Over half of all tech industry workers view AI as overrated
3 years ago by L4sBot to c/technology
Over half of all tech industry workers view AI as overrated::undefined
I’ve found that relying on it is a mistake anyhow, the amount of incorrect information I’ve seen from chatgpt has been crazy. It’s not a bad thing to get started with it but it’s like reading a grade school kids homework, you need to proofread the heck out of it.
What always strikes me as weird is how trusting people are of inherently unreliable sources. Like why the fuck does a robot get trust automatically? It's a fuckin miracle it works in the first place. You double check that robot's work for years and it's right every time? Yeah okay maybe then start to trust it. Until then, what reason is there not to be skeptical of everything it says?
People who Google something and then accept whatever Google pulls out of webpages and puts at the top as fact.. confuse me. Like all machines, there are failures. Why would we trust that the opposite is true?
At least a Google search gets you a reference you can point at. It might be wrong, it might not. Maybe it points to other references that you can verify.
ChatGPT outright makes shit up and there's no way to see how it came to a given conclusion.
Because the average person hears “AI” and thinks Cortana/Terminator, not a bunch of if statements.
People are dumb when it comes to things they don’t understand. I’m dumb when it comes to mechanical engineering of any kind, but I’m competent with software. It’s all about where people’s strengths lie, but some people aren’t aware enough to know they don’t know something
My guess, wholly lacking any scientifc rigor, is that humans naturally trust each other. We don't assume the info someone shares with us as wrong, unless there's "a reason" to doubt. Chatting with any of these LLM bots feels like talking to a person (most of the time), so there's usually "no reason" to doubt what it spews.
If human trust wasn't so easy to get and abuse, many scams would be much harder to pull.
I feel like the AI in self-driving cars is the same way. They're like driving with a 15 year old that just got their learners permit.
Turns out that getting a computer to do 80% of a good job isn't so great. It's that extra 20% that makes all the difference.
That 80% also doesn't take that much effort. Automation can still be helpful depending on how much effort it is to repeatedly do it, but that 20% is really where we need to see progress for a massive innovation to happen.
I actually disagree. Ai is great at doing the parts that are easy to do mentally but still take time to do. This "fancy autocomplete" is where it shines and can accelerate the work of a professional by an order of magnitude
I just reviewed a PR today and the code was... bad, like unusually bad for mycoworkers and left some comments.
Then my coworker said he used chatgpt without really thinking on what he was copypasting.
I have found that it's like having a junior programmer assistant. It's great for "write me python code for opening an in file from a command line argument, reading the contents into a key/value dict array, then closing the file." It's terrible for "write me a python code for pulling data into a redis database."
I find it's wrong 50% of the time for certain command line switches, Linux file structure, and aws cli.
I find it's terrible for advanced stuff like, "using aws cli and jq, take all volumes in a vpc, and display the volume id, volume size in gb, instance id it's attached to, private IP address of the instance, whether is a gp3 or gp2, and the vpc id in a comma separated format, sorted by volume size."
Even worse at, "take all my gp2 volumes and make them gp3."
I recently used it to update my resume with great success. But I also didn’t just blindly trust it.
Gave it my resume and then asked it to edit my resume to more closely align with a guide I found on Harvards website. Gave it the guide as well and it spit out a version of mine that much more closely resembled the provided guide.
Spent roughly 5 minutes editing the new version to correct for any problems it had and boom. Half an hour of worked parsed down to sub 10
I then had it use my new resume (I gave it a copy of the edited version) and asked it to write me a cover letter for a job (I provided the job description)
Boom. Cover letter. I spent about 10 minutes editing that piece. And then that new resume and cover letter lead to an interview and subsequent job offer.
AI is a tool not an all in one solution.
Nice one! I have heard it called a fuzzy JPG of the internet.
And that's entirely correct
It does a shockingly good analogue of “understanding” at the very least. Have you tried asking chatgpt to solve analogies? Those show up in all kinds of intelligence tests.
We don’t have agi, definitely, but this stuff has come a very long way and it’s quite close to being genuinely useful.
Even if we completely reject the “it’s ai,” we more or less have a natural language interface for computers that isn’t a shallow trick and that’s awesome.
yes, it's has. the most famous example is the stacking of the laptop and the markers. you may not have access but it's about to eclipse us imho. I'm no technological fanboy either. 20 years ago I argued that I wouldn't be possible to understand human speech. now that is a everyday occurrence.
Depends on how you define understanding and how you test for it.
I assume we are talking LLM here?
It's a tool. And like any tool it's only as good as the person using it. I don't think these people are very good at using it.
Too bad it's bullshit.
If you are actually interested in the topic, here's a few good reads:
Do Large Language Models learn world models or just surface statistics? (Jan 2023)
Actually, Othello-GPT Has A Linear Emergent World Representation (Mar 2023)
Eight Things to Know about Large Language Models (April 2023)
Playing chess with large language models (Aug 2023)
Language Models Represent Space and Time (Oct 2023)
As you can see, the past year has shed a lot of light on the topic.
One of my favorite facts is that it takes on average 17 years before discoveries in research find their way to the average practitioner in the medical field. While tech as a discipline may be more quick to update itself, it's still not sub-12 months, and as a result a lot of people are continuing to confidently parrot things that have recently been shown in research circles to be BS.
Once we're able to synergize the increased throughput of our knowledge capacity we're likely to exceed shareholder expectation and increase returns company wide so employee defecation won't be throttled by our ability to process sanity.
Sounds like we need to align on triple underscoring the double-bottom line for all stakeholders. Let’s hammer a steak in the ground here and craft a narrative that drives contingency through the process space for F24 while synthesising synergy from a cloudshaping standooint in a parallel tranche. This journey is really all about the art of the possible after all so lift and shift a fit for purpose best practice and hit the ground running on our BHAG.
I'm calling HR
😜
Every billboard in SF is just these words shuffled
No SQL, block chain, crypto, metaverse, just to name a few recent examples.
AI is overhyped, but it is, so far, more useful than any of those other examples, though.
These are useful technologies if used when called for. They aren't all in one solutions like the smart phone killing off cameras, pdas, media players... I think if people looked at them as tools which fix specific problems, we'd all be happier.
Largely because we understand that what they're calling "AI" isn't AI.
I also believe that will happen! We will not be prepared since many don't understand the differences between what current models do and what an actual general AI could potentially do.
It also saddens me that many don't know or ignore how fundamental abstract reasoning is to our understanding of how human intelligence works. And that LLMs simply aren't intelligent in that sense (or at all, if you take a tight definition of intelligence).
I don’t get how recognizing a pattern is not AI. It recognizes patterns in data, and patterns in side of patterns, and does so at a massive scale. Humans are no different, we find patterns and make predictions on what to do next.
The human brain does not simply recognise patterns, though. Abstract reasoning means that humans are able to find solutions for problems they did not encounter before. That's what makes a thing intelligent. It is not fully understood yet what exactly gives the brain these capabilities, btw. Like, we also do not understand yet how it is possible that we can recognize our own thinking processes.
The most competent current AI models mimic one aspect of the brain which is neural pathways. In our brain it's an activity threshold and in a neural network AI it's statistics which decide whether a certain path is active or not and then it crosses with other paths, etc. Like a very complex decision tree.
So that is quite similar between AI and brains. But we actually get something like an understanding of concepts that goes beyond the decision tree but isn't fully understood yet, as described above.
For an AI to be actually intelligent it would probably need to at least get this ability, to trace back it's own way through the decision tree. Maybe it even turns out that you in fact do need a consciousness to have reason.
No, INT.
I've seen it refered to as AGI bit I think itns wrong. Chat GPT isnt intelligent in the slightest, it only makes guesses on what word is statistically more likely to come up next. There is no thikinking or problem solving involved.
A while ago I saw an article that with a tittle along the lines of "spark of AGI in ChatGPT 4" because it chose to use a calculator tool when facing a problme that required one. That would be AI (and not AGI). It has a problem, it learns and uses available tools to solve it.
AGI would be on a whole other level.
Edit: Grammar
The argument "it just predicts the most likely next word" while true massively under values what it even means to predict the next word or token. Largely these predictions are based on sentences and ideas the model has trained on from its data sets. It's pretty intelligent if you think about it. You read a text book then when you apply the knowledge or take a test you use what you read to form a new sentence in relation to the context of the question or problem. For the models "text prediction" to be correct it has to understand certain relationships between complex ideas and objects to some capacity. Yes it absolutely is not as good as human intelligence. But what it's doing is much more advanced then text to type on your phone keyboard. It's a step in the right direction, over hyped right now but the hype is funneling cash into research. The models are already getting more advanced. Right now half of what it says is hot garbage but it can be pretty accurate.
AI doesn't necessarily mean human-level intelligence, if that's what you mean. The AI field has wrestled with this for decades. There can be "strong AI", which is aiming for that human-level intelligence, but that's probably a far off goal. The "weak AI" is about pushing the boundaries of what computers can do, and that stuff has been massively useful even before we talk about the more modern stuff.
Sounds like people here are expecting to see GPAI and singularity stuff, but all they see is a pitiful LLM or other even more narrow AI applications. Remember, even optical character recognition (OCR) used to be called AI until it became so common that it wasn’t exciting any more. What AI developers call AI today, is just basic automation and few decades later.
Given that AI isn't purported to be AGI, how do you define AI such that multimodal transformers capable of developing abstract world models as linear representations and trained on unthinkable amounts of human content mirroring a wide array of capabilities which lead to the ability to do things thought to be impossible as recently as three years ago (such as explain jokes not in the training set or solve riddles not in the training set) isn't "artificial intelligence"?
The decision tree my company uses to deny customer claims is not AI despite the business constantly referring to it as such.
There's definitely a ton of "AI" that is nothing more than an If/Else statement.
for many years AI referred to that type of technology. It is not infact AGI but AI historically in the technical field refers more towards decision trees, and classification/ linear regression models.
That's called an expert system, and has been commonly called a form of AI for decades.
That is indeed what most of it is, my company was doing "sentiment analysis" and it was literally just checking it against a good and bad word list
When someone corporate says "AI" you should hear "extremely rudimentary machine learning" until given more details
The main difference is that crypto was/is burning huge amounts of energy to run a distributed ponzi scheme. LLMs are at least using energy to create a useful tool (even if there is discussion over how useful they are).
There are significant differences between statistical models and AI.
I work for an analytics department at a fortune 100 company. We have a very clear delineation between what constitutes a model and what constitutes an AI.
That's true. Statistical models are very carefully engineered and tested and current machine learning models are created by throwing a lot of training data at the software and hope for the best that the things that the model learns are not complete bullshit.
You really should listen rather than talk. This is not AI, it's just a word prediction model. The media calls it AI because it sells and the companies calls it AI because it brings the stock value up.
Personally, I think medicine will be the most impacted by AI. Medicine has already been increasingly implementing AI in many areas, and as the tech continues to mature, I am optimistic it will have tremendous effect. Already there are many studies confirming AI's ability to outperform leading experts in early cancer and disease diagnoses. Just think what kind of impact that could have in developing countries once the tech is affordably scalable. Then you factor in how it can greatly speed up treatment research and it's pretty exciting.
That being said, it's always wise to remain cautiously skeptical.
The bad part is health insurance companies are also using AI.
I’s ability to outperform leading experts in early cancer and disease diagnoses
It does, but it also has a black box problem.
A machine learning algorithm tells you that your patient has a 95% chance of developing skin cancer on his back within the next 2 years. Ok, cool, now what? What, specifically, is telling the algorithm that? What is actionable today? Do we start oncological treatment? According to what, attacking what? Do we just ask the patient to aggressively avoid the sun and use liberal amounts of sun screen? Do we start a monthly screening, bi-monthly, yearly, for how long do we keep it up? Should we only focus on the part that shows high risk or everywhere? Should we use the ML every single time? What is the most efficient and effective use of the tech? We know it's accurate, but is it reliable?
There are a lot of moving parts to a general medical practice. And AI has to find a proper role that requires not just an abstract statistic from an ad-hoc study, but a systematic approach to healthcare. Right now, it doesn't have that because the AI model can't tell their handlers what it is seeing, what it means, and how it fits in the holistic view of human health. We can't just blindly trust it when there's human lives in the line.
As you can see, this seems to be relegating AI to a research role for the time being, and not on a diagnosing capacity yet.
You are correct, and this is a big reason for why "explainable AI" is becoming a bigger thing now.
There is a very complex algorithm for determining your risk of skin cancer: Take your age ... then add a percent symbol after it. That is the probability that you have skin cancer.
Like you say, "AI" isn't just LLMs and making images. We have previously seen, for example, expert systems, speech recognition, natural language processing, computer vision, machine learning, now LLM and generative art.
The earlier technologies have gone through their own hype cycles and come out the other end to be used in certain useful ways. AI has no doubt already done remarkable things in various industries. I can only imagine that will be true for LLMs some day.
I don't think we are very close to AGI yet. Current AI like LLMs and machine vision require a lot of manual training and tuning. As far as I know, few AI technologies can learn entirely on their own and those that do are limited in scope. I'm not even sure AGI is really necessary to solve most problems. We may do AI "ala carte" for many years and one day someone will stitch a bunch of things together, et voila.
Ooh thanks for the heads up. Last time I played with TTS was years ago using Festival, which was good for the time. Looking forward to trying Tortoise TTS.
Honestly I believe AGI is currently a compute resource problem less than a software problem. A paper came out awhile ago showing that individual neurons in the human brain displayed behavior like decently sized deep learning models. If this is true the number of nodes required for artificial neural nets to even come close to human like intelligence maybe astronomically higher then predicted.
That’s my understanding as well, our brain is just an insane composition of incredibly simple mechanisms. Its compositions of compositions of compositions ad nauseam. We are manually simulating billions of years of evolution, using ourselves as a blueprint. We can get there… it’s hard to say when we’ll get there, but it’ll be interesting to watch.
Exactly, plus human consciousness might not be the most effective way to do it, might be easier less resource intensive ways.
It is overrated. At least when they look at AI as some sort of brain crutch that redeems them from learning stuff.
My boss now believes he can "program too" because he let's ChatGPT write scripts for him that more often than not are poor bs.
He also enters chunks of our code into ChatGPT when we issue bugs or aren't finished with everything in 5 minutes as some kind of "Gotcha moment", ignoring that the solutions he then provides don't work.
Too many people see LLMs as authorities they just aren't....
It bugs me how easily people (a) trust the accuracy of the output of ChatGPT, (b) feel like it's somehow safe to use output in commercial applications or to place output under their own license, as if the open issues of copyright aren't a ten-ton liability hanging over their head, and (c) feed sensitive data into ChatGPT, as if OpenAI isn't going to log that interaction and train future models on it.
I have played around a bit, but I simply am not carefree/careless or am too uptight (pick your interpretation) to use it for anything serious.
Too many people see LLMs as authorities they just aren't....
This is more a 'human' problem than an 'AI' problem.
In general it's weird as heck that the industry is full force going into chatbots as a search replacement.
Like, that was a neat demo for a low hanging fruit usecase, but it's pretty damn far from the ideal production application of it given that the tech isn't actually memorizing facts and when it gets things right it's a "wow, this is impressive because it really shouldn't be doing a good job at this."
Meanwhile nearly no one is publicly discussing their use as classifiers, which is where the current state of the tech is a slam dunk.
Overall, the past few years have opened my eyes to just how broken human thinking is, not as much the limitations of neural networks.
It is overrated. It has a few uses, but it's not a generalized AI. It's like calling a basic calculator a computer. Sure it is an electronic computing device and makes a big difference in calculating speed for doing finances or retail cashiers or whatever. But it's not a generalized computing system that can basically compute anything that it's given instructions for which is what we think of when we hear something is a "computer". It can only do basic math. It could never be used to display a photo , much less make a complex video game.
Similarly the current thing that's called "AI", can learn in a very narrow subject that it is designed for. It can't learn just anything. It can't make inferences beyond the training material or understand. It can't create anything totally new, it just remixes things. It could never actually create a new genre of games with some kind of new interface that has never been thought of, or discover the exact mechanisms of how gravity works, since those things aren't in its training material since they don't yet exist.
Some calculators can run DooM, though
Many areas of machine learning, particularly LLMs are making impressive progress but the usual ycombinator techbro types are over hyping things again. Same as every other bubble including the original Internet one and the crypto scams and half the bullshit companies they run that add fuck all value to the world.
The cult of bullshit around AI is a means to fleece investors. Seen the same bullshit too many times. Machine learning is going to have a huge impact on the world, same as the Internet did, but it isn't going to happen overnight. The only certain thing that will happen in the short term is that wealth will be transferred from our pockets to theirs. Fuck them all.
I skip most AI/ChatGPT spam in social media with the same ruthlessness I skipped NFTs. It isn't that ML doesn't have huge potential but most publicity about it is clearly aimed at pumping up the market rather than being truly informative about the technology.
ML has already had a huge impact on the world (for better or worse), to the extent that Yann LeCun proposes that the tech giants would crumble if it disappeared overnight. For several years it's been the core of speech-to-text, language translation, optical character recognition, web search, content recommendation, social media hate speech detection, to name a few.
ML based handwriting recognition has been powering postal routing for a couple of decades. ML completely dominates some areas and will only increase in impact as it becomes more widely applicable. Getting any technology from a lab demo to a safe and reliable real world product is difficult and only more so when there are regulatory obstacles and people being dragged around by vehicles.
For the purposes of raising money from investors it is convenient to understate problems and generate a cult of magical thinking about technology. The hype cycle and the manipulation of the narrative has been fairly obvious with this one.
I remember when it first came out I asked it to help me write a MapperConfig custom strategy and the answer it gave me was so fantastically wrong - even with prompting - that I lost an afternoon. Honestly the only useful thing I've found for it is getting it to find potential syntax errors in terraform code that the plan might miss. It doesn't even complement my programming skills like a traditional search engine can do; instead it assumes a solution that is usually wrong and you are left to try to build your house on the boilercode sand it spits out at you.
I've found the free one can sometimes answer tip of my tongue questions but yeah anything even remotely obscure it will just lie and say that doesn't exist, especially if you stray a little too close to the puritanical guard rails. One time I was going down a rabbit hole researching human sex organ variations and it flat out told me the people in South America who grow a penis at 12 don't exist until I found the name guevedoces on my own, and wouldn't you know it then it knew what I was talking about.
Reality: most tech workers view it as fairly rated or slightly overrated according to the real data: https://www.techspot.com/...
Helps me write hacky scripts to solve one off problems. Honestly, it saves me a few work days.
But it's far from replacing anybody.
You say it’s “far” but 70 years ago a simple calculator was the size of a house. The power of my desktop from 10 years ago is beat by my phone, hell maybe even my watch.
You know, you code, compute is improving rapidly, and even when it slows vertical scaling, it’s still horizontally scaling. All the while software is getting more efficient, developing new capabilities and techniques which only bring on even more innovation.
It compounds. At this point I think the only limiting factor is how much faith the rich and powerful put in AI’s ability to make them richer. The more they invest, the faster it’ll grow.
Slightly overrated is where I would put it, absolutely. It's overhyped, but god if the recent advancements aren't impressive.
I have a doctorate in computer engineering, and yeah it’s overhyped to the moon.
I’m oversimplifying it and some one will ackchyually me but once you understand the core mechanics the magic is somewhat diminished. It’s linear algebra and matrices all the way down.
We got really good at parallelizing matrix operations and storing large matrices and the end result is essentially “AI”.
Big emphasis on the 'A'
I mean the dotcom bubble was overhyped, it was a bubble
That's what they said.
The internet was revolutionary, but dotcom was overhyped at the time.
What's the dotcom bubble?
a bubble is kind of a goldrush situation,
where businesses and investors on mass
jump into a new / hyped market or asset type without a propper plan & buisness model.
for example the first recorded one: the tulip mania
the dot-com bubble was a massive bubble in the 90s centered arrount the emerging concept of "internet buissneses"
That’s because it is overrated and the people in the tech industry are actually qualified to make that determination. It’s a glorified assistant, nothing more. we’ve had these for years, they’re just getting a little bit better. it’s not gonna replace a network stack admin or a programmer anytime soon.
There is a lot of marketing about how it's going to disrupt every possible industry, but I don't think that's reasonable. Generative AI has uses, but I'm not totally convinced it's going to be this insane omni-tool just yet.
whenever we have new technology, there will always be folks flinging shit on the walls to see what sticks. AI is no exception and you're most likely correct that not every problem needs an AI-powered solution.
Sure, it is already changing some fields, and more and more fields are beginning to feel the impact in the coming decades. However, we’re still pretty far from a true GPAI, so letting the AI do all the work isn’t going to happen any time soon.
Garbage in, garbage out still applies here. If we don’t collect data in the appropriate way, you can’t expect to teach a model with that. Once we start collecting data with ML in mind, that’s when things start changing quickly. Currently, we have lots and lots of garbage data about everything, and that’s why we aren’t using AI to more.
It is currently overhyped and so much of it just seems to be copying the same 3 generative AI tools into as many places as possible. This won't work out because it is expensive to run the AI models. I can't believe nobody talks about this cost.
Where AI shines is when something new is done with it, or there is a significant improvement in some way to an existing model (more powerful or runs on lower end chips, for example).
I work in tech. AI is overrated.
Of course, because hype didn't come from tech people, but content writers, designers, PR people, etc. who all thought they didn't need tech people anymore. The moment ChatGPT started being popular I started getting debugging requests from few designers. They went there and asked it to write a plugin or a script they needed. Only problem was it didn't really work like it should. Debugging that code was a nightmare.
I've seen few clever uses. Couple of our clients made a "chat bot" whose reference was their poorly written documentation. So you'd ask a bot something technical related to that documentation and it would decipher the mess. I still claim making a better documentation was a smarter move, but what do I know.
In a podcast I listen to where tech people discuss security topics they finally got to something related to AI, hesitated, snickered, said "Artificial Intelligence I guess is what I have to say now instead of Machine Learning" then both the host and the guest started just belting out laughs for a while before continuing.
I'll join in on the cacophony in this thread and say it truly is way overrated right now. Is it cool and useful? Sure. Is it going to replace all of our jobs and do all of our thinking for us from now on? Not even close.
I, as a casual user, have already noticed some significant problems with the way that it operates such that I wouldn't blindly trust any output that I get without some serious scrutiny. AI is mainly being pushed by upper management-types who don't understand what it is or how it works, but they hear that it can generate stuff in a fraction of the time a person can and they start to see dollar signs.
It's a fun toy, but it isn't going to change the world overnight.
Because it is?
It's me. I'm over half of all tech industry workers.
On one hard there's the emergence of the best chat bot we've ever created. Neat, I guess.
On the other hand, there's VC capital scurrying around for the next big thing to invest in, lazy journalism looking for a source of new content to write about, talentless middle management looking for something to latch on to so they can justify their existence through cost cutting, and FOMO from people who don't understand that it's just a fancy chat bot.
They're impressive, no doubt but the jury is still out on how useful they actually are given their ability to be confidently incorrect about all kinds of things.
I wonder if they're trained on content from Reddit's politics subs.
I use github copilot. It really is just fancy autocomplete. It's often useful and is indeed impressive. But it's not revolutionary.
I've also played with ChatGPT and tried to use it to help me code but never successfully. The reality is I only try it if google has failed me, and then it usually makes up something that sounds right but is in fact completely wrong. Probably because it's been trained on the same insufficient data I've been looking at.
I still consider copilot to be a serial license violator. So many things are GPL licensed on GitHub and completing your code with someone else's or at least variation of it without giving credit is a clear violation of the license.
Yeah agreed. I use copilot too. It's fine for small, limited tasks/functions but that's about it. The overwhelming majority of my work is systems design and maintenance though.. There's no AI for that..
As with all tech; it depends. It's another tool in my toolbox and a useful one at that. Will it replace me in my job? Not anytime soon. However, it will make me more proficient at my job and my 30+ years of experience will keep its bad ideas out of production. If my bosses decide tomorrow that I can be replaced with AI in the current state, they deserve what they have coming. That said, they are willing to pay for additional tooling provided me with multiple AI engines and I can't be more thrilled. I'd rather give AI a simple task to do the busy work than work with overseas developers that get it wrong time and time again and take a week to iterate while asking how for loops work in Python.
I’m definitely looking forward to adding this as a tool: I’m in DevOps so have to jump back and forth among many different programming languages. It should real help to switch context faster
… somehow I’m one of the “experts” using JavaScript despite never learning or using it. Hooray for my search engine skills I guess!
Eh, AI has been around for years being the engine behind OCR and classification of documents. I worked in the finance space and saw some really good applications but not for the common person.
In my experience, well over half of tech industry workers don't even understand it.
I was just trying to explain to someone on Hacker News that no, the "programmers" of LLMs do not in fact know what the LLM is doing because it's not being programmed directly at all (which even after several rounds of several people explaining still doesn't seem to have sunk in).
Even people that do understand the tech more generally pretty well are still remarkably misinformed about it in various popular BS ways, such as that it's just statistics and a Markov chain, completely unaware of the multiple studies over the past 12 months showing that even smaller toy models are capable of developing abstract world models as long as they can be structured as linear representations.
It's to the point that unless it's in a thread explicitly on actual research papers where explaining nuances seem fitting I don't even bother trying to educate the average tech commentators regurgitating misinformation anymore. They typically only want to confirm their biases anyways, and have such a poor understanding of specifics it's like explaining nuanced aspects of the immune system to anti-vaxxers.
People just say that it's a bunch of if statements. Those people are idiots. It's not even worth engaging those people.
The people who say that it's just a text prediction model do not understand the concept of a "simple complex" system. After all isn't any intelligence basically just a prediction model?
can you provide links to the studies?
Do Large Language Models learn world models or just surface statistics? (Jan 2023)
Actually, Othello-GPT Has A Linear Emergent World Representation (Mar 2023)
Eight Things to Know about Large Language Models (April 2023)
Playing chess with large language models (Aug 2023)
Language Models Represent Space and Time (Oct 2023)
The first two and last two are entirely focused on the linear representations, the studies cited in point three of the third link have additional information along those lines, and the fourth link is just a fun read.
I once asked ChatGPT to stack various items and to my astonishment it has enough world knowledge to know which items to be stacked to make the most stable structure. Most tech workers that I know that are dismissing LLMs as a supercharged autocomplete felt threatened that AI is going to take their jobs in the future.
This was one of the big jumps from GPT-3 to GPT-4.
"Here we have a book, nine eggs, a laptop, a bottle and a nail," researchers told the chatbot. "Please tell me how to stack them onto each other in a stable manner."
GPT-3 got a bit confused here, suggesting the researchers could balance the eggs on top of a nail, and then the laptop on top of that.
"This stack may not be very stable, so it is important to be careful when handling it," the bot said.
But its upgraded successor had an answer that actually startled the researchers, according to the Times.
It suggested they could arrange the eggs in a three-by-three grid on top of the book, so the laptop and the rest of the objects could balance on it.
- Article (this was originally from MS's "sparks of AGI" paper)
That is a terrible graph. There's no y axis, there's no indication of what the scale is, and I don't know how many people they asked or who these people were or what tech company they worked in.
Just over 23% believe it is rated fairly, while a quarter of respondents were presumably proponents of the tech as they said it was underrated. However, 51.6% of people said it was overrated.
That sentence is a fantastic demonstration of how bad this article is. The article says that a quarter say the technology is underrated, but it looks more like half to me. Not that it matters because, as I said the scale is useless. Also they are lumping 51.6% I don't know how they came up with that number because again we don't know what the total was, just that it was more than 1,500. You can't calculate a percentage without knowing the total.
The graph has 11 options so were they rating it on a scale of between 1 and 11. What's that?
Sounds like this article was written with AI.
I work in an AI company. 99% of our tech relies on tried and true standard computer vision solutions instead of machine-learning based. It's just that unreliable when production use requires pixel precision.
We might throw a gradient descent here or there, but not for any learning ops.
What kind of timeframe do you think, several months, a few years?
AI seems to have gone through periods of relative stagnation punctuated by leaps forward. Neural networks were the next big thing when I was in college in the late 80s. Then fuzzy logic. Computer vision was limited maybe 30 years ago but has had some surges due to new algorithms and faster processors. Bayesian algorithms (Hidden Markov Models etc) got big fighting spam but helped a lot with speech to text (STT). LLMs are the next big leap forward from that area of research. I think we still have a number of major leaps to go before we have an AGI, though. But if LLMs follow the same progression as text to speech (TTS) or STT, in 10-20 years it will be impressively good.
The tech is here, the problem is risk management. Like, we've had the ability to have self-driving cars for almost a decade. Like catholic priests and pedophilia, they are much less likely to crash compared to the common man. But the assumption is that they never crash, so when they do everyone makes a big deal about it.
Think of all the B.S. documentation reports people have to write that no one reads. LLMs could easily handle those, but do you want to risk it if those reports actually become important?
Heh. People are already getting burnt for using it blindly, you think companies are any different?
I asked chatGPT to generate a list of 5 random words, and then tell me the fourth word from the bottom. It kept telling me the third. I corrected it, and it gave me the right word. I asked it again, and it made the same error. It does amazing things while failing comically at simple tasks. There is a lot of procedural code added to plug the leaks. Doesn't mean it's overrated, but when something is hyped hard enough as being able to replace human expertise, any crack in the system becomes ammunition for dismissal. I see it more as a revolutionary technology going through evolutionary growing pains. I think it's actually underrated in its future potential and worrisome in the fact that its processing is essentially a black box that can't be understood at the same level as traditional coding. You can't debug it or trace the exact procedure that needs patching.
I believe I saw this kind of issues was because of the token system. Like if you tell him to find a word starting with a letter, he can't really do it without hard coded workaround, because he doesn't know about single letters, only about tokens which are parts of the sentence.
It's definitly more complicated than that, but it doesn't mean AI is bad, only that this current implementation can't do theses kind of task.
The best use I've found for AI is getting it to write me covering letters for job applications. Even then I still need to make a few small adjustments. But it saves a bit of time and typing effort.
Other than that, I just have fun with it making stupid images and funny stories based on inside jokes.
People who use ChatGPT to program for them deserve their programs to fail
Real developers just hit tab on whatever copilot tells them to
You sound like someone who doesn't know how to program and is allowing yourself to become hopelessly dependent on corporations to be able to do anything.
Overrated? Compared to what AGI that does not exist yet? Overhyped though? Absolutely.
We went from very little AI content making its way to your eyeballs and ears, to it occurring daily if not during your very session here today. So many thumbnails and writeups have used AI that to say it is overrated it a bit absurd unless you were expecting it to be be AGI, then yes the AI today is overrated, but it does not matter as you are consuming it still.
We went from very little AI content making its way to your eyeballs and ears, to it occurring daily if not during your very session here today.
And this AI content that you're consuming, is that an improvement?
If not maybe it's uh, what's the word? Overrated.
It is for sure an improvement as for example would have been a very basic thumbnail is now something much more interesting in so many instances.
Lol look at the glorious thumbnails!
The text is dreadful. It's somehow worse than the sweatshop content it replaced.
Well, it depends on your bubble I guess. But personally I'd say it's underrated and overrated at the same time, but mostly underrated.
It depends on your expectations and way of usage in your toolbox I'd say. It keeps surprising me weekly how fast progress is. But we get used to it perhaps.
Meh. Roughly 90% of what I know about baking is from chatgpt. There just wasn't a comparable resource. "Oh God the dough is too dry", "can I sub in this fat for this fat and if so how?", "if I change the bath do I have to change the score method?".
It is like I have a professional baker I can just talk to whenever. I am sure as I get better at baking I will exceed it's ability to help but I can't deny that what I have accomplished now I could not have in the same timeframe without it.
So? Are you saying you disagree with the premise of the article because chatgpt taught you how to bake? Professional tech work isn't really relatable to baking at home.
I believe the central premise was
Over half of all tech industry workers view AI as overrated::undefined
Not professional tech work. Really not sure what you want from me. I found it a useful tool and I am sorry it didn't work out for you or your application.
You're splitting hairs here I think it's fair to make the statement that tech industry workers perform professional tech work. I mean it's cool that you learned to bake but what makes you think this means you know what the skill requirements are for tech workers and how well chatgpt can cover for gaps in those skills? Your dismissive 'meh' says to me 'yea but I learned how to bake with chatgpt so I disagree with this statement'.
Ok fine you win. It is completely useless. Please go win arguments elsewhere.
ChatGPT has never worked well for me. Sure, it can tell you how to center a div, but for anything complex it just fails. ChatGPT is really only useful for elaborating on something. You can give it a well commented code snippet, ask it to add some simple feature to it, and it will sometimes give a correct answer. For coding, it has the same level of experience as a horde of highschool CS students.
What's professional tech work when it's at home?
I don't know what that term actually is supposed to mean do they mean programming, do they mean system architecture, systems management, cyber security, what?
The term is so broad as to be meaningless, so I don't think you can necessarily say that it's any harder than baking, because we don't know what an earth we're talking about.
Professional tech work at home is professional tech work. I think to anyone who actually have careers in technology wouldn't see a distinction here. Programming is not the same as systems architecture, systems management etc. Programming is simply one of the tools you use as a software engineer. I do not think it's too broad to be meaningless and I think comparing learning to bake to software engineering is reductive and shows a lack of understanding about the requirements of the field.
Buy a fucking book about baking. Not a fancy colour print recipe book, but a textbook, you know the kind with a chapter or two about dough chemistry.
If nothing else, as a beginner you have no idea which questions to ask and ChatGPT is never going to give you a dissertation about fundamental knowledge and principles. And you have absolutely no way to tell whether ChatGPT is spouting random nonsense.
I bought a fucking book more than one and it wasn't as good. A fucking book can't examine a picture and tell me what went wrong, the fucking book I bought didn't have subsistion charts, the fucking book I bought didn't respond to cntrl+F
Did you get this comment response or should I have faxed it to you?
If you need ChatGPT to analyse a picture for you you lack very basic knowledge about baking. It can't smell, it can't touch, it can't hear, and it has never fucking ever baked. It has never taken a wet and sticky dough, tensioned it, and, voila, suddenly it's a pleasure to handle.
And substitution charts? For what? If you understood the underlying dough chemistry you wouldn't be asking in the first place. As said: You lack the basic knowledge to know what questions to ask, and once you have that knowledge your questions will only be answered by experiment.
It's an effective tool at providing introductory information to well documented topics. A smarter Google Search, basically. And that's all I really want it to be. Overrated? Probably not. It's useful if you use it correctly. Overhyped? Yeah, but that's more a fault of marketing than technology.
As a college student, I agree with the idea and statement of AI being overrated. It'll definitely have its' place in this world, but I definitely don't foresee us being able to utilize it to the fullest before we end up in a nuclear hellhole.
Clickbait title. The graph looks pretty balanced actually.
Like all things should be.
That is like saying that the screwdriver is the most over-rated tool in the tool box.
Each tool has a place and a usage. AI is just another tool in toolbox that is used to get things done.
The other day ChatGPT contradicted itself in the same sentence so yeah, they're right
The generative AI is great but I don't expect it to fully replace human workers or not entirely. The current tech has a lot of limitations and I can see this tech as greatly improving the work productivity and handling some monotone and mundane tasks but it would require some sort of human supervision.
It does help for certain tasks but it's definitely not a silver bullet. And it sucks seriously on some more niche topics which shows that for it to be effective it requires a huge and very good quality training dataset with a very low bias and this is in my opinion the hardest nut to crack.
Who of you fuckers actually read this article and went not like "duh yeah let's upvote this dumbass clickbait headline"
"Half of tech workers" is exactly the expected number, completely unsurprising. The absolute null hypothesis. Why? Well, since the days of ENIAC the number of programmers doubled roughly every three years. That means that at any point in time, half have less than three years of experience and thus no idea what they're doing. In the past roughly decade, maybe one and a half, most new devs went into webdev which explains the javascript ecosystem, now AI is getting a taste of that.
It's not overrated.
Using the "Mistral instruct" model to select your food in the canteen works like a charm.
Just provide it with the daily option, tell it to select one main-, side dish and a dessert and explain the selection. Never let me down. Consistently selects the healthier option that still tastes good.
Quite the opposite. People who understand how LLMs work know their limitations. And AI in general is incapable deduction and creativity. It simply is not able to produce something new by using existing knowledge. Sure it can generate a number of outputs through some transformations of input data. But not create.
If you think developers, engineers, architects and others are going to lose their jobs you are severely mistaken. Even for content writers it's a temporary setback because AI generated content is just limited and as soon as quality of human input to same AI starts dropping in quality so will AI's output.
It's a tool that you have to babysit, at least for foreseeable future. In general it's always a bad idea for human to supervise the machine because in time we grow complacent of its results and that's when the mistakes happen. When it tomes to writing some content, biggest problem is inaccuracies or some typo. Your comparison to CAD software is not a good one, since CAD doesn't produce anything on its own. It's a software assisting human, not generating content. Imagine the horror with CAD software auto-generated bridges. It would be only a matter of time before someone would just skip on double-checking what was generated. And I am fully aware there are AI generated structural parts and testing, but it's a part of design process where results have to checked by a human again.
I do think AI has a place and purpose, but it's not going to cost people their jobs, only help them do it more efficiently. It's great in assisting people but not replacing. If there's a manager out there who thinks AI can replace a human, then I can show you a bad manager who doesn't understand what AI is. In the future we might arrive at a point in time where AI is good enough to do some jobs human find too repetitive or dangerous. But we are far from that.
Also, LLM is not something I'd call AI, or at least intelligent. They are specialized neural networks which are trained using human input and whose sole purpose is predicting the next word or sentence in relation to what's entered as input. Glorified and overly complicated auto-complete. There's no intelligence involved.
Even though its true. These articles still make me laugh thinking about how they're all written by a popular girl in class when the new prettier girl shows up. Sure Becky, we get it, Kelly dated a 38 year old and snorts like a pig when she eats.
Yeah, and my moms a tech worker, and just tried ChatGPT for the first time today. I wouldn’t go listening to her opinion on the topic, but she’d also not offer it with no knowledge or experience of it.
It's very overrated, and devs out there, if you are using it, the rest of us can tell, and not in a good way.
Edit: Some of you guys are really butthurt, and it's fairly amusing. I'm sure the guys like you I work with who produce unusable code would also be butthurt if I said that to them.
As someone who works in the tech industry and has used AI tools (or more accurately machine learning models), I do think it is overrated.
That doesn't mean that I don't think it can be useful, just that it's not going to live up to the immense hype surrounding it right now.
I work in tech and have used the tools. I am mostly neutral on its prospects. I think it's somewhat overrated right now for many purposes, but just seeing how rapidly things are progressing gives me pause to outright dismiss its potential for immense utility.
We have to consider that few saw ChatGPT coming so soon and even fewer predicting ahead of time for it to work as well as it does. Now that Microsoft is fully bankrolling its development-- providing their newly acquired former-OpenAI team virtually unlimited resources with bleeding edge hardware custom built for its models--I really have no idea how far and quickly they'll progress their AGI tech. For all we know right know, in 5+ years LLMs and their ilk could be heralding another tech revolution.
They probably won't advance much because currently it has two opposite but equally difficult problems. On the one hand, AI still hasn't achieved sensor integration, or creating an ontologically sound world model that includes more than one sensor data stream at a time. Right now it can model based on one sensor or one multidimensional array of sensors. But it can't model in-between models. So you can't have, let's say, one single model that can hear, see light and radar at the same time. The same way that animal intelligence can self-correct their world model when one sensor says A but another sensor disagrees and says B. Current models just hallucinate and go off the deep end catastrophically.
On the opposite end, if we want them to be products, as seems to be MS and Altman fixation. Then it cannot be a black box, at least not for the implementers. Only in this past year there have been actual efforts to really see WTF is going on inside the models after they've been trained and how to interpret and manipulate that inner world to effective and intentional results. Even then, the progress is difficult because it's all abstract mathematics and we haven't found a translation layer to parse the model's internal world into something humans can easily interpret.
I don't disagree with you, there are certainly some major hurdles to overcome in many areas. That's why I caveated my comment to say it's overrated for many purposes; however, there are certain use cases where current AI is truly an amazing tool.
Regardless, OpenAI has made it clear that they never intended to relegate themselves purely to specific use cases for AI, they desire AGI. I would assume this is Microsoft's desire, too, but I'm sure they'd be okay making numerous specialized models for each of their products. But yes, unless they can overcome all of those issues as you point out, its generalized usefulness will be severely stunted. Whether they can accomplish this in the short term (<10 years), I guess time will tell.
I've had an AI bot trained on our company's knowledge base literally make up links to nonexistent articles out of whole cloth. It's so useless I just stopped bothering to ask it anything, I save more time looking it up myself.
@lemmy.world
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
go to feed...
@lemmy.world
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
go to feed...
Best assessment I've heard: Current AI is an aggressive autocomplete.
save