I read "the AI con" recently. Interesting book about how this AI hype cycle is no different from the rest.

Now, here are my thoughts that include a significant amount of ideas from the book. However, i'm not an LLM so these are my own thoughts and not a regurgitation or a rehash (qualified later).

Book + brief thoughts on book

So on the things that I agree with the book (well at least how I interpret the book to say): 

1. LLMs are a dead end, no artificial general intelligence is coming out of LLMs

2. Much of what AI purports to be able to do, it cannot actually do (especially in the area of replacing workers)

3. Corporate leadership jump on the AI bandwagon too much and think that it can do much more than it actually do. In other words, AI appears better to leadership than to workers.

What this implies then is that:

4. The current capital allocated to AI is inefficient and there are opportunities to profit by betting against, especially, pure play LLMs - not that I am intelligent enough to figure out how. I tried for maybe 30 mins or so and gave up.

Ok. The book also seems a bit woke to me and talks about racism, environmental damage, oppression of poor and marginalised communities and other such inequalities. I'm not particularly convinced that AI is an exceptional harm on that front, any more so than the current structures already in place. At most it is a tool used by the powerful, but if not AI, I reckon any other tool would be as ably wielded. The fact that AI biases are embedded within AI is just a net neutral... no? I mean racist AI judge is the same as racist judge / jury. Wrt the environmental damage due to high energy cost, well I do believe that the world will reallocate capital away so it seems to me that this is just a short term hit, if any. You can't be right on both fronts that AI will doom the world and that AI doesn't work as people think it will... right? I doubt people are THAT deluded.

The book also goes into the history of AI hype cycles and covers the impact of AI in work -> sucks, doesn't work. Social services and the creatives -> sucks AND not human (not just sucky because not human) and how fearing AI that will takeover the world or that AI will lead to a better world is both sides of the same kind and AI will do none of that sort - well I suppose the analogy is quite apt that it will be something like maybe the steam engine over horses and wind, relatively and significantly change the world but not in a way that is like THAT big a break.

Theres also a large bit on AI being theft and stealing other people's intellectual property and such and such. To me if it is crawlable or on the internet then it is usable for AI training theres no SHOULD or SHOULD NOT, it just is and if you think it is a SHOULD NOT then it is on you to figure out how to enforce the should not -> sabotaging some datasets sound really fun and again can be done. This is the new reality. Theres also a bit on how it exploits people in the third world to do a lot of the manual training of AI models... well, nothing different from regular exploitation IMO.

AI - at best a useful next-rung tool

My main working theory is that this current version of AI and versions for the foreseeable future will suck and most jobs will be well protected from AI. And we should define AI here. The book actually goes to say that AI is such a shit term and is purely used for hype making people think there is intelligence when there is actually none - together with hallucination and other anthropomorphic terms. Within the current field of AI, there are quite a few distinct things that are being done, each of which have their own use cases. Examples of these would be categorising, decision making (or recommendation), text and image generation.

To me, I like the non-text based AIs a lot more because I think their use cases are clearer by far and there is no false impression that it will change the world apart from their specific use cases. I am a firm believer in AI as a large data model, with the purpose of predicting by extrapolating from past data except that the dataset is too large and complex to set explainable and deterministic rules. This would be, of course, why I own lmnd stock. But digression aside, this is actually what AI (or actually this deep neural net thing powered by massive compute which is what i'm going to mean when I say AI) is doing in all the forms be it categorising, decision making (or recommendation), text and image generation. I think if we think of AI as prediction machines from training data rather than possessing any understanding that would be quite useful. Here we have to recall the chinese room thought experiment from the philosophy archives and know that there is really NO ONE in the chinese room and hence no understanding DESPITE it being very very intuitive to us humans that proper wielding of language = understanding.

Honestly, regarding work, I feel like we are are so far away from using deterministic rules automations (i.e code) that we really don't need to try this AI shit for a large majority of the tasks because simply put, the data isnt that large or complex that it cannot be put into deterministic rules. The obvious benefit of deterministic rules over AI are that they are explainable and, well, repeatable which the AI isnt. Insofar as the rules are upkept (as I presume the AI's training data will be updated), code automation is more useful and practical any time.

Well then, obviously it follows that the more narrow the use case and the more tailormade the data supporting the use case, the better the results. This is because there will be insane amounts of noise if insane amount of data is put in, as any data person can tell you. Bespoke AI to do certain tasks will have an advantage if useful data is in large quantities and very snap decisions have to be made such that they may outperform humans on that very particular task, and these are seldom text outputs (LLMs) i'm thinking of but more like e.g classifying whether a particular transaction is likely fraudulent out of many many.

Thoughts on LLMs - stuck in the text

LLMs through some impressive amount of training essentially are almost able to "rule based" most languages and are able to output sensible sentences and paragraphs. This is already an incredible achievement, as many people seem still unable to speak in proper english. In my opinion, teaching proper language skills / rewriting in a style is probably one of the best use cases for LLMs because that is exactly what the training data is for (assuming there is some cleaning in the quality of the english or at least a large majority is good quality english hence doesn't matter). 

However, when you start using LLMs to make predictions or things on the real world, there is this very very obvious gap in information between what the LLM is trained on (i.e text) and the real world. Text, however useful it is, is but a limited representation of the real world and most definitely not the real world. If you do bible study with any amount of rigour, you would know that any text assumes a lot of hidden common shared non-textual information and is actually a means of communication between a reader and the writer. By chucking in a massive amount of text, presumably it is stripped away of its context AND the information stored in the "metadata" e.g this blog post is written by me at this time. Even if this "metadata" were included, the next step of deducing whats the significance of me and this time (e.g 2026 when AI is at such and such stage) would be very difficult for the LLM to discern which are the relevant facts (assuming that it has the relevant facts somewhere in its large database). 

For example, I was playing around in some LLMs, writing a story about Singapore and I noticed that it could spit out things that are commonly written down e.g common singlish, place names (especially web famous), weather .etc. Nonetheless, it made quite some rather absurd assumptions e.g that 25-35 characters stay alone - even more absurd, in a HDB, that random Chinese name is Muslim hanging out with a bunch of Muslim girls, that teachers start work at 9am. Presumably it knows that most people stay in HDBs, that these are common names in Singapore, that Singapore has a significant Muslim population .etc. Presumably if I prompt it to think about that particular issue or fact specifically it would output the right answer but if it is unprompted, the prediction would be too general and thus wrong. As such, in theory, you could prompt engineer everything deeply but then am not sure that the marginal gains are worth the very lengthy prompts and also it seems to me that LLMs don't work too well with very lengthy prompts. Also simple continuity issues really make writing more than say 3000 words a very big pain. I don't even know why that is given that the context frame for newer models are now very large, seems like the LLM is unable to correctly identify emphasis or need for continuity and get it from within the context frame and sometimes just plugs it with generic values from training base. Side note: I also find the writing style of LLMs for the narrative really repetitive and boring, like its boring prose, very clichéd. 

Putting aside the fact that it has randomness inbuilt, the very fundamental project of having LLMs give correct statements about reality seems to me to be fundamentally futile if we understand that LLMs are text prediction machines trained on text because text in itself is not a 1. full and complete 2. current and updated 3. POV-less representation of reality. This is made worse by garbage in garbage out and the fact that language is always evolving, as is reality.

In theory then LLMs should have a decent use case in areas of abstraction whereby language or symbols are reality by definition e.g math. I think a purpose built math LLM would do quite well and perhaps be able to potentially challenge some frontiers in math (i dunno for sure). But general LLMs seem to quite suck at even basic math at times due to randomness while math is deterministic. Furthermore, unless it is a very obvious application I highly doubt the LLM's ability to predict a novel application outside its training dataset - and our math as of now seems to require very novel applications of disparate observations, long gone are the days when you can just think hard about something recursively and come up with new designs (though to be fair to the pioneers, theirs were probably very novel applications as well and a LLM trained on text at that time would suck incredibly.)

So what LLMs have going for them is a massive amount of text that has some verified claim to reality or at least a communicated representation of reality and massive computing power to process everything and lots of poor people to train it to be "better" manually. Insofar as it is a massive corpus of text that is somewhat organised, I guess there is some value to be had there although the next best alternative (not even sure that it is a worse alternative) is a search engine that can find you the base text for you to read, interpret contextually and make judgements by yourself. Instead you get this generalised output from the LLM that might include low quality crap and reasonably quality crap all muddled together such that you can't tell the difference and without sources to boot. So you end up with really slop, low quality trash that might be passable in many cases but you don't know which is full of trash because it is stripped of all context and your brain becomes dumb because it just eat the slop.

Adding something on about work

I also fully agree that managements love AI a but too much with too little understanding and shoving it down employee's throats will bring no happiness to anyone, not least the employees. So far all the LLM solutions that I see in my workplace all really cannot make it. If not for this hype and oh yeah look at us do AI, i'm quite sure that they will all be thrown out. The problem is really that the AI solution is not bespoke enough, uploading a few documents into a "knowledge base" and slapping on a generic LLM is WORSE than just ctrl-f ing the documents.

Especially with a particular agency that thinks there are a million useful use cases for bots that can be created in 5 minutes. Wtf seriously how do you think that you can create anything useful in 5 minutes??? There are just not enough parameters adjustable for it to give any level of differentiation with just talking to a generic LLM. Honestly they are super lame, grok "projects" probably is their entire model and the value add is... suspicious at most just yknow you dont need to repeat the prompt all the time i.e templated prompt. Its crazy why its still being funded and pushed.