It is weird that AI is solving hard problems for you. I can't get it to do the most basic things consistently, most of the time it's just pure garbage. I'd never pay for "AI" because it wastes more of my time than it saves. But I've never had a problem wrapping my head around a problem, I solve problems.

I'm curious what kind of problem your "brain cant wrap around", but the AI could.

  I'm curious what kind of problem your "brain cant wrap around", but the AI could.

One of the most common use cases is that I can't figure out why my SQL statement is erroring or doesn't work the way it should. I throw it into ChatGPT and it usually solves it instantly.

Is that a "hard problem" though? Really?

Yes. To me, it is. Sometimes queries I give it are 100-200 lines long. Sure, I can solve it eventually but getting an "instant" answer that is usually correct? Absolutely priceless.

It's pretty common for me to spend a day being stuck on a gnarly problem in the past. Most developers have. Now I'd say that's extremely rare. Either an LLM will solve it outright quickly or I get enough clues from an LLM to solve it efficiently.

You might be robbing yourself of the opportunity to learn SQL for real by short-cutting to a solution that might not even be correct one.

I've tried using LLMs for SQL and it fails at exactly that: complexity. Sure it'll get the basic queries right, but throw in anything that's not standard every day SQL into it and it'll give you solutions that are not great really confidently.

If you don't know SQL enough to figure out these issues in the first place, you don't know if the solutions the LLM provides are actually good or not. That's a real bad place to be in.

Usually the term, "hard problem", is reserved for problems that require novel solutions

Have you ever read Zen and the Art of Motorcycle Maintenance? One of the first examples in that book is how when you are disassembling a motorcycle any one bolt is trivial until one is stuck. Then it becomes your entire world for a while as you try to solve this problem and the solution can range from trivial to amazingly complex.

You are using the term “hard problem” to mean something like solving P = NP. But in reality as soon as you step outside of your area of expertise most problems will be hard for you. I will give you some examples of things you might find to be hard problems (without knowing your background):

- what is the correct way to frame a door into a structural exterior wall of a house with 10 foot ceilings that minimized heat transfer and is code compliant.

- what is the correct torque spec and sequence for a Briggs and Stratton single cylinder 500 cc motor.

- how to correctly identify a vintage Stanley hand plane (there were nearly two dozen generations of them, some with a dozen different types), and how to compare them and assess their value.

- how to repair a cracked piece of structural plastic. This one was really interesting for me because I came up with about 5 approaches and tried two of them before asking an LLM and it quickly explained to me why none of the solutions I came up with would work with that specific type of plastic (HDPE is not something you can glue with most types of resins or epoxies and it turns out plastic welding is the main and best solution). What it came up with was more cost efficient, easier, and quicker than anything I thought up.

- explaining why mixing felt, rust, and CA glue caused an exothermal reaction.

- find obscure local programs designed to financially help first time home buyers and analyze their eligibility criteria.

In all cases I was able to verify the solutions. In all cases I was not an expert on the subject and in all cases for me these problems presented serious difficulty so you might colloquially refer to them as hard problems.

It is not. It’s relative to the subject.

In this case, the original author stated that AI only good for rewriting emails. I showed a much harder problem that AI is able to help me with. So clearly, my problem can be reasonably described as “hard” relative to rewriting emails.

If you have 200 line SQL queries you have a whole other kind of problem.

not unless you are working on todo apps.

TODO: refactor the schema design.

What happens when these "AI" companies start charging you what it really costs to run the "AI"? You'd very likely balk at it and have to learn SQL yourself. Enjoy it while it lasts, I guess?

Problem with this is people will accept tech debt and slow query's so long as the LLM can make sense of it (allegedly!).

So the craft is lost. Making that optimised query or simplifying the solution space.

No one will ask "should it be relational even?" if the LLM can spit out sql then move on to next problem.

So why not ask the LLM if it should be relational and provide the pros and cons?

Anyway, I'm sure people have asked if we should be programming in C rather than Assembly to preserve the craft.

Surely you understand the difference between not knowing how to do anything by yourself and only knowing how to use high-level languages?

That is like using the LLM like a book. Sure do that! But human still needs to understand and make the decisions.

I work with some very complex queries (that I didn't write), and yeah, AI is an absolute lifesaver, especially in troubleshooting situations. What used to take me hours now takes me minutes.

In my case, Learning new stuff is one place I see AI playing major role. Especially the academic research which is hard to start if you are newbie but with AI I can start my research, read more papers with better clarity.

Which model are you using?

Sorry, I'm not taking a comment like this from a 2-hour old account seriously. You don't know me at all.

"It can't be that stupid, you must be prompting it wrong!"

Sigh.

Can you give some examples??

Calculate the return on investment for a solar installation of a specified size on a specified property based on the current dynamic prices of all of the panels, batteries, inverter, and balance of system components, the current zoning and electrical code, the current cost of capital, the average insolation and weather taking into account likely changes in weather in the future as weather instability increases due to more global increase of temperature, the chosen installation method and angle, and the optimal angle of the solar panels if adjusted monthly or quarterly. Now do a Manual J calculation to determine the correct size of heat pump in each section of that property, taking into account number of occupants, insulation level, etc.

ChatGPT is currently the best solar calculator on the publicly accessible internet and it's not even close. It'll give you the internal rate of return, it'll ask all the relevant questions, find you all the discounts you can take in taxes and incentives, determine whether you should pay the additional permitting and inspection cost for net metering or just go local usage with batteries, size the batteries for you, and find some candidate electricians to do the actual installation once you acquire the equipment.

Edit: My guess is that it'd cost several thousand dollars to hire someone to do this for you, and it'll save you probably in the $10k-$30k range on the final outcomes, depending on the size of system.

My God, the first example is having an AI do math, then he says "Well I trust it to a standard deviation"

So it's literally the same as googling "what's the ballpark solar installation cost for X in Y area" unbelievable, and people pay $20+ per month for this

$200 :)

Any way to tell if the convincing final numbers it told you are real or halucinated ?

Solve the same task with ChatGPT, Gemini and Claude. If they agree, you can be reasonably sure.

I'm not opposed to experimenting, but that's a a recipe for false confidence in a final decision.

Where they agree it shows the data supports that answer - not necessarily that it is true, where they disagree it shows you need to hedge. That's useful.

This is so wrong!

e.g., if you had a heart condition, you can't just poll three LLMs and be "reasonably sure" you've properly diagnosed the ailment.

I checked them carefully myself with various other tools. It was using python to do the math so I trust it to a single standard deviation at least.

Standard deviation of what

I'm lost too. Financials are technology agnostic.

They probably meant that they could read (and trace) the logic in Python for correctness.

I wouldn't publish it as statistically significant but it's within the error bounds for a real human accomplishing the same task, to reword.

I see your edit.

I would recommend spending that "couple thousand" for quote(s). It's a second opinion from someone who hopefully has high volume in your local market. And your downside could be the entire system plus remediation, fines, etc.

To be clear, I'm not opposed to experimenting, but I wouldn't rely on this. Appreciate your comment for the discussion.

No I'm not relying on it in the sense of going out and running the entire project through it, but as an accurate screener for whether it's worth doing, there's nothing comparable available.

>Weird because AI has been solving hard problems for me.

Examples or it didn't happen.