24 Seriously Embarrassing Hours for AI | Hacker News Reader

24 Seriously Embarrassing Hours for AI

garymarcus.substack.com

28 pointsheadalgorithm25 comments

"Real AI wouldn't need vast amounts of data"

Errrrr yes it would. That's why we're born as dribbling blobs who can't fend for themselves, and then spend years absorbing data from our environment and those around us.

ChatGPT-3 uses way more linguistic data than any single human is ever exposed to before displaying similar levels of intelligence.

And far less data in total. Linguistic data is only a tiny fraction of what we humans are exposed to.

I don't think we have good estimates of the total amount of data that factors into human learning. The total amount of raw sensory data we receive over a lifetime is large, for sure, but only a tiny fraction of that is relevant to learning. Just think about how much bigger ChatGPT's training corpus would be if it consisted of .WAV files of people speaking an equivalent amount of text.

Not saying that LLMs are particularly data efficient, but in their defence, they don't only learn what a human learns in a lifetime but throughout the whole evolution. There may be information encoded in our genes that LLMs need to learn from the training data

For sure a human baby has lots of initial structure that ChatGPT is lacking. Skinner made the point that operant conditioning and evolution are abstractly very similar processes. If this were the case then it would make sense to think of learning as a process that takes place both within individual human lifespans and over the course of evolutionary history. In fact, on almost anyone's account these days (including that of proponents of LLMs), learning is not very much like operant conditioning. Thus the analogy breaks down, and one can't excuse the amount of data that ChatGPT requires by hand waving about how this is just the equivalent of the 'learning' that a human baby got indirectly via evolution.

I used to think like this too, but I realized recently that actually the human brain doesn't train over ~20 years of data, it trains over ~200 million years of data. Human babies aren't an empty neural network with no weights, the entire network architecture and the entire set of checkpoint starting weights that the brain begins training from is encoded into the human genome. It would be silly to expect evolution to do anything other than this, actually.

Also at this point I'm not even sure if "general" intelligence is a thing that exists or can exist. Humans have a specific intelligence which is very specifically designed for the world and universe that we live in. Intelligent machines will have to be similarly designed.

I think it's remarkable that one can't help but use language like 'designed' when referring to highly ordered and complex engineering marvels in nature but what they actually mean is that they believe this robust, highly ordered information management system was haphazardly built one bitflip at a time until we have the most successful organism in the known universe.

The further we push the boundaries of what we know about engineering, the more far fetched that paradigm becomes. I predict that mankind will see more successful breakthroughs when they start looking at organic systems through the lens of deliberate system design rather than spending so long kidding ourselves that it was all achieved by chance against impossible odds.

I think the way evolution operates is quite a bit more complicated than "one bitflip at a time". The way evolution operates is a meta-process that itself was shaped by an evolutionary process which took place (mostly) long ago.

IMO, the genome is like an abstract system which was "designed" for evolution. It can move in one direction or another on the phenotypic plane actually extremely quickly under strong environmental pressures (on the order of only a few hundred or thousand years). And the winds of environmental pressures are frequently shifting, producing what was likely a very thorough and aggressive search of the state space.

I agree there is probably a lot to be learned from further study of the human body and brain, but we will have to re-invent a lot ourselves too since the machine capabilities we are looking for are really a lot different than the constraints that humans were shaped under.

We have an environmental adaptation system which records state, replicates data with efficient parsing, integrity checking and hot / cool storage.

Even the behaviour that most people call 'evolution' relies on our genetic information system.

We have never in our history, ever, seen a system work with such and so many deliberate intentions in managing encoded information that wasn't the product of deliberate intention.

The burden of proof is on those who say that the work of all the software engineers on this website could all be done if you just leave a computer on for enough time. By this logic, the computer left alone will design its own hardware and supply chain, its own OS, its own backup system too, etc etc.

We create systems, we know enough about them to know how hard they are to maintain. And how much effort it takes to create a working system. Even a python hello world is built on >thousands of hours of software engineering. Why should something like the cell be considered any differently?

The universe is so utterly vast that there's enough interactions over billions of years that the likelihood of not life is less then that of life.

You're quite welcome to believe that, but that doesn't satisfyingly account for entropy.

Also, the longer any complex but lifeless structure in that scenario exists, the more likely it is to be destroyed by natural forces that we know and operate with every day.

It's not likely, in fact, it takes more faith to believe in those odds than it does to believe that humans are the deliberately engineered.

you can calculate the odds of life arising from interactions in the universe. you can't do that for some extra-universal being, constant across time and space, creating life on their fancy. i suppose you could if you are trying to estimate the likelihood that the universe is a simulation and you assume the simulation is somehow driven by some intelligent force that set the initial and boundary conditions of the universe.

Science has no agreed consensus on what the definition of 'life' is. If we can't agree on its definition, how can we accurately calculate the odds of it arising ab initio?

Also, your second point is the wrong question being asked. The first is "how likely is it that life arose by chance in a lifeless universe?" The second should be "how likely is it that this ordered and organised system in front of me is the product of engineering and intelligence?"

Based on what we know about information management systems, it's extremely likely that any information management system that we stumble across is the product of intelligence and intention.

I don't think it's reasonable to represent our brains as similar to that of artificial neural networks. They don't work the same. And the metaphor ends up being unfair to both the human and the mathematical and engineering accomplishment of an ANN.

There are a lot of differences certainly, but there have been papers written about the similarities, such as https://news.ycombinator.com/item?id=26697892 (arguments against are included there too).

The main reason why I draw parallels is because my main objective is progress towards machines achieving human-level intelligence. We already have an "AGI" which works incredibly well; the human brain. We should be studying it as deeply as possible and borrowing its means & mechanisms for use in our machine intelligence systems. Trying to re-invent the brain from scratch is surely going to be far harder.

why start from the assumption that a computer and your brain work the same? there is little evidence supporting this. As an example, memory in a computer recalls specific information and stores it for further use. "memory" in a person is reconstructed as a more or less new experience assembled out of bits and pieces of previous experiences. It is why we are creative but also why we are terrible eye witnesses.

Yeah the bit at the end was kinda silly, eh? Sure, we don't have magic AI yet, but what do we do about what we do have?

Tesla's things are embarassing for Tesla, not AI as a field.

> Real AI wouldn’t need vast amounts of data.

And yet he expects an encyclopedic knowledge of the plot to Lost (a thing few humans who learned by watching it as it came out have) as part of a "Real AI".

I think it’s misleading to say Tesla faked the driving.

They were running one off software, purpose trained for the route, showing the golden run that was successful.. yes. But it was computer vision and self driving that they demonstrated. It just wasn’t their production system.

It’s certainly highly deceptive, if not exactly fake.

Not uncommon for software demos to have a golden path that you must follow of crashes will result, but that’s not a good thing.

Surprised people are upset about the OpenAI/Kenya thing. I thought it was common knowledge that this is how all the humans-required grunt work is done. Whether its dataset tagging work or facebook content moderation...outsourcing it to poor countries for peanuts is the SOP.

It's important to remember that peanuts in the US can be life changing in developing countries.

ChatGPT had possible options....then I saw the politics were more important....so yeah. Of course it's crap...it was a smoke screen like FSD a la musk....stinks

Gary Marcus is a turd. He just fights straw-men and drums up conflict on Twitter for engagement. He's a leech.

Reply on news.ycombinator.com