sbeve

Reporting benchmark results as a scalar number, e.g. "75% on XYZ" is completely meaningless at this point. You should always report efficiency scores, e.g. "75% at a cost per task of $10."

The way we will create a future where powerful AI is open-source and available to all is by making AI radically more efficient, both in terms of inference compute and (more importantly) in terms of training data requirements. This is what symbolic learning will achieve.

@TrueAIHound It'll take them another 70 years to realize DL won't get them there.

Nowadays, it seems that every AI luminary is talking about the need for continual learning in the real world. So it took the AI community 70 years to realize that there is no intelligence without continual learning? And somehow, during all this time, they pretended to be working on artificial intelligence and got away with it? I feel cheated. I'm not amused. 😠

@TrueAIHound Fair enough.

@48psx It is to me for the reasons that I mentioned.

My thoughts on symbolic and neurosymbolic AI The problem with symbolic AI, as I understand it, is that it's a top-down approach to intelligence. Before one can have symbols, one must have a perceptual system that learns the generalizable percepts that the symbols represent. The system must be situated in the real world and must include specialized sensory mechanisms to capture information from the world. Symbolic AI proponents don't provide any solution for perceptual learning. Some, such as Gary Marcus and other neurosymbolic AI proponents, want to create a hybrid model using deep learning for perception. It will never work because DL can't generalize on the fly and can't learn continually in the real world. So, neurosymbolic AI is a waste of time in my opinion. Sorry.

@TrueAIHound Right. But from my understanding, Neurosymbolic AI cannot be a waste of time since it's an extremely broad and isn't limited to any specific architecture.

@48psx I have seen him and others use the word 'hybrid' before. I don't think they use it in a genetic sense, more like a mix.

@TrueAIHound No worries. But don't you think it's the other way around? Can you solve generalization before addressing credit assignment?

@48psx Sorry I missed your comment. In my opinion, the credit assignment problem cannot be solved unless generalization is solved.

Neuroscience I love this illusion because it's proof that human visual perception is based on edge movement detection by retinal ganglion cells (RGCs) in the retina. Deep learning experts have no clue what true vision is. It will revolutionize computer vision. I estimate that the data efficiency of pattern recognition in the brain is at least 2 orders of magnitude higher than DL. This alone will change everything. Work in progress.

@AgentRev @TrueAIHound Likely so indeed

@48psx @TrueAIHound I'm a different guy, but it's likely a combination of associative memories and dopamine cells: cell.com/cell/fulltext/…

@xeophon Why don't you do it since seem so sure?

i don't know why cOnTinUaL LeArNinG can't be just a .txt file and at this point i am too afraid to ask

@redtachyon This highlights a flaw in how the AI community approaches benchmarks in general. It is not an issue specific to ARC-AGI.

It really seems like ARC-AGI is only derailing the quest for AGI, mainly due to its popularity. I actually think it's good for what it is - a necessary (but not sufficient) condition for AGI. That is, if your model can't solve it, it's probably not AGI. But so much effort is now being poured into approaches designed to solve colored grid puzzles, and I can't see most of those methods being good for anything else.

@GregKamradt @LiaoIsaac91893 I'd say a problem is defined by the work required to get to the solution, not the solution itself

@GregKamradt @LiaoIsaac91893 Wouldn't be better to measure similarly of two tasks based on the "reasoning" that leads up to the solution instead of the length of the solution itself?

I asked @LiaoIsaac91893 his take on if we could measure how similar two ARC-AGI tasks were Not by pixels, but by transformations and solution programs He had an interesting idea involving minimum description lengths and code golf For those that don’t know, code golf is the gamified act of creating the *shortest* possibly program to complete a goal. Usually by counting # characters. Let’s say we have 2 tasks (Task A and Task B) and we want to know how similar they are Isaac proposed that you do 3 things: 1) Find the minimum description length of Task A and then 2) find the same thing for Task B Now here’s the cool part Finally you 3) Find the minimum description length of the program that solves *both* Task A and Task B Then combine the lengths of those programs len(task A) + len(task B) - len(Task A + B) Let’s assume you had identical tasks. The minimum description length (mdl) would be the same for both A and B. Let’s say it has the value 1. The same program could solve A + B so that als has a length of 1. So 1 + 1 - 1 = 1 (very similar) Opposite tasks with no similarity would have a low result (not similar) Thanks for the fun conversation Isaac

@ChombaBupe "Intelligence arises from severe constraints on access to resources like data, energy etc because intelligence makes up for lacking those resources." That goes so hard bro

TL DR: Intelligence arises from severe constraints on access to resources like data, energy etc because intelligence makes up for lacking those resources. It enables an intelligent entity to make best use of whats available. Intelligence is about efficiency.

People crying that this hinders artificial intelligence (AI) development don't understand that this creates an evolutionary pressure for AI to become truly intelligent, to not depend on tons of high quality work. Only lazy & shady tech companies sees this as an obstacle.

@TrueAIHound @RichardSSutton Worst part is they're gonna come up with some shitty patch and say "we've cracked continual learning" same way they did with "chain-of-thought"

For the past 6 months or so, I've been hearing about how AI is moving from pretrained models to continual learning models, or as @RichardSSutton calls it, "experiential AI". Where is it? My bet is that the AI community has no idea how to do continual learning. They've been collecting data to "train" deep neural nets from the beginning. Once your mind has been poisoned by the deep learning bug, you're done for. They are like crack addicts. They can't quit. Too bad. So sad. 😁

@JadeCole2112 You'd think all the mental gymnastics would help make them less dumb.

"Seems to know" is doing all the lifting here.

@TrueAIHound They have made progress in the sense that they've explored what NOT to do when trying to create intelligence, so they're like a cautionary tale

I completely disagree that the AI community has made any progress toward solving the intelligence problem in the last 70 years. All they have accomplished is the creation of computer automation tools, thanks primarily to deep learning, the availability of fast computers and the massive source of data afforded by the internet. Yes, automation is a good thing but calling it AI has strangled critical research in actual intelligence. Intelligence is not automation.

@MatthewBerman @dwarkesh_sp Claiming LLMs are a "different kind of intelligence" is cope for the fact they fundamentally lack intelligence. All general intelligence has common core properties that are invariable, properties these systems plainly don't have.

.@dwarkesh_sp published an article arguing that AI progress is slower than people expect, but the core of his case rests on treating AGI as something that must learn and develop like a human. Something about that argument has been rubbing me the wrong way. It frames everything around human-style cognition instead of recognizing that AI is a different kind of system. Forcing AI into a biological template underestimates how fast models can improve when they use their real strengths like scale, parallelism, and rapid iteration. I’ve quoted parts of the article below, followed by my thoughts. “I’m confused why some people have short timelines and at the same time are bullish on RLVR. If we’re actually close to a human-like learner, this whole approach is doomed.” It’s entirely possible AI will never be “human-like” in its learning ability. RLVW is not how humans learn, and that’s fine. RLVW works precisely because it doesn’t rely on human-style learning. It removes the need for humans to label data, create lesson plans, or supervise each step. AI learns by interacting with an environment and optimizing against a reward signal. That loop is fundamentally different from how humans learn, and that’s not a bug, it’s a feature. “Humans don’t have to go through a special training phase where they need to rehearse every single piece of software they might ever use.” Humans don’t rehearse every piece of software because our learning architecture develops slowly over years and generalizes across domains. But AI doesn’t need to follow that path. Different substrates learn differently. Neural networks have advantages humans never had: massive parallelism, near-instant iteration, and the ability to process billions of examples at superhuman speed. "Currently the labs are trying to bake in a bunch of skills into these models through “mid-training” - there’s an entire supply chain of companies building RL environments which teach the model how to use Excel to write financial models or navigate a web browser." He’s right that labs are stuffing skills into models through mid-training and custom RL environments. But the real bottleneck isn’t the existence of RL itself. It’s the lack of a universal way to generate clean, verifiable rewards across any domain. Once we have a general method for producing reliable reward signals for arbitrary tasks, bespoke training pipelines aren’t necessary. You get a system that can improve itself across countless domains without hand-crafted scaffolding. "It’s totally reasonable to look at this and say, “Oh actually there’s more to intelligence and labor than I previously realized. And while we’re really close to (and in many ways have surpassed) what I would have defined as AGI in the past, the fact that model companies are not making trillions is revenue clearly reveals that my previous definition of AGI was too narrow.”" Saying “models aren’t generating trillions in revenue, so my AGI definition must be too narrow” ignores the actual timeline. We’re only about three years into the generative AI wave. It takes time to convert raw intelligence into enterprise deployment. You need tooling, integration layers, workflow redesigns, and organizational adoption. The lag isn’t because the intelligence isn’t present. It’s because distribution and integration always trail capability. (am I coping?) Dwarkesh quotes Steven Byrnes: “New technologies take a long time to integrate into the economy? Well ask yourself: how do highly-skilled, experienced, and entrepreneurial immigrant humans manage to integrate into the economy immediately?” The immigrant analogy breaks because immigrants use human-style intelligence to plug into a world built for humans. They already understand our norms, interfaces, incentives, and communication styles. AI doesn’t. Even a smarter AGI must translate its very different style of intelligence into systems built around human mental models. The delay comes from the mismatch between how AI thinks and how our institutions expect workers to think, not from a lack of capability. “Human workers are valuable precisely because we don’t need to build schleppy training loops for every small part of their job. It’s not net-productive to build a custom training pipeline to identify what macrophages look like given the way this particular lab prepares slides, then another for the next lab-specific micro-task, and so on.” That point is fair. As mentioned, the right answer is to develop a general method for applying RLVW across any domain without manual pipelines. Once we can generate reliable rewards for arbitrary tasks, you no longer need thousands of tiny hand-curated training loops. You get a system that can RL itself into any environment on demand. The goal might not be knowledge transfer, but rather giving AI a general mechanism to adapt quickly to whatever domain it’s dropped into. “All these scenarios neglect what I think will be the main driver of further improvements atop AGI: continual learning.” He keeps returning to continual learning, but here are my two counters. First, pre-training may already be the continual learning. It’s just slow and expensive today. As the cost and speed of pre-training fall, the value of a separate continual learning loop changes. Second, we may not need human-style continual learning if we build generalized RLVW systems that can make a model an expert in a new domain on short notice. Continual learning is one option but not necessarily the natural one for non-biological intelligence. “The agents themselves could be quite specialized - containing what @karpathy called “the cognitive core” plus knowledge and skills relevant to the job they’re being deployed to do.” This seems likely. Models may end up with a strong cognitive core and then acquire specific skills on demand. Instead of storing every skill up front, they’ll have core intelligence plus general scaffolding, and use RLVW to gain any domain expertise they need in real time. “There’ll also probably be diminishing returns from learning-from-deployment. Each of the first 1000 consultant agents are each learning a ton from deployment. Less so the next 1000. And is there such a long tail to consultant work that the millionth deployed instance is likely to see something super important the other 999,999 instances missed? In fact, I wouldn’t be surprised if continual learning also ends up leading to a power law, but with respect to the number of instances deployed.” This is basically the 80/20 rule. The first chunk of expertise comes fast, and the last 20 percent takes most of the effort. For humans, that long tail usually isn’t worth chasing. AI doesn’t share that constraint. Even the one-in-999,999 edge case has value, because compute can dig into those rare scenarios cheaply and in parallel. That makes the long tail economically viable and often where the biggest capability gains hide. The whole debate changes once we stop forcing AI into the mold of human intelligence. Humans learn slowly, sequentially, and only when it’s worth their limited time. AI learns differently. It can brute-force the long tail, adapt to new domains through generalized RLVW, and re-train at industrial speed as the cost curve falls. The real question isn’t whether AI can mimic human learning. It’s whether AI can exploit its own alien advantages to surpass what humans can do. And once you judge AI on its actual substrate instead of on a biological one, the ceiling looks far higher and the timeline looks far faster than the article suggests. Written with the help of AI 😊

@fchollet Intelligence is about the process that leads to the compression. Not the compression itself.

To perfectly understand a phenomenon is to perfectly compress it, to have a model of it that cannot be made any simpler. If a DL model requires millions parameters to model something that can be described by a differential equation of three terms, it has not really understood it, it has merely cached the data.

@TrueAIHound There's likely many more low hanging fruits that haven't been discovered yet simply because people have been too focused on scaling

Scaling is the lowest hanging fruit in AI. Why? Only because it doesn't require any scientific breakthrough in understanding intelligence. It's also the least important fruit but, amazingly enough, all the investors, grifters and scammers are focused on it. They will be sorry. 😁😂