Rise of AI Conference 2019 · Berlin

Live, love, die, repeat – From worm to AI

Wagner Steuer Costa · Data Scientist for Neuroscience and Biochemistry

At Rise of AI 2019 in Berlin, Wagner Steuer Costa, a data scientist working in neuroscience and biochemistry, presents research on the nematode C. elegans, a one-millimeter worm with exactly 302 neurons, as a model for decision-making mechanisms that current AI systems lack. Using optogenetics, proteins controlled by light, his lab activates specific neurons and observes behavior and synaptic ultrastructure changes, showing that faster neural signaling increases both the release probability and the size, or information content, of synaptic vesicles. He describes a single neuron that decides whether the animal flees or freezes based on conflicting signals from head and tail, and shows that a distinct sleep neuron stays active for extended periods while the animal is unresponsive, only ceasing when the worm is woken by an oxygen pulse. He argues that neuromodulators, slower-acting chemicals like neuropeptides, help maintain the neural network in a state of criticality that allows fast, accurate responses to changing conditions. Steuer Costa closes by arguing that current AI approaches lack compartmentalized processing, sleep-like memory consolidation and, most provocatively, death, which he argues could be a necessary driver of intelligence through evolutionary pressure across many AI iterations.

Key takeaways

  • Wagner Steuer Costa studies C. elegans, a worm with exactly 302 neurons, as a minimal model for decision-making circuits missing from current AI.
  • Optogenetically speeding up neural signaling increases both vesicle release probability and vesicle size, showing modulation works by changing signal content, not just frequency.
  • A single neuron integrates conflicting head and tail signals to decide whether the worm flees or freezes when threatened.
  • A dedicated sleep neuron stays active for long periods during unresponsiveness and deactivates when the animal is woken by an oxygen pulse.
  • Steuer Costa argues AI should be built to die, proposing mortality across iterations as an evolutionary driver of genuine intelligence.

Detailed notes

C. elegans as a minimal model for decision-making

Steuer Costa studies C. elegans, a one-millimeter hermaphroditic worm with exactly 302 neurons whose position is identical across individual animals, making it a tractable, genetically modifiable model organism. His lab works with optogenetics, using light-sensitive proteins to activate specific neurons, and the animal's roughly 7,000 synapses form a connectome comparable in relative scale to a small region of the human brain.

Optogenetics reveals how faster signaling changes synaptic vesicles

By photoactivating neurons to speed up the animal's swimming by about 30%, his team found faster movement came with more errors, and examined synaptic ultrastructure by electron microscopy. Faster signaling increased both the probability that synaptic vesicles release their contents and, unexpectedly, the size of dense-core vesicles carrying neuromodulators, meaning the modulation partly works by changing how much signal each vesicle carries, not just how often it fires.

A single neuron that decides whether to flee or freeze

He shows video data of a single neuron that determines whether the worm flees or waits when threatened, for example by a predator worm. The neuron activates briefly before a reversal in movement; if it also receives a hyperpolarizing signal from the tail indicating a threat from behind, the animal freezes instead of fleeing forward. He calls this the animal's decision-making circuit, tuned to be fast and highly specific because an error in either direction can be fatal.

A dedicated sleep neuron and network criticality

A different neuron stays active for long stretches while the animal is behaviorally unresponsive, a state Steuer Costa identifies as sleep; when the worm is stimulated by a pulse of oxygen, the neuron deactivates and the animal responds quickly. He argues this kind of rapid responsiveness depends on keeping the neural network in a state of criticality, and that neuromodulators, including more than 100 different neuropeptides and over 1,000 receptors identified so far, likely help maintain that state, though this has not yet been proven directly in C. elegans.

Population-level behavior: mating and individual variation

At the population level, he describes how harsh conditions increase the proportion of male worms, which mate to increase genetic variation and can live up to seven days versus the animal's normal three-day lifespan, and how males that fail to mate release neuromodulators that make younger males chase their own tails. Despite being genetically identical from the same mother, individual worms show large behavioral variation; about 40% of synaptic connections are consistent across animals, while roughly 10% are unique to a given individual, for reasons not yet understood.

What AI is still missing: compartmentalization, sleep and death

Steuer Costa argues current AI architectures already use some biological ideas, such as network topology, but lack the fine-grained compartmentalization seen in biological neurons, where different regions of a single cell process information differently; he cites an insect neuron with about 1,400 distinct compartments as an extreme example. He also argues AI systems lack a sleep-like mechanism for consolidating memory and, most provocatively, argues AI should be built to die, since mortality across many iterations of an AI population could be the evolutionary pressure needed to drive genuine intelligence.

Quotes

The moment where an artificial intelligence is capable of dying, it will be the point where we will truly achieve intelligence.

Wagner Steuer Costa

Questions answered in this talk

Why does Wagner Steuer Costa study a worm to understand AI decision-making?

At Rise of AI 2019, Steuer Costa explains that C. elegans has only 302 neurons in fixed positions across individuals, making it possible to trace exactly how a single neuron drives a decision, such as fleeing or freezing, in ways current AI models do not replicate.

What did the sleep-neuron experiment show?

Steuer Costa describes a neuron that stays active for long stretches while the worm is behaviorally unresponsive, a sleep-like state, and deactivates when the animal is woken with a pulse of oxygen, suggesting sleep exists outside mammals.

Why does Steuer Costa argue AI should be capable of dying?

He argues that mortality, tested across many iterations of an AI population in an evolutionary process, could be the mechanism that drives genuine intelligence, comparing it to how death has shaped efficiency in biological evolution.

Mentioned

C. elegans, optogenetics, connectome, neuropeptides, network criticality, sleep neuron

Keep exploring

More from Rise of AI Conference 2019

More on AGI & Future of Intelligence

Transcript

Auto-generated transcript, not edited. · 3,885 words

Show the full transcript

but decision-making processes that's quite an interesting aspect and that is where I wanted to go into AI because it is a decision-making process that we want to generate and now what can we learn out of biology okay we had already to talks about it and now I'm going to focus even more into very small minutiae of details with a very small animal one millimeter in size and only 300 neurons are 302 to be exact on a hermaphrodite we will come to it so um the first thing that I wanted to ask you is how many of you are aware of the work from Harmon and Cara hi yeah that's great so these are yours right are how we see them and we saw it also and Moritz talk so you know neurons great next part how many I know how neurons work so action potentials and all the muscles we do know right the muscle activation and so but are you aware of this talk astok activation our single sign synaptic vesicles fill in and

we have a little bit of information flow that we don't know why it is there so is that aware no yes thank you - yeah thank you thank you very much okay background through myself I work with chemical synapses the pointer doesn't work that well but Kanaka synapse is in a sense have our our neurotransmitters that we use furnace is a Sudoku liner right and there is also medulla Tori vesicles for dopamine serotonin and things that we hear chemical synapse is one of the types there are other types of sanity synapses and when I worked with for 10 years now optogenetics so I'm using proteins that are modulated by light to control how the neuron behave so if you wish you could depolarize in your room and then you have the neuron activated right if you would do that to all my motor neurons what would happen we know that it's something called tetanus tannic and that's it okay that's not good so we want other methods of modulating how neurons work and these is

organism to apply these is on ceiling and ceilings is the small worm here one millimeter complete this is the head with a mouth of it there are some eggs that you can see down there it's transparent genetically tractable so we know the genome it's easy to genetically modify those and it's always with 302 neurons all cells are always in the same position so it is easy to check between animals what is the effect that you're generating okay so one of the things that I performed in the lab was to apply a new optogenetic tool and to it and here you can see the nearest act could tell me oh these are all the neurons the head is there this is the feeding organ that I said them in the mouth so to say and here are the neurons and you can hear nerve ring if you want to talk about brain for the neuron for this animal then it would be there it's nerve ring where most of the connections are there are seven thousand

synapses in there so about the same number of connectors that's seven neurons in our brain to write that we picked from the other talk then if we now activate these two genetically could you press please on a movie that's a movie no that doesn't work so if if you would see whether this movie is playing that it was playing before what you would see us upon photoactivation the animal is swimming around and one of the questions was can we make these connections faster in the last talk right well not substantially faster but about 30% faster so we can improve how fast the neurons are working and the animal will move faster and it will swim faster in there it comes with a little hinge and that is it will have beam ox accessible to noise so the will do more errors while it is performing its movement and we went into it to see what what are the effects actually in the ultra structure of the neurons and if you now see one single synapse

and cut it in half and look at these what you would see is something like that these are two hundred nanometers here at this bar so it's it's really small and you can observe that there are many synaptic vesicles these are the ones that carry the transmitters that we use you also say blue one here this is modulator so dense core vesicle carrying modulators and you see the red ones in here these are the ones that are ready to deploy and send the information through so what we observed is that when we when we make the animal faster set up vesicle release all these green ones they talk preferentially and then we have higher transmission of the synaptic vesicles but we also induce modulation by releasing these blueish they're the ones with the dark content the modulator and a further one and that was quite surprisingly the size of these ones were increased so if you pick a synaptic vesicle we would say that well there is a content in it and if we now

increase the size it is because it's swelling because the content is increased so the modulation of the neuron works by changing quantal size of information inside of a single synaptic has shown that for all so long story short what it happened to this optogenetic activation there was that we increase the probability of release and then by a second method we also increased the information content in this single synaptic vesicle so that is that is interesting because we can modulate how the activity of the neurons will be over time and now we come to these modulators and the connectivity that we talked about so if we now look at the connectome of C elegans so all the connections that the animal does the green one in here these are all synaptic connections and chemical synaptic connections this one in here and dark gray quite different C is the one for the electrical kind synapse the ones that we see in the electron micrograph but the ones from dopamine and serotonin we have in here there are

different connections that go across and they modulate everything and now I have these bluish dense core vesicles that I showed you with neuro peptides and they have a complete different in their pathway and another network that they generate we and disentangled about 10 percent of the neuropeptide air jet Konnect own and c elegans and what you're seeing here is just a few of those it's not as fully ten percent we're talking about over 100 different neuro peptides and over a thousand different receptors they're all intermingled so the connectome itself is also very very specific in there and modulation happens in a very interesting way so for instance this is one of their numerous risk it's a very smaller and this should rotate so that you can see better but this is the nerve ring as you could see this is the cell body of the animal of the neuron sorry and this one is a small branch characteristic for this neuron and it is in the head of the animal when please work no well

if this movie would be running what you would see is that the animal does not move so when you photo activate at this neuron and you depolarize it what happens is that the animal does not move anymore so it doesn't move around you turn it on animal stops waits waits waits waits waits until lights love and then it will move again and continue moving and ah it works so this is the best video okay so do you see how the colors change and then if the color change it is because this neuron was active okay and you can see that the neuron gets activated before it stops and then it stops and takes an action so this is the decision-making in the animal this this neuron is there to make a decision should I stop should I go what is happening in here okay so it it turns out that is this activation here if we would see this is the neuron right this is the cell body this branch region the nerve ring that

we talked about we see an activation briefly before the animal there's a reversal right this is the way that we plotted it in there well interestingly enough if before - that there is a hyperpolarization the neuron gets information from somewhere else coming from the tail then it knows oh there is from something bad also behind me I shouldn't move to there I should better just wait and see what happens best right so this is unfortunately neuron that is taking care of this small well it's actually very big another worm that predates on our poor C elegans this this one being eaten up is unfortunately my C elegans animal and it wasn't happy about it but what what this is doing this single neuron there is picking up the information is something bad ahead of me or behind me and what should I do okay and if things go south it will wait and hope for the best and in this case it was trapped and it couldn't just go away but this neuron is extremely

specific for fleeing around and that should happen as fast as you can do this decision should be as specific as possible right and you better air make an error in the sense of I flee one time more than not and this neuron is there doing exactly that in a very consistent and fast method but it's doing another thing too that in us humans we are in all of the mammalian animals it's divided into multiple different neurons and I hope that this movie works and this red button here is the cell body forest in long stimulation alright so there it is active for a long period of time and it does work fantastic and what you see is that the neuron is active for a long time and the animal is essentially sleeping and if we now wake him up using a pulse of oxygen so the animal things on what's going on and there it will respond quite fastly and you can see the other neurons up lighting up in a second and now animal

is woken up and it's trying to move away or forward or backward but it's a restraint that's why I can't move sort of you can fill it and yes unfortunately well it did try a lot of things it didn't manage to get away and then we get the oxygen levels down again and what we see is that the wrist neuron goes up for a long period of time and the animal sleeps again okay so sleep is also important for other animals was not something that is exclusive for mammalian items and the point is you need to be very fast if there are changes in environment and in order to be fast when there are changes in environment you you must secure that you're on neuronal network isn't in a criticality State so system criticality is the way that a system can best react to outside information right but the moment that it reacts to this outside of information it will by definition go away from the criticality State and in order to maintain this criticality

you need another another modulation of your network and these are what we expect to be required or the requirement of neuropeptides for instance our neuromodulators they are probably there or we think they are probably there in order to maintain this criticality state so that the animal can react as fast as possible as accurate as possible to external stimuli it hasn't been proven yet for C elegans mainly because it is really hard to get information from neurons that are moving and changing their the how they look like in the video as you saw in the movie where the size of the neuron is also changing but there are a few pointers that show that also the neuronal code in C elegans is in a criticality stage okay and this could be one way of joining everything to maintain a fast and accurate representation of the world and then also modulating your own activity to it so now going out from the single animal and going to something completely different and this is all the animals in

a population so ceilings as I said is hermaphroditic animal hemorrhage the animal means it lays its eggs Facundo by itself so it doesn't need actually to copulate to maintain its generations right but there is an interesting aspect that when things go bad and an environment is harsh thence elegans will produce males and these males by mating will produce different offspring and that the rate of change in the genome will be higher in C elegans so it's easier to adapt if things get really bad okay but there is a way that they found or that the salience has control on the population of males and this is by for instance if the animal does not mate the male one what that happens is that it can live up to seven days perfectly fine and remember this animal lives only for three days it takes three days to go from egg to adult so this male is capable of a couldn t not only its own generation the next generation and the next generation there thereafter would

still be possible to change information genetic information while mated animals after three days are almost dead and do not mate anymore so it is a way of controlling that and then there is a second way if you have too many males I don't have a video for that excuse me but if you have too many males these males will release our neuromodulators that make the younger males turn around in circles change it chasing their own tail so this is a tail chase behavior and by that they can mate okay yeah so these animals can control also other animals in there that there are a little bit younger than their in themselves if things are required and then we think about well if the animals are always in heaven always the same near in the same position and all the cells are equal will the change only be in if we have a mating profile no that's unfortunately not like that easy so um doesn't work well okay what you would see in here are two

different well actually almost 100 animals they are all from the same progeny from the same mother haven't been mated so they are pretty much genetically identical but they behaved extremely differently so the red one would be running around all the time while the blue one would be sitting down and mostly not moving at all so the the interesting part of it is although they are genetically identical and we know the connectome and we know where all these seven thousand synapses are they're not always there so we have about 40 percent of the synapses are pretty much robust across animals half of it is not and some I mean animals do have these silences while others do not and ten percent of those are unique per animal and we still don't know why they are unique in these animals and what makes them unique in there so there is a variation in there by itself so with that what I wanted to convey is that in in neurobiology from last year we have a lot of

new information that could be used in order to improve the AI systems that we are talking about in this conference and further so things that we already apply there are topology between neurons right so we call these artificial neural networks but also the topology inside of some nerves for instance in the HTM machines are its but what we don't have is this compartmentalization that you saw for instance in this restaurant where depending on where it is active the behavior will be completely different this is not specific for the elegance of in well insects there is a single neurone amacrine cell that has almost thousand four hundred compartments there are unique to each other right and they do a completely different processing depending on their location of the cell and this is possibly also the case for our own vision systems so the neurons themselves they are extremely compartmentalized and we do have some algorithms that go into this exploration exploitation that the animals need to do in order to survive that are quite effective but

what we don't really have as a way of modulating activity depending on the size so we could use a system like metal reinforcement learning within lsdm and behind it to model what would happen if the oxygen level will change right but if we do that this our owl part will be very very slow most likely while in reality these things are extremely fast right you need to be extremely fast in order to resolve these problems and this is mainly because in biological neurons what we have is a very different pathways from information change so we have very fast depolarizations through gap junctions for instance and then we have the chemical ones that are a little bit slower and then we have the neuromodulators like dopamine and serotonin which are even slower and neuropeptides can take up to seconds to work perfectly fine in a network and with that we end up having a system that possibly is working on a criticality environment and that's a that has been shown for AI furnaces in this self-organizing

recurrent neural networks so it is I think that we should keep in mind and think about it and train them also or use them and one of the things that we don't really use is sleep so sleep as a method that we think think it's a used in their own networks biological ones to improve or the feedback and use that and create them the memories that we see so this is probably one thing that we could apply in order to improve our current methods nowadays group dynamics have been shown to work furnaces and multi-agent systems right but the and reproduction and and genetic algorithms but they are all specific parts of it we have tried to join these but the interesting part will be when all of these to work together and in the end one thing that is really missing is death unfortunately so we all speak about artificial intelligence but not about when will it die so I would plead that artificial intelligence should die in order to actually work out because if

not all of the effort that we are applying to create an artificial intelligence won't work exactly as or I won't be reliable enough so I think that the moment where an artificial intelligence is capable of dying it will be the point where we will truly achieve intelligence because this is the motivator to maintain its life and with that I'd like to thank you all and thank for the Prince yeah okay thank you very much now does anybody has a question we've got 10 minutes for Q&A hi you spoke about you were peptides networks within the neural network so I didn't know that existed before today so our artificial neural networks only simulate the course overall neural network structure would you say that the information processing in the sub networks of neuro peptides is more important than the neurons in the overall larger neural structure or as important or do you think we can get away with not simulating that what do you think we have to go deeper right I can speak to two C

elegans what we saw in the neuro peptides that we looked at and you can do what neuro peptides are doing were using chemical synapses for instance so the neuro peptides will require this long time for exposures from released to a long time exposure neurotransmitters are usually cleared much faster so they can't maintain their activity for long periods of time neuro peptides have another effect you so you can have a single neuropeptides that acts on multiple receptors and these multiple receptors can have different concentrations to which they work and by that you can further modulate the effect of a single nerve peptide and so in in a sense I don't think that it is possible to to adapted these with the current methods okay there's another question come on guys I'm pretty sure this was like they are some open questions monster girls as well thanks again for your presentation just lucky University more of a request for elaboration than a actual question in that you quickly went over the death aspect at the end could

you explain more about your actual logic about why you're in favor of death by default simply because from my own I'm generally not but it's from a different disciplinary perspective so I'd be interested in your logic behind your reasoning thank you so they why I think that a death by default is interesting is in an evolutionary sense that if we are going to program the algorithm that is going to create AI then the AI itself will have multiple iterations of it and I think that if you allow multi agent ai's to cooperate or to be competitive to each other then death is a very good benchmark to improve the efficiency of of an AI in a sense of how in evolution and things worked out I'm not saying that this is the only way but it is one that worked out quite well at least in our planet hopefully yeah okay does anybody has a question okay last roll hello I was just wondering if you are associated with the open worm project at all

no I'm unfortunate enough associated with those okay yeah are you no I was just wondering yeah found it interesting so I I talked to them and we met multiple times but I am not associated with them okay so anybody who's not into worms what is the open worm a project so the open worm project tries to emulate C elegans in in silico so they have generated all the connections and all the properties of the neurons and the muscles and silicon and try to try to simulate how the animal will move in a certain environment considering the restraints of the biology

Rise of AI 2027 · Berlin · May 2027

Be in the room for the next one.

Ten years of talks started as conversations in one curated room in Berlin. The next edition brings 300 selected leaders together in May 2027.