Is General Artificial Intelligence a Breakthrough or just a Financial Hype?

Unfortunately right now the tech industry is heavily investing in large language models, even when they are not the kind of AI that is helping science. As far as I understand, the cost of running those hallucination machines it still is way above the value they provide, so I don't expect that to be a "trend" that lasts. Unless they manage to massively scale down the resources required to run them, it risks becoming just another piece of hyped tech used to make a quick buck from investors and early adopters before it dies (like what happened to NFT and crypto).

Investors invest to make money. Material science is definitely worthwhile if it solves a problem.

AI investment does surprisingly subtle things - like channeling advertising to those with a certain product interest. AI is learning all about us online, to leverage and manipulate us for someone else's profit and power. Its no doubt that power is worth investing by those with the $$$ to fund it. 

IMHO it's not hype. Over the past 30 years I've seen various forms of software being developed by various teams. It's typically a case (generalising here) of you get about 90% done quickly and then it takes you about 9 times longer to close out that last 10% as best you can. It's somewhat messy, inherently buggy and pretty slow. I'm experiencing first hand that with the help of AI you are getting to about 98% done in about 50% of that initial 10% of time and then the final 2% is done in about 3 times longer than that initial phase. So overall, quality improved and time saving massive. The amount of resource required is also substantially reduced although you definitely need experienced resource to achieve. Junior engineers and AI is a toxic mix at this stage simply because they don't recognise the problems... and AI at the moment needs direction. It's power at the mo. is comprehension (reading and interpreting 100 page docs in seconds).

People (including some tech experts) thought the cloud was hype and a fad, and it took less than a decade for it to become crazy to not use the cloud for most businesses. Others assumed IoT was useless, and yet there are now hundreds of devices all around you wherever you go, being monitored or controlled over combined wireless and Internet, from energy meters to vending machines. Probably around 60% of my search engine queries have now been replaced by LLM interaction, it just works better. I use AI to help me with all manner of tasks (non-technical and technical).

Do you remember how scientists were able to use DNA as a storage,1 gram could store many petabytes but it wasn't popularised even though it was some same level crazy stuff like Chatgpt and all, I personally think it's not extremely populariser

Personally, I still think cloud is crazy without homomorphic encryption. Businesses are risking their IP, at least, by putting control over it in the hands of others.

To be fair for every example you have of a tech that people doubted, but ended up becoming crucial, there's an example of one that was touted as the next best thing but failed. Ruby on Rails was supposed to be the future of web apps, 3D displays were the future of TV but were never widely adopted. AR/VR never really took off even when Facebook(now Meta) believed in it so much that it shaped its new identity around it; the blockchain was hyped to hell and back, and now is quietly in a corner, only used for the few crypto outlets that survived, etc.

Now, I'm not fully against LLMs as a technology. I'm very well involved in the development of AI-based systems and I deal with different LLMs on a daily basis. I've also tested them in different tasks that are part of my job, and there's only been a very small subset of tasks that they perform adequately, or where they are actually a help, and not a hindrance.

Right now they are struggling to make sense as a business model, and I don't see them being as widely adopted as investors/tech bros want them to be unless they solve several problems with them (from ethics/morals to financial, technical and practical). So, repeating a bit of what I said earlier, unless they find a way of solving the shortcomings and cutting their running costs down, they will slowly stop being used everywhere, and it will become tools only used in a handful of contexts, for a limited number of tasks (like summarizing reports and meetings in the corporate world, auto translation of videos on social media, etc).

Part of their problem is they keep overpromising and overselling the tech, hoping to increase adoption and then "solve the issues" once the user base increases, but that's been a very bad strategy so far.

Luckily, "regular" AI (i.e: Not gen-ai) continues helping science and will do so even if the LLM hype dies. What could be a problem, however, is that maybe at one point people will get so "sick" of hearing the word AI that actual AI-based research will receive less funding because of this poisoned association.

battlecoder said:

What could be a problem, however, is that maybe at one point people will get so "sick" of hearing the word AI that actual AI-based research will receive less funding because of this poisoned association.

Back to the 90s and the first "AI Winter"...

I don't know enough to comment with much authority, but I think that most orgs consider the encryption enough of a solved problem. The orgs get to control where their data resides, and almost all orgs have to trust other orgs to transport their data anyway (I.e. even if they were not using cloud, but just moving data to a redundant site they themselves own - and most of it even used to be unencrypted links if it wasn't on the public Internet), until Snowdon era.

From what little I've seen, orgs with a tremendous amount to lose, such as banks, trust the cloud (not all PaaS features, but a large enough subset) for implementing services, and they trust the public Internet for their customers to perform online banking and commerce. (Of course, as mentioned we know since Snowdon that all that data flows through certain rooms, so nothing is 100% certain).

The cloud providers cannot decrypt (or claim they cannot) the data belonging to businesses who are using their secure technologies. Governments also use the cloud (admittedly only specific data centers owned by the cloud providers).

I'm not sure how homomorphicness helps, I'd need more input because I'm no subject-matter expert at all. I had a job interview one time for a firm that wished to do something like that (they didn't call it that at least to me) for medical data (since it's highly desirable to look for health trends and anomalies if only it could be done without providing health records that are traceable to individuals, except once it is known for sure that some of those individuals need to be contacted in a matter of life or death, and then it needs to be reversible by the original data owners), but I think they will have difficulty marketing it, because to me it seemed high risk that something would eventually leak, let alone the issues of getting governments to allow such transfer of data to orgs for data mining, and talking to their mathematicians, didn't give me a warm feeling (granted they were talking to a non-expert, plus I was not an employee). 

I wish I knew more than this vagueness of my current knowledge, I'd like to learn more through self study, but I suspect I'd end up stuck as soon as I hit any reasonably difficult maths.

Not heard of that! Maybe it's plausible, I don't know (my knowledge of chemistry and biology can be written on a postage stamp).