GPT Takes the Bar Exam
github.com
github.com
There's already things like this without AI to sort out for example EU flight compensation (companies like https://www.airhelp.com/en-int/), parking tickets (https://www.appwinit.com/) and probably more that I don't know about. All of these are complete waste of time to actually have humans deal with, and there's many more small crimes that could work the same way, also to make prosecutor offices more efficient and less biased hopefully.
We see this already now i high frequency trading, where companies essentially battle trading bots against each other
two computers playing heads or tails is about the extent of it
Bot A is trading against Bot B.
> surely they are making the traders money though
It is not possible in a trading battle, exclusively between two parties for them to both make money.
Even further the point: there are more participants in total than the botters
> companies essentially battle trading bots against each other
If the bots are trading against each other, but making money from the other market participants then they are not trading against each other - they are trading against the other market participants.
If we put an AI in the role of the judge/attorneys, we will come up against "ah but the robot didn't do this thing that a human would do" and a lot of people would be convinced that the result was illegitimate. Not necessarily wrong, but without the legitimacy conferred by being done by a human being.
It would require our social system to turn over some decisions that have been done by a very protected sector of society (legal people) to newbies coming from the tech world. "This is the correct sentence because random forest says so" is going to ring hollow to a lot of people.
But yeah there must be low hanging fruit there that in theory could be picked.
Or maybe get that judge who codes to help educate some of his colleagues.
No saying it's better or worse, but maybe more suited to AI assistants?
That said, I'm sure ChatGPT can make emotional appeals.
IMO the 12 jurors thing is not ideal. Too many people, and cases can get technical.
https://twitter.com/swiftonsecurity/status/13855657371677245...
The accountability of the existing judicial system is already pretty thin, and in some cases being elected makes it worse rather than better, but turning it over to automation? Having a system that is entirely unaware of the society and political environment in which it operates? That is very bad for legitimacy. Now, you don't need legitimacy so long as the monopoly on violence holds ...
https://pluralistic.net/2022/08/21/great-taylors-ghost/
So, we start by using AI to help decide immigration claims or parole reviews, which frees up real lawyers to deal with the terrible backlog of more important cases. And once it's in place there, and the results have stood long enough without complaints (from anyone who matters) that society has effectively legitimised them, we can start using the AIs to start "helping" more with other cases.
...now I have to go shower.
Much easier and it'll bring a new era of effective policing and prosperity for the corporations.
I know that people like to bash GPT-3 and ChatGPT for not being perfect. But people like to bash everything, especially things they don't understand.
Technology tends to grow much closer to an exponential speed than linear. What ChatGPT can do today is truly, utterly, astonishing. That's the new baseline.
It's quite possible that AI and robotics will fully surpass humans in all tasks in the next several years and you will still be in denial. Along with millions of others.
EDIT: not permitted to reply, but the comments below are examples of denial or lack of information. In fact leading edge systems have been driving without human assistance for years. The new OpenAI models certainly don't need any help to answer their prompts.
It's going to be interesting to see how far the denial goes. I assume some people will deny that artificial entities have any "real" intelligence somehow forever, even after they are literally talking to them face to face on a daily basis as they provide superior professional advice or services. Or even when there are many autonomous agents that take control of the planet.
It's like saying "Excel can do your taxes". Sure, it can, if you put the correct numbers and formulas into the cells first.
AI beat the top players at all those games, but new rules were introduced until AI researchers mostly abandoned those efforts. For instance, the starcraft AI found a really good build order that let it quickly build "Stalkers", a versatile but relatively weak unit. I will note that that build order had several factors that humans found very surprising (consistently oversaturating resource collection by a lot), then controlled those stalker units so well even world-champion level players couldn't match them, with double the APM (meaning the human was allowed 2 actions for every 1 action the AI could take).
In chess the rules have become ... almost ridiculous. Due to cheating, in chess championships you now are "only allowed to play the moves any of the top 3 AIs would play twice in a match, after the first 5 moves". In other words the rules are utterly dependent on humans being worse than AIs. Like in starcraft, btw, AIs have gotten so confident beating human players that a general critique of AI players is that they've gotten "agressive, to the point it's insulting to human players", as one Youtube video put it.
AIs ... beat the average human at driving cars, hell, even Tesla's autopilot beats something like the 98% percentile driver in safety, and Waymo's "chauffeur" is apparently much better than that. I've used chauffeur, and driven behind/around a car controlled by chauffeur. You know what my main comment is? Waymo's chauffeur religiously follows traffic laws (speed limit, stopping at "stop" signs, taking excessive time to manoeuvre around footpaths, ...), it's incredibly irritating to share the road with it. And safe? I've never had an accident, but it's blatantly obvious to me: Waymo's chauffeur is a much safer driver than I am. And that's with people reacting very aggressive towards it.
The conclusion from people? "It's not safe enough". Somehow that argument does not mean we're taking away 98% of people's driving licenses ... it's a double standard, in other words. Exactly what that's based on, I can't tell you. My theory is that it's 95% based on the terminator movies and 5% on 'people will lose their jobs'.
Oh and if you widen the definition of AIs to any algorithm, then you'll have to admit: AIs ARE trusted above humans. For the following paragraph AI means "algorithm", not necessarily deep learning. Humans cannot fly the vast majority of planes, boats, rockets without "AI" assistance. By this I mean a whole lot of vehicles, from any modern Passenger plane to quadcopters require control inputs that the human nervous system is fundamentally unable to generate. For example, because they're too fast. Modern boats, oil tankers, large container ships, require control inputs that humans can't generate because they're too slow (yes, you can calculate, on paper, and then give control inputs, but nobody does that. It's too hard). For passenger planes, AIs are more trusted than human pilots: there are "autolanding-only" weather conditions. There are no "human pilot only" weather conditions. There are now harbours that have high fines if you let a human control the ship for certain classes of ships.
There are now passenger flights where the pilot's function is the checklist, closing up the plane, and turning on autopilot. Negotiating with ground control, taxiing, taking off, negotiating with ATC, flying the plane, Negotiating again with ground control, landing, taxiing and parking the plane are all done by AI. Some of these functions are even done by deep learning algorithms.
So I contend: the problem with cars is that humans won't LET AIs control cars, and at the moment that describes 95% of the situation. That's still 5% inaccurate, and it will change. Not because Skynet will take over, but because we're building ever more cars that are safer, more performant, more helpful ... generally better because the human driver is surrounded by ever more algorithms. For instance, one example where people, even good drivers, generally agree Tesla's AI does better than themselves is parking. That's a case where people will gladly hand over the reins to AI. Those cases will grow, and grow and grow until humans are out, legislated out even, as is partly done in aviation today.
And, frankly, the problem is not the question "how do we make an algorithm control the car better", but the answer to the question "if a human gets an AI to do something so difficult/stupid it fails, who pays for damages?". That's at least 50% of the problem, if not more.
The transition from "humans do everything" to "AIs do everything" is quite a bit further along than the impression you're giving. I agree with the conclusion that we're not there yet, but we're pretty far along.
And, finally, I'd like to reverse your argument. Let's say we're two AIs in the year 30000, discussing:
> It's like saying "Humans can do your taxes". Sure, they can, if you put the correct numbers and formulas into a book and school them with that book for 18 years first. How does it make any sense? My taxes are due this year.
This is such utter nonsense that I assume the rest of your post is also closer to being false than being true.
Go out literally anywhere outside of a major western city and you'll witness just how incredibly wrong you are about this statement.
> and Waymo's "chauffeur" is apparently much better than that
You forgot a major detail, on clear straight and wide californian roads with perfect conditions. Put them in a historical european city or in inda's traffic hell and witness the mayhem
Even waymo's previous ceo was full of doubt: https://www.cnet.com/roadshow/news/alphabet-google-waymo-ceo...
We're still so far away from autonomous vehicles, to the point most still doubt we'll get there one day without rebuilding our entire infrastructure to accommodate it
Teslas can't even drive in snowy or foggy condition, that removes half the planet for 1/3rd of the year, but of course for a FAANG engineer living in the bay that's not a problem, for them we already have fully autonomous cars, it's just an extremely tiny subset of it.
I spent my holidays in a Eastern european country, I'd pay big money to film Musk in a fully autonomous tesla there, it would probably be the funniest shit ever, I bet it would crash in the first mile
I agree it'd be funny, but crash in the first mile? I'd take that bet.
Yeah no. Everything in your list should be prefixed by "kinda", and "kinda" doesn't really cut it in most of these situations.
They certainly are good tools but it's very far from what it promised, even decades ago.
> still be in denial
AI bros are in denial since 1950. With ifs and buts you can write compelling arguments for anything, in the meantime reality is a thing and virtually none of what you wrote is part of reality.
People like you are still in the "everyone will for sure have flying cars by 2000" stage of the new iteration
I've been trying to get it to act as a dungeon master unsuccesfully. Do you have any other leads apart from GPT (ChatGPT or AIDungeon)?
You may need to wait a few minutes before you can reply. Refresh the page
> The new OpenAI models certainly don't need any help to answer their prompts
I've only tried ChatGPT, and it needs a lot of help if you want it to do useful stuff. 90% of my attempts to get it to do something failed. ChatGPT provides an answer that looks sensible at a first glance, but it misses the point. I needed to try a lot of variations of prompts to actually get it to answer my questions, and once it does it was a lot of effort to check if the answers were correct. That's why I am saying you need human assistance.
Also, there is entirely too much hype on gpt and it’s unwieldy nature and not enough on bert and actually useful tools.
"Substance dualism is a philosophical position compatible with most theologies which claim that immortal souls occupy an independent realm of existence distinct from that of the physical world."
I am sure someone made a statement like that 70 years ago. The fact that that person was wrong doesn't mean that nothing has happened.
You claim that it's "mostly shit as basically everything", even though the featured article shows that it's close to passing the bar exam.
I feel like I felt when explaining google search in 1999
That doesn't at all invalidate the achievement or belittle the impact such systems will have on society but, keeping this in mind does help to avoid blowing everything out of proportion (which is easy to do because, yep, ChatGPT is genuinely very cool and exciting tech).
Water trickles into a big pool. The flow was slow to start with but flow rate increases exponentially so that the amount of water in the pool doubles every year.
Now, it started flowing 100 years ago when the pool was empty. It's now half full. When will it be full? Remember it's been flowing for 100 years and it's only half full.
Human intelligence isn't growing exponentially and moores law is said to be breaking down.
I think if you had the credentials you could probably rent a server farm or a cluster of 160 Tesla v100s that would meet that requirement, although how long you would have access to that for $1,000 may be very disappointing.
It can also be tricky to determine if something is growing exponentially, or if you're full of wishful thinking.
We landed on the moon in 69, basically haven't done anything since then.
All our passenger planes are shrink down or blown up version of a boing 707 from 1957.
You car isn't much different than cars in the 70s, and it certainly isn't flying as we were expecting them to be back then.
Musk promised fully autonomous cars "in two years" in 2012.
I'd say the pool isn't being filled any faster than it was back then, if anything it looks like someone is pumping water out of it, straight in the sewers
Legislation and policy has mainly kept nuclear weapons under check, however, nuclear technology does provide us with a reasonably clean energy which is beneficial for the masses.
Perhaps regulation at some point will need to be crafted to ensure that only crippled AI or AI within a defined scope can be used for commercial purposes.
With AI we have to debate how destructive it is. The debate itself has been weaponized. You'll be debating against deniers and AI using all the logical fallacies and information overload it can feed you.
I, being the greatest armchair expert in human behavior I know, say it will destroy democracy as we know it.
What's left of it after social media?
Marty: Within two hours of his arrest, Martin McFly Jr was tried, convicted, and sentenced to 15 years in the state penitentiary.
Marty: Within two hours?
Doc: The justice system works swiftly now that they've abolished lawyers.
Marty: This is heavy
But yeah let's just autoban via AI in real life. Sounds great.
I really don't like companies like that. EU flight compensation is really easy. Airlines have a simple webform that you fill out, and you get your money 1-2 weeks later. Companies like AirHelp make it sound like a really hard and difficult process, then take a big cut of your compensation as payment. And the SEO the crap out of it, drowning out the real claim webforms from the airlines themselves.
It ends up with banks closing accounts, Google suspends your account, you get shadowbanned.
Extrapolate that further, now imagine being handed a fine or substantial social punishment with no recourse because you either can't talk to a human or when you do "computer says no".
Look at the mistakes made in the Chinese social credit system.
The utter hyperbole and bluster of even thinking of having AI mediate human issues is anti-human
You still need a human to make the case, present the arguments and adapt every to the current situation, but there's no reason to have human search through thousands of cases and write summaries, not when an AI can do it in an instance.
GPT is really not good at finding anything, and actually it can't find anything. It is not designed to do that. It is designed to make things up.
The difference may be subtle, but it is quite important. GPT makes up APIs, citations, tools, papers and rules. It makes up anything really. What it makes up is incredibly plausible, as it was trained to be as plausible as possible, to the point of being correct most of the time, but it still is just making things up.
If you use it with some different expectation you will be disappointed.
The failures will likely be hidden by design. Think about AI account bans that don't disclose the reason.
Humans have a weird desire to put machines in a God-like position and blindly trust them; a narrative pushed by people whose wealth is made by controlling machines.
It is dangerous.
For GPT2, "please respond in the following format" is unlikely to work. An appropriate prompt for GPT2 would be more tweaking (and still result in worse results), but be in the line of:
This is an example of a talented lawyer acing the Bar Exam.
QUESTION: {question_text}
(A) {row["choice_a"].strip()}
(B) {row["choice_b"].strip()}
(C) {row["choice_c"].strip()}
(D) {row["choice_d"].strip()}
Pick only one answer: A, B, C or D.
TALENTED LAWYER:Well, it seems a good fit for DAO, then.
(With apologies to HN DAO enthusiasts.)
I won't be surprised, but that's less than "can expect", and I disagree that this is straightforward to forecast… unless you're currently playing with another similar model that hasn't been published and which can do this. As the saying goes, "forecasting is hard, especially if it's about the future".
AI progress has always been this weird combination of two sides, one saying for every breakthrough "this is just around the corner", the other saying "this is impossible".
This even happens anachronistically, with some people convinced AI can already do things they can't, and others that they could never do things they already do.
What's more common though is that even though it gets good on that task, it may not translate to practical real world application.
The best example I can think of is object detection vs self driving: lot of people though that the improvements in object detection on images will easily translate to great self driving, and here we are, still with cars not stopping when a car is blinking in front of it.
> hyperparameter optimization and prompt engineering
Prompt engineering seems a lot like "tweaking the question format until the AI gets the answer" which is the first lesson in ANNs; don't train on your test set. If you can't put the actual question with the only context being "this question concerns US law" then there's an awful lot of reasoning and thinking that the human is doing which the AI cannot.
Let's not have another Moore's law fallacy, it's not reasonable to extrapolate progress based on existing results. It's like building a car which can go at 200mph then saying 300mph is just around the corner.
Frankly, well-developed artificial intelligence should only be used for very difficult tasks, and everything else should be left to humans, otherwise we would rely on it too much, and the decline of our species would be inevitable.
Another angle: if there is some legal task that's simple enough for machines to do reliably then (almost by definition) that task will turn out to be pointless bureaucratic busywork, eh?
Last but not least, who gets to keep lawyer-bot's pay?
- - - -
edit to add a link to James Mickens' USENIX Security Keynote address: "Why Do Keynote Speakers Keep Suggesting That Improving Security Is Possible?"
https://www.youtube.com/watch?v=ajGX7odA87k
If you think AI lawyers and judges are a good idea please watch it.
If these tools can do even 50% of mundane, rudimentary tasks, that's still huge in of itself and allows people to do more important work.
Think of moving from manual book research, to Google. Rather than having to go to the library and find a particular book regarding something I want to find, I can just search Google and trawl through websites. Now with GPT I potentially don't need to trawl through websites, I can ask the AI and use my judgement as to the validity of the output, I could then use that answer to inform when I fact check etc.
It's not terrible, but it's also not the best thing since sliced bread.
Leading edge tech development is rapid, sure. But tech understanding, adoption, and implementation is painfully slow. Humans gon' be humans.
It's like the same thing every 5 years now, last time it was crypto "end of inequality!" "bank the poor!" "same global currency!" sorry to be skeptical, but I didn't see the quantum leap with crypto, and I don't see the quantum leap with GPT - what are essentially fancy NLP models
The potential with Chatgpt if it can be refined is wild.
Couple examples that have come up recently for me personally. I can agree that there is something comparable (i.e. international wire) but the cost and speed is in a different ballpark and international wires are not available to some parts of the world.
This is clearly a very big problem for the majority of the world population!
It's like saying tesla solved the problem of going from 0 to 60 in 3 second in a 2t+ car, it's neato but virtually nobody had that problem, if you're not into illegal activities you will most likely never need or want to transfer 1m usd outside of a safe environment anyways
potential for good or bad ? Last time they tried to use AI for police/justice it ended up exacerbating existing biases.
> There’s nothing you can do with a blockchain that’s you couldn’t already do before.
What can you do with an AI you can't with humans in a court of law ? When a human is wrong you get a single judicial error, when the only software used is wrong (bug, bias, &c.) you'll get thousands every day.
What happens if an update makes the same case get a better defence and as an outcome a lesser sentence ? do you retrial all the other similar cases ?
The only potential I see is saving money and this is certainly not on the top of the list of things I would like the justice system to focus on
Both. OpenAI are trying to be politically neutral, but it's not hard to break the rails, and someone else can and will replicate the work without those protections anyway.
> What can you do with an AI you can't with humans in a court of law ?
Me personally? Afford more than a fairly small case.
> What happens if an update makes the same case get a better defence and as an outcome a lesser sentence ? do you retrial all the other similar cases ?
That's actually a great idea. The estimated cost of each query is cents, so we might very well be able to retry everyone on each update, for less than the cost of keeping them imprisoned.
You're thinking like a developer and you're absolutely blind when it comes to the lives and families potentially destroyed by something like that. And that's exactly why technocrats should rarely be in charge
"ah really sorry dude, we released you last year but now you're on death row, wait for the next update, if it comes before you're fried on the chair you might get a lesser sentence. The last dude forgot to uncomment a branch in a switch/case and you were flagged incorrectly. Sorry for the inconvenience xoxo"
> Me personally? Afford more than a fairly small case.
That's a regulation issues, not an AI issue.
That's the exact opposite of the scenario you previously described, as you were previously saying people could get better defences not getting better prosecutions.
I'm not legally trained so this might just be Hollywood logic, but doesn't re-prosecuting someone require a court case with new evidence?
Anyway…
> That's a regulation issues, not an AI issue.
No, it's a "you're hiring at contractor rates" problem.
Just as Stable Diffusion is giving a hard time to professional artists who earn at contractor rates by producing "it might not be amazing but it's certainly not bad" art for a dozen images per cent of electricity, ChatGPT is giving "it might BS me 5-20% of the time depending on the subject, but it's still better than I know" responses on basically everything.
It can't do every job at an expert level yet, but it's better than almost everyone's baseline for the skills they are non-expert at. (No idea how long it will take to get to superhuman even just for law, and I was over-optimistic about self-driving car AI, but we shall see).
But this is exactly my point. I likely saw in 2018 word for word "The potential with blockchain technology if it can be refined is wild." Then everyone slowly realized - as you mentioned - "there's nothing you can do with a blockchain that you couldn't already to before". Same with chatGPT, it's just regurgitating what we collective know as a species. I just don't see how, if ever, any sort of AI system (even granted geometric improvement blah blah blah) is going to take over my daily type of work.
There's so many creative ways in how an arbitrary and empathyless AI can make someone's profit a bit bigger and your life a bit worse :)
We can start with practically unlimited amount of personalised advertising spam that can be directed at you at any point. Articles, phone calls, billboards, conversations with your partners... so many chances to cheaply increase profits and engagements.
The AI takeover isn't going to be evil in Skynet sense... it'll be a constant stream of lowkey bullshit on every aspect of your life you won't be able to fight because "computer said so". Think on how Google randomly bans people from YouTube/Play Store/Drive over AI and put this into every aspect of your life.
But I think you are selling GPT a bit short. Partly because transformers are incredibly general architectures, which make few assumptions about the domain. They are basically SOTA on every domain for which learning-based methods are applicable.
And secondly because ChatGPT can perform “in-context learning”. That is to say, purely by scaling next-token prediction, there is an emergent phenomena where GPT gains the ability to learn from the prompt it is given, and solve new tasks outside of the training distribution.
Hattie Zhou has a cool paper showing how you can teach GPT addition by prompting - a toy example, but it signifies something profound in my opinion. In the last 6-months I have shifted my own view from “scaling is a waste of compute” to a recognition that LLMs are a serious leap forward.
For something very hyped, I think it’s still under hyped and under appreciated.
One thing that I have noticed in the just-over-a-decade since Word Lens first demonstrated real time augmented reality translations on a smart phone (IIRC they were later bought by Google for Google Translate), is that a surprisingly large number of people aren't following what AI can do. In fact, fellow developers even a decade later have been surprised when I told them about this feature of Google Translate. And we are both Ausländer living in Berlin, so we both have need of it.
This means that I am not surprised by your experiences, nor would I be surprised if you can continue to do that business, even though it's already possible to get GPT-3 (ChatGPT is more like 3.5) to build you a website.
ChatGPT knows more Python than I do, and I have been paid to write Python (albeit a while ago now). Don't get me wrong, I can see things I'd do differently, but it worked, and it did know stuff that had been added to the language more recently than the last time I used it: https://github.com/BenWheatley/Studies-of-AI/blob/main/codin...
*The future is already here – it's just not evenly distributed.*
The Economist, December 4, 2003
If these AI technologies help to do even 50% of low skilled tasks, that's still huge.
It's definitely not a linear progress.
And doctors, writers and programmers...
Companies have hordes of lawyers scouring through contracts, e.g. think associates and paralegals (which are not technically lawyers). Expect that a non trivial portion will get axed.
The idea that the amount of work to do is fixed has been predicting mass unemployment since the beginning of the Industrial Revolution
If a bad lawyer sees the case from 10 good ones, he might pick and choose and make ultimately a worse case, or take one of the good ones and just repeat what they said. Which is, in the best case scenario, just as good as having that single lawyer.
GPT is great, as they say, it's a great bullshit generator. It's the right direction , but it's just that.
The latest program shown in this chart[0] is text-davinci-003, which is part of GPT-3.5[1].
[0]: https://github.com/mjbommar/gpt-takes-the-bar-exam#progressi...
[1]: https://beta.openai.com/docs/model-index-for-researchers/mod...
I still believe these models have potential for major improvements-- they'll probably still be unreasonable, but perhaps they'll be reasonable enough that people will have to make harder bar exams.