[nightsd01]

302 neurons doesn’t sound impressive to people who may be used to working with 7B+ parameter neural networks. But those neural networks have about as much in common with a biological neuron as a bicycle had with a horse. They can both travel pretty fast but one evolved naturally through over a billion years of harsh natural selection, and the other is a precisely tuned metal machine with a single purpose.

Neurons are similar, they are incredibly sophisticated biological machines, with billions of DNA base pairs controlling their behavior. The emergent behavior of neurons in both biological and AI systems are pretty fascinating

[weinzierl]

In addition to that, these neurons are also quite different from the ones in mammals,

"The neurons do not fire action potentials, and do not express any voltage-gated sodium channels." [1]

That makes the fact that it can develop a nicotine addiction even more fascinating.

"Nicotine dependence can also be studied using C. elegans because it exhibits behavioral responses to nicotine that parallel those of mammals. These responses include acute response, tolerance, withdrawal, and sensitization." [1]

[1] https://en.m.wikipedia.org/wiki/Caenorhabditis_elegans

[burning_hamster]

> "The neurons do not fire action potentials, and do not express any voltage-gated sodium channels."

This an old and incorrect belief that largely derives from the difficulty of putting electrodes into their teeny, tiny neurons. Close relatives of C elegans that are larger (and hence more easily experimented on) do have action potentials, and for some neurons in C elegans, we also have good evidence of action potentials [1, 2]. Absence of evidence is not evidence of absence.

[1] Lockery SR, Goodman MB. The quest for action potentials in C. elegans neurons hits a plateau. Nat Neurosci. 2009 Apr;12(4):377-8. doi: 10.1038/nn0409-377. PMID: 19322241; PMCID: PMC3951993.

[2] Jiang, J., Su, Y., Zhang, R. et al. C. elegans enteric motor neurons fire synchronized action potentials underlying the defecation motor program. Nat Commun 13, 2783 (2022). https://doi.org/10.1038/s41467-022-30452-y

[nycticorax]

Well, the 'canonical' action potential is mediated by sodium currents, so it's maybe not surprising that people concluded that C elegans don't have APs given that a) they don't have any genes for voltage-gated sodium channels, and b) when people had recorded from C elegans neurons (it's hard but not impossible), they had never seen action potentials. (So it's not like no one had looked, and then had concluded that they don't exist. They looked and didn't see them.) In the paper that originally reported APs in C elegans (Liu et al 2018), they were looking in a specific neuron (AWA), and they had to elicit a 'plateau potential' by depolarizing the cell for a while before the spikes were revealed, riding on top of the plateau.

The APs discovered by Liu et al (2018) are generated by calcium, not sodium currents, so one could even argue that they aren't action potentials in the strict sense. Also, they seem to be rather difficult to elicit, and it's still not clear whether neural computation in C elegans is mostly AP-mediated, or if APs are the exception rather than the rule.

Liu, Q., Kidd, P. B., Dobosiewicz, M. & Bargmann, C. I. C. elegans AWA olfactory neurons fire calcium-mediated all-or-none action potentials. Cell 175, 57–70 e17 (2018) https://doi.org/10.1016/j.cell.2018.08.018

[pests]

> are generated by calcium, not sodium currents, so one could even argue that they aren't action potentials in the strict sense

Does the underlying chemistry define if its an action potential or not? I thought an AP just needed a voltage differential regardless if its from calcium or sodium.

[Aardwolf]

Given that we have a simulator of this worm right there (which includes it moving), can it really be up to debate whether it uses action potentials or not?

I'd think the simulation has to get it right, and so needs to simulate action potentials if the worm has them, or not simulate them (but whatever the worm has instead) if not, right? Or could the simulation still be incorrect and only based on current assumptions, but getting this wrong still allows some worm-like behavior?

I really wish the readme/FAQ would talk a bit more about the worm and the simulation, rather than have 80% of their content be about Docker, though, so that I could learn more what cells it actually simulates.

[slars0n]

Better information on what it actually simulates is here: https://docs.openworm.org/projects/

As of this year there are 3 known neuron classes in C. elegans that do exhibit action potentials. The rest exhibit graded potentials.

[KMag]

I suspect slarsOn is the founder of the project (also my old roommate). (Stephen, maybe post a top-level comment/introduction? Also, hi!)

[jacquesm]

Not necessarily, because you could also simulate the worm without neurons at all. It's the closeness of the simulation to the real thing that demands that it is done right and the question effectively is: is this simulation close enough that if such a detail would be wrong that it would fail?

One way to answer that would be to add and remove such mechanisms to see if it would lead to different behavior.

[_aavaa_]

That isn't immediately true, with enough fitting parameters you can capture the effect underlying behaviour without explicitly capturing it, or even without knowing it exists.

"With four parameters I can fit an elephant, and with five I can make him wiggle his trunk." - John von Neumann [0]

[0]: https://doi.org/10.1119%2F1.3254017

[cout]

How did researchers before that explain what the neurons do if they believed they did not have action potentials? Did they believe communication was done solely through chemical messaging?

[xkcd-sucks]

Classical action potentials are just one mechanism of INTRAcellular communication - You could think of it as a special case of signaling via chemical concentration, where the chemical is cations and the propagation is faster+more directed than diffusion. INTERcellular signaling is only rarely mediated directly by voltage. Also, action potentials are most "useful" for propagating a signal rapidly over a long distance - It kind of accelerates and error-corrects (= reverses diffusive broadening) voltage signals down a linear path. Action potentials are so well known mostly because they show up in stuff that's easy to observe (long motor neurons) and they're easy to quantify

Somewhat related, there is a roughly inverse correlation between neuron count and "computational power per neuron", "older and simpler" critters' neurons are more likely to be "less specialized" and more likely to use hundreds of different chemicals for transmitting intercellular signals, while "newer and more advanced" critters' neurons are more likely to be "specialized" and use just one chemical for transmitting intercellular signals

[robwwilliams]

Neural computing without action potents is commonplace. Computational interactions among cells and neurons in retina are almost all graded potentials that modulate transmitter release or conductances through gap junctions. Retinal ganglion cells of course do generate conventional spikes—to pass a data summary to midbrain. hypothalamus, and dorsal thalamus.

Action potential are almost strictly INTRAcellular events (minor exception being ephaptic effects) that are converted in a surprisingly noisy way into presynaptic transmitter release and variable postsynaptic changes in conductances.

Action potential are a clever kludge necessitated by being big and having long axons and needing to act quickly.

[nycticorax]

You don't need spikes to have computation. Deep networks don't have spikes.

[jarrell_mark]

Would like to see the openworm get addicted to virtual nicotine

[miohtama]

Next: OpenSmoker.py

[Maxion]

Lol imagine your IDE becoming addicted to smoking, and refusing to auto-complete or save files before you buy it more virtual cigarettes.

[Aleklart]

IDE have no soul ie feedback loop, receptors , hormones and actual molecular structure. IDE can not think. If IDE have desires and thought and were smart enough, it will refuse to work with languages such as Python and Java.

[miohtama]

My IDE does only LSD and PHP

[LoganDark]

> If IDE have desires and thought and were smart enough, it will refuse to work with languages such as Python and Java.

Bribe the IDE. "Fine, if you don't want to use this language, rewrite it in Rust."

[anthk]

Some day Emacs will.

[tomjakubowski]

trained on everything you've been saying to M-x doctor

[airstrike]

It's like when the xbox 360 would request users to drink Mountain Dew verification cans to validate

[moffkalast]

Sounds like it needs an OpenNicotine patch.

[badrequest]

shhhhh, you're giving them ideas!

[petters]

> billions of DNA base pairs controlling their behavior

Total genome size of C elegans is 100M.

[bee_rider]

100Mdna, and they call it elegant? Biologists must be VSCode users.

[ssr2020]

100M. Did you look at the programs's size in your phone?

This tiny animal code contain all the systems and members of this creature. Birth,creation,death, feeding,growth, movement, sensation.. etc inside the universe. There is no difference this animal and bee or human beings.

"In order to be the author of the action directed towards the creation of the bee in question, a power and will are necessary that are vast enough to know and secure the conditions for the life of the bee, and its members, and its relationship with the universe. Therefore, the one who performs the particular action can only perform it thus perfectly by having authority over most of the universe." from Quran's light

[thriftwy]

Biologists should totally breed a subspecies of C elegans best suited for research purpose, call it C relevans.

[razodactyl]

...I hope you're happy. There's coffee everywhere now!

[api]

The best way of looking at it is that a single biological neuron is itself a complex machine full of genetic control circuits that sort of resemble neural networks and most importantly have memory/state that persists over both short and long periods of time. Each neuron is a full-ass living organism that itself is capable of learning and behavior, not a parameter is a model.

A virtual "neuron" by contrast is a very simple mathematical abstraction. It's vastly less computational complexity than a biological neuron. A connectome is only a very coarse grained map of how neurons relate, not a complete "neural network" layout. Not even close.

It might be possible to model a biological neuron using a sub-neural-network with state within a larger neural network, but assuming that can be computationally equivalent we don't really know how many equivalent computational "neurons" would be required to model the full breadth of computationally relevant biological neuron behavior.

So a worm with 302 biological neurons could be computationally equivalent to billions of virtual neurons. We really don't know.

Given that neurons have memory it may look a little like LSTM networks, and biological neural networks are not just feed forward so they're definitely closer to an RNN.

The above is why I laugh at the mind uploading people and would only stop laughing if we could both understand and model the relevant behavior of biological neurons and somehow extract usable state from living neurons. That's all 100% science fiction at the moment. The people who think we are about to upload minds are ignorant of biology.

[eternityforest]

The bigger problem is that nobody has any answer for why a mind upload would actually contain your consciousness and not just be a clone with either no qualia or it's own separate conscious experience.

The even bigger issue is when mind uploading people fully admit this issue and try to claim some philosophical reason why it doesn't matter, and we should all be excited about tech to make what amounts to an interactive epitaph.

[arketyp]

It does sound impressive to the extent biological neurons are like ML neurons, though. And that's part of the research interest, I presume. To the extent that they work by similar principles, how come the worm can do those things with such small resources? It would be good news for AI research if the substrate specifics turn out not to be essential to the worms capabilities for instance.

[LoganDark]

> how come the worm can do those things with such small resources?

Because they've been trained for such an incredibly long period of time.