I'm finding the detractors worse than the hype, because it seems like a certain subset of detractors [0] formed their opinion on AI in late 2022/early 2023 when ChatGPT came out (REALLY!? Over 3 years ago!?) and then never updated their opinions since then. They'll say things like "why would I want to consume X amount of energy and Y amount of water just to get a wrong answer?"
In other words, the people who think generative AI is an absolutely worthless and useless product are more annoying than the ones that think it's going to solve all the world's problems. They have no idea how much AI has improved since it reached center stage 3 years ago. Hallucinations are exceptionally rare now, since they now rely on searching for answers rather than what was in its training data.
We got Claude Desktop at work and it's been a godsend. It works so much better to find information from Confluence and present it to me in a digestible format than having to search by hand and combing through a dozen irrelevant results to find the one bit of information I need.
[0] For the purpose of this comment, this subset is meant to be detraction based on the quality of the product, not the other criticisms like copyright/content theft concerns, water/energy usage, whether or not Sam Altman is a good person, etc.
Detractors, doomers, and techno-pessimists have got to be the most consistently wrong group in history. https://pessimistsarchive.org/
But I do think humanity is worse off because of it. So I'm a detractor in that way. :)
Well, I wouldn't go that far, but the hallucinations have moved up to being about more complicated things than they used to be.
Also, I've seen a few recent ones that "think" (for lack of a better word) that they know enough about politics to "know" they don't need to search for current events to, for example, answer a question about the consequences of the White House threatening military action to take Greenland. (The AI replied with something like "It is completely inconceivable that the US would ever do this").
I mean, you can get mad at people you made up in your head, that's a thing people do, but this caricature falls in the same comforting bucket as "anyone who doesn't like <thing I like> is just ignorant/stupid" and "if you don't like me you're just jealous".
Maybe non-straw people have criticisms that aren't all butterflies and rainbows for good reasons, but you won't get to engage with them honestly and critically if you're telling yourself they're just ignorant from the start.
For example, I will bet that non-straw people will take issue with this, and for good reasons:
> Hallucinations are exceptionally rare now
In contrast, what harm do those detractors cause? They don't generate as much code per hour?
The "harm" (if you can call it that) is clear, detractors slow the pace of progress with meaningless and incorrect hand-wringing. A lack of progress harms everyone (as evidenced our amazing QoL today compared to any historical lens.)
Considering our climate, political and economic situation, I'd say not only is slowing the pace of progress not harmful, it's actually imperative for our long-term survival.
Also we need detractors because if we race into any technological advance too quickly we may cause unnecessary harm. Not all progress is without harms, and we need to be responsible about implementing it as risk-free as possible.
This is such a perfect example of the mania behind this rollout.
There's no way you can make the financials work here compared to JetBrains spending the same millions spent on AI infrastructure and instead building better search in Confluence. Confluence search SUCKS, but that's just a lack of focus (or resources) on building a more complex, more robust solution. It's a wiki.
Either way, making a more robust search is a one time cost that benefits everyone. Instead, you're running a piece of software that goes directly to Anthropic's bank account, and to the data centers and to hyper scalers. Every single query must be re-run from scratch, costing your company a fortune, that if not managed properly will come out of spending that money elsewhere.
It seems ridiculous right now because we don’t have hardware to accelerate the LLMs, but in 5 years this will be trivial to run.
Even with an LLM agent getting cheaper to run in the future, it's still fundamentally non-deterministic so the ongoing cost for a single exploration query run can never get anywhere near as cheap as running a wiki with a proper search engine.
If I could pay a world class architect $1.50 to give me tips on how to maximize sunlight in my loft I would.
Would it be nice if confluence just had a robust search that had a one time cost and then benefited everyone thereafter? Sure, but that's not the current reality, and I do not have control over their actions. I can only control mine.
I honestly am finding Codex considerably better, as much as I despise OpenAI.
On the contrary. I update my opinion all the time, but every time I try the latest LLM it still sucks just as much. That is why it sounds like my opinion hasn't changed.
I say all of that to establish that I'm not a reflexive critic when I tell you, hallucinations are absolutely not exceptionally rare now. On multiple occasions this week (and it's only Tuesday!) I've had to disprove a LLM hallucination at work. They're just not as fun to talk about anymore, both because they're no longer new and because straightforward guardrails are effective at blocking the funny ones.
AI is alright. It's moderately useful, in certain contexts it speeds me up a lot, in other contexts not so much.
I also think that the economics of it make no sense and that it is, generally, a destructive technology. But it's not up to me to fix anything, I just try to keep on top of best practices while I need to pay bills.
The economics bit is not my problem though. If all AI companies go bust and AI services disappear I can 100% manage without it.
We're in "too big to fail" territory, if we handled the recession we were heading towards/in years ago, instead of letting AI hype distract and redirect massive amounts of investment, attention and labor from elsewhere, we might have been in a better position.
Though, my cynical take is that the investor class seemed dead-set on forcing us all to weave LLMs deep into our corporate infrastructures in a way that I'm not too sure it will ever "disappear" now. It'll cost just as much to detangle it as it was to adopt it.
The way we talk about "hallucinations" is extremely unproductive. Everything an LLM outputs is a hallucination. Just like how human perception is hallucination. These days I pretty much only hear this word come up among people that are ignorant of how LLMs work or what they're used for.
I've been asked why LLMs hallucinate. As if omniscient computer programs are some achievable goal and we just need to hammer out a few kinks to make our current crop of english-speaking computers perfect.
