Saturday morning cleaning out my browser by sharing links and closing tabs

Happy Saturday! Today, I'm decluttering my browser by sharing a curated mix from this week's open tabs and browsing. Topics will include the psychological twitches triggered by AI, the freshest slices of memory tech, and the ever-evolving economic landscape shaped by generative AI. I’ll also dip into the world of mysterious new chatbots, the ripple effects of AI in academia (including its impact on publishing in prestigious outlets), and how AI is reshaping legal and video landscapes. But if AI isn’t your thing, I’ll also tell you about new tech gadgets, local versus national burger joints, and home security in the age of squatters. Most of it is behind the paywall, as it’s a special thing I do for the paying subscribers. But you might like becoming a paying subscriber so think about a 7-day free trial!

Memory and the Hazards of Lifelike ChatBOTs

I continue to be interested in OpenAI’s quiet updates to ChatGPT. It seems like it’s everyday there’s some subtle new thing, and I can’t quite keep up. If you aren’t a developer of AI tools, I think it’s easy not to see these things, and a lot of them I don’t understand well, but the one I am experimenting a lot with this week is the memory update I mentioned recently (see below).

The Pinocchio Effect, Economies of Scale and Market Concentration in AI

Memory is something that ChatGPT-4 is doing automatically now, though you can toggle it off. ChatGPT-4 will get triggered by something in the prompting you’re using to “remember” something. Maybe it will “remember” that you are working on a paper, or that you are trying to find some luggage for your trip, or whatever. It seems like something you’ve been saying in the prompting will cause it to go into memory mode, and then store something into like a second set of "custom instructions” (which was itself an earlier update), and then with both custom instructions (which are things you record) and these new “memories”, it interprets the prompts with them in mind. They say it is being remembered in something called “like a notepad” and it can be forgotten whenever you want allegedly. You can even tell it to forget those memories.

Well, what I decided to do was take more control of that process — rather than simply relying on ChatGPT to use its own triggers, I began to tell ChatGPT-4, “I want you to remember something for me about me”. Not really sure where it’s going, but I’m convinced that memory will be a major part of the difficulty that consumers will in the long run have to switching away from OpenAI's products. If you read the fine print, closing your account with OpenAI will cause your chatbot to forget everything.

So, what happens as these chatbots converge closer and closer to something like companions? The other day in the NYT, the headline read “Meet my AI friends”. A NYT columnist spent a few weeks with 18 chatbot “companions” which I guess is some more formalized AI created friends you can do somewhere (I didn’t read the article closely, hence why it’s on the “open tab” list). Well, again, I continue to maintain that I wouldn’t want to be in competition with OpenAI for any AI service, because if I wanted to make an AI friend, it would take me all but a day to do it with ChatGPT-4. Although I’m sure that OpenAI has built into its LLM a ton of guardrails to make that hard, as I get the distinct sense that OpenAI is a very risk averse company unwilling to allow users to do a whole lot with it that could be compromising, and what’s more compromising than a large language model so realistic that people begin to develop feelings for it? I honestly have to believe that is a risk. Again, if we anthromorphize pets and plants, or we name our cars and boats, then how much more is it a risk that LLMs become so life like that people end up … gasp, dare I say … falling in love with them? Probably should put into your custom instructions and memories sooner than later, “ChatGPT please remember that I do not want you to do anything that would invite me to fall in love with you”. I think we are moving into uncharted territory with many hazards that we simply aren’t ready for at all.

In essence I don’t think the risk is that chatbots “become” sentient. I personally don’t think that is happening or will happen. I think the real risk is internal and psychological and nearly unavoidable, almost like we are hard wired to do something else. I think the real risk is that we will feel that they are alive. We will in essence become like Dr. Frankenstein, not to be melodramatic, and create living creatures out of things that are not living creatures. And the reason is, I think people do this already with far less compelling inanimate objects than ChatGPT-4. And the effect that this will have on OpenAI’s market power is probably unexplored, as I wrote about earlier in that substack above on “The Pinocchio Effect”. Not sure if these are dumb ideas or whatever, but I personally think as the chatbots become realer and realer in our minds and feelings, our willingness-to-pay grows, and switching away becomes astronomically expensive. All of which favors OpenAI becoming more and more like a natural monopoly given the high fixed costs of training these models and the implied economies of scale that come with that. Although for all I know the variable costs of a chatbot are actually gigantic too.

Mystery ChatBot

There’s been a bunch of stories of a “mystery” large language model that goes by the name of “gpt2-chatbot”. It showed up at this site maybe 1-2 weeks ago and had all the signs of being an extremely advanced LLM. The weird thing, though, was it showed up, then disappeared, then showed up again. But weirdly when it returned, it had two names: “im-a-good-gpt2-chatbot” and “im-also-a-good-gpt2-chatbot.” Sam Altman made some cryptic tweet about it and now there appears to be confirmation that is indeed an OpenAI LLM. But you can only get it randomly on this site which I’ve never used but is some kind of roulette arena thing where LLMs I guess fight each other? I don’t know. It’s an open tab, like I said. People are speculating it is ChatGPT-4.5 or even 5, but I think it’s hard to say much because you can only get access to it in that arena.

Economic Impacts of Generative AI Report