I suspect the harness matters more than the model. The same model, dropped into different systems of context, memory, and tool access, produces wildly different results — the orchestration around a model does more work than another few points of benchmark accuracy. [1] The enterprises getting value from AI have built the scaffolding — retrieval, evaluation, guardrails — to make an ordinary model trustworthy enough to put in front of a real decision. [2]
That suspicion assumes models are all roughly alike — that the interesting variation happens around them, not inside them. Liquid AI’s research program bets the other way: not on scaling, but on what a neuron and a synapse actually do.
Liquid AI is a foundation model company, building a diverse portfolio that includes small, efficient models built to run locally, on a phone or a Raspberry Pi or inside a car, especially where network latency and bandwidth make backhaul to cloud impractical. They also train models on data that isn’t language at all — protein sequences, for example.
Liquid AI’s research started with a one-millimeter-long worm called C. elegans. It runs its entire nervous system on three hundred and two neurons [3] and still manages to forage and navigate a complicated world. Lechner’s team asked what those neurons and synapses were doing differently, and built liquid time-constant networks [4] — architectures where the connection between two neurons isn’t a fixed, linear weight but something that varies over time, the way a real synapse does.
The conversation opens with what Liquid actually ships, and the real trade-off between large and small models — not the trade-off people assume. Small models are mostly invisible, Lechner says — they run quietly in the background of products you already use. Maybe that’s why I haven’t thought much about small models?
From there we get into the liquid time-constant network itself, and then into memory — why a language model still needs something like a transformer’s perfect, growing memory even when a liquid architecture does more with less everywhere else. The KV cache discussion explains why memory, not compute, is the real constraint on model architecture.
Lechner also has a read on what separates enterprise leaders who are getting value from AI from those who aren’t: the ones who succeed understand they’re investing in a moving target, not a finished product. The conversation closes on whether we can build biology up from physics, or whether some phenomena only exist at a higher level of abstraction. Maybe that matters for how we model the world?
What Liquid AI Builds
In this section:
Liquid AI is a foundation model company whose research began with C. elegans, a worm that runs its entire nervous system on 302 neurons.
Foundation models are pre-trained by a company like Liquid AI, then customized by the customer through fine-tuning or prompting; large language models are only one type.
Liquid AI builds a diverse portfolio, including large language models scaled down for edge deployment — able to run on a Raspberry Pi — and non-language models like protein-sequence generators.
James Kaplan: All right. Hi, it’s James Kaplan here with another Prosaic Times podcast. It’s a gorgeous day in New York, and I’m here with Mathias Lechner, the CTO of Liquid AI. Do you wanna introduce yourself and tell us a little about your journey and how you came to Liquid AI? And then we’ll get into it.
Mathias Lechner: Thanks for having me, James. My name is Mathias. I’m CTO and co-founder of Liquid AI. I come from a research background — our research was about looking into neuroscience and seeing where there are advantages if you bring some of these models into machine learning. It’s been a journey from research to now, at Liquid AI, the company.
James: What does Liquid AI do?
Mathias: We are a foundation model company. In machine learning, the classical approach was: you collect your dataset, you train a model from scratch on that data, and then deploy it. That’s changed a lot in the last ten years with foundation modeling, where somebody — some entity, in our case Liquid AI — pre-trains a model on a much larger set of data, crawled from the web, all types of images, text, and so on.
Then the customer fine-tunes this model for their use case, or customizes it in some way. Even prompting ChatGPT is still using a foundation model that’s not fine-tuned — it’s just prompted; there’s a conditioning layer there. So instead of training from scratch, foundation model means somebody else pre-trains it, and the customer customizes it either with fine-tuning or prompting.
James: But it’s not a large language model. What type of model is it?
Mathias: Exactly — the term foundation model is agnostic to any type of model. Large language models are just one type of foundation model. It could be image —
James: But the type of models you build — what type of models does Liquid AI build?
Mathias: We build a diverse set of models. Our most prominent ones are large language models, but at small scale, for edge deployment. We can run a large language model on a Raspberry Pi at around forty tokens per second [5] — that’s roughly what you get from some of the ChatGPT models. We can run a smaller, more efficient version on a Raspberry Pi, which enables a lot of edge use cases. We also train models not just on text — for instance, we have protein models. Think of a generative model, like an image-generation model that you prompt, but instead of an image it generates a protein sequence.
Large vs. Small Model Trade-offs
In this section:
Large general-purpose models still win for productivity workflows like coding; small, specialized models win where sub-100-millisecond latency or limited bandwidth make cloud calls impractical.
Small models are mostly invisible to users — they run quietly inside voice assistants and other embedded products rather than being something people deliberately choose.
Enterprise adoption is uneven, with some companies fully AI-native and others just touching a few models; small-model use is driven as much by capability and bandwidth constraints as by token cost.
James: How do you guys think about the trade-offs between large language models versus small language models? How should people think about the trade-offs between a large general-purpose model and a more specialized model?
Mathias: That’s a good question. There’s always use cases for both. There are productivity workflows where you just want the best model out there — when we code, for instance, we want the latest model from Anthropic or OpenAI, irrespective of size. And then there are use cases — say, certain cloud e-commerce companies or automotive companies — that have latency requirements. End-to-end request-to-response needs to happen within about 100 milliseconds before the user experience degrades. You can’t run that with a large language model. You need a specialized small model for that.
James: You want it to be small. In many cases you want it to be local, so you don’t have to deal with network latency, correct?
Mathias: Exactly. We have an e-commerce company we deploy with where our portion of processing the request is tiny compared to the network latency. That’s very interesting.
James: And the large-language-model-versus-small-language-model debate — some people might be surprised that we haven’t seen more small language models evolve. If I think back a year or eighteen months, there was more assumption that we’d see more specialized models, and that hasn’t, at least in my experience, been the case.
So a couple of questions: are you surprised by that? Why do you think that’s happened? And how might that change in the future?
Mathias: That’s an interesting observation. Part of it is that a lot of these small or specialized models are actually running in the background, so people aren’t used to noticing them. When you have a voice assistant on your laptop — that’s a local model, and you don’t really notice it. Same with other companies we work with — a lot of that is hidden use cases: assistants, intelligence, getting deployed on all of these devices. It’s not that you go to a website or ask a certain tool — it’s more that it’s happening in the background, in hidden use cases.
James: Got it. Do you think people in the enterprise are using smaller models as thoroughly as they should? How do you see this evolving? There’s one world where most enterprises, broadly speaking, use large general-purpose models, and product companies use smaller language models for specific use cases — voice, whatever’s built into a product.
Do you think that bifurcation continues, or do you think enterprises will start using more smaller language models? How do you see things evolving?
Mathias: Looking at the customers and engagements we have, we see both gaining traction — companies adopting more of the large models, but also more of the small models rolling out.
In the three years since starting the company, I’ve observed a vast difference out there on adoption. There are some companies — I’d call them older companies — that are completely AI native, that have flipped their entire workflow upside down. And there are companies you’d assume are heavily using AI but are just getting started.
They’re just touching a few models here and there, which is quite interesting — the diversity I’ve seen in terms of AI adoption.
James: How does this intersect with concerns about token costs? Everybody in this category is now talking about token costs in the enterprise. Sometimes I think that’s a little early, because most places — even very large companies — are spending single digits of millions of dollars on tokens at this point. [6]
So I don’t think it’s a crisis yet, or will become one shortly, but certainly token costs have gotten people’s attention. Do you think small specialized models running locally, in conjunction with larger models running remotely, will be part of the answer for managing token costs?
Mathias: Yes, I’d say so. But it’s not only about cost — it’s also about capabilities.
We have some prototype applications with our own models — imagine an agent, a voice assistant, running twenty-four-seven on your — on your computer, that then communicates with the cloud models, the large language models, to send or pull certain results, to orchestrate together. We see much more orchestration of agents, and small agents will be part of that equation — because, as you said, token cost is one factor, but also just being able to run without transmitting everything to the cloud.
James: And latency, obviously, is a big issue, right? You’re saying it’s functionality as well as latency. Can you give an example? When you say functionality or capability, do you mean some combination of reducing hallucination, more precision, more relevance?
Can you give an example — in a use case, how does that play out, and technically speaking, why does that tend to be the case?
What’s the better answer you get, and why, if you’re using a specialized small language model as opposed to a general-purpose large language model? Do you get a better answer?
Mathias: I’d say it’s everything — token cost, infrastructure cost. Some companies want to deploy multimodal models — audio and voice — on their devices, and the bandwidth to transmit those to the cloud is prohibitive. It’s actually less about the token cost.
The token cost is another story, but it’s also about the connection — the bandwidth of, for instance, video, which is quite expensive. So there are advantages to running locally, and the token cost is part of the combination.
That’s why this agent is running twenty-four-seven, basically, on your laptop. We see more and more of those applications coming, where you need to be local — not just because of token cost, but connectivity and other factors.
The Liquid Time-Constant Network
In this section:
Liquid AI’s research started with C. elegans, whose nervous system runs on 302 neurons and roughly 8,000 synapses yet still supports foraging and social behavior.
The key difference from artificial neural networks isn’t the neuron — it’s the synapse: biological connections vary in neurotransmitter type, latency, and saturation, rather than acting as a fixed linear weight.
Liquid time-constant networks bring that variability into machine learning, making a neuron’s time constant liquid rather than fixed, for more expressivity without more size.
James: Tell us what a liquid time-constant network is, or a model built using a liquid time-constant neural network. Tell us about the innovation and the benefit of it.
Mathias: Our research started, as I mentioned, from neuroscience — there’s a worm called C. elegans, about one millimeter long, with three hundred and two neurons. Every worm has almost the same nervous system, wired the same way, and yet it can forage for food and show social behavior. [7] That’s remarkable for a nervous system with only three hundred and two neurons and about eight thousand synapses, compared to artificial neural networks, which have thousands of neurons and do far simpler tasks.
When we studied how the worm’s nervous system handles these interactions, we found the neurons themselves are similar to the ones in artificial networks. What’s different is the connection between two neurons. In the worm, the synapse has different neurotransmitter types, latency, delay, and saturation — both at the neuron level and the synapse level. In an artificial neural network, a synapse is just a single connection: a linear, proportional element.
We brought that difference into machine learning as liquid time-constant networks. The idea is that the time constant — how a neuron behaves — isn’t fixed. It’s liquid. It varies over time, and the connection between two neurons isn’t statically wired. It has more complexity, more expressivity, in how two neurons communicate.
James: So — I’m going to oversimplify massively here — what you’ve said is that the size and density of the neural network isn’t always correlated with the level of capability provided. That’s the point of your neurological analysis with worms, I guess. And the implication is that there’s a certain level of capability you can achieve without the processing overhead of other types of models.
Is that correct?
Mathias: Yes, at least at a high level, that’s correct.
James: Not bad. I was a history major.
Mathias: Yes.
Memory, Context, and the KV Cache
In this section:
Liquid networks excel where context is small and time-bound, as in robotics, where only recent state matters and old, irrelevant history can be filtered out.
Language modeling is different: it still needs a transformer’s growing, perfect-memory KV cache, because some information must always be exactly retrievable rather than filtered.
Liquid AI’s own language models pair that transformer memory with a short-term, convolution-based liquid component, letting attention specialize on the longer horizon and easing KV cache pressure.
James: How does this relate to memory and context window? As I’ve thought about model performance — let me back up. My long history in enterprise technology is: it’s never the compute, it’s always the memory.
That’s the bandwidth. And if you look at the history of large-model evolution — the GPU evolution — compute exceeds memory bandwidth, right?
Mathias: Yes.
James: And as a result, people who build these things have to use aggressively lossy compression around context windows, because you can only access so much context, and that’s where hallucination and imprecision come from. The question for you is: what’s the intersection between the type of models you build and the scale of the context window?
Can you use a bigger context window versus a smaller one? How do you think about context and what that means for model fidelity or precision? Does that make sense?
Mathias: To go back a bit — we started this research on liquid time-constant networks mostly looking at a robotics problem, because in a way it’s a biological robot. There, we observed really nice behavior, because there’s a temporal component — you need to remember, you need to filter out disturbances, but you don’t need to remember what action the robot took twenty minutes ago. So the context window, in a sense, is precious but typically very small. You need something time-bound.
James: What happens fifty milliseconds ago matters a lot more than what happened fifty years ago.
Mathias: Exactly. That’s the robotics case, and it’s exactly where the liquid time-constant network helps a lot. In language modeling we have a different problem. There we really need the transformer architecture — multi-head attention — because certain things need to always be perfectly remembered.
Instead of a fixed memory system, the transformer’s memory grows with each additional token. That’s a massive disadvantage for scalability — it becomes quadratic in compute, linear in memory size — but it has the advantage of perfect memory: it can always go back and look exactly at what happened.
When we tried applying our liquid networks to language modeling, we found we always need some of that component — the transformer’s growing KV cache — so the model always perfectly remembers what happened in the past. That’s always necessary, at least in language, but it can be enhanced with a short-term, liquid, gated component running alongside it.
James: And your KV cache has constraints, right? You can’t have an infinitely large KV cache, correct?
Mathias: Exactly. That becomes prohibitive for several reasons. But at least with the short-term interactions in our own model — the LFMs, the liquid foundation models we’ve published — our language models have an additional short-term time component in the form of convolutions. That lets the attention layers specialize more on the longer horizon, which again helps a bit with KV cache compression.
James: Mm-hmm. Good. What does this mean as we move from language models to world models? I’ll admit — having thought a lot about language, I sort of understand language models, or think I do. But world models are based on physics, fundamentally different, with a different set of dynamics.
How does that change the architectural trade-offs? How does it change the experience? How does it change the decisions between a large model and a small model?
Mathias: That’s a good question. In terms of world modeling, what we’ve observed that really benefits is cross-modality world models. If you train only a video model, a video-generation model, it’s quite tricky — if you have the scale, it will work. But for us, the most interesting component — that’s why we’re refocusing mostly on this part — is the cross-modality aspect.
Imagine you have a text input and generate audio for it, or the other way around — a video input, and how does the world then look in text space. This cross-modality, bridging the gap between modalities, is for us much more interesting than a world model that just predicts the next frame from previous frames. That’s where I think most of our own research is going.
Advice for Enterprise Technology Leaders
In this section:
The enterprise leaders getting the most value understand model capability is a moving target, not a finished product — treating today’s model as the end state is the mistake that leaves companies behind.
Lechner expects multimodality — audio, video, and image in and out — to become inherent to models rather than a special-case feature, paired with new NPU-equipped hardware from Qualcomm, Intel, and AMD.
Liquid AI’s own research is shifting toward cross-modality — translating between text, audio, and video — over single-modality world models that just predict the next video frame.
James: Imagine you’re the CIO or CTO of an enterprise — not a technology company, but a bank, a pharma company, an insurance company, a consumer packaged goods company, a manufacturing company. What do you think people need to know more of, in terms of model choice? What do you think people get or understand, and what do you think they don’t?
Mathias: What very few people understand — but the ones who do, I have deep respect for — is that we don’t live in a static world. Model capabilities aren’t static. When we approach some of our customers or partners, some of them think of it in a static world.
This is the final product, the final model, nothing comes afterwards — this is it, the end of the game. And then there are others who really see, hey, this is just the current point in the evolution. It keeps going further —
with better data, better training algorithms, better hardware that’s available and feeds more data through the process — all working together, algorithmically, hardware, modeling, data, to improve the model. It’s interesting to talk with people who think about it very statically, versus enterprise leaders who really understand that the model is going up, and there’s enough evidence this will continue for at least the next couple of years.
James: What, how will the landscape be different two years from now? What predictions would you make about the mix of models people will use, and how they’ll use them?
Mathias: I can only speak with confidence about our own model-development roadmap. Even if you freeze compute and everything else, capabilities will keep increasing — better data, better algorithms, more research into different ways to model these components. One thing happening on the small-local-versus-large-model side is multimodality becoming an inherent part of it — audio, video input, the way we think of text right now — that’s just standard, it’s always supported. Same with output.
If your model speaks back to you, writes back to you, or sends back an image, that will be indiscriminate — there won’t be a difference in it being a special case triggering a special function; it will just be inherent within the model. And if you combine that with advancements in hardware — the new chip manufacturers, Qualcomm, Intel, AMD and so on, all shipping new chips with NPUs, neural processing units, packing a lot of flops at very low precision — and building models to leverage that, the combination of better hardware and better models will definitely lead us to more capable models. Where a cloud model was a year ago, we’ll have locally maybe next year.
James: Terrific. What did I neglect to ask about? What else should people know, or what else would you like to talk about before we wrap up?
Mathias: One area I’m most excited about is how it connects the large-language-model, multimodal domain with other data modalities. We’ve talked about text, audio, and vision. But there are many more modalities out there — time series, biological signals.
As I mentioned, we’re training protein models. I think that becomes very interesting — do you model this by mapping it into text, which seems to be what a lot of big companies are currently doing? They map every modality to text, process it with a large language model, and map it back. Or will there be something smarter out there that already models the data itself?
James: Let me ask this: some scientists say everything is physics — ultimately physics is the substrate, and I guess the math is under the physics. But if you build a world model that’s physics-based, does that imply everything else goes on top of that? Does that make sense at all?
That chemistry is physics, biology is physics, and so on.
Mathias: Yes and no. If you model —
James: I realize that’s something physicists say, which other people may not agree with, but —
Mathias: Yes — a lot of physicists say everything is —
James: They say that, right? They say everything is physics, correct?
Mathias: Physics, yeah. In some sense that’s correct. If you model everything with — if you train a world model with some quantum interactions, whatever — sure, it will work in a way. But the problem is the interesting interactions actually happen at a much higher abstraction scale. Even biology, I would argue, is very bit detached from physics in a way — how the immune system responds to certain things. It’s connected to physics and molecular chemistry, obviously, but it’s already so much detached that if you model it —
James: You could argue there are only two types of things in the world: physics and semantics. There’s language, video, and so on, which are sort of human-constructed artifacts. At some level you could argue the English language doesn’t describe physical reality — it’s not grounded in physical reality. And then you have everything else, which is physics. But as I said, that’s something physicists say.
Mathias: Yes —
James: And then, math describes them both, I suppose.
Mathias: Yes, in some way. And I think that’s totally fine, totally valid. But the question is whether it’s the right abstraction level to think about. If you model it with an ML model, isn’t it better to go one abstraction level higher — talk about chemical interactions, biological interactions, social stuff? At the moment, words have basically emerged as the standard unit — but you could also model language at the character level instead of words, and those aren’t that great.
James: Yeah.
Mathias: They really struggle. So again, character-level is too low-level — for some reason, at that resolution, it doesn’t work as well. Maybe in the future they’ll come, but at the moment character-level large language models aren’t at the same scale as word-level. Again, it’s about the abstraction level at which you model them.
James: I think your point about abstraction is a good one — tell me if I’m being naive. Inorganic chemistry isn’t that far abstracted from physics. But once you get to biology, you’re pretty separate from the underlying physics, and that would be hard.
We may not be able to draw a straight line from the underlying physics to the biological interaction, so we need to model biology separately, independently, rather than building it up from the physics. Does that make any sense?
Mathias: Exactly. At least I have a belief this might be the case, based on this tokens-versus-words-versus-character-levels point.
James: Mm-hmm.
Mathias: But again, there’s some dynamics — so much has changed, there’s a new approach coming out there. Maybe if you have enough data, this abstraction will emerge within the model. But yeah, that’s interesting, curious to watch what happens in machine learning going forward.
Physics, Language, and Epistemology
In this section:
Lechner is skeptical that intelligence can be built up cleanly from physics: biology, like language modeling itself, seems to work better at a higher level of abstraction than the most granular one.
The word-versus-character-level modeling gap is offered as evidence — character-level language models consistently underperform word-level ones, suggesting the right abstraction level matters more than raw resolution.
James closes by proposing generative AI as a form of “entropy transformation,” a way to learn from the world’s raw entropy that deterministic systems never could — an idea Lechner doesn’t take a position on.
James: There is part of me that wonders if we as a civilization have fully grasped the epistemological change in front of us — that we can model things we never imagined we’d be able to model before. Things that were just too entropic to model, we can now model.
Is that too broad a statement, or do you think that’s a generally correct statement?
Mathias: I think it depends. There’s a lot of people who —
James: Don’t go out on a limb there.
Mathias: — yeah.
James: Depends on what? Talk to me.
Mathias: I think there’s some people who predicted this — that artificial intelligence emerges, whether it was through large language models or different paradigms, is more the nuance, the technical detail, that’s not relevant for them.
So that’s, in some sense, predictable — a lot of sci-fi —
James: Thank you.
Mathias: — turned out to be the case. There are probably a lot of people who were right that this was going to win out.
James: There’s a book I’ve been meaning to read for years called The Dreams Our Stuff Is Made Of, which I think makes the argument that science fiction has much more influence on technology innovation than we’ve ever been willing to admit to ourselves — that the people who do the technology innovation grew up reading science fiction, and it shaped how they think about...
It both raised their aspiration levels and shaped how they think about innovation. Now, I haven’t read the book yet, so I may not be describing it that well, but it’s on my list of things to read.
Mathias: Yeah, I think it’s to some extent probably true.
James: And I suppose the corollary, just as an aside, is: absent science fiction, we would have a much lower pace of technology innovation.
Mathias: Yes, probably. It sparks the imagination.
James: Yeah, it sparks the imagination. You were saying?
Mathias: Yes.
James: Okay. Now, one way I think about this, I’m curious if you agree with this: generative AI is a way of doing entropy transformation. Deterministic systems can’t process entropy — they need deterministic, non-entropic information.
And for the first time, we can ingest data about the world in all its entropy and start to learn from that. I was wondering if that resonates with you at all.
Mathias: I’m not sure.
James: Well, come back, think about it a little bit. When you come back for the follow-up at some point in the future, you could think about that, or you can shoot me an email. Tell me what you think. Hey, thanks so much.
This was a terrific discussion. I very much appreciate you coming on and talking to us.
Mathias: Thanks so much, James. It was great speaking to you, and thanks for the invitation.
James: All right, terrific. Thanks.
Mathias: Awesome.
Footnotes
[1]: Epoch AI’s analysis of SWE-bench Verified found the same model’s score moves purely on scaffold choice, holding the model fixed: Claude 3.7 Sonnet rose from 62.3% to 70.2% under a custom scaffold; GPT-4o rose from 23% (SWE-Agent scaffold) to 33.2% (Agentless scaffold); DeepSeek R1-0528 scored 33% under Epoch’s own tooling versus 57.6% reported by DeepSeek under the Agentless scaffold. Separately, Cursor’s engineering team reported taking the same underlying model from 46% to 80% on an identical benchmark, and from Top 30 to Top 5 on Terminal-Bench 2.0, by changing only the harness.
[2]: IBM Institute for Business Value, “AI in motion: Orchestrating AI at scale for sovereignty and resilience” (2026, with the Dubai Future Foundation, 1,000+ senior leaders across 20 countries) — organizations with an orchestration-led AI governance layer (coordinating retrieval, evaluation, and guardrails across the AI lifecycle) were roughly 13x more likely to scale AI successfully and reported significantly fewer losses tied to AI errors and misfires, versus peers without one.
[3]: White, J.G., Southgate, E., Thomson, J.N., & Brenner, S. (1986). “The Structure of the Nervous System of the Nematode Caenorhabditis elegans.” Philosophical Transactions of the Royal Society B. The original C. elegans connectome paper — source for the “three hundred and two neurons” figure.
[4]: Hasani, R., Lechner, M., Amini, A., Rus, D., & Grosu, R. (2021). “Liquid Time-constant Networks.” AAAI. The original LTC paper referenced in the C. elegans/liquid-time-constant-network discussion.
[5]: The “around forty tokens per second” Raspberry Pi figure is Mathias’s spoken claim. No published benchmark shows a genuinely large model hitting 40 tok/s on a Raspberry Pi 5 — independent testing (arXiv, “An Evaluation of LLMs Inference on Popular Single-board Computers,” 2025) puts 7B+ models at 1-3 tok/s, 3B models at 2-5 tok/s, and 1-1.5B models at 5-15 tok/s on that hardware, all memory-bandwidth-bound. The one figure that does land at “around forty” is Liquid AI’s own LFM2.5-230M — a 230-million-parameter model, not a large one by any standard measure — reported at 42 tok/s in under 1GB of RAM. Likely what Mathias means by “large language model” here is relative to what people expect an edge device to run at all, not large by parameter count.
[6]: The “single digits of millions of dollars” enterprise token-spend estimate is James’s spoken estimate. It sits at the low end of the 2026 range: CloudZero’s State of AI Costs 2025 report put average enterprise AI spend at roughly USD 1MM/year (USD 85,521/month), while Menlo Ventures’ 2025 Mid-Year LLM Market Update put average enterprise AI budgets at roughly USD 7MM/year — both consistent with “single digits of millions” as a typical figure at the time of recording. Agentic workloads are starting to push some companies past that band, though: one healthcare enterprise reportedly burned through more than USD 6MM in six months once agentic AI use took hold, per the FinOps Foundation’s State of FinOps 2026 report.
[7]: de Bono, M., & Bargmann, C.I. (1998). “Natural Variation in a Neuropeptide Y Receptor Homolog Modifies Social Behavior and Food Response in C. elegans.” Cell. Primary source for the social/solitary (npr-1) distinction underlying the “social behavior” reference. See also: de Bono lab et al. (2012), “A Modular Library of Small Molecule Signals Regulates Social Behaviors in Caenorhabditis elegans,” PLOS Biology, on the signaling mechanism.
