I get "it's not good enough". I get "even occasional mistakes are unacceptable in my field". I get "it's not really intelligent" even though I think that's a question of terminology.
I don't get "worse than crypto" or "unable to find a single use".
I didn't say it's worse than crypto. Just a biggest waste of resources by tech companies.
Crypto never got investment from tech companies, but those have already diverted millions of hours of work from other projects to shoehorn GenAI garbage everywhere they can.
Similarly, I didn't say I wasn't able to find a use. Just not a legitimate use. Something that is a net positive to society and couldn't be done better without GenAI.
> I didn't say it's worse than crypto. Just a biggest waste of resources by tech companies.
Currently that comment says "the biggest waste of resources of the tech world. Even before crypto." — even if it was not your intent, you wrote something which absolutely can have the meaning I read.
> Just not a legitimate use
No true scotsman. For any value of X, at least one person will claim that X is not "legitimate".
I, like everyone else here going "WTF?" at you, find it totally legitimate despite its flaws.
I literally just used OAI to walk through my financial plan and suggest my next steps towards retirement. It pointed out all sorts of things like conversion ladders, my cash on equity for my rental properties, average cost of health insurance and more which would have taken me a ton of time to research on my own.
I don't agree, but the current usages aren't exploiting its full potential.
GenAI is the best that we currently have for parsing natural language, in a way that is multilingual, tolerant 2 tipoz n slang and swearing users. It helps by being a bridge between unestructured data and structured.
It's actually terrible at parsing natural language. So bad that on a long enough text (or even short if you're unlucky) it will 100% of the time come up with tokens that are not present in the original text.
This sort of rethoric is exactly the same as with crypto "yeah ok it's bad now but think of the future".
Sorry you have had such bad experiences we won't be able to convince and nobody can see the future but there are exciting things happening at an amazingly short scale.
Really? ChatGPT 3.5 and beyond models are fairly capable of understanding PoS and doing text analysis. I have never seen that issue yet with the more advanced models, although smaller/older ones tend to imagine fmthings about the text.
Last year I wrote a paper about using LLMs for definition generation for unknown words based on context, and the models did a fairly good job. https://ieeexplore.ieee.org/abstract/document/10346136/ if someone is curious.
I would like to read prompts where the models are failing in such way. The field is moving quite fast.
I think one of the techniques underexplored in all the hype is guiding the evaluation process depending on the context. I.e. if you're generating code, it has to satisfy the parser for the given language. If the token is unsatisfactory, throw it out and try another one. Thought chains could be generated in a similar way (you can do so with special tokens, see "Recursion of Thought").
But yeah overall GenAI tends to remain hype-over-substance.
The main blocker for this is that LLMs are slow. Imagine waiting 3 seconds for your output in a pretty happy case, only for it to be invalid and have to wait an extra 3 seconds, with again non negligible chances of being wrong.
We envisioned doing this for an SQL query generator at work but with our constraints a single query already takes 15 seconds.
While there is definite hype and buzz among executives to jump on the bandwagon, from my experience and perspective massive amounts of value to be obtained. We are doing work with LLMs that would have previously taken teams of people to do. Many legitimate uses but overshadowed by the hype.
I think the way the corporates are applying it are not very useful, specifically the Googles and Microsofts of the world. I am overall bullish on the niche applications of LLM though. Google and Microsoft are just throwing it at everything and I don't think much of it is sticking. The Google search experience has definitely downgraded with their AI implementation. Kagi on the flip side has done a much better job imo, it does not get in the way and it is generally answering the question I asked.
I am bullish in areas outside of RAG/chatbots that gain so much of the hype. Classification, extraction, summarization and similar natural language workflows. My work is net positive but again anecdotal to my corner of the world and the others I interact with. Could be different from your distant side of the world.
To me the most clear-cut benefit of AI is automated content moderation. People literally get PTSD from moderating Facebook. It's no different from replacing the most hazardous factory work with robotic automation. You will still need a human in the loop for the more difficult cases, of course, but by any reasonable utilitarian calculation AI content moderation is a win.
(I'm not implying anything about whether or not LLMs are good for Google search.)
> Maybe someday but not with LLMs, which by nature do not understand who's talking, who is being quoted, and who is being falsely quoted.
Exact opposite. Humans don't have time to work out any of those things in practice. Machines do have time, and LLMs make a much better job of those things given the real limitation on human labour that actually exist in practice.
The genie left the bottle. The Large Bullshit Confabulator revolution is here and either google does it or someone else will.
Seriously though, Shannon's theory of communication is highly applicable here - AI noise is reducing available bandwidth (in the entropy sense, not in the advertised Mbps sense) of the internet because the noise floor is raising faster than the error correction advances.
> I am still unable to find a single legitimate use.
Really? Nothing? Not even really clear-cut use cases that are already in production at a bunch of companies like rapid document templating or surfacing esoteric yet relevant internal documents and knowledge?
Those use cases alone have saved seven figure sums at companies where I've seen them implemented. And those savings allowed the positions to be repurposed in more useful ways e.g threat hunts rather than document prep.
---
Lots of adjusted goalposts in the comments below. The fact that I can simply ask an LLM to bake me a template and adjust as needed — or ask it to fetch me an internal policy document describing compliance requirements I might need for a new mobile app — and get either of these answers in literally seconds makes them the best tool for the job by a wild margin. And the fact that I'm seeing hiring managers rework their positions for different roles as a result of implementing GenAI to obviate mundane job responsibilities supports this.
Many practitioners across many fields have their blinders on with regards to the risk of disruption to their own disciplines. Surprised I'm seeing those blinders on here too.
It's true, LLM AI is really great at creating content as well as spam and other "slop". While this is useful in some circumstances, I think it's a net negative in the long run. Once enough slop is created, will AI be able to sort through it?
"Slop" is a good way to put it. AI seems really useful for creating a huge sludge of "filler" art and words on a particular topic. The quality may or may not be good, but it can produce endless quantity of this slop.
So, if you need to fake a blog by generating 10,000 posts on a topic that are each, say, 1,000 words long, AI is great. If you need to make a web forum that, at a surface level, looks active, AI can generate all the fake posts you need. If you have something you can say in one sentence, but need to make your text 10,000 words in order to appease an algorithm or "look credible" then AI slop is the way to pad it out.
The way I see it, many of the "good" uses of LLMs boil down to easily counterfeiting traditionally-trusted indicators of human expertise, time investment, or intelligence.
For example: "Hey LLM, generate a smart sounding cover letter for a job at AcmeCo as a sales representative, highlighting my diligent work ethic and ability to generate new sales."
When the counterfeiting becomes widespread, those writing indicators become debased as receivers start to realize how easy they are to fake.
I wouldn't be entirely surprised if some of those prose-heavy documents started to move (culturally) toward raw bullet points.
> All GenAI is good for at best is prototyping to show what COULD be done if you invested the time to do it without GenAI.
That is the argument used against mechanical freezers replacing natural ice and power looms replacing artisnal weavers and coffe shops replacing real human baristas with identikit machines.
None of those use cases actually require AI, and each of them is potentially undermined by the tendencies of AI to hallucinate, or as yet unresolved legal questions about the copyright status of AI content.
Can you name a use case for which AI was the only or best possible solution?
Very quickly finding texual needles in long-document haystacks. Can't generally do that with a plain text search unless you already know what the needle looks like, and doing it as a human is expensive — see Terry Pratchett's experience with German soup adverts.
Real-time translation within the price constraints of tourists.
