It's a bloody calculator and a damn good one (but it can and does go mad). Use it like that and you'll be fine.

Have you actually engaged with it in any way? Its free to sign up.

By the way, what on earth is an AI booster? Do you really mean "enthusiast"?

"My mental model seems to clash with that of the boosters." Please tell us about your mental model, we can't read your mind.

It is very nearly the opposite to a calculator.

A calculator gives you exact results for a very small subset of problems (equations).

Chat-GPT will give you approximate (and occasionally incorrect) solutions to an extremely wide range of problems.

The main problem with AI is that there are only so many problems where an approximate solution is fine, and an incorrect solution is easy to work around. In order for AI to be broadly useful, it has to get to the point where it's correct more often than an expert human. It may well get there within the next couple of years, but this isn't a space LLMs are designed to be able to solve so it's not as guaranteed as a lot of people seem to think.

Disagree. For AI to be useful it needs to be better at the job than how the job currently being done.

We humans rely on thousands of things where an approximate solution is fine. Especially when it's about interacting with other humans.

Programming is an exception, but what % of the population are programmers and engineers?

Programmers, engineers, accountants, lawyers, doctors, nurses, builders, farmers, manufacturers, drivers (including pilots and ships captains), and scientists all need to be very precise in their work and mistakes will cause deaths (I have almost certainly forgotten some). This may not be the majority of workers by number, but these are the jobs civilization is _built_ on, they are the jobs that need to keep going no matter what or the whole house of cards collapses.

The majority of jobs that require less precision falls under the umbrella of artists (I will include advertisers here), people managers, and service workers.

I would put cooks into that too, allergic reactions can kill, badly cooked food can kill, and in fact people can be penalized for it too.

Ture but does it matter if the chicken is cooked for 10 minutes, 10 minutes 5 seconds, 10 minutes 10 seconds,... programming requires far more precision.

It does matter if you put some peanut in when the customer is allergic to peanuts.

That's the kind of thing you can solve with deterministic algorithm guardrails though.

In order for a system to be useful it doesn't need to match an expert human, it's sufficient if it can do most of the job. If companies can replace an entry-level person with something that's slightly worse but much cheaper, they're going to be widely used. In any profession where you have some sort of supervisor handing off tasks to unskilled humans and verifying their work, it will be broadly useful to hand those tasks to some AI model instead even if it needs a bit more handholding than humans, since it's going to be much cheaper than those humans and that is what makes something economically useful.

A "booster" is one who promotes something, not just an enthusiast. See "civic booster" or "athletic booster" for similar uses.

"Booster" can also now mean technologists who have correctly predicted the number and size of executive erections once teased with the idea of automating most of their office. Put on something black and tight, go up on stage and twirl around that notion until balled-up Jacksons grease the creaking floorboards. You know you'll be riding in a Cayenne tonight. Maybe a Boxster.

It actually sounds pretty familiar for those of us in the side shows for other Miracle Technologies That'll Fix Your Business, but it's a hell of a lot easier to set up a convincing model than something like SAP. Also, unlike ERP, GPT is a real thing, and it does real stuff, so that will make it easier to take the money. Real science is happening, which is always more fun, even if Homo Executivus uses it as a marital aid. I guess this makes me a Booster, but one of those old wrinkled Boosters who sounds like a carton of Newports.

Upvoted for being the kind of comment AI can’t yet write.

> a half-working system may be tremendously exciting to AI enthusiasts, but it's not that useful to society.

A half-working system might actually be worse than useless if people believe it 100% of the time.

There’s a simple solution to that problem.

And a 98% working system even worse due to it being treated as 100%.

Well that depends. If a hypothetical emergency room doctor is right about patient diagnoses 90% of the time, then treating the even more hypothetical 98%-correct AI diagnoses as 100% will still probably give you better outcomes.

We're still far off from that, and I'm skeptical we'll get there anytime soon, but humans aren't infallible either.

The real problem is how humans react. Imagine a doc comes in, and is very wrong about a diagnosis.

Say... you have a broken arm, doc says it is sprained. You say No! and point out why.

AI says this, you say no, and the minumum wage person who works with the AI, shrugs and says the computer says so.

Can you give a single example of a "system" that works 100% well, 100% of the time?

This is the self driving car argument applied to things it doesn’t fit. Sure, unless a self driving car is several 9’s reliable, it’s not terribly useful for society. That’s because there’s no way for a human to be safely in the loop at highway speed.

But a LLM assisting a wide rage of professionals at even a low accuracy could be part of an efficient workflow, allowing for be doctor or lawyer to give quality service to more people.

The author touches on this in the very article linked. His point is that the low accuracy answers provided by ChatGPT risks reinforcing the doctor's biases and then not thinking outside the box. In other words it risks the doctor blindly trusting the AI too much.

The inherent problem with chat bots is you can't investigate why they're telling you something. When I search Google, I'm critically evaluating the information, and it's frustrating when the results it gives are low-accuracy. If I ask a chat bot a question, what would I do with the answer if it might be wrong? There's no context for me to evaluate.

You can ask it to explain its reasoning. And it's even more useful if you ask it to show its reasoning when you ask a question, as it improves the answer.

True. Though, I said investigate, not interrogate. In the context of the pre-bias example I was responding to, it's fine to integrate it into a search engine to assist interpreting queries and improving search results, but it shouldn't be talking back at you in the first place. The machine manipulation of information is inappropriate.

I my mind, the AI only needs to be more reliable than the human expert, and doctors and lawyers and other experts are "only human".

There is only so much a human can read and remember. There is only so much they can be convinced to change their mind in the face of new evidence. They get tired. They get cranky. They have biases, and fears.

Some doctors are just terrible. I think the AI is probably more reliable than them already.

I don't think we are very far away from being able to correct an AI, and for it to remember that correction forever.

I totally agree, but we're going to need to figure out better ways to verify correctness of what the AI tells us. This will be pretty hard since we can't hardly do that with human-produced stuff either. And further if the AI is trained on human-produced stuff that contains lots of errors... I don't really know how we're going to do it.

If I were to hazard a guess, I would say, at first, companies are going to pay human experts to sit down with AI, ask it questions, and correct any mistakes.

It might look like a conversation where a panel of human experts and the AI agree on whats true and whats not. Like scientific consensus.

Eventually the AI will just run its own experiments and test its own hypothesisis.

I suspect an ER doctor who misdiagnosed patients 50% of the time would not be an ER doctor for very long.

Also consider that, in this case, the 50% the AI got right was defined such somewhat generously. For each patient, the AI presented several possible diagnoses, only one of which was correct. If the AI were making all the decisions, would the AI also decide to treat the correct ailment in all those cases? In an emergency situation, you may not have the opportunity to go back and try other options if you're wrong the first time.

But it's also a ridiculous assessment. How much training material for ER doctors was in the training set? I wonder what my rate would be (having never done a single course in medicine)... It should be totally unsurprising that it produced bad results. That it got as much right is an indication there's a good chance you can do reasonably well by training a model with a lot of actual case studies.

Even providing it with a tool to search and load its context with relevant training textbooks / papers.

Apparently much of the usefulness of ChatGPT comes from human guidance via reinforcement learning. The base model seems to be quite useless, but with only a little bit of human input it becomes deceptively smart.

The key purpose for the reinforcement learning part (RLHF) is so that it would be socially acceptable to provide the model to the general public without getting into a PR nightmare like Microsoft Tay; that post-training does not make it "deceptively smart", it trades off a bit of smartness in order to ensure some alignment with certain restrictions. It decreases the performance on some tasks and e.g. the GPT-4 paper, which is very light on other details, provides some experimental evidence that this post-training significantly hurts the model's confidence calibration, which decreases its usability for tasks where you want to know how certain the model is.

The goalposts people usually talk about were just the relatively low expectations of the past - low at least by the standards set by ChatGPT.

Before ChatGPT there wasn't anything quite like it, so I think that the goalposts were mostly about passing the Turing test, or at least "not sucking" after 4 or 5 phrases. The leap from whatever existed before it was enormous, and I think comparable to the first iPhone versus, let's say, the Blackberry or Windows CE. Maybe even more than that. Go back to the first iPhone, it had a lot of issues but it offered so much as it was that people wouldn't look back.

It's too much to expect ChatGPT 3 to be better an experienced emergency doctor or neurosurgeon. But the fact that it can be a competent assistant to a doctor instead of a toy assistant - that by itself is amazing.

> ...I think that the goalposts were mostly about passing the Turing test...

I mean, if that was ever the goalpost, ChatGPT does not clear it: the Turing test isn't about merely tricking someone who isn't paying attention, or making someone seriously doubt their belief system... it involved a careful discussion with a chat agent by someone who actively was trying to determine whether or not it/they were human. If you sit down with ChatGPT--even 4--and start discussing complex topics about the world with it, you absolutely can quickly get it starting to adamantly defend things about the world which a human would never assert... and, even more damning, if you just ask it "are you human?" it will reply "no", because we actively told it to fail the Turing test.

If we instead merely re-define Turing's test to be "it is able to converse with a human and manages to occasionally trick a human into thinking it is also human" then I had a friend (Roy Keene) who managed to pull that off years ago with a chatbot that was literally just "every time someone messages it, wait a small random period of time and then reply with a random response from the following set: Yes, No, Maybe, I dunno, :), :(, ;P (and like one other thing I forget)". He had it set up to reply when he was not available and when I finally caught him (I asked about a task I gave him the day before which he claimed to have no memory of, so he admitted to me he had this bot) he gave me a log file and OMG I had had long conversations about personal topics with this thing and had never noticed... but only because I wasn't trying.

Sounds more like you failed the Turing test rather than your friends chatbot passing it :)

The point was that people mis-use the term "Turing test" as "manages to trick people sometimes into thinking its a human". To quote the description from Wikipedia:

> Turing proposed that a human evaluator would judge natural language conversations between a human and a machine designed to generate human-like responses. The evaluator would be aware that one of the two partners in conversation was a machine, and all participants would be separated from one another. The conversation would be limited to a text-only channel, such as a computer keyboard and screen, so the result would not depend on the machine's ability to render words as speech. If the evaluator could not reliably tell the machine from the human, the machine would be said to have passed the test. The test results would not depend on the machine's ability to give correct answers to questions, only on how closely its answers resembled those a human would give.

ChatGPT would not manage to pass an actual Turing test anymore than Roy's auto-responder; and, while we might get there frighteningly soon, that goalpost stands.

Turing was more far-seeing than many of his successors - he set up the first goalposts, and the first moving occurred in the late 1970s or 80s when crude hand-crafted rules engines were dubbed 'AI'. It was so bad that we had to invent a new term, AGI, to recover Turing's vision.

A half-working system is cute and useful when it recommends a pizza place you might like. But the same system is a menace to society when you tell it to drive a car or prescribe a drug to treat an illness or decide a person's prison sentence or protect a perimeter with lethal force.

It's absolutely amazing that so many AI scientists don't seem to understand this distinction.

A half working system is good enough for real life tasks when combined and used by and in conjunction with attentive half working human collaborators.

What does it mean to be attentive and half working?

You can be attentive while not giving your 100% to the task. For example while writing a piece of code you might know what you want to accomplish but you aren't straining your brain to type out the actual code. So you type the name of the method and GitHub Copilot gives you an implementation of that method.

Have you never corrected your phones autocorrect before? That's what they're referring to.

Translators who've been picking up after machine translation for the last decade will tell you that it is not, in fact, useful to them. They get paid half as much as before, but correcting the machine output is as much work as translating from scratch, and sometimes more, since the machine can give results that look good at first glance but are fundamentally incorrect.

I don't know much about professional translating, but wouldn't you be able to track the error rate of the machine translation by looking at the number of corrections? And then translators could use this as leverage to negotiate better pay if it's obvious they need are translating from scratch in, say, 50% of cases anyway.

Pay isn't really decided by how much work you do or even completely by how much they need you to do it. What you're describing is a reason to ask for more but it's not leverage to make them give it to you.

I think of machine translation as a tool for people who are not translators, but admins, writers and so on, but need cheap translations that are good enough.

I'm a translator, and that might be true for Google Translate, but not GPT-4. GPT-4 produces translations that are generally accurate, and when instructed to copyedit its own work, well-written. In most cases, all I need to do is fix the occasional error and do some light copyediting. This is much less work than translating from scratch and results in superior quality, since it combines my own skills with the AI's.

> It is also useful for researching equivalents to obscure terms across languages.

I'm super curious to learn how you use AI to do that. Thanks!

"They get paid half as much as before" equals "this technology has helped the customers of translators to get the same thing done for half the cost".

To be clear, a lot of them aren't "AI scientists," just hypemen, many of them who came into the "tech" world chasing crypto.

Came here to say this. Just like with crypto.

What about helping correct the grammar and style of an article? Writing out boring code? Creating a test for code you already have? These are all tasks that have important economic consequences. AI is not half-bad at them, it is actually approaching human levels of performance.

If we continue to improve AI, maybe one day its performance in these critical tasks you describe will also surpass humans.

Is it actually approaching human levels of performance, or are we continually reducing the commercially acceptable skill level of all of these things?

I don't think it's anything approaching human-level performance unless the humans you're hiring are inappropriate for the role in the first place.

Just wait a few years.

Today: "Senior" devs by virtue of knowing the cookbook of, at most, one stack.

2028: "Senior" devs by knowing when GPT-12 output POST when it should have been PUT.

The subject at hand is today, not tomorrow.

Two points make a trend.

Many, many AI scientists are very clearly aware of this problem. AI safety researchers have been predicting exactly the failure modes we see in the various incarnations of GPT for years. Issues like alignment and over-fitting are incredibly hard problems to solve, and there's still a lot that isn't known, but they are recognised in the field. The problem is business leaders, investors and governments are utterly clueless on any of this, hence the Google Bard and Microsoft Sydney debacles when they were rushed into the public view.

There is also the risk that the recent massive success of LLMs will bring a lot of excited bandwagon seekers into the field that don't have or appreciate the background on AI safety.

As with all safety rules, those for AI will be written in blood. That blood just hasn't been spilled yet. It will take the eventual benefit of hindsight to look back at the current era and estimate how much danger we are putting ourselves in.

I’d take a half-working system over what we have now, which is a system I’d put at around a third-working.

People pushing the curve get to the middle of the sigmoid and want to believe it's gonna be all vertical now. People observing from the outside heavily discount the hype because they know limits will be reached soon. It happens every time.

Not sure why the outside view is so wise? Predicting the future in complex systems is hard, regardless of your prediction.

It's unwise to hype up technology for profit. Those who can see it should point it out. Forecasting is hard and pricing is an useful signalling system for allocating resources socially. Hype creates noise to seek rent.

The half-working AI is like 95% of the progress to a better-than-the-best-human AI.

And then what? Do you just think that humans are just going to cede that part of existence to that AI, in a fit of self-fulfilling prophecy? Cutting off the source of novel training data, thereby stagnating any progress of humanity at "we made a 1% better than the best human on a bad day AI once"?

This is the central conceit of AI maximalists; and tech maximalists as a whole imo. Someone wants something that fits in a box that everyone else must eventually delegate to. Given capitalism's focus on creating owned systems, and the tendency to centralize and maintain asymmetry of access to rent extract, this is a most concerning direction to keep pushing development in.

We should be building better people; but we've abandoned that in favor of making a better mantrap.

A half-working system that hasn't been trained specifically for this purpose is a strong indicator that a system that can handle this use-case far better is within reach. That is what is exciting about it. I'm constantly mystified by those who expect a system trained only on publicly accessible data pulled from the web to be an expert in everything rather than seeing it as an indication of what can be possible if enough specialised training data is added.

As an AI booster that falls into the regular folks view -- YES! THIS exactly!

People will use and fund AI tools that have usefulness to them.

AI boosters love to see the capabilities because it unlocks a whole new verdant valley in the continent of useful ideas, but until a street is laid down and a general store is available it remains the domain of early adopters.

LLMs are genuinely impressive. They aren't yet AGI, but they are fantastic tools. More exciting to me as a practitioner is the Segment Anything foundational Model (SAM) since I can put that into immediate use for computer vision in a variety of domains. LLMs still require some baking for broad usefulness (though there are many places they are already useful, its not quite broad enough).

When folks use AI focused on augmenting capabilities rather than replacing people, you get a great return. Until AGI is clear (and we'll know, the AI will tell us without being prompted I reckon), let's treat AI as an exoskeleton rather than a new species.