From the text, it sounds to me like they're selling multi-agent systems of some sort. They sounded really promising in the 1990s and then fell silent. Much like neural networks. But neural networks are back, so I guess they think multi-agent systems might be ready for a comeback too. They could be right.
Multi-agent-setups with LLMs (AutoGPT for example) were hyped some weeks ago. And OpenAI with their specialized Bots(?) or what they introduced recently goes in the same direction.
So what exactly are these guys building? For a product that’s supposedly 1.0 it sure lacks even the most basic of examples. I also think the title should reflect the product name because it’s clickbait as it is right now.
Early in my career, I wrote sales funnel optimization algorithms for online-shops and travel websites. We've used an evolutionary algorithm approach - basically, a number of algos implementing competing strategies and aspects that could mutate, crossover, reproduce and die.
We ran this over two years and saw impressive results - especially for the time. I feel, this very much was a "thinking system" in the same sense. And I concur - there are many architectures and approaches to built this sort of thing.
Yeah - I've been toying with a similar approach in my banking days. Basically, we've used "Genetic Programming" (very pompous term) to optimize order routing to minimize slippage (the difference between the expected price of a trade and the actual executed price).
Basically, you create a "population" out of possible placement strategies for orders. You then cycle through "generations" that quickly adapt to the changing market conditions.
Ultimately, you end up with something that - on average - provides execution closer to the target price than traditional, more static strategies. But the problem for us was that the outliers where really far out. Basically, if it got it wrong, it got it really wrong and you were often stuck with a set of half executed orders that had to be cleaned up manually.
It seems the value of this type of approach is not the "thinking" (i.e., contextual understanding) but rather the computational speed of assessing random trials. Compared to evolutionary timescales, it's very, very fast but I wouldn't characterize "thinking" to the randomness of evolution, despite its success.
- uses the information it already has combined with new inputs to form new information based on both (aka reasoning)
- retains learned information over time (aka memory)
- contains feedback loops to eventually eradicate "wrong" information (aka learning)
- converges to similar conclusions in similar circumstances (aka reasoning or possibly instict)
At least several of the most important parts of thought are there, though it is obviously very different than our own mind. That's OK, both submarines and penguins can move underwater but they do it in very different ways. No reason to think that thought wouldn't have multiple ways in which it could be implemented.
Evolution is also almost certainly not self-aware but neither is a dog and we consider dogs to be capable of thinking too.
I agree, and I didn't mean to imply it's not useful. But I do think it's different and important to understand the limitations regarding those differences. In one important aspect, contextual understanding allows for better decisions in novel environments. The "random trial and error" approach has much less additional benefit in that regard.
I think that in a few decades after we've had a lot more experience with various NN-based AIs, someone will come up with a more general quantification of "intelligence" that can integrate all these approaches into some vaguely orderable classes. After all a genetic algorithm seems clearly smarter (in some ways) than a rock but also it's clearly not as smart (in some way) as a dog. In that sense you could also say a physical lock is smarter than a rock (in that it processes a little bit of information) but not as smart as a human because it cannot learn.
Similar to how we have things like NP-hard > NP > (maybe) P, you could have a classification based on how many (types of) information can be processed or something like that. Maybe a similar but separate scale for learning capacity?
I think there's a meaningful difference between adaptive systems and thinking systems. I would say that the latter have to build internal models of the system they're analyzing and adapting to in order to qualify as "thinking".
Evolution by natural selection is adaptive in the sense you describe, but no one would consider to be "thinking".
> But that leaves large and important areas that GPTs are entirely unfit for: Real-time problem solving in dynamic environments, understanding and reacting to current events or spatial reasoning and coordination in the physical world.
It is important to recognize the limitations of these deep neural networks and especially GPTs which both have been hyped to the point where most see it as the solution to everything; since many deep-learning supporters continue to compare it to the human brain, despite the severely limited explainability in these models to the point where not even researchers know what it is doing.
It's one of the reasons why there's very low trust in deep neural networks in general. Humans can be held to account for their actions but with almost all these neural network systems cannot be held to account when something goes wrong and accept the output as the answer whilst 'it is thinking' or 'reasoning'.
> The mechanism behind real world complex systems requires both simple rulesets and a communicative fabric that allows for real time feedback loops. For flocks of starlings this can be as easy as keeping eye contact with the next bird. For Termites, it is pheromone trails. For societies, it is language, reputation and status. For capital markets, it is money. And for the internet, its the physical wiring and network architecture that makes it all possible.
As long as this solution provides transparency and explainable results rather than giving an answer from a black box system then I'm looking forward to this approach.
Human explainability/interpretability is terrible -- it takes years to understand a person well enough to model their actions well in novel situations, and there's tons of literature that backs up that people consistently invent bullshit explanations when asked why they did things. Humans can only be held to account by telling them to change, which we can do with neural networks too.
> Humans can only be held to account by telling them to change, which we can do with neural networks too.
Even if humans invent bullshit explanations, in serious cases of accountability is done in the courts system which humans investigate about the whole timeline of events of a dispute which there is very little room for perverting the course of justice and making everything up.
Hence this scenario, lawyers would liked to have known as to why did an AI system give hallucinated citations when it was used in a legal proceeding? It's even worse that legal experts knowingly trusted it and failed to reason with the results from this AI system; because fundamentally it cannot explain why that issue happened. [0] It even goes beyond basic citations, with autonomous cars without humans behind the wheel [1] with the company (Cruise) being unable to convince the regulators or even explain the crashes and had to pull the vehicles off the road due to this high amount of risk.
So yet again, explainability in AI with neural networks is still far worse than humans, even when these systems cannot be trusted in high risk and novel situations.
It's interesting how when discussing neural networks "inspired by the human brain" always comes up when the brain, as far as we know, is emergent behavior from many little things coordinating, not architected.
"Top down" never really works for complex systems, the economy being an obvious example. But we tend to ignore that when thinking about neural networks.
If you look at the actual process for building these deep neural networks, it's actually far more emergent and bottom-up than you might think just reading the news.
I took a deep learning class before things really took off [0], and what we were taught was that a lot of deep learning research was just trying different structures to see what worked. Even the people who were good at it didn't base their architectures on strong theories for how and why things would work, they just had developed an intuition through lots of trial and error.
All the explanations for what the various structures in a deep learning model are doing are post hoc—they're observations made of the behavior of the systems after we build them.
[0] Edit: Not just a random deep learning class either. The TA for that class created AI Dungeon that semester. The professor came in one day with a story of how they accidentally racked up tens of thousands of dollars in Google Cloud bills in 24 hours when that went viral.
Early architecture/structural engineering (figuring out arches, up to building cathedrals) was done by "just trying different structures to see what worked". Calling this "emergent" is not what is typically meant by the word -- we don't consider cathedrals to be "emergent structures". It's fair to say that many people imprecisely use the term, but then you are contributing to its drift into becoming (or perhaps always having been) a floating signifier.
But Cathedrals still had central planning, blueprints, ropes stuck in the ground to outline what should be built. Thermites on the other hand use a process called "Stigmergy" in which work allocation and building structures emerge organically without central coordination.
They justified their point by saying the research was "emergent". My point is that almost all pre-modern research was "emergent" by their usage, and a large amount of contemporaneous research is "emergent" as well.
That would be slightly annoying, but they did this to then justify points about how the resulting model is "emergent". The research process, and the resulting output, can and (my main point) almost always do, differ as to whether they are truly emergent.
There is nothing emergent about building a specific topology and hoping it works. Emergent would be providing N number of neurons with no topology and letting the topology arise by itself. No current ANN architectures do this, at least not any widespread ones, and trying several different topologies without knowing which one will work isn’t emergent either.
The concept has a long history going back to the 1940s (Pitts and McCulloch) and was based on their ideas at the time about how neurons in the brain functioned. That's why the association with the brain is still with us, even if our biological and neurological understanding has evolved.
I wish the industry would be more transparent about where all these brilliant ideas come from (spoiler alert: academics from the 1930s-70s, not "genius founders").
> when the brain, as far as we know, is emergent behavior from many little things coordinating, not architected
I wouldn't characterize it like that ... The brain has a specific learning architecture / dynamics that has been created via evolution under selection pressure to learn (i.e predict outcomes) better.
As a product of evolution I wouldn't want to call it "architected" or a top-down design, but for the time being it's the only example we have of such a successful learning system so it would make sense to copy what evolution has done, which means copying how it works on all levels.
A simpler example is convolutional neural nets for vision which were explicitly designed to simplistically mimic some of the behavior of our visual system with it's multi-layer (V1, V2, etc) architecture and local learning rules. Sure there's emergent behavior occurring when data is fed into such a system (both brain's visual system or CNN), such as the pattern detectors we see emerge at lowest level, but this emergent behavior is a result of the overall architecture.
Actually the link between economy and neural network is quite strong historically. Friedrich Hayek is an economist who invented the word catallaxy to describe "the order brought about by the mutual adjustment of many individual economies in a market" [2]. Then he went on to work on a global theory of the brain, and was cited by Frank Rosenblatt in 1958 as an inspiration for the model of the perceptron [3].
It's insightful to observe the difference between the emergent behavior of natural systems and the architectured approach often used in AI development. The focus on neural networks, while beneficial for certain applications, might overlook the potential of emergent, decentralized systems that could offer solutions for complex, dynamic challenges. Hivekit's work is a step towards exploring these alternative approaches, emphasizing the importance of communication and interconnectedness in creating effective complex systems.
Sure - one only needs a sufficiently powerful learning architecture to extract the signal from the data, and assuming that is the case then what is learnt depends on the data not the architecture.
That said, I don't think that LLMs are yet close to fully utilizing the training data. e.g. they are prone to hallucinating due to not thinking ahead and backing themselves into a corner where they have to say something. One obvious improvement for that one is more forward looking prediction (i.e. "thinking ahead" - engage brain before opening mouth), which for LLMs can be addressed by tree of thought rollouts and RL learning.
So, while architecture is not important if all architectures are equally powerful (in which case you'll learn what's available to learn), it certainly does matter if not all are fully up to the job, as it would appear none currently are.
A quote from the link you posted: "Then, when you refer to “Lambda”, “ChatGPT”, “Bard”, or “Claude” then, it’s not the model weights that you are referring to. It’s the dataset."
Yet all four of these examples use the same model architecture (transformers).
We don't know that. No one has demonstrated it. It's very likely that at larger scales, for a given amount of compute you cannot train a traditional RNN to be as a good as a transformer.
We are saying the same thing. Transformers are more compute efficient than RNNs. Nobody is denying that but the switch from RNNs didn't precede some performance wall(i.e it's not like we were training bigger RNNs that weren't getting better).
We use Transformers today in large part because they got rid of recursion and in effect could massively parallelize compute.
Going on a tangent sorry, you started it!
So... in a private company the employees are very often a lot more beholden to the managers/owners, so top-down command and control can be applied a lot more than in an administration where job security is often much higher. For better or for worse, this can result in a lot more bottom-up decision-making...
In view of general wastefulness, lack of longer term (environmental) planning, I would challenge the "working fine" a bit.
The retina, dopamine system, basal ganglia, cerebellum, entorhinal cortex all have very specialised coarse grained structures for their specialisations.
The eye is part of the brain for instance. The details are emergent but there is most definitely strong top down architecture encoded in DNA
Building a platform for spatially aware communication and coordination seems pretty valuable. But it seems pretty unrelated to AI. Their only other blogpost is also a vacuous AI buzz piece. I can only guess that someone not involved with the product has decided that putting out AI-related content for non-technical audiences is a good way to build brand awareness right now?
Swarms in nature are composed of nodes (organisms) that each have a neural network, and which have light and ephemeral connectivity (sensory input) to adjacent nodes.
I guess if you wanted to be pedantic, there is no swarm, or in other words the swarm is just a shorthand way of describing the collective decisions of all the nodes.
Where exactly does the emergent behavior come from if the nodes are dumb, e.g. non-autonomous?
Well, if it's so useless why is it on the HN front page? Are there "PR" companies behind promoting items to the HN front page? I'm sure there are because sometimes an article like this comes up at #3 and everyone says it's got no substance, clickbait, etc
I'll take the "PR Company" thing as a compliment. But no, afraid not. I'm one of the two founders and I wrote this one - sorry if it came across as click-baity, but I just wanted to outline the basic, long term idea that Hivekit is going for.
> Isn’t Tesla using GPTs at least for vision? What do you call the AI technologies employed by self driving projects and robotics?
It is certainly not a generative pre-trained model (GPT) used in FSD or autopilot. Most likely advanced deep learning models for object detection or segmentation with very low latency.
Whatever they are using, it still easily can get confused with similar looking objects on the road, which is why Tesla requires drivers to keep their eyes on the road and hands on the wheel. Not even Tesla, trusts its own AI systems.
