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.

Imagine a life or death situation where a calculation held the key. Would you rather use a calculator that is right 50% of the time or 99% of the time. The obvious choice highlights that the utility is not exactly zero. But I see your point.

>I don't want to have a calculator (or say bookkeeping software) that gives correct results most of the time but not always

And yet Excel is widely popular, despite several footguns and inaccurate calculations!

Excel has consistent (I guess coherent) flaws/bugs, unlike GPT.

Don’t rely on floating point accuracy!

Those are predictably correct and incorrept, in a very formal way!

So what about large asteroid strike levels of improbability? I have to reject your premise that there is no probability where you can stop worrying about the problem.

(This might be a moot point, because I'm not sure current methods can ever get to this level of accuracy, due to limitations of the training data. Needs an entirely new method or a clever insight to optimize for truthfulness with unclean training data, and InstructGPT hasn't made much progress on this, and it might not even be possible)

Confining an LLM to the very narrow domain of "calculators" is a mistake, I think.

You wouldn't say "a programmer that is 99% correct is worthless, I need 100%". I'm pushing it, but for a more fair comparison I'd say measure it against a programmer. How often are we wrong? 75% of the time? :) being generous here. It's the tools that make us productive.

I don't know about you specifically, but I don't think you'll be very productive with a bare terminal lacking any modern IDE-like or even REPL facilities. I'll ask you to come up with instantly working code every time, all the time. It doesn't work like that. You need iteration and I believe these kinds of AI have the same issues as us. There are wrong sometimes (often) and need feedback.

> You need iteration and I believe these kinds of AI have the same issues as us.

It's funny how we resort to humanizing the machines when their results are inaccurate. We don't do that with the calculator, because it's expected to be 100% bug free. When there's a bug in the calculator code we expect it to be fixed, not gradually improved.

Speaking of bugs: mistakes in code is one thing, wrong output because of a fundamental flaw in the algorithm is another. The statistical machines we are dealing with work as intended, or at least the wrong output the top comment here brings up is not a bug, it's a feature. That's the difference.

That's (IMO) too narrow view of what a "machine" is. Complex machinery of any kind never is 100% correct and needs constant correction and maintenance. I still think approaching this as a "calculator" is awkward at best.

> Complex machinery of any kind never is 100% correct and needs constant correction and maintenance

Computers are extremely close to 100%, we generally expect a CPU to never make errors even after years of working. If it starts making any errors at all we throw it away and make a new one.

>Computers are extremely close to 100%

We must work in extremely different industries!

Do you code in checks to check the calculations made by the CPU? I've never ever seen anyone do that. If a CPU starts making errors we throw it away. A typical CPU will make many quadrillions of correct calculations before its first error, I'd say that is basically 0 errors.

This is a very weird statement that's failing based on logical category.

My computer will pretty much add 1+1 correctly forever never making a mistake.

My computer will perform an 'error' every time I put bad code into it, and some of those logic chains and error conditions are not very obvious.

The issue here is you think the LLM is performing a category 1 error, when the problem we are seeing is a much more human like category 2 error.

Literally LLMs get much better with chain of thought, feedback, and/or consensus.

Gpt-3 performance on MultiArith goes from 18% to 92% with all three. This isn't some hackneyed anthropomizing. Countless research papers showing massive improvement with these processes.

Is the typing the hard part? I’ll look up libraries and apis, pretty regularly. maybe an algorithm every few years.

Figuring out what’s wanted from me takes forever though.

He is comparing it to a calculator and CHatGPT doesn’t measure up in some aspects.

The good things about reliable tools is you can offload the cognitive burden onto them and know they won’t screw you over.

Almost every single post here about using ChatGPT mentions checking through its output. People don’t check though the output of their calculators.

Surely this is "just" a matter of teaching the LLM to recognize this is a job for say Wolfram Aplha and generate a query to it, then feed the response back to you?

I suspect that humans have an accuracy lower than 99.999%, and are similarly capable of producing confidently incorrect results.

GPT has a lot of hype and hysteria around it, but demanding 100% accuracy from it is a bit over the top imo. It doesn't need to have 100% accuracy on any arbitrary prompt in order to be a useful and valuable tool.

"It doesn't need to have 100% accuracy on any arbitrary prompt in order to be a useful and valuable tool."

Yet anything less than represents a danger. Never underestimate stupid.

Especially don't underestimate automated stupid

Yes. WE have a higher standard for computers. But humans produce a ton of bull shit answers. Are we trying to produce an 'answer machine' or a 'human mimic'. Because the bull shit makes it more human. And a wrong answer does not mean it's broken, what goes on in its neural net to produce a confident wrong answer might be similar to our own.

have you ever got on an airplane? are you aware that they are not 100% safe?

airplanes do not execute in a vacuum eg: computer. There are external factors that might cause an airplane to fail. A fully working airplane all things equal in terms of weather and a good pilot, will be safe.

you're missing the point. my point was: if everyone waited til something is 100% safe/complete/satisfaction guaranteed, there would be no progress in human civilization. we need to be willing to take a chance on a promising technology in order to allow for further progress. imagine if everyone said i'll wait for planes to be 100% safety guaranteed before i'll get on one - we wouldn't be flying today.

Do you have colleagues, bosses or reports that are correct 100% or the time? 99.999%? I would love colleagues that are 99% accurate, I certainly am not unfortunately.

> Do you have colleagues, bosses or reports that are correct 100% or the time? 99.999%? I would love colleagues that are 99% accurate, I certainly am not unfortunately.

I see this analogy all the time in these comment sections but it's not a very good one. A person is not a tool. One of the great achievements of humanity, and in computing in particular, is that we make tools that are more accurate than we are.

I expect a hammer to deliver a hard forward blow 100% of the time. If one out of one hundred times it delivers a hard backward blow, I cannot use it on a job site due to risk of injury to the user. The same is true of a calculator being used for financial transactions. And the same is true of a LLM that would be used for drug interactions as discussed in this thread. We already have 100% accurate ways of pulling data from a drug database—it's called SQL. A tool that is not as accurate is in at least some ways a step backward, even if it's easier to use due to its natural language interface.

Except hammers already do not work as expected 100% of the time, as evident by them painfully hitting the hands of the workers in mishaps. Yet we still use them.

How is a hammer "not working as expected" if you hit your hand with it? This makes no sense. If I hit my hand with a hammer, what am I supposed to expect except causing pain and injury?

If I try to light a cigarette and accidentally set my beard on fire due to my own clumsyness, did the lighter malfunction? No, the lighter did exactly what it was supposed to do, it was my hand that didn't do what I expected.

If you want to continue with semantics, the worker did not expect to hit his thumb until the hammer bounced off of it, hence the tool did not work as expected. Until regular hammers are able to put nails into wood by themselves, we are talking about the success of the complete system that includes both the hammer and its user.

Either way, when working with humans we already deal with plenty of misses and mistakes. Programmers create 10 to 20 bugs per 1000 lines of code, 9 out of 10 businesses fail, accountants make detrimental blunders, etc.

The point is that in the end the ML systems need only to replace these already non-perfect systems. I'll refrain from judging the consequences of this as I think it's out of scope.

Nice packpedal, except now what you said makes no sense anyway because the discussion was about tools doing what you expect them to do, not "tool + human systems" doing what you expect them to. These are two different things.

Do you think a carpenter who hits themselves with a hammer blames the hammer or themselves? Or are you going to unironically tell me they would blame the hammer + human system?

Do you have a more specific point you're trying to make?

The discussion is about AI partially replacing human colleagues. In practice, people already are not 100% reliable. You make a reasonable request and someone makes a stupid blunder instead. That's the "hammer" hitting your thumb. Maybe you were not specific enough or maybe they didn't listen but the damage is done.

Our work processes already take mistakes and iterative refinement into account. If AI, in some specific niche, is cheaper and makes no more mistakes than humans do, it gets the job.

It doesn't need to be perfect or perfectly reliable. Some guardrails will be built into it, and we'll come to trust over time.

>not "tool + human systems" doing what you expect them to. These are two different things.

Please explain how?

>Or are you going to unironically tell me they would blame the hammer + human system?

As your tooling gets more complex, yes it is very easy to have a non-zero blame assignment to each party. Look at any human+machine system where complex failure conditions can occur.

You want me to explain the difference between a hammer and a human wielding one?

One is a hammer. The other is a human wielding a hammer.

> Except hammers already do not work as expected 100% of the time, as evident by them painfully hitting the hands of the workers in mishaps. Yet we still use them.

The hammer did work 100% as expected. It's the human, who is fallible, that hit their hand with the hammer. My analogy stands. We make mistakes, we want tools that do not. LLMs should not be compared to humans, they should be compared to other tools.

To a company a person is a tool. "Human resources" means exactly what it says.

Gpt is not a calculator. I suspect that it could solve calculator problems if it was given a calculator though.

So connect GPT4 with calculator using LangChain, problem solved!

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".