Hacker News Reader

Comment by solidasparagus | Hacker News Reader

The author seems confused about what artificial general intelligence is. People have not meaningfully moved towards AGI - it's still a distant pipe dream.

The closest we've gotten is probably a Dota bot that's pretty good as long as you give the bot a huge advantage. Which is an incredible piece of technology, but about as close to AGI as an ant is to a human.

View on HN

The ant to human analogy is surprisingly apt in a way you might not consider though - evolutionarily speaking the ant and humans diverged relatively recently, if you count things from the beginning of life. In that way, we might also be closer to AGI than some might think..

Ants are from the other branch of creatures with bilateral symmetry. That fork would have occurred before the Cambrian explosion, so that's at least a half a billion years ago. That's a long time!

Which is relatively new compared to geological or cosmological time.

It's a while even on a geological time scale. By comparison, Pangea was a mere 175 million years ago. Half a billion years is 1/9th of the age of the Earth. Another half billion years in the future, geological changes are expected to have ended the carbon cycle, killing most plant life on Earth.

Agreed. This paragraph (in an otherwise insightful essay) was particularly jarring:

"Remarkable things are happening in the field of artificial (general) intelligence. Deep learning algorithms have proven to be better than humans at spotting lung cancer."

This is very emphatically narrow AI.

Pattern matching is not intelligence. Calling it AI rather than ML is disingenuous.

Tell me when a generic algorithm can solve many different games on different environments at the very least.

Pattern matching is certainly intelligence - classical AI was focused on identifying patterns and reacting to them with techniques like decision trees.

You can make a fairly convincing argument that intelligence is nothing more than a hierarchical system of pattern matchers.

But that's precisely the point. A microtargetting solution for McDonalds isn't A[G]I, and as the article states - everyone knows it. But the brand power of calling something AI is too strong to resist it.

Not even an ant. If AGI is a human then what we have is the equivalent of synthetic RNA molecules.

What? Ants don’t have general intelligence, and even ant colonies’ decision making (= simple swarm intelligence) is readily replicable in a programmed system, and has been, for a while.

I don’t think a gradual scale is very helpful because I don’t think that the progression from current-generation AI to AGI is going to be gradual (it will require at least one paradigm shift). That said, if you want to compare AI progress to actual animals then our current-gen AI way beyond ants. Note that, while we haven’t fully mapped the neurons/connectome of ants yet, this is unnecessary to emulate their decision-making power. And we have mapped (and can simulate) the full connectome of simpler animals (e.g. C. elegans, P. dumerilii) so we’re definitely a long way beyond single molecules.

If you are referring to the open worm project, then the conclusions you have drawn are exactly the opposite of what I have drawn.

As I understand it open worm is a hodge podge of statistical and numerical methods to try and replicate the sensorimotor behavior of c elegans. Open worm is neither complete, accurate nor elegant, despite knowing c elegans connectome and having mapped the some 900 cells in the worms body.

I wasn’t explicitly referring to that, it’s just one of many efforts. Anyway, you’re certainly right that none of the existing efforts are “elegant” but that’s hardly relevant. What matters is that the connectome is fully mapped, and that we can accurately simulate arbitrary behaviour. The issue with projects such as OpenWorm is that they have so far not been successful in generating new insight (this may be connected to your issue with lack of elegance) but this is distinct from being able to accurately simulate behaviour. Another issue is that of simulating the physical environment because — surprise, surprise — simulating the worm neurons without any realistic external stimuli is a pretty pointless exercise for most purposes.

But pick any set of stimuli you like, feed it into the models and you get a response that corresponds exactly with empirical observation. I’d therefore definitely call the neuronal model itself accurate and complete.

No actually we can't do arbitrary simulation of c elegans. Can you link me towards a publication which contains validated results supporting your assertion?

For RNA molecules they are doing quite well at trashing us at go, chess, dota and starcraft.

> a Dota bot that's pretty good as long as you give the bot a huge advantage

Excuse my ignorance but what is the huge advantage? And what happens if you don't give the bot that?

The details require a little knowledge of Dota, but essentially the bot only knows how to play a much simplified version of the game.

They play with a reduced number of playable heroes (5 of the 100+). Each hero changes the dynamic of the game and many heroes have unique interactions with each other, so this is a very substantial simplification.

Additionally, Dota is a game where mechanics (the ability to quickly and precisely click the thing you intend to click) play a huge role and bots have a natural advantage there.

Another important skill in Dota is watching everything that's going on, positioning your screen in the right place and paying attention to the minimap. As bots don't interact with the game via a screen (I believe the game exposes variables that describe the full state of the game that the bots can see), this is another advantage they have.

Thanks, that's interesting. I don't recall those caveats being explained very prominently when the "machines beat humans" articles came out about that.

The DotA bot is such an impressive accomplishment that the caveats aren't usually worth highlighting.

The DotA bot is a very specific subproblem. Look at the open worm project for progress towards AGI

There are broadly two approaches to AGI. Behavior-driven or biology-driven. That's a cool biology-driven project, but for now, OpenAI Five is much closer to AGI than OpenWorm is.

Do you care to cite your assertions? The dichotomy you are referring to doesn't exist. Do you have a reliable source that says that there are two broad approaches to AGI, behavior driven and biology driven?

https://en.m.wikipedia.org/wiki/Embodied_cognition

There is widespread consensus that embodiment is significant for AGI. Modern approaches to AGI tend to be biologically inspired.

OpenAI five is not even remotely in the same class as an AGI system.

I doubt I can find something to cite - it's an observation that I didn't think was particularly questionable. We have researchers like OpenAI trying to move towards AGI by advancing RL through specific applications. Others like OpenWorm are trying to mimic the building blocks of life. Of course biology inspires all AI, but given the nascent state of AGI, it seems like you have to choose between those two approaches.

And if we're being pedantic, can you cite your claim that there "is widespread consensus that embodiment is significant for AGI"? I find it hard to believe that there is widespread consensus about anything around AGI.

Reply on news.ycombinator.com