Damn you've been busy today! I feel proud to have poked the bear. Seriously, this is really good, in a stream of brilliant consciousness sort of way. Thank you. This is, to use a word I don't like, progress. Onward.
Huh. You have made a very good argument that I am cool. I'll take it.
Re: the big LLMs and their fiefdoms, this is why I insist on my little LLM project being able to run on consumer hardware or bust.
I like the idea of Superfly. But mapping a fruit fly's connectome and understanding how it *works* are two different things. In over to transfer it to a car it will have to be reverse engineered and reworked. I still have strong doubts that the simulated synapse model is the complete picture. The mapped connectome may turn out to be about as useful as the mapped genome, which is to say, not nearly as interesting as we had hoped.
Did you send me something on your LLM project once? I can't recall, and I have a bad habit of letting things drop off my stack. Superfly was my attempt at disclosing a real research agenda while pointing out that we don't like stuff that look dead and plastic and doesn't connect in any meaningful way to our culture, save "efficiency."
The core idea of Superfly stems from recent progress in neuroscience. Now that the connectome of the fruit fly has been mapped—down to every synapse and neuron—we have an opportunity to develop bio-inspired models based on real neural pathways. In the fly, sensory input is processed directly into motor outputs, without complex intermediary layers for object recognition or higher-level processing. Ablation studies on a virtual connectome allow us to investigate motor responses directly from input patterns, essentially running experiments on a simulated fly brain to understand how it navigates and reacts. (To your point about connectome research and reductionism, yes I get it. There are also chemical signals (not visual or auditory or tactile) that flies use to navigate. It's not clear we can isolate their action at the level of the neuron connectome. Still, my take here is that something is better than nothing; we are looking for inspiration for different NN architectures.)
Liquid Neural Networks (LLNs) here are key. LLNs could allow Superfly to process data dynamically, continuously updating as new sensory information arrives. They can handle unstructured, real-time data with less computational overhead than traditional AI models, which means Superfly can quickly adapt to changes in the environment without relying on predefined, rule-based programming.
I'm seeing something like: the fruit fly model handles instinctual navigation—rapid responses to movement, optic flow, and collision avoidance—while the LNNs allow it to learn from these interactions, improving response times and efficiency without compromising on adaptability. So while Superfly operates with the fluidity of biological brains, it also harnesses the processing power of advanced AI, maintaining a seamless interplay between efficiency and adaptability.
Plus, Superfly would be so "pimp." By the way, I neglected to mention the autonomy conditions. Level 5 could be activated, or it can be manually driven. I'm wondering if there are cases were even die-hard car enthusiasts (my dad was a part-time race car driver) would opt for automatic control--on the phone, or going 120 miles an hour with sudden loss of human confidence, or?
It does sound very cool (Superfly) and I've been very interested in LNNs and what can be done with them. Targeting something relatively small like a fruitfly seems like a good research project. I wonder if there's an even smaller thing that is minimally viable. Like maybe a nematode is too simple, but a cockroach? Something that functions on an effectively 2d plane but has to solve the same kinds of problems in that space.
Re: my project I've mentioned it a few times here before but no, have not sent it to you. This is a pseudonymous account so I try to keep a few degrees of separation from RL projects.
It's nothing special anyway, just an LLM client for small models with some extra tricks to squeeze a little more utility out of them.
Main point of mentioning it was, I think the only way forward if we don't want to reproduce in AI the very-uncool walled garden rent seeking of the modern web we need to find ways to make useful AI tools that run locally on our own hardware. At least, that is my take.
I'm the most uncool guy ever, but like most of us I think I can recognize it when I see it. Hanging a computer on your head is rough stuff. Not sure. But the principle with technology is always a trade-off logic. If in the scheme of things we're better off (all things considered) wearing Orion because it augments us in ways desired, then it'll get adopted. And whether it's cool or not is relational. If cool people were them, they're cool!
Possibly the best Substack that I've read this whole year, maybe ever.
I don't care about being cool, which may or may not be the defining factor of coolness. Who cares? The who cares girl is still crying (Tom Wolfe Acid Test reference). GenX grey hairs showing. Poetry is rusty. Western culture dawn to decadence (Burzum reference).
Drunk posting. No filter. Who's with me, man? Come on come on (loose an endless longing) Silver Mt. Zion reference.
We can be more. All we have to do is realize the potential we have buried along with her name nobody came (Beatles reference). There's a sweet spot between individualism and communalism that we have to find and stick with. I hate what I do for a living.
We are the pinnacle of evolution and civilization. What are we gonna do with it? Throw the gauntlet down. Rise or fall?
Lol. Right on Cynocephalus. I appreciate your opening remark very much. Thank you. I think what I was trying to do with this piece is show more, rather than tell, what I mean by humanism. It looks like we're on the same page. Best, Erik
Damn you've been busy today! I feel proud to have poked the bear. Seriously, this is really good, in a stream of brilliant consciousness sort of way. Thank you. This is, to use a word I don't like, progress. Onward.
Thanks, Bert. Appreciate you! See in Colorado some day :)
Huh. You have made a very good argument that I am cool. I'll take it.
Re: the big LLMs and their fiefdoms, this is why I insist on my little LLM project being able to run on consumer hardware or bust.
I like the idea of Superfly. But mapping a fruit fly's connectome and understanding how it *works* are two different things. In over to transfer it to a car it will have to be reverse engineered and reworked. I still have strong doubts that the simulated synapse model is the complete picture. The mapped connectome may turn out to be about as useful as the mapped genome, which is to say, not nearly as interesting as we had hoped.
Hi Fukitol,
Did you send me something on your LLM project once? I can't recall, and I have a bad habit of letting things drop off my stack. Superfly was my attempt at disclosing a real research agenda while pointing out that we don't like stuff that look dead and plastic and doesn't connect in any meaningful way to our culture, save "efficiency."
The core idea of Superfly stems from recent progress in neuroscience. Now that the connectome of the fruit fly has been mapped—down to every synapse and neuron—we have an opportunity to develop bio-inspired models based on real neural pathways. In the fly, sensory input is processed directly into motor outputs, without complex intermediary layers for object recognition or higher-level processing. Ablation studies on a virtual connectome allow us to investigate motor responses directly from input patterns, essentially running experiments on a simulated fly brain to understand how it navigates and reacts. (To your point about connectome research and reductionism, yes I get it. There are also chemical signals (not visual or auditory or tactile) that flies use to navigate. It's not clear we can isolate their action at the level of the neuron connectome. Still, my take here is that something is better than nothing; we are looking for inspiration for different NN architectures.)
Liquid Neural Networks (LLNs) here are key. LLNs could allow Superfly to process data dynamically, continuously updating as new sensory information arrives. They can handle unstructured, real-time data with less computational overhead than traditional AI models, which means Superfly can quickly adapt to changes in the environment without relying on predefined, rule-based programming.
I'm seeing something like: the fruit fly model handles instinctual navigation—rapid responses to movement, optic flow, and collision avoidance—while the LNNs allow it to learn from these interactions, improving response times and efficiency without compromising on adaptability. So while Superfly operates with the fluidity of biological brains, it also harnesses the processing power of advanced AI, maintaining a seamless interplay between efficiency and adaptability.
Plus, Superfly would be so "pimp." By the way, I neglected to mention the autonomy conditions. Level 5 could be activated, or it can be manually driven. I'm wondering if there are cases were even die-hard car enthusiasts (my dad was a part-time race car driver) would opt for automatic control--on the phone, or going 120 miles an hour with sudden loss of human confidence, or?
Anyway, thank you.
It does sound very cool (Superfly) and I've been very interested in LNNs and what can be done with them. Targeting something relatively small like a fruitfly seems like a good research project. I wonder if there's an even smaller thing that is minimally viable. Like maybe a nematode is too simple, but a cockroach? Something that functions on an effectively 2d plane but has to solve the same kinds of problems in that space.
Re: my project I've mentioned it a few times here before but no, have not sent it to you. This is a pseudonymous account so I try to keep a few degrees of separation from RL projects.
It's nothing special anyway, just an LLM client for small models with some extra tricks to squeeze a little more utility out of them.
Main point of mentioning it was, I think the only way forward if we don't want to reproduce in AI the very-uncool walled garden rent seeking of the modern web we need to find ways to make useful AI tools that run locally on our own hardware. At least, that is my take.
Brilliant insights, Erik!!! I wonder if you think smart glasses (Orion specifically) will be 'cool' technology?
I'm the most uncool guy ever, but like most of us I think I can recognize it when I see it. Hanging a computer on your head is rough stuff. Not sure. But the principle with technology is always a trade-off logic. If in the scheme of things we're better off (all things considered) wearing Orion because it augments us in ways desired, then it'll get adopted. And whether it's cool or not is relational. If cool people were them, they're cool!
Haha... brilliant! Thanks Erik 😎👍
Possibly the best Substack that I've read this whole year, maybe ever.
I don't care about being cool, which may or may not be the defining factor of coolness. Who cares? The who cares girl is still crying (Tom Wolfe Acid Test reference). GenX grey hairs showing. Poetry is rusty. Western culture dawn to decadence (Burzum reference).
Drunk posting. No filter. Who's with me, man? Come on come on (loose an endless longing) Silver Mt. Zion reference.
We can be more. All we have to do is realize the potential we have buried along with her name nobody came (Beatles reference). There's a sweet spot between individualism and communalism that we have to find and stick with. I hate what I do for a living.
We are the pinnacle of evolution and civilization. What are we gonna do with it? Throw the gauntlet down. Rise or fall?
Lol. Right on Cynocephalus. I appreciate your opening remark very much. Thank you. I think what I was trying to do with this piece is show more, rather than tell, what I mean by humanism. It looks like we're on the same page. Best, Erik
Printing ...
Will read soon ...
Thanks Contarini. I tried to keep it under 500 pages. The trees, the trees….
Tonight I am bashing away on my tree-slaughtering novel, and I don't want break the flow.
It will shamelessly incorporate some of your ideas from your book, btw.
Article later tonight or tomorrow.