Good lord! #gpt4 is a workhorse! Most humans will have difficulty generating a table like this:
The same table, but pivoted!
I can slice and dice comparisons of cognitive theories. Discovering alternative perspectives on the fly! This thing is definitely a superpower! It's like having several tireless graduate assistants on your becking call 24x7.
Need to start thinking in Aspect Oriented design!
I'm looking forward to seeing the API. I'll need to know how to get new knowledge into this system. For lesser-known theories (like @drmichaellevin TAME), it hallucinates the concepts.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
1/n Wolfram's physics (@wolframphysics) has many rich metaphors that I find useful for reasoning about deep learning AI. Here's a short digest from ChatGPT explaining Wolfram's physics:
@wolframphysics 2/n Deep learning is a multi-way system with many parallel derivations taking place. In Wolfram physics (is this what it is called?), one can derive quantum physics and relative through the projection of the multi-way space to different reference frames. It's like a lens.
1/n The private biotech industry perhaps has at least 10-20 years more accumulated knowledge than the academic world. A majority do not realize this. This is why nation-states like China and Russia cannot create vaccines like in the West. AI will end up similarly.
2/n This large disparity in knowledge and know-how is a consequence of the evolutionary nature of biology. This nature also exists in deep learning AI. Evolution's creativity is a consequence of frozen accidents; these accidents cannot be uncovered through first principles.
3/n Genentech, an early pioneer in genetic engineering, has specially engineered organisms (i.e., mice) that the rest of the world cannot access. Many biotech companies have secret sauce that can only be discovered through experimentation.
1/n Time to explore Alexander's Pattern Language and GPT-4. By default, GPT-4 appears to be familiar with Alexander's ideas.
2/n You can even ask it to do the reverse query. What forces are mentioned in which pattern.
3/n I'm not an architect so I don't have a deep intuition on these patterns. So let me explore instead something I'm more familiar with. The GoF Design Patterns:
1/n Humans invent models of reality so that they can reason about it. The ancients used gods as proxies to model natural phenomena. Today we use abstractions. But it takes effort to understand how models compare to each other. GPT-4 helps you explore models through many lenses.
2/n It is often the case that we conflate models that appear the same when in fact, they are slightly different. In the above table, we can see the commonalities and differences. It's the differences that aren't as obvious.
3/n When we compare models, we do so through features that are shared by these models. This implies a meta-model where all models belong. What is amazing about GPT-4 is that it has a meta-model of those features. Where does this meta-model come from?
When you first play around with txt2image generators like Dall-E and Midjourney, you come to a realization that (1) you don't know what to generate and (2) you don't know how to prompt in a good way. This kind of revelation hits you later when you use GPT-X.
ChatGPT worked around this problem through an improved conversational interface. It's analogous to incremental disclosure in UIX design. Don't demand that a user conjure up everything up front, but instead gradually ease the user into a conversation.
This may be fine for a user of ChatGPT, but as an application developer, I'm trying to find out how to make this tool jump through hoops. But the problem is, I don't know what a jump is or what a hoop looks like. What I mean to say is, are their incantations that can't be… twitter.com/i/web/status/1…
GPT-4 is extremely good and can perform tasks beyond the average human. Even if GPT-4 remains unchanged, we cannot predict the kinds of applications that will be invented. But for many, it's difficult to see what GPT-5 would look like.
I have a AI capability maturity framework that I formulated a few years ago. It needs a few small revisions, but it's remains informative as to what the next steps in AI evolution will look like. medium.com/intuitionmachi…
The maturity framework is based on seven levels (image below needs to be updated). GPT-4 is what I would label as an artificial fluent systems. It's limitation is that of its limited modality (text and static imagery), but for the domain that it has been trained on, it extends… twitter.com/i/web/status/1…