Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

What it feels like to work in AI right now

robotic.substack.com

Post reply on HN
Print view View original

691–700 of 768 posts

Re: What it feels like to work in AI right now

#691

Post by notahacker »

PeterStuer wrote: post #623
notahacker wrote: post #521

Earlier quoted context omitted.

> Most of the people asking the question "to do _what_, exactly" seem to have not much of a "knowledge worker" experience. From personal experience I'd say the opposite: GPT lacks the specialist knowledge to produce useful writing or yield accurate answers in any of the markets I've worked in (I'll grant that less niche markets exist, and that GPT is pretty good at fixing the writing of people that lack English langu…

As is already a cliche now: GPT will not replace your job. Your colleague using GPT will.

As I already suggested: in my experience, certainty that GPT will be transformational in a field of knowledge work seems to actually be inversely related to experience of that field. A response whose certainty that GPT could be transformational to the stuff I worked on exceeded only by ignorance of what any of that stuff was is quite a good demonstration of that point...

(and really, there's nothing particularly special about any of the stuff I've worked on, it's just GPT doesn't have relevant knowledge or a path to acquiring it so doesn't generate remotely adequate responses, struggles even more with novel concepts and would be terrible at real time discussion even if suitable interfaces to it existed and were unobjectionable, and that's before we get started on the privacy implications)

Re: What it feels like to work in AI right now

#692

Post by int_19h »

creata wrote: post #473
Al-Khwarizmi wrote: post #450

Earlier quoted context omitted.

Some people are just bad at this kind of thing, and sometimes saying nothing is not an option if there is a social expectation that you say something. Which can create a lot of anxiety.

I don't care how bad it is, I just want something from your heart. And I'm sure you can still make something heartfelt with the help of a chatbot, but I'm afraid that people won't .

Many people won't with or without a chatbot. A chatbot might at least make it less cringe.

Re: What it feels like to work in AI right now

#693

Post by int_19h »

bmitc wrote: post #207
nr2x wrote: post #158

Earlier quoted context omitted.

Natural language is now a fully functional user interface out of the box. This is bigger than the mouse.

Fully functional in what way? As far as I can tell, ChatGPT is a box that I put sentences into, and I get grammatically correct sentences that contains topics or words loosely statistically correlated to what were in my sentences, that may or may not be correct and often are not. The box has little to no memory. I honestly don't see what's so useful about this box.

You can tell it what you need to do in terms of data processing, and ask it to write Python code that does that. E.g. ever had to write a convoluted ImageMagick command to do something complicated? GPT will write it for you.

Think of it as a natural language interface to anything that has an API.

Re: What it feels like to work in AI right now

#694

Post by int_19h »

snowe2010 wrote: post #412
AussieWog93 wrote: post #343

Earlier quoted context omitted.

I've not used GPT-4, so it could be different, but regular old GPT-3.5 gets a _lot_ of things wrong.

GPT 4 is quite astounding. It might be wrong on occasion, but it will easily point you in the right direction most of the time. It still messes up, but like a twentieth of what 3.5 did. Honestly it is like an incredible rubber ducky for me. Not only can I just talk like I’m talking to a rubber duck but I can get fast, mostly informed, feedback that unblocks me. If I have a bunch of things competing for my attention I…

My favorite part about GPT-4 is that if it generates code that is wrong, and you ask it to verify what it just wrote - without telling it whether it's wrong or not, much less pointing out the specific issue - more often than not it will spot the problem and fix it right away.

And yes, it does indeed make an amazing rubber duck for brainstorming.

Re: What it feels like to work in AI right now

#695

Post by int_19h »

brendamn wrote: post #98
redox99 wrote: post #47

Earlier quoted context omitted.

I don't know about GPT4, but GPT3.5 I'd bet is pretty traditional and boring. It's power comes from a really good, properly curated dataset (including the RLHF). GPT3.5 turbo is much more interesting probably, because they seem to have found out how to make it much more efficient (some kind of distillation?). GPT4 if I had to make a very rough guess, probably flash attention, 100% of the (useful) internet/books for i…

> I'd say with GPT4 they probably reached the limit of how big the dataset can be I’m curious about this too; not just on the dataset size, but also the model size. My hunch is that the rapid improvements of the underlying model by making it bigger/giving it more data will slow, and there’ll be more focus on shrinking the models/other optimisations.

I don't think we're anywhere close to the limit of sheer hardware scalability on this. Returns are diminishing, but if GPT-4 (with its 8+ k context window) is any indication, even those diminishing returns are still very worthwhile.

If anything, I wonder if the actual limit that'll be hit first will be the global manufacturing capacity for relevant hardware. Check out the stock price of NVDA since last October.

Re: What it feels like to work in AI right now

#696

Post by int_19h »

chillfox wrote: post #150
UncleOxidant wrote: post #61

Earlier quoted context omitted.

Possibly, but wouldn't Google and Meta have access to way more compute resources and data than OpenAI? Google has been touting their TPUs for several years now.

Google has the compute, from the comparisons I have seen Bard smokes GPT-3.5-Turbo on response times. So my guess is that internal politics prevents them from putting out something better. There would have to be immense pressure from the search division to not make them obsolete.

Bard is also a fair bit worse than GPT-3.5, though, so that can be a function of model size.

Re: What it feels like to work in AI right now

#697

Post by bestcoder69 »

wzdd wrote: post #11
bestcoder69 wrote: post #5

Earlier quoted context omitted.

Popular things = crypto ? I’d say the differentiator is that in addition to hype scammers, you have people like Stephen Wolfram excited.

The absolutely-not-hype-susceptible man who has been claiming for the past 20 years to have produced a fundamental theory of everything based on his research into finite-state automata?

Alright man if one guys obsession of his own work is hype, too, then hype now encompasses short-lived popular things as well as long-lived unpopular things, so in that case yes AI is hype and only hype, everything bad to you is hype and I’m hype too.

I’ll continue to use AI where it’s helpful, at very low cost, and know deep down that it’s hype.

Re: What it feels like to work in AI right now

#698

Post by sashank_1509 »

ModernMech wrote: post #673
sashank_1509 wrote: post #642

Earlier quoted context omitted.

Again I did in fact open pornhub when you point it out and that’s Tesla Autopilot not FSD. I do find it beyond the pale that people drunk use FSD, because there are very strict controls for people using FSD. Autopilot is advanced cruise control, apart from regulatory authorities saying it’s safe compared to competitors, Tesla releases quarterly safety reports since 2019 counting the incidents with Autopilot. It is sh…

> Autopilot is advanced cruise control, apart from regulatory authorities saying it’s safe compared to competitors, Tesla releases quarterly safety reports since 2019 counting the incidents with Autopilot. So what did you do to fix the issue? Why did a second person die in the same exact way as the first person after 3 years? Were you working on the fix at all? Or did you do nothing? Just admit it if you did nothing…

I feel like I’m repeating myself to your constant emotional attacks of trying to make me feel guilt for 2016/ 2019 case so let me try to end it with: you’re right about one thing, AP is similar to other tech out there, it is cruise control with some extra features. If you think it is ethically wrong to put AP in the public, then every car since 2005 at least are a moral hazard, if you believe that fine I don’t. Cruise control is regulated, Tesla passed them with the highest ratings and releases safety reports since 2019 as I already said. Why the distinction is so important, most of Tesla AP doesn’t even use new tech, very little deep learning (just some CV stuff), AP still uses radar like every other car manufacturer out there and AP has a very simple state space planner which most cars use now though most Toyotas use older PID tech.

FSD beta is different, FSD is filled with new DL tech. DL are black box models that work surprisingly well but are not interpretable. They can suddenly output something nonsensical (like bing Chat did, ChatGPT surprisingly hasn’t) and you won’t even know why. There is a risk involved with putting DL based FSD out there, because you don’t know when it will fail. Tesla took that risk. Tesla however to date has had no FSD crashes that involved injuries, had 1 crash that involved the front of a Tesla being significantly damaged (which is being investigated by NHTSA as I already said), and several smaller collisions that have caused scratches on Tesla cars (at which point we promptly ban that user for life, you can see YouTube videos of this). Uber self driving killed a pedestrian, (though the paid QA driver should have been paying attention, it was not really Ubers engineers fault), Tesla actually handled the risk of using DL tech pretty well. It was a real risk, we still have no injuries and the tech keeps getting better. So yes your tiresome moral attacks don’t affect me and I prolly won’t respond again if I just have to repeat myself.

Re: What it feels like to work in AI right now

#699

Post by chopete3 »

espe wrote: post #332
chopete3 wrote: post #167

At least, people working on ML models that handle these tasks must be feeling terrible. They know their models will be abandoned sooner or later and composed on top an LLM. 1. Classification 2. Named Entity Recognition (NER) 3. Dialog Engine 4. Sentiment Analysis 5. Tone Analysis 6. Language Translation 7. Summarization 8. Tokenization 9. Simple NLP Tasks (part-of-speech tagging, dependency parsing, lemmatization, mo…

actually that might not be the case. don't underestimate the value of older, better understood and much smaller models. also, why not call bert-style (encoder) models LLMs as well. i would expect last-gen models to give us an edge in controlling the effects of the latest ones (cf. the alignment discussion).

BERT models are also LLMs. I referred to LLM more as an API based access, hosted by Microsoft, Google or AWS for large scale isolated/production consumption, like RDS (MySQL, Postgres).

There will always be custom models, with controlled training data and specific use cases.

Re: What it feels like to work in AI right now

#700

Post by girvo »

rrrrrrrrrrrryan wrote: post #629
girvo wrote: post #500

Earlier quoted context omitted.

One worry I have for the long term is: how will we learn or adopt new programming languages in the future, when LLM du-jour knows nothing about Language X? Will we be stuck with what we’ve got because LLMs make us too productive in them?

I thought about this as well. As long as the new language is well-documented, it should be simple to teach the LLM to use it.

But for a new language, there isn’t a large body of work to train the LLM on what idiomatic code looks like. It really worries me, that we’re going to be stuck at this local maximum
Post reply on HN