremely impressive).
Step 2, energy production.
Okay, so municipal waste generators are large. But have you ever thought about compost heaps? These are, in a sense, mother nature’s municipal waste generator. By just dumping a load of grass, food, or whatever organic material you like into a pile, naturally-occurring organisms start to break it down into nutrients, gases, and heat. Synthetic biology, which I often refer to as a modernised version of genetic engineering, could not only make this process more efficient, but could even see it produce a direct electrical output. No need for burning off the gases to power a turbine, like in a full-sized generator. Just skip that step entirely and leech off some of the electrons from the various electron transfer steps that make up the biochemical processes behind fermentation, and direct those electrons into electrodes. Sounds farfetched? Hardly – we’re already doing it. These genetically-engineered electrogenetic E. Coli produce electricity from wastewater through a process called Extracellular Electron Transfer, which is possible thanks to the expression of various cytochromes for improved electroactivity.
Now this may not be sufficient to power a MacBook just yet. We’d be lucky to set up a machine that can even do this on the milliwatt scale reliably (this other wastewater microbial fuel cell seems to produce about 8 milliwatts). But that’s the beauty of synthetic biology – with a tool like directed mutagenesis at our disposal, we could produce hundreds of mutants of our fermentation microbes at scale and test them all in parallel, selecting the most promising mutants and then repeating the cycle. A bit like traditional Darwinian evolution, except set to overdrive. And even that technique is a bit old-hat these days: the latest in AI biological language models can come up with thousands, if not millions of custom-designed proteins for the job, which could also be tested at scale, and iteratively redesigned for optimisation.
With synthetic biology, if there’s a will, there’s a way.
Step 3, size.
Oh! Would you look at that – by cutting out the turbine and sticking to direct energy production, we’ve already cut down the size of our generator; with a minimum size of just our electrodes. So long as we’ve got them, as well as a healthy culture of our optimised fermentation organism, we’ve got our generator.
Now imagine all that but set up in something that looks a bit like a bin: all you’d need is custom bin bags to accommodate the electrodes, and you could continue using the trash like normal. Attach a battery with a USB-C output to the output of the electrodes and you’re set.
If we managed all of this… if we had a means of converting food waste into energy in our own homes, we could prevent it being burned in municipal waste facilities. We could also power our own electronics. Even our own LLMs.
Bye bye, guilt-inducing leftovers. Have fun being in my MacBook’s stomach. And hello, local sustainability.