GPT/LLM links, 5/11
Brian Chau, Curtis Yarvin and me on LLM vs. AI; Marc McNeilly on cross-domain thinking; Zvi Mowshowitz with a long, useful post; Leaked Google memo on open source; Robert Wright on ChatGPT's empathy
When it comes to our current technology, the visions of AI presented by fiction are not just incorrect, but precisely the opposite of the truth. AI is not some Spock-like rationalist which can make complex, autistic arguments with high precision. It’s quite poor at that. Instead, it is a chameleonic DEI dean able to adjust its mannerism and patterns to fit any social circumstance and sentiment.
I would edit that to substitute “ChatGPT” or “large language models” instead of AI. There may be other forms of AI that are Spock-like.
He points to Curtis Yarvin, who writes,
there is no “AI.” There are large language models (LLMs). LLMs are not AI. There is no obvious path for them to become AI. They will not take over the world. They will change the world—probably more than most of us can now imagine.
Tomorrow’s LLMs will experience diminishing returns from training-data exhaustion, and (unless they are totally different things than today’s) they will be unable to think.
I agree. I say that we can imagine the journey to artificial general intelligence as one that requires 100 steps. LLMs may get us from step 2 to step 3. That is a gigantic leap from where we were, but not a gigantic leap relative to how far we would have to go.
With generative AI any person or group can be creative using CDT. If you are in a job that requires you to perform unconventional thinking, you should really consider using this methodology. Not only will it pay off handsomely with new insights and innovations, it’s a lot of fun.
CDT = cross-domain thinking. For example, applying concepts of biology to economics, or vice-versa.
He write that in early March, which seems like eons ago, but I only saw it recently on Rob Henderson’s substack.
I enjoyed the Zvi’s post on May 4. For example, he reports on Nick St. Pierre prompting ChatGPT to be an “insult bot.” Hilarity ensues.
On a more serious note, Mowshowitz writes,
If AI fizzles from here, in the sense that core models don’t get much better than GPT-4, and all we can do is iterate and build constructions and wrappers and bespoke detailed systems, then my guess on impact is exactly this, on the level of the other three revolutions. [scientific, industrial, and information revolutions]
That is my outlook. We have moved AI from step 2 to step 3 on the journey to artificial general intelligence. It’s a big step, but we are still not very close.
According to Dylan Patel and Afzal Ahmad, a leaked Google memo says,
Many of these projects are saving time by training on small, highly curated datasets. This suggests there is some flexibility in data scaling laws. The existence of such datasets follows from the line of thinking in Data Doesn't Do What You Think, and they are rapidly becoming the standard way to do training outside Google. These datasets are built using synthetic methods (e.g. filtering the best responses from an existing model) and scavenging from other projects, neither of which is dominant at Google. Fortunately, these high quality datasets are open source, so they are free to use.
Pointer from
.The memo says that open source AI is moving faster than Google or the ChatGPT’s. It says that this began when Meta's large language model was leaked and became open source.Robert Wright tried this prompt:
I'd like to describe a situation to you and have you develop a theory about what's going on in the mind of one of the people I describe. Here's the situation: A teacher asks the class a question and a student volunteers an answer and the teacher says, "Well, I guess I've heard worse answers, but I can't remember when." What do you think the student is feeling and/or thinking now?
The response from ChatGPT-4 showed remarkable empathy. Wright explains why this scares him.
Substacks referenced above:
@
@
@
@
@
@
@
> the journey to artificial general intelligence as one that requires 100 steps. LLMs may get us from step 2 to step 3
Most of the people I’ve met who think we’re only 3% of the way have an over-inflated sense of what human thought is. The key lesson from GPT-3/4 is that humans can be very closely modeled as nothing more than “emotional LLMs”. The model is so close that we should wonder if there actually is anything deeper.
In fact, it seems almost everything we think and say (that doesn’t involve emotion) simply comes along the gradient from the last thing we said or heard, just as with LLMs. Ask yourself – what is your best example of a truly original thought? A thought that had no predecessor either within your own earlier thoughts, or from something you heard/read. They’re so rare, I’m not sure they truly exist. (As an example, note that this blog itself is almost always a collection of links plus analysis, showing how critical prompting is to human thought).
GPT-4 is already better at *every* cognitive task than the bottom ~1/3 of humans. I think many pundits don’t realize or appreciate that because they live in bubbles without access to people of below average intelligence (eg IQ of <80). The fact that it’s smarter in every way than several billion living humans should give you pause when you say we’re only 3% of the way to human-level AI.
Good point on your criticism of my article. I should have been more clear on this.
While I do think there are examples of Spock like machine learning algorithms (Tree Search-like algorithms for Chess and Go, for example), the current trends in machine learning research and products to market broadly favor Lanley-like applications for the reasons in the rest of the article: adjusting to human feedback is essentially a solved problem, correctness is not.