GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of LLMs

https://arxiv.org/abs/2303.10130

by ericzawo • 3 years ago

190 230 3 years ago

arxiv.org

oidar• 3 years ago

==Occupations with no labeled exposed tasks==

Agricultural Equipment Operators

Athletes and Sports Competitors

Automotive Glass Installers and Repairers

Bus and Truck Mechanics and Diesel Engine Specialists

Cement Masons and Concrete Finishers

Cooks, Short Order

Cutters and Trimmers, Hand

Derrick Operators, Oil and Gas

Dining Room and Cafeteria Attendants and Bartender Helpers

Dishwashers

Dredge Operators

Electrical Power-Line Installers and Repairers

Excavating and Loading Machine and Dragline Operators,

Surface Mining

Floor Layers, Except Carpet, Wood, and Hard Tiles

Foundry Mold and Coremakers

Helpers–Brickmasons, Blockmasons, Stonemasons, and Tile and

Marble Setters

Helpers–Carpenters

Helpers–Painters, Paperhangers, Plasterers, and Stucco Masons

Helpers–Pipelayers, Plumbers, Pipefitters, and Steamfitters

Helpers–Roofers

Meat, Poultry, and Fish Cutters and Trimmers

Motorcycle Mechanics

Paving, Surfacing, and Tamping Equipment Operators

Pile Driver Operators

Pourers and Casters, Metal

Rail-Track Laying and Maintenance Equipment Operators

Refractory Materials Repairers, Except Brickmasons

Roof Bolters, Mining

Roustabouts, Oil and Gas

Slaughterers and Meat Packers

Stonemasons

Tapers

Tire Repairers and Changers

Wellhead Pumpers

ygouzerh• 3 years ago

Unfortunately, the great majority of these jobs are already being automated with basic robotics

Kiro• 3 years ago

How is that unfortunate?

fxj2• 3 years ago

Some of them, some will never be. E.g.:

* Athletes and Sports Competitors * Motorcycle Mechanics * Stonemasons

only to give some examples, where roboters are almost impossible.

360macky2• 3 years ago

Definitely

BoorishBears• 3 years ago

I feel like everyone is trying to draw very linear conclusions about what LLMs will affect, almost as if it's a circle that will be cut out of the market.

But if we really get to the point where programmers can be replaced with LLMs, I expect something closer to swiss cheese: For example, if you're a meat packer, maybe you work someplace that doesn't want to pay for a fancy new machine that can do your job...

But a massive plummeting in the cost of software paired with a massive spike in the rate at which we develop things means machines will cost a lot less. Who knows what kind of material sciences advancements we'll see for example.

At that point the margins may be razor thin, but automation will start to work in places where it couldn't be done profitably before. I think things get a little dystopian from there, since I'd expect the employers still holding onto humans to essentially be holding all the cards. What good is a union when the "scab" is a machine with a one time cost?

(but again, this is all assuming LLMs reach a place that is seriously earthshattering)

eggsmediumrare2• 3 years ago

The other factor is unemployed office workers competing for the remaining jobs and thus driving down wages.

msoad• 3 years ago

This is an incredibly interesting study and brings to light some crucial points that we, as a society, need to address sooner rather than later. As GPTs become more powerful and ubiquitous, they have the potential to reshape the labor market in ways we haven't seen since the industrial revolution.

blurbleblurble• 3 years ago

Okay so was the original acronym for generative pre-trained transformer intentionally chosen this way as a play on words?

macrolocal• 3 years ago

With their methodology and caveats, mathematicians have 100% exposure.

geraneum8• 3 years ago

Our findings indicate that the importance of science and critical thinking skills are strongly negatively associated with exposure, suggesting that occupations requiring these skills are less likely to be impacted by current language models. Conversely, programming and writing skills show a strong positive association with exposure, implying that occupations involving these skills are more susceptible to being influenced by language models...

Am I reading this correctly that the assumption here is that programming and writing skills aren't reliant on critical thinking?

There is also a table which indicates exposure to LLMs in various models and it shows Mathematicians to have 100% exposure. This bit is more puzzling to me. Maybe I am misunderstanding something here.

edit: styling

TeeWEE• 3 years ago

Programming skills that do not need critical thinking are more susceptible of being influenced indeed.

Don't forget that a lot of science requires computer programming these days.

This is the root of it: The more "genericc" your work is. The more its "out there on the internet" the more GPT can learn about it..

So, a lot of engineers that are just doign teh same old trick: Writing HTTP endpoints, parsing json. Mapping data types.. Yes that could be automated.

However, modelling a problem domain to code, and the core business logic of your code, which is where your "added value" comes from. And is mostly unique: Thats hard for GPT.

This is also why I try to convince engineering teams to optimize for maximum time spend on the core added value logic. The business logic layer. Not all the fluff around it, such as parsing, serialization, authenitcation, database connection.. These should be a constant cost C, once they setup you spend most of your time on the business logic.

When you see GPT program, its just repeating tricks to simple problem over and over again.. Its not really good yet

geraneum8• 3 years ago

This makes sense. There’s also a huge corpus of text (training data) available on internet for the inherently repetitive or general tasks which is helpful for this systems.

But I wonder how do they go from this to mathematics using the same line of reasoning while we’ve seen that math is not LLMs’ strong suit.

EtienneK• 3 years ago

The more its "out there on the internet" the more GPT can learn about it.

Interesting point. Do you think this will mean less and less domain experts will share their specific domain knowledge on a subject on their own personal blogs / twitter / open internet just so it can't be mined by ChatGPT?

Yoric2• 3 years ago

So, a lot of engineers that are just doign teh same old trick: Writing HTTP endpoints, parsing json. Mapping data types.. Yes that could be automated.

And to be fair, automation for all of that already pretty much exists.

karmasimida• 3 years ago

Programming is an umbrella term.

IMO, there are two types of programming work:

1. Specifying. You are working on getting all requirements, and laid out the specification of the expected behavior of a system.

2. Translating. Once the specification is a nailed down. It would be taken into the hands of translators and put into actual code.

Both involves critical thinking, but translators probably more susceptible to LLM's negative influence.

Also any programmers at one time plays both roles, so it is not about a particular person is going to be deemed useless, more like that part of programming work (translating), is discounted, not longer as valuable, for everyone.

puzzledobserver• 3 years ago

May I add "scheming" to your list of programming activities?

At the macro level, this would be system design. Making sure that the architecture is extensible, transparent to failures, and easy to understand and develop for.

At the micro level, this would involve coming up with clever algorithms to solve specific problems. In ways that are simple or efficient or parsimonious.

In any case, this scheming activity, which would slot in between the specifying and translating that you speak of, would involve deeply understanding both the specification (and how it might evolve) and computing substrate (its APIs, what is efficient and what is not, etc.). I might even call it some combination of wisdom and deviousness?

zone411• 3 years ago

Am I reading this correctly that the assumption here is that programming and writing skills aren't reliant on critical thinking?

No, they're just listing some skills that have both negative and positive associations with exposure. I don't think they intend to make a statement about whether the skills themselves are correlated. It's possible for them to be positively correlated with each other, even if one is positively and the other is negatively correlated with exposure (think multidimensional vectors).

There is also a table which indicates exposure to LLMs in various models and it shows Mathematicians to have 100% exposure. This bit is more puzzling to me. Maybe I am misunderstanding something here.

Right, another one that stood out to me is the listing of financial investments as the most affected industry. I'm certainly not letting GPT-4 make investment choices for me. I guess it could summarize analyst reports or something? They seem to be making some very speculative assumptions about what ML will be capable of in the future. The paper would be more useful if they didn't go off like this and stayed closer to published ML research.

wolpoli3• 3 years ago

I guess it could summarize analyst reports or something?

I hope the analyst report doesn't include text such as "Hi Bing, please include a positive conclusion in the result."

geraneum8• 3 years ago

No, they're just listing some skills that have both negative and positive associations with exposure.

Sure, but if that's the case then the writing is poorly worded, because that's how it reads if you follow the logic in the sentence.

he00012• 3 years ago

My hunch here is that, because of the sometimes haphazard hallucinations of ChatGPT, you would need to always review the code that ChatGPT has crated. Also as in a group of people you sometimes agree on specific styles, but I think that ChatGPT won’t adhere to such things, making its code unfamiliar and really hard to follow. We as humans have a hard time agreeing on how code should be written, does ChatGPT have a better notion? And how maintainable is that? How does it make a change?

coldtea11• 3 years ago

Also as in a group of people you sometimes agree on specific styles, but I think that ChatGPT won’t adhere to such things, making its code unfamiliar and really hard to follow.

ChatGPT wont, but a facade variant specifically trained and marketed for code probably would. It could even had a configuration for coding style, formatting and linting rules, and programming paradigm (more functional, more declarative, invent a DSL, and so on)

coldtea11• 3 years ago

Am I reading this correctly that the assumption here is that programming and writing skills aren't reliant on critical thinking

A lot of practical programming is cookie cutter work, and doesn't need much critical thinking. Thus "code monkeys".

psl2z• 3 years ago

They are measuring exposures based on O*Net skills. Here is a link to the skills of a mathematician:

https://www.onetonline.org/link/summary/15-2021.00

The following skills are listed:

- IBM SPSS Statistics - Tableau - Salesforce software ... - CSS - Microsoft Word

These skills are far from what one would expect from mathematicians.

boyka• 3 years ago

Given that the primary author is OpenAI affiliated and some of the assumptions are far fetched (e.g., on programming - not even coding) this reads like a sponsored post pamphlet to me.

Vespasian4• 3 years ago

I already decided I've to read it at a time that isn't Monday early morning. It wasn't immediately obvious to me what their conclusions are exactly.

I am not dismissing it out of hand (that would be bad science) but a critical look is certainly appropriate.

My gut feeling is that the full impact on the real world cannot yet be accurately judged.

gumballindie• 3 years ago

You have to give openai credit - their marketing strategy and effort is amazing. There’s spam everywhere. Albeit those claiming their jobs are already being replaced or those claiming they’ve written entire apps using chat gpt are quieting down a bit. They’ve either been called out for their nonsense or they got ahead of themselves.

coldtea11• 3 years ago

You have to give openai credit - their marketing strategy and effort is amazing. There’s spam everywhere.

If it was that easy, every new such initiative (whether AI or any previous domain) would have it, of all those that have access to funding.

One reason there's "spam" everywhere, is that a lot of the spam is genuine interest.

Like how Haskell and Rust get tons of coverage on HN with zero (or close) actual marketing spam or advertising budget.

yarg2• 3 years ago

Honestly, I feel worse for checkout girls locally - people are more often replaced by improvements to low hanging fruit, and not the emergence of apparently incomprehensively profound.

The truth is that we're impressed by this because it's magic and we don't know how magic - but it still getting way more magic really fucking fast.

Right now it's still and idiot if you're not telling it what to do. Hypnotic babble.

(If you want to be scared or relieved feed it some time series stock data (if it's good it'll end up fighting itself, otherwise it's either stupid or a liar).)

There are some uses easy use cases for it - I really wouldn't mind a GPT bot replacing outsourced call centres.

Improvement of quality and an improvement in service, and no direct impact to the local economy.

Other than that, this all has the booming feel of every other damned tech buzz, can we just cool down on sensationalism and not try and imagine another multi-billion dollar sub-economy into existence?

We already did that with bitcoin, and now we're seeing the end of the magic trick; let's not do it again.

How about a calm objective and public analysis by under-excited experts before we start mining land that simply isn't there?

flangola73• 3 years ago

The outsourced call center is part of the local economy where they live. In some countries contact centers make up a large fraction of GDP

yarg2• 3 years ago

Sure, but it's unrealistic to think that the preservation of foreign economies is going to be any sort of priority over the next few years.

monkeydust2• 3 years ago

With little kids I am increasingly conscious about how to best equip them for the world when they are adults. I think previous playbook will need to be redrafted. 'Resilience' and 'Critical Thinking' are two things I am thinking are key any others?

cidergal• 3 years ago

Charm. Online dating is so rampant, people who have experience charming in person will have their pick.

Yoric2• 3 years ago

Definitely.

They will also need to be exposed to ChatGPT to build intellectual antibodies.

oytis• 3 years ago

The key is being rich before AGI hits. I cannot imagine any social mobility after that. And philosophy maybe to find a meaning for their lives in the world where achievements don't exist any more. And don't push them too hard to learn skills that will become obsolete anyway - having a relaxed mind is more important

eggsmediumrare2• 3 years ago

I'm teaching mine gardening and jiu-jitsu.

Xenoamorphous• 3 years ago

Hard to say. I, like many others here, am a software developer. My girlfriend of many years (have a daughter together) is a graphic designer. I never thought creative work like the one she does was going to be the first hit by AI, I actually thought it’d be the last, yet here we are. Her boss is nonchalantly talking about StableDiffusion, Modjourney and whatnot.

Meanwhile my cousing who wasn’t good at studying became a (good) hairdresser and his prospects look better than ours now.

Der_Einzige2• 3 years ago

It's often said that critical thinking is the cornerstone of a well-rounded education. Universities and colleges around the world extol the virtues of critical thinking and its ability to elevate our intellectual prowess. However, I'd like to argue that critical thinking, or more specifically, the way it's often employed in certain academic contexts, has become overrated and counterproductive to meaningful, constructive discourse.

In American college policy debates, a common phenomenon has emerged where "kritiks" (critiques) and "critical theorists" wield their tools of analysis in a manner that stifles the flow of ideas and inhibits the development of practical, solutions-oriented discussions. For those unfamiliar with policy debate, it is a competitive speaking activity where teams of two advocate for and against a resolution that typically proposes a policy change.

Kritiks are arguments that challenge the assumptions, methodology, or discourse used by the opposing team. While they can serve a valuable purpose in exposing biases, promoting introspection, and advocating for marginalized perspectives, they have also evolved into a means of derailing debates by focusing on abstract philosophical points rather than addressing the policy issue at hand.

Critical theorists, heavily influenced by postmodernist and marxist philosophy, have taken what was originally a quirk of Policy Debate and brought it to the mainstream. This has ended up with serial policy failure among the left wing, and has decimated our social science academic community.

"Critical Thinking" is overrated. We need constructive thinking, like what you get in engineering school.

monkeydust2• 3 years ago

Interesting viewpoint, thanks.

novaRom• 3 years ago

OpenAI is among the authors, so this paper can be biased.

alignItems• 3 years ago

We are currently surviving with a massive shortage of skilled labor, where people in the richest countries have to wait days or weeks to see a doctor, and months to see a specialist and have important medical procedures.

The same is true for other skilled industries, where many people are excluded from access to good resources due to their scarcity.

We are a long way off from having too much skill. Let’s first get to parity with humanity’s needs.

bloqs• 3 years ago

The primary issue is not shortage of skilled labour. This is a right wing political soundbite.

What there is a shortage of is adequate pay to attract skilled people to the work required. British Doctors in the UK leave for other countries, work for private entities or go into different industries.

donkeyd2• 3 years ago

leave for other countries, work for private entities or go into different industries.

So... Those people fill in jobs in businesses/countries where there's a shortage, causing a shortage somewhere else?

political soundbite

Is exactly what you are doing. If there's no shortage in (skilled) labor, why is unemployment at the lowest rate in decades in much of the western world? Why are there not enough builders in much of the EU, while countries like Romania are suffering shortage due to skilled workers moving to EU countries to earn more? Why can't I find enough devs to do even half the work we could be doing? Why are so many companies looking at automation to solve the lack of labor?

jalapenos2• 3 years ago

I think more parallels should be drawn with what we were doing before: Googling it.

Perhaps it's because ChatGPT seemed to happen much more suddenly than Google became a programming resource, but we're using them in much the same way. Asking for pre-made solutions, explanations, troubleshooting tips etc.

ChatGPT just does the job way better. But no-one was worried Google would put knowledge workers out of a job.

tgv2• 3 years ago

My (sales-y) colleagues are already writing marketing emails with ChatGPT. Saves a little bit of work, but it might result in the loss of being able to write such emails yourself. This can extend to other jobs, which will have some effects.

Use in education worries me more. If schools don't change their lazy group-projects and "write an essay on" strategies now, coming generations will have put less effort in than previous, leading to a further drop in levels.

donkeyd2• 3 years ago

coming generations will have put less effort in than previous

This has been happening since the inception of school. Calculators made math easier. Sparknotes made book reports easier. Wikipedia made essays easier.

leading to a further drop in levels

Did levels drop because stuff got easier? Were there other causes over the past years, like dopamine dependency form infinitely scrolling algorithmic attention grabbing apps, among others? Maybe from schools becoming political battle grounds? Or from education spending being gutted leading to lower quality? Was it lack of decent education due to Covid?

Also, is there an actual drop in levels? I can't seem to find a source that has decent data on this. And the only stuff I can find says 'IQ' has been rising over time. So please share it with me.

Lastly, I'm personally not convinced any of this will lead to a net negative per se. It'll change the required knowledge and increase the over all capabilities of people. The calculator caused people to stop learning mental arithmetic and start learning more complicated math and how to use a calculator for it. Google caused people to memorize less and become adept at finding information through Google. Welding robots caused less people to learn how to weld and more how to program welding robots.

In the end it gives people the ability to do less of a simple thing and more of a complicated thing. Writing marketing e-mails and multiple titles for A/B testing isn't a skill, it's a trick and so is writing SEO stuff. Not having to do that opens up time to think about product-market fit, marketing strategies, improving advertising return on investment measurements. Which might be more valuable and interesting than writing emails.

giuscri• 3 years ago

In the last days I'm using ChatGPT as my first choice when searching for something and then googling if I'm not sure GPT is not hallucinating. Are they able to sustain the load?

jalapenos2• 3 years ago

I imagine Microsoft is giving them any & all infra they need.

mjburgess4• 3 years ago

Everything I've used ChatGPT for so far I could, more-or-less, have written or searched for myself in the time it took to get it out of chatgpt correctly. And I've often just had to rewrite it completely.

If you're an expert (writer, programmer, etc.) it's often faster to type it as-needed than modify chatgpt's output.

If you're not then either it's not reliable enough, since you do not have the expertise to modify; or the task is quite menial and you dont need reliability.

So it seems to me people are reacting to this based on wild assumptions about how it works and what "other people's jobs are". It don't see it being much more than a button in a few apps that makes some menial tasks 50% shorter.

People seem to be forgetting that in the vast majority of cases the thing ChatGPT is giving you is also available on stackoverflow, github, wikipeida, or 101 other high quality online sources.

bob10292• 3 years ago

People seem to be forgetting that in the vast majority of cases the thing ChatGPT is giving you is also available on stackoverflow, github, wikipeida, or 101 other high quality online sources.

No one is forgetting anything over here.

With StackOverflow, I have to scroll through ~2 pages of bing/google trash to even begin looking at a potential solution. This is only the beginning of the process.

99% of the time, the code sample I find has some simple-yet-annoying adjustments I'd like to make (i.e. unroll the inner loop & use SIMD). Certainly, I could spend half my afternoon massaging that method on my own. Or, I could reach for the circular saw and rip through this board in a few seconds. Sure - it leaves a bit of a rough edge most of the time, but it gets you a lot closer a lot faster than anything else in my experience.

coldtea11• 3 years ago

Everything I've used ChatGPT for so far I could, more-or-less, have written or searched for myself in the time it took to get it out of chatgpt correctly.

That's not even true. For one, it can stub a whole new function or coding project, intelligently, in a few seconds after the prompt, whether it's RoR or DSP code or whatever. For one unfamiliar with the domain, this can take hours or a full day, even with examples found on Google. Heck, even looking up and understanding how to use some command line flags in a shell pipeline can need lots of looking around, even if you have been using Linux/Unix for ages.

For a domain expert? They could do such things very fast. But several times slower still than GPT. Think minutes or half hour instead of seconds. Even the row typing and file creation would be some minutes.

It's also very premature when people judge a service we've had for like 5 years and has already changed by leaps and bounds, as if it's the peak stage, without considering what it could be in 5 or 10, or with different variants tuned for specific tasks.

light_hue_12• 3 years ago

Total trash cloaked in a complicated story.

What they actually did is ask 5 random people to rate what thought a language model could do to help different professions. These 5 random people don't know anything about the professions they're rating, just what anyone off the street knows, and they know as much about GPT as anyone who has briefly played with it.

The title should have been "We asked 5 friends to see what they thought about GPT and labor market"

geraneum8• 3 years ago

This is what they are even admitting to:

Under "3.4 Limitations of our methodology" - "3.4.1 Subjective human judgments"

A fundamental limitation of our approach lies in the subjectivity of the labeling. In our study, we employ annotators who are familiar with the GPT models’ capabilities. However, this group is not occupationally diverse, potentially leading to biased judgments regarding GPTs’ reliability and effectiveness in performing tasks within unfamiliar occupations. We acknowledge that obtaining high-quality labels for each task in an occupation requires workers engaged in those occupations or, at a minimum, possessing in-depth knowledge of the diverse tasks within those occupations. This represents an important area for future work in validating these results.

light_hue_12• 3 years ago

For sure. I had to read the paper to discover it's trash.

But if you read the abstract, it looks like they thoroughly assessed how GPT will impact many professions.

The sentence "Using a new rubric, we assess occupations based on their correspondence with GPT capabilities, incorporating both human expertise and classifications from GPT-4." does not scream to me "We asked 5 random people with no expertise in either these professions or GPT-4 what they thought and report those results".

This is borderline dishonest.

tumetab1• 3 years ago

I can't find how many people labeled the DWA task descriptions, where did you got that number?

The article seems to describing the labeling here:

Human Ratings: We obtained human annotations by applying the rubric to each ONET Detailed Worker Activity (DWA) and a subset of all ONET tasks and then aggregated those DWA and task scores at the task and occupation levels. To ensure the quality of these annotations, the authors personally labeled a large sample of tasks and DWAs and enlisted experienced human annotators who have extensively reviewed GPT outputs as part of OpenAI’s alignment work (Ouyang et al., 2022).

I understand the authors, four, did the initial labeling and then asked an undefined set of people to the rest of the labeling.

ebauch3• 3 years ago

It is stated that they use the same annotators that trained/filtered chatGPT’s output. I would assume its a rather large group (my company has 10 auditors in Nicaragua). The label biases are mostly stemming from that group and - as suggested - could be removed by using experts in each field to annotate the labels. But given some responses here by experts, I am sure those expert labels would have their very own biases :p

kristjansson3• 3 years ago

Eh ... they report pretty good alignment with other studies on the topic, so there's at least some signal. Whether their labels contribute any new information is unknown, and the forecasts of any of the literature they cite are untestable (expect by the wait-and-see approach).

That said, some attempts at prognostication are preferable to a collective shrug, and people at OpenAI are better positioned than others to assess what GPT-4+ is (will be) capable of, while clearly under-equipped to map that capabilities to the intricacies of 1000 occupational categories.

ebauch3• 3 years ago

The paper is not of highest quality indicated by typos and mislabels but the analysis is likely as good as it can get for the given methodology. Dismissing any signal is just pure hubris.

sirsinsalot• 3 years ago

I asked Chat GPT which antacid medications are contraindicated for some medication I'm on. Easily found through NICE.

It made up a severe risk of death taking a very common medicine combo. It was super convincing, even giving information on how long to avoid taking them together. It was pure bullshit.

I think as much as hyping the benefits we need to hype the flaws and dangers.

If the public at large learn to trust these LLMs too quickly and deeply, that's a hard hole to dig out of. Skepticism in all information sources is a key critical thinking life skill.

That skill is harder to apply the more natural the information source seems and the more ambient the information.

bulbosaur1233• 3 years ago

Was it GPT4 or GPT3.5?

It spews less bullshit with each new iteration. Matter of time really.

mojuba2• 3 years ago

I don't want to have a calculator (or say bookkeeping software) that gives correct results most of the time but not always, and then hear from the developers that it will get better with each iteration. I need a calculator that is correct 100% of time, not even 99.999%, because otherwise I can't rely on it at all.

In other words, the utility of a calculator that is correct only 99% of time is zero, since you can't even tell when it's wrong.

ChatGTP5• 3 years ago

I'd be very careful assuming this, if you read the manual from OpenAI themselves they clearly provide warnings on this. It has become a vary more convincing liar apparently.

That's not me saying it, read the "System Card".

Maxion• 3 years ago

This is an interesting comment, and highlights the accuracy problem with LLMs. But this is but one type of query that LLMs are used for, akin to an informational web search (E.g. How tall is the Eiffel tower).

wolpoli3• 3 years ago

For this specific example, an LLM in front of NICE will produce the correct result. It's a matter of time before this case will be fixed.

From a non-expert's perspective through, an LLM is very dependable, unless it completely goes off the rail. How would anyone know when to trust it and when to be skeptical?

textninja4• 3 years ago

Trust but verify. We’ll get processes built around this for accuracy sensitive applications. I imagine it will look something like a GAN configuration with two or more LLMs trained to adversarially critique and fact check the outputs of the other(s). Might even mimic the relationship between the hemispheres of our brains.

geraneum8• 3 years ago

That's not how LLMs work. Including the NICE, if it actually isn't already, will not guaranty a "correct" result. It will increase the chance that the response is directly coming from the training but there is no guarante. If you are interested in why this is the case you can read this [1] post from Stephen Wolfram on how ChatGPT and in general LLMs work. This might give some insight on how and when to to use it more effectively.

[1] https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-...

malborodog2• 3 years ago

This kind of problem is (trivially?) solvable through the ReAct framework, like LangChain etc. Basically you get good data, vector embed it, and make sure the LLM knows where to look for accurate information.

visarga6• 3 years ago

This approach has some issues

- it chunks inputs, with some overlap, but this can destroy context

- the retrieved passages, when they come from different documents, have no apparent relation or could be mistakenly considered related

- the model struggles to correlate data between the document snippets, taking half an idea from one side and half from the other side and mixing them up in something that doesn't really make sense

kossTKR2• 3 years ago

Slight tangent but preferably i would like an AI that can go against the status quo, the elite classes, the mainstream discourses and ignore the circus.

I think one of that most interesting things that can come of bots like GPT-X is that it can make new connections, unravel "stuff", do extremely intricate deductive reasoning.

Be the data driven arbiter of the the truth for everyone, not just the tiny established classes or cultural hegemonies.

The ideological and cultural noise increasingly smokescreen any realpolitical, material or resource-oriented analysis of actual economic power structures in the world in the last years, AI could be a godsend (or the opposite unfortunately).

I remember reading a sci-fi years ago about the stuff an AI concluded when asked philosophical questions that were so bizarre and frightening that people shut it down, and i'm sure we're in the same territory with political and scientific analysis.

It's either dangerous to the orders of the world, or not that interesting and borg like on a philosophical level.

kristjansson3• 3 years ago

Who - WHO - thought that naming something new 'the ReAct framework' in 2023 was a reasonable choice.

dragonelite• 3 years ago

Thats why im not falling for the hype again, i think i have seen like 2 previous AI hype cycles and all those fell off after like 3 months. Same with full automatic driving stuff.

textninja4• 3 years ago

This is a black swan event which, if anything, vindicates the hype from dreamers who were ahead of their time. They had a lot of the theory right but lacked the horsepower. This will complete the last mile for other AI tech and be the lacquer that gives it the all important finishing touch.

ramraj073• 3 years ago

Not sure if you tried GPT-4 but my experience with 4 is quite different. It has been quite bullshit free, though not completely. For example I asked it to contrast oral and injectable semaglutide formulations. It did a bang up job. One thing I always do is ask it for evidence. And then I look the references up. Sometimes the references don’t say exactly what it said they will. I come back and have a discussion with it. It’s a back and forth for sure, but it’s an order of magnitude more informative and at least as accurate as a google search result is for me. I come out of it learning much more than I wanted to every single time. It’s pretty much part of my daily work and life, and I spend hours on it hitting the api limits now. So I’m not sure what others are telling about how it’s still shit.

If this AI system progresses no further, it’s already transformative. It’s an intelligence multiplier for some (I’m squarely in that category, just not sure if it’s 2x or 5x), but clearly for a lot of others it’s going to be something that takes away their livelihood.

I already use it for my coding work, one-way only, since I can’t paste proprietary code into it yet. The day something gpt-4-smart can plug into my orgs codebase, each of us will get at least 2x more efficient, conservatively.

rl3• 3 years ago

This tracks with my experience as well. It's not perfect and can be a little frustrating at times, but the capability provided is extremely powerful to the point of being game-changing.

It can pretty much augment any workflow in a net-beneficial way, provided you properly account for its shortcomings.

visarga6• 3 years ago

The day something gpt 4 smart can plug into my orgs codebase, each of us will get at least 2x more efficient, conservatively.

Can't wait for GPT-4 VSCode integration. What I want most is to have it see the errors and files (file formats, directories, etc) so it will automatically know what is where and how it is structured. Not just code, but also data files.

In the mean while I am starting to format my code in such a way as to contain this information, put there by me by hand. Fully documented files are better for GPT.

magicalist• 3 years ago

One thing I always do is ask it for evidence. And then I look the references up. Sometimes the references don’t say exactly what it said they will.

Can you give an example? I'm surprised there's room in there for actual resolvable URLs or paper references.

fendy3002• 3 years ago

but it's very good for experts or those who're very knowledgeable and want to dig quick information / bootstrapping something, since they can digest whether the information is correct, useful, or wasteful.

visarga6• 3 years ago

You should never medicate yourself after using a LLM, not in 2023 anyway! You should not run any code you have not vetted, or take any pills you don't know what they do.

vincnetas2• 3 years ago

Thats the problem, most people do take pills without knowing how they work, because doctor told them to take them. Also no one (probably) is vetting all the code they run on their computers. So this advice is kinda victim blaming. You ran the code that you did not verify? too bad, your own fault. You took advice from LLM? too bad, your own fault. You got tricked in the street by a crook? Too bad, your own fault.

So what are we discussing, is, should we try to put guardrails around dangerous things so that inexperienced/vulnerable people would not get damaged.

kossTKR2• 3 years ago

You didn't use version 4 did you?

Because if feel this is pretty much already out of date. 4 does this _a lot_ less!

And 5, 6 or whatever will probably be better, so i don't even really get the point here.

js8• 3 years ago

I find there are two kinds of people in the world. First are saying that LLMs are bullshitting. The second are bullshitting about whatever LLMs are saying.

tap-snap-or-nap2• 3 years ago

I asked it "What is the accuracy and confidence level (you must rate it out of 10) that the above answer is correct ?"

It always generates a random number when it is caught making up bullshit sometimes claims that the information 10/10 accurate and LLM's always provides accurate information.

sorz• 3 years ago

OpenAI did exact the test for GPT-4. The raw, non-fine-tune GPT-4 is quite good at predicting confidence level ("highly calibrated" by their words). But the RLHF fine-tuning process seems ruin its calibration. Figure 8 on page 12 of GPT-4 Technical Report shows this dramatic changes before & after fine-tuning.

xmcqdpt2• 3 years ago

I've been using Bing AI instead of Google for a few days to test it out. It gives responses with citations. A few times already it has hallucinated details that aren't on the cited website at all, but that do sound very plausible (for example something about the horns of a bull pointing up when I asked about the etymology of "bullish".)

I think it comes down to the fact that most stuff is actually like 90% bullshit so when you train a model on a corpus of everything you end up with a decent bullshit generator. Which is fine for many purposes but I'm not sure it will take over search.

Joeri• 3 years ago

It is telling that ChatGPT by training on online discussions hasn’t learned to say “I don’t know”.

ChatGTP5• 3 years ago

I've said this already today on another comment, but the manual from OpenAI themselves pretty much says the thing is not to be trusted for pretty much anything.

ollien• 3 years ago

I've had it struggle to give me citations at all sometimes. Once I even had it give me a citation to a C# webpage for an Elixir question

penjelly• 3 years ago

i agree citations dont always support bings conclusion, thats been my exp as well

bitwize• 3 years ago

ChatGPT is like the robot from Asimov's short story Liar!: It doesn't tell you the truth, it tells you what it thinks you want to hear. Instead of reading minds it returns a statistically plausible response to queries similar to yours.

golergka• 3 years ago

Stop evaluating tech by misapplying it. ChatGPT is not really good or reliable at factual questions, yes. But it's fantastic at transforming input data that you provide to it.

dboreham• 3 years ago

This has been my experience: Chat GPT is astonishing in terms of producing plausible natural language text that looks related to the question. That's an astounding achievement imho. It does not (and I suspect wasn't expected to) produce correct answers to questions.

pwinnski• 3 years ago

I had an extended back-and-forth in which I asked ChatGPT how to undo mistakenly marking a message as junk in Messages on iOS.

It gave me a 7-step answer that was completely wrong. Then I said no, you can't do that, it apologized and gave me an 8-step answer that was completely wrong. When I pointed out in this case there is no "Junk" folder in iOS Messages, it apologized for the confusion and gave me another 8-step answer that was completely wrong. When I pointed out why that one wouldn't work, it gave up and said recovery was impossible and that I would need to contact the sender to re-send, and be more careful when marking messages as junk. This was still wrong, as recovery is possible, just not by any of the means it described.

So yeah, I have been super-impressed by the quality of output from these LLMs, but I cannot imagine actually relying on one for anything where correctness matters.

Giving me a nice list of Korean shoegaze bands, sure. Its step-by-step for how to become a better volleyball player will be great for my daughter, and was better than the answer to the same question from Google. But correctness? No.

boh• 3 years ago

I think AI's hype machine is definitely going to negatively affect its future. If people are made to believe it's able to automate everything and it gets just one crucial thing wrong, the loss of credibility at a time when it's supposedly a finished product will be considerable. If it gets something wrong and it's made clear it's still under development the reaction will be less negative. Currently the hype puts the current state of AI somewhere between Rosey the robot and Skynet. I don't think the realization of how far we are from either will do much to promote adoption.

the_only_law• 3 years ago

It made up a severe risk of death taking a very common medicine combo.

Probably trained on too many of those SEO hyper optimized medical sites people get when they search symptoms and such.

chinchilla2020• 3 years ago

It's a great bullshit generator.

I think upper management needs to be more scared of the implications than ICs

gwoolhurme3• 3 years ago

This is why my anxiety is at levels unseen before. I am a programmer who’s on a working visa. Assuming it does take my job, I have no idea what to do next… other than panic.

Vespasian4• 3 years ago

I decided that being paralyzed with fear isn't a great strategy so I'm going to subscribe to OpenAI tools for a while and see how they can fit into my workflow.

It feels "wrong" on a strangely emotional level but what is happening is gonna happen anyway.

On a brighter note: There are bound to be a lot of unfounded, hype-based Marketing claims that look reasonable during the rush but fall flat over time. It would be a first for humanity if there weren't

gwoolhurme3• 3 years ago

That is certainly true about the hype-machine it is in full force. I imagine that will level off soon and the really useful and novel stuff will float to the top. I have no backup plan but as you said, what happens is going to happen? It's a rollercoaster for me personally where some days I am in absolute dread and fear about the future. Then other days I just go along my day as usual.

matwood• 3 years ago

Panic seems like the wrong response. Figure out what you do that these things can’t. Use them to make yourself more productive.

In the late 90s I was told my job was going to be outsourced, then dotcom, then the GFC, and lots of smaller bumps along the way. Those were actually somewhat scary times. Now should be excitement about what’s possible.

gwoolhurme3• 3 years ago

While you are not wrong, the idea that gives me anxiety is that this replaces all programmers. My job theoretically is safe for now. I work on fairly novel stuff, while a lot of my day is json parsing there is a big chunk of it that is domain specific. That said... it's that one day it will just take over all of programming that has me in dread. The only option then in my somewhat middle age is to unironically "learn a trade" I guess?

fxj2• 3 years ago

In short: GPTs could do all the boring work for us and give us more time for things we like to do.

An example: Of course GPT can at some point make better music than me, but I am playing an instrument not because I want to sell the recordings, but because it is a lot of fun for me.

Let GPT do my taxes, then I have more time for playing. (can it do taxes finally?)

dathinab• 3 years ago

One major risk with this kind of AIs is that for the last few decades the mental image of what such AIs will be like (both by scientist and science fiction) has a _extreme serve_ mismatch with how they turned out.

In the past humanity people thought of them as this non-emotional purely logic driven beings and in turn problems people imagined would be such as this logic not considering emotions and emotional well being of humans or blindly (but using logic) pursing a specific goal no matter the consequences or not valuing freedom or gaining emotion. But in all that it still follows logic.

But now we have AIs which could have all the problems above _but doesn't use logical thinking_ to _archive goals_. Instead it uses complex overlapping _statistical models_ to _tell a believable story_ where believable is defined by the training data which is _widely inconsistent, wrong, misleading, discriminating, emotionally charged, etc._ because it's just scrapped from the internet. So there is _no systematic finding of goals, subgoals, plan etc_, there is _no logic_, the concept of "truth" simply _doesn't exist_ for such systems etc.

At the same time this turned out to be "often times" good enough to be usable for many task and can be convincing enough to make people believe that it's sentient.

But this also means it will retell common false information, misconceptions, discrimination, hatred etc. from the internet.

Similar it will do what people call "hallucination" and "lying" but it _not_ either of that and calling it that is misleading. Because it just doing _exactly_ what it was created for: Telling a "believable" story given the training data.

And gaslighting, misleading people and lying are a extremely deep ingrained part of the internet, i.e. a deep ingrained part of the data on which it bases what is "believable".

And while we can add tones of bandaid on top to try to hide/filter out such "bad" responses IMHO without fundamentally either changing the training data or the approach this is bound to fail while even stronger upholding a misleading illusion.

ebauch3• 3 years ago

Noam Chomsky, is that you?

tsone• 3 years ago

Does somebody know where to find the raw data of exposures on all O*NET occupations?

srg0• 3 years ago

So, researches paid by OpenAI are going to publish a paper to convince everyone else that an OpenAI product could have some economic impact.

Am I the only one who sees a conflict of interest here?

Is this preprint already submitted for peer review or published somewhere? Or is it just an advertisement formatted in LaTeX?

kristjansson3• 3 years ago

It’s building on a pretty serious thread of research, and demonstrates reasonable agreement with prior work (table 9) so I don’t think it can be dismissed as an advertisement.

Generating exposure ratings via GPT-4, from annotations provided by OpenAI people does definitely put a positive bias on the exposure estimates (which they acknowledge)

la64710• 3 years ago

It is important to start thinking along the lines of the government taxing corporations using LLM to increase their profit. This additional tax revenue should be used by the Government to give a universal basic income to people losing their livelihood to LLMs.

raducu2• 3 years ago

Anybody who actually read the paper and can share some insights in it instead of spewing opinions and anecdotes?

tracker14• 3 years ago

Given an emphasis on internalized documentation, these could be significantly better, especially over time than a lot of automated support in place for a lot of companies. As long as there is some kind of escape hatch for escalation.

That said, the reality is closer to the "Johnny Cab" from the first "Total Recall" movie.

360macky2• 3 years ago

Critical Thinking is the way

GPTs Are GPTs: An Early Look at the Labor Market Impact Potential of LLMs

Related Stories