I’m impressed
For the last couple of years, my colleagues and myself have been playing around with AI. I can’t speak for them but from the perspective of a casual user I am very much impressed by the capabilities on display. If you dedicate some time to it you can get a lot of mileage out of AI, it can help you in the way a highly interactive book would help, with vast amounts of knowledge at the tips of your fingers. As long as you are wary of the inevitable hallucinations and the AI’s tendency to get stuck in loops. Because the one thing that we have not yet managed to program in there is the ability to be humble: to realize that no matter how much knowledge you’ve got you don’t know everything. So “I don’t know” is not on the menu. It is not the kind of thing I have seen as output from the various AI offerings.
If you had told me ten years ago that we would be having conversations with a machine at this level today I would have thought it to be impossible, and yet, here we are. The ‘Attention is all you need paper’ combined with Hinton’s persistence paid off immensely, probably more than even the people behind that effort had expected. Current generative AI is able to converse on an immense variety of topics at a reasonably high level.
Professional use
Even so, I would never use its output in a professional setting. When we are hired to do work we are hired to do work, not to come up with prompts and then to let the AI do the work. Anybody can do that, it would feel like fraud to use AI as some kind of shortcut, then we may as well give our customers the prompts.
As a search engine and uniform interface to documentation though, it is unparalleled, as long as you check the output for factual correctness.
Commercializing AI
Because we like playing around with new tech we put together a monster computer with a lot of GPUs and a very large amount of memory. This gives us the capability to run these models in-house and to evaluate their abilities without accidentally leaking our data to the AI providers. What struck me is that in any conversation longer than a couple of lines you end up with one or more times where you have to correct the AI on some point, after which the conversation may stay on the rails a bit longer. But there is always a point where the growing context becomes problematic and of course when you start a new session all of those corrections are forgotten. That plus any kind of formatting instructions regarding the responses have to be repeated over and over again at the beginning of every new session, a kind of ‘system prompt’ at the user level. Running a system like this commercially requires a lot of power, memory, computation and storage for every a single user. I’m fortunate in that I have a lot of solar panels but if not for that my power bill would be astronomical.
And that’s where the data centers and the price of RAM and GPUs comes in. The price of a stick of DDR5 memory must have the smell of onions baked in because when I look I start crying. This will force the AI companies to commercialize it but their avenues to do so are limited, and I’m not sure how they think they are going to survive without doing some kind of rug-pull.
Confidence
What bugs me about the interaction with the various AI models is that they are so incredibly over confident. They are jump-to-conclusion machines that will answer a one line prompt with six pages of plausible nonsense as often as they get it right. The signal-to-noise ratio can be very low and if you’re not careful you will end up reading a lot of self congratulationary bullshit that does not move the needle at all in terms of your understanding or insight. The number of times I’ve seen the word ‘working’ for something that could not work and ‘no handwaving’ followed by a lot of handwaving is pretty high, at least as high as the number of times that I came away with the idea that I had genuinely learned something. It would be great if the current crop of AI used font size or levels of gray to denote the things they are actually confident about and the things that are speculative. This alone would significantly increase its value to me.
When it does not work
The most use I seem to get out of it is when I’m studying some subject that I already know the basics of to increase that knowledge and give hints on fruitful avenues of reading or research. It rarely happens that the interaction with the AI itself gives direct insight on things that are non-obvious. Over the course of the year that has happened probably a handful of times. The more clueless I am about a subject the less use I get out of AI. Conversely, the more I know about something the more I’m irritated at the basic stuff that it - repeatedly - gets wrong and you end up with endless corrections. It is able to pull a massive amount of information up at a moments notice but it will happily forget something that came up only a few minutes ago. This can be quite annoying, but then, whenever I’m annoyed I am also impressed: I am annoyed at a machine for not being able to so something as if it were a person. And it clearly is not, despite all of the language hints that it is.
The Sleaze
Some of the models that are in widespread use are extremely slick at giving the user the impression that they are smart. They will endlessly congratulate and reinforce that the user is a smart person even when they’re clearly not. The paternalizing attitude is so ingrained that even after multiple reminders it will often just find new ways to effectively do the same thing.
The Hive Mind
After playing with something for a while I always wonder: What’s next? Where are we headed? I think the current crop of AI products is useful, but the push towards commercialization will most likely destroy a lot of that utility. The same has happened with all other media that preceded it and I think it is fair to consider AI not so much a new block of interactive content but to see it as a new medium, a way to access information. It got jumpstarted by violating copyright on a massive scale (violating copyright law is what will get individuals into the poorhouse and will get companies very large valuations). But once that was done and the obvious content mountains exhauste the race was on to mine the remainder.
Content that was stored off-line or behind closed access (what are the chances that the likes of Google and Microsoft will be able to resist the lure of your mailbox and your file storage?) and content that was stored in inconvenient forms (books, for instance) all got digitized and gobbled up. The end of that is in sight as well.
But there is one source of fresh, and mostly un-tainted content that will remain: the direct interaction with the AI’s themselves. And this is where the future lies. I strongly believe that endless re-training AI on the conversations it has with actual humans is what will give the AI leaders their edge. This is content that only they have and that gives them a strong signal to help fix their confidence issues and the hallucination problem. Humans are correcting AI, effectively giving very high grade labeled training data every day. And lots of other useful information as well. By moving from a batch learning model with a release every so many months to a continuous learning model where your conversations are first stored as a shell around the original AI release and then later become incorporated into the inner core we will most likely end up improving the AIs we interact with the most.
It’s funny because we have to pay to access the AI which recycles a lot of stolen content and then we have to pay again to be allowed to teach it! Effectively the AI will become a crossbar between the brains of all of the people interacting with it and the closer that gets to real time the closer we will be to an actual hive mind. And deciding to ‘opt out’ will at that point probably make you effectively unemployable, you will be either augmented and expected to deliver or you will be on the dole.
The next decade will be interesting, and if there one thing I’m grateful for then it is that this wave hits me this late in my career, mostly because I wonder if there will still be room for dinosaurs like me, who refuse to have anything to do with AI in a professional setting. No matter how much Google wants to ram it down my throat.
AGI
Another often seen theme is that of AGI, the bigger brother of today’s systems. There is this hope amongst AI proponents that it will be the new frontier, super human intelligence. But I don’t think that that will matter as much. People are already using these systems as though they are AGI, and in that sense the narrow academic or techno definition is irrelevant. What matters is how people perceive it and when you look around you and see people talking to AIs about what they should cook for dinner and relationship advice then it is clear that even if AGI never actually happens it isn’t going to be a blocker to further adoption. We have to be careful that we’re not accidentally creating a new religion as well, the AI worshippers. I think I’ve already met some.