On 2018-04-09, OpenAI said[1]:
OpenAI’s mission is to ensure that artificial general intelligence (AGI) [...] benefits all of humanity.
In contrast, in 2023, OpenAI said[2]:
[...] OpenAI’s mission: to build artificial general intelligence (AGI) that is safe and benefits all of humanity.
This archived snapshot is from 2023-05-17, but the document didn’t get much attention until November that year.
Another example is risk compensation: You make an activity safer (yay) and participants compensate by taking more risks (oh no).
Interesting, it felt less messy to me than, say, rationalist-adjacent research retreats.
lsuser says that as a result of his spiritual journey, “now if there is so much as a cardboard box on my kitchen counter, it bothers me”. Has your spiritual practice changed your tolerance of clutter?
In other words, the zero-information oblivion that produced you once can produce you again, maybe in a different form.
Huh, that’s Epicurus’s argument against fearing death. But while Epicurus assumed there is no afterlife, you’re using it to argue there is one!
Re: safety, it depends on exactly where you are, your skill in assessing strangers’ intentions from a distance, and probably the way you carry yourself.
Speaking of which, I’d be interested in playing some improv games with you at less.online, if you want to do that!
I’d like to know what Holden did while serving on the board, and what OpenAI would have done if he hadn’t joined. That’s crucial for assessing the grant’s impact.
But since board meetings are private, this will remain unknown for a long time. Unfortunately, the best we can do is speculate.
Of course, Karpathy’s post could be in the multimodal training data.
12 years ago, in The state of Computer Vision and AI: we are really, really far away, Andrej Karpathy wrote:
The picture above is funny.
But for me it is also one of those examples that make me sad about the outlook for AI and for Computer Vision. What would it take for a computer to understand this image as you or I do? [...]
In any case, we are very, very far and this depresses me. What is the way forward? :(
I just asked gpt-4o what’s going on in the picture, and it understood most of it:
In this image, a group of men in business attire are seen in a locker room or a similar setting. The focus is on two men, where the taller man is standing on a scale. The shorter man, who appears to be playfully pressing down on the taller man’s shoulders to increase his weight on the scale, is creating a humorous situation. Both men and those observing in the background are smiling or laughing, indicating that they are enjoying the lighthearted moment. The man pressing down seems to be taking part in a playful prank or joke, adding a sense of camaraderie and fun to the scene.
That does look like a rough commute, the kind that can use up the mental energy you want to spend on learning. One thing you could consider is staying in a hotel overnight near your school sometimes.
Also, consider wearing ear protection on the Transbay Tube. I wish I had done that when I commuted that way for a year.
I suppose if you had more hidden states than observables, you could distinguish hidden-state prediction from next-token prediction by the dimension of the fractal.
If I understand correctly, the next-token prediction of Mess3 is related to the current-state prediction by a nonsingular linear transformation. So a linear probe showing “the meta-structure of an observer’s belief updates over the hidden states of the generating structure” is equivalent to one showing “the structure of the next-token predictions”, no?
The subject of this post appears in the “Did you know...” section of Wikipedia’s front page(archived) right now.
I’m saying “transformers” every time I am tempted to write “LLMs” because many modern LLMs also do image processing, so the term “LLM” is not quite right.
“Transformer”’s not quite right either because you can train a transformer on a narrow task. How about foundation model: “models (e.g., BERT, DALL-E, GPT-3) that are trained on broad data at scale and are adaptable to a wide range of downstream tasks”.
I agree 100%. It would be interesting to explore how the term “AGI” has evolved, maybe starting with Goertzel and Pennachin 2007 who define it as:
a software program that can solve a variety of complex problems in a variety of different domains, and that controls itself autonomously, with its own thoughts, worries, feelings, strengths, weaknesses and predispositions
On the other hand, Stuart Russell testified that AGI means
machines that match or exceed human capabilities in every relevant dimension
so the experts seem to disagree. (On the other hand, Stuart & Russell’s textbook cite Goertzel and Pennachin 2007 when mentioning AGI. Confusing.)
In any case, I think it’s right to say that today’s best language models are AGIs for any of these reasons:
They’re not narrow AIs.
They satisfy the important parts of Goertzel and Pennachin’s definition.
The tasks they can perform are not limited to a “bounded” domain.
In fact, GPT-2 is an AGI.
Maybe the right word for this would be corporatism.
I’m surprised to see an application of the Banach fixed-point theorem as an example of something that’s too implicit from the perspective of a computer scientist. After all, real quantities can only be represented in a computer as a sequence of approximations — and that’s exactly what the theorem provides.
I would have expected you to use, say, the Brouwer fixed-point theorem instead, because Brouwer fixed points can’t be computed to arbitrary precision in general.
(I come from a mathematical background, fwiw.)
For reference, here’s the Gears of Aging sequence.
This article saved me some time just now. Thanks!
Scaling temperature up by a factor of 4 scales up all the velocities by a factor of 2 [...] slowing down the playback of a video has the effect of increasing the time between collisions [....]
Oh, good point! But hm, scaling up temperature by 4x should increase velocities by 2x and energy transfer per collision by 4x. And it should increase the rate of collisions per time by 2x. So the rate of energy transfer per time should increase 8x. But that violates Newton’s law as well. What am I missing here?
This is the perfect time to start an AI + education project. AI today is not quite reliable enough to be a trustworthy teacher; and in the near future generic AI assistants will likely be smart enough to teach anything well (if they want to).
In the meantime, Eureka Labs faces an interesting alignment problem: Can they ensure that their AI teachers teach only true things? It will be tempting to make teachers that only seem to teach well. I hope they figure out how to navigate that!