It is mind-blowing. This has absolutely no bearing on whether anyone should be critical of it or not, though.

>For such a "technical" and "smart" cohort

You are attempting to imply that HN should de-facto align with AI by appealing to ego, but that doesn't really work here - it is in fact that "technical"/"smart" cohort that leads to the healthy criticism.

I would wager there is an issue wherein healthy criticism has started to get drowned out, and if your point was that - rather than trying to say it's just denial - you might have an interesting discussion.

I will try another approach:

I typically would label the HN crowd as above-average when it comes to understanding technical progress. I would not call myself that; I'm just a software dev that enjoys tech and happens to also enjoy learning about the human brain.

And my question stems from the fact that I see all these smart folks initiating conversations, like "it doesn't understand anything", and then whenever a follow-up is presented ("what does it mean to understand?"), it's silent.

That's a common theme I've seen these last 3-4 months. And I truly feel like I'm missing something.

This is a common issue that arises in tech communities, where some are willing to sit with the tech and probe it, while others are effectively psychologically locked out of it by default; they spend their time looking at technology as something more pattern- and period-bound, e.g. the technology curve and cultural timeline-positions relative to that curve, relative to a type of product, relative to a single product.

IOW the big-picture scope & associated viewpoints naturally enforce some limits on the tech which are not as easily observed by interrogating / examining / analyzing the tech itself. Those who view life _through those viewpoints_ by default will find it a bit strange that others even have to ask why a thing has the described limits.

This can also be described as a systems-viewpoint on tech, in which individual movements (be that AI w/ LLMs or whatever) are part of a whole and this whole must transact with other wholes. At such a level there is a tractability of the hand-wave that is frustrating to those seeking to understand the broader concepts by examining the item at hand (it doesn't understand? So what does it mean, to understand? Let's discuss its property of understanding), as opposed to the scope in which the item exists (it doesn't understand in the broader sense of X or Y; let's discuss what understanding means in broader terms like the scope of its design, rather than the conversational effect of understanding).

Further, I think it's important to acknowledge that there's sometimes substance behind the "no one owes you a debate" perspective. Especially with regard to a conflict of natural viewpoints on things. If you tend to insist on a debate that is constrained to local limits like "define understanding," and this is in effect too subjective or too local for the other party, then you are essentially asking the parties engaging in debate with you to "think like me", or else justify why they won't do that. So now the problem set has an extra layer of translation-fatigue.

This is an area of interest for me personally, and I see it all the time in communications within tech communities. It can really be frustrating due to some additional, fairly effective/interesting theories about how big-picture thinkers can be perceived as "grandiose" by "I'm just average"-perspective-preferring folks, and so on. So the arguments will even take on an emotional tone due to an implication of impropriety of thought process, so to speak.

Further I'd add that some tech communities are naturally much more _about_ bringing forward the best parts of the technological spirit of the past, and less about celebrating what's just arrived.

Such communities may be more likely to celebrate the 8-bit era, or the migration of a humble form of enterprise-level tech to the home environment than they are to take the higher-dopamine perspectives on how new tech is taking the world by storm. In fact, when the latter arrives, it is greeted with exhaustion. Oftentimes it's less a logical position, especially if you consider that it may not even be conscious to the subject taking the position. So naturally they may not feel comfortable sticking around to debate, but I'd point out that sticking around or not doesn't necessarily have a bearing on the validity of their opinion, though I'm sure it's frustrating to experience sometimes.

Just some thoughts though, it's an interesting q in a lot of ways.

Also, great bit here:

"some tech communities are naturally much more _about_ bringing forward the best parts of the technological spirit of the past, and less about celebrating what's just arrived."

I try to be wary of this as well. I'm often reminded of the famous Dropbox HN post, where one of the top comments said "why Dropbox? Just make it yourself".

Not sure how relevant that is, but some AI discussion does indeed remind me of that.

As for ChatGPT: I just don't use it that much anymore. I played with it alot. and it was kinda fun, but when I tried to do anything much more productive with it, it just wasn't that much help.

LangChain and AgentGPT looked more interesting, but when I watched some videos about them on youtube they didn't seem to add accomplish much and got stuck in loops.

The biggest problem is it's just not that reliable.

I think there's an last mile type problem with it.

Also, there's a governance issue with it. Who gets to decide how to use it or how it works. I've seen no democratic options for it. And a free un-encumbered version seems more dangerous than a corporate controlled one.

I think it's a false assumption that because YOU can't use it for productivity-efforts, no one can.

A lot of smaller marketing-agencies would be able to automate the generation of bullshit ad-copy for all their smaller clients, seeing as quality in that area is low in general.

There are so many jobs what could benefit from the use of ChatGPT. I know enough coding to formulate a prompt that helps me build small scripts. I got ChatGPT to build me a Scrapy-spider the other day.

Yes, AI is progressing. Yes, it's good and getting better. But 'AI' is the 'fad of the day' at present, just like all the previous fads-of-the-day were everywhere in the media for a while, and then just faded out of sight. 'Faded out of sight' doesn't mean they went away, they merely took their rightful place in the hierarchy of science and were no longer so much 'in your face'. Hint: When did you last hear about one of history's previous fads, 'Fuzzy Logic'? It's still around, but not everywhere in the media.

If you need a 6th-grader to write some copy for you, then go ahead. But I'd choose an actual 6th-grader instead. They could use the practice, and they will get better with time.

But that's beside the point:

How do people not extend the timeline a bit? It often feels like the release of the internet, but instead of people going, "Holy crap, this is going to be huge!" we have a majority saying, "meh, it only connects like 12 different universities."

As though progress somehow deadstops right where we are.

    so dismissive -- I truly question if it's some form of denial.

> "It's a parrot."

Analyzing the argument why or why it's just a parrot is important. If you are not interested in the argument itself, just upset about the conclusion you are just engaging some kind of techno hype mini-religion.

Turning discussions into general sentiments or mental states is not beneficial unless you are in marketing.

Yes. This is an incredible new development much in the way the atomic bomb was. Except everyone saw the atomic bomb as very, very dangerous and should be handled very very carefully. For LLMs they're just sprouting up on every website like it's no big deal

Neatness is not always gamechanging, though. We've had glasses-free 3D displays available at low-prices for decades, but... it's not a great interaction paradigm. It's neat, to be sure - but sucks for usability.

That's where I'm at with LLMs right now. I've watched this field evolve since Talk-To-Transformer came out, and while AI is "better" it has still failed to address it's serious pitfalls (cannot cite sources, is often wrong, will confidently take extremist positions, etc.). It's genuinely feeling like a re-hash of the cryptocurrency situation where the whitepaper is nerd porn, but the implementation is a horror story. Judging from the overall negative public sentiment of AI (even from fearmongering) I'd wager most people aren't really interested.

When you zoom out and look at it from a business perspective, paying $0.02 to get your fortune told by ChatGPT is not worth it. It's barely more useful with a less-accurate, dumber local version that's a 4gb download. Might pan out to something interesting in the future, but so-far both language models and image generation models have disappointed me. They also haven't really changed society that much, at least from where I'm standing.

Are gamechangers known to be such upon arrival? (That's a real question -- I might argue they aren't known immediately. )

The ability to generate plausible language is impressive.

The fact that so much of the generated language is inaccurate and unreliable is significantly less so.

Basically, we have automated the art of building a talking database ... along with the art of bullshit and mimicry.

Depends on what your background is. Replace "AI" with "information compression algorithm with semantic language lookup capability", and the endeavor still sounds cool, but not as sexy as AI.

Also, the current "control" of how these LLMs are being trained and deployed is not really that exciting, its even disheartening on some levels tbh.

What needs to happen is that AMD and Nvidia (and possibly others) should be racing to the bottom to create the best bang/buck ML accelerators, where people can realistically buy racks to run in their house and transfer train these architectures on their own data sets.

I don't personally use it or have a need at the moment, but await the day it's matured enough I can set it on an old scanned pdf, and it'll work out how to code a new pdf with the same type font, either by discovery or create a new font, typeset the same position, if of duplicator (ditto machine) nature, work out any miss formed letters ... also clean up any images within, such it is as good as if not better than what was originally scanned if it were printed, and greatly reduce my disc space required to store my collection of pdfs.

Sadly I already know that won't become the norm for the average user, unless for some weird instance, the AI is totally under their own control - and that situation would be doubtful due to potential misuse.

That's would be an example of well placed concern ... unseen misuse slowly addressed.

Then there will be the monetisation and I imagine it'll creep in just like search engines, start, get better, become great and then ... fuck everyone over so those that really need a good search, pay to try others. (Sadly for any potential new web crawlers, in that time span from the end of 90s where it was open to all, there's no simple crawling all the web any more ... )

The next I see as an issue will be early adopters trying to wrangle a customer base around some inappropriate use of it. New tech is nice and some people are scared of things which haven't been well tested, but sometimes that fear is well founded - I only have to think back to voice identification technology which was really interesting in those times, but I know of at least one govt organisation that inflicted it on a wide user base despite being told it wasn't that secure - assuming most would be hackers would not have large enough systems to mimic a voice on the fly, isn't a good start IMO.

Also yesterday I asked Siri "what's 1300 mils in ounces". The text-to-speech worked properly, but instead of giving me the answer it opened a web search for "1300ml in oz".

These are the same companies in the same market as those gearing up to dive my car and interpret my blood tests.

We've had the technology for a decade to correctly process my AI queries from yesterday. But there's a chasm between technology and helping users, and crossing that gap is rarely the way to maximise shareholder value.

I understand it to be some thing that understands communication and can answer questions based on the communication it has.

There are some questionable capabilities like an appearance of logic. But I am going to uphold a belief that it is borrowing some standard information that it already has and manipulating it to fit the question.

There are areas that are logically assessable, but I do not think an LLM alone can do that. The next frontier of AI will be trying to prioritize which piece of logic is more applicable, assuming all pieces of logic are already invented.

The "inside" of an LLM is a black box in that we don't know how an input arrives to an output. We know some general ideas ("it predicts the next thing"), but not the detailed route to get to the output.

Similarly, we do not know how human minds work. How does a thought arise? How does a new idea arise?

I would not argue that an LLM is identical to a human brain. But I do think there's some interesting discussion to be had on that topic.

It's as though we have two black boxes, and so many folks are willing to say, "One of these is definitely, certainly not like the other black box!"

And I don't see where that confidence comes from.

That comparison, to my silly mind, is astonishing to think about. And I've yet to see why I should (or could) dismiss the notion of similarity.

Or, simply using GPT-4; that's quite astonishing to me. I'm not a stellar coder by any means, but I've quite literally 5x-ed my productivity with GPT-4.

"Well, you must not be working on novel problems." You're right! But is "solving novel problems" a requirement for discussion now?

It sounds like you're conflating multiple different things here which makes it hard to respond to your question as-written.

When scientists say "We don't know" they mean "It would take more time than I care to spend, to investigate the particulars to answer that question." The public jumps straight to "It's magic, Jesus or aliens!"

It's a parrot. That's not in question.

Given that the output essentially is similar to what human posters online are doing in many many cases the term isn't useful to distinguish between the two.

It's certainly not working like Eliza where there are specific programmed response trees to choose from. It doesn't seem to be just copying sentences and pasting them into the output.

But also, I'm not so sure scientists "don't have the time" but rather, "literally cannot" know with regard to an LLM's "insides".

With both of them, I'm very underwhelmed with "use it like a forum" or "ask it a tech support question". Maybe because my "tech support questions" tend to be much harder than my "scripting questions", but it's not super useful. Anything it finds I actually find "better" because of context and screenshots just using a traditional search and clicking through to the blog or whatever that's talking about the issue.

Ok, now to the "It's just a parrot". I disagree strongly with that. It's no more a parrot than wikipedia authors are. Many of the Neeva AI search results read like an autogenerated Wikipedia page, complete with the source links.

Philosophically - I think a lot of people either aren't bringing philosophy and some sci fi knowledge to these arguments or don't know of them. Many of the "it doesn't understand anything" seem to be somewhat substrate chauvinist to me - I'm not sure if Data from Star Trek showed up in front of them that they'd admit he's intelligent or a person, simply because he's not organic. And you're also right - people claiming it "doesn't understand" can't or won't define what "understand" means in this context. Other terms, like "intelligence" also have changed over time. I'm pretty sure if in 1800 you asked if a machine could do calculus if it was intelligent - people would say yes. However by the 1960s computers were enough of a thing doing math that we decided math was no longer a marker of intelligence for a machine to do. Now we claim that GPT is just advanced autosuggest. Well, I guess an iPad is just an advanced calculator - but I'd argue it's transcended being a calculator in practical use. And I think GPT has transcended next word suggestions. My phone keyboard suggests 3 options for the next word, but it can't write the next paragraph for me coherently, nor does it take directions or questions and output pretty reasonable responses.

So I think I'm in the middle here - I don't discount GPT3.5 and 4, but I don't think the're Star Trek level AIs either. I don't think they are magic as it were.

I will agree that it's by no means Star Trek level, but it is about the most fascinating piece of technology I've seen in my lifetime.

And yes, that sounds silly! But therein lies my reason for writing this post.

However, I know that changing the game, inevitable as it is, does not mean making the world better.

We're already seeing AI being adopted by the people who ruin things. LLMs are bullshit machines, and bullshit people are using it to make bullshit faster, cheaper and more convincing.

Yet AI must be trained on original art and original research. Our work. It just does so without giving credit, or a way to opt out. My work is used to train bullshit machines that will be turned on me, to profit someone else.

Talk-to-Transformer has existed for 5 years now. If the game is to be changed, it has already happened.

And retention stats relative to the normal/non-AI way of doing it (e.g. learning a wikipedia page) with these models for most use cases have looked like: GPT-2: 2% GPT-3: 10% GPT-3.5: 40% GPT-4: 65% GPT-5: ?

The change has came noticeably because benefits of LLM compound with continued use-case usage.