So... How does one actually use AI?

In My Humble Opinion
101

I was curious what IBM was doing with Watson currently, having sold off Watson Health some time ago. They now have something called Watsonx, which has 6 variants tailored to different business purposes, which include generative AI and machine learning capabilities, a coding assistant. and a general-purpose natural language front end. This appears to be a major initiative to provide businesses with a diverse set of AI-powered capabilities. It’s been 15 years since Watson won the Jeopardy challenge and 11 years since IBM introduced Watson Health, so the new tech is undoubtedly vastly different and certainly structured very differently.

Here’s a blurb about what IBM Research is currently up to in the realm of AI:

102

Machine learning is what we tried out, but it was too difficult to work with at the time. I’m sure it’s much better now.

103

Not sure if this is the best place, but I’m interested in impressions to a phenomena my wife experiences. She teaches business law at a community college. A recent assignment (in a fully on-line class) asks students to provide specific data from a company’s official reports. For example, the number of oustanding shares of stock. It tells them what site to go to, where to go on the page to open the correct report, and where in the report to find the requested info.

At least 1/3 of her students do not follow those steps. Instead they apparently type the question into either google or AI, and provide the non-specific or inaccurate information that results.

I’m not suggesting AI has NO utility. But how do you make sense of this apparently widespread reliance on AI, when very straightforward and simple instructions are provided? Does a significant percentage of young people believe a general AI query is better than a specific search?

I suggested that, in addition to her clear instructions, she ought to include in her instructions: “DO NOT USE AI TO COMPLETE THIS ASSIGNMENT, AS IT WILL RESULT IN INCORRECT INFORMATION AND A GRADE OF ZERO.”

104

Everyone believe the AI is easier. Easy wins over thorough every time.

Next question. :wink:

105

But IS AI EASIER than following simple instructions? How many people in what demographics generally assume that AI results are identical to/just as good as the actual source material?

Just seems so different from how my mind works - if a teacher told me to get the answer from page 231 of the textbook, I’d turn to that page rather than ask some stranger what that page says.

106

Yup. That’s my issue in a nutshell.

Of course, Google intentionally broke its own search function to facilitate AI use. It’s useless for most things now. I pay to use a different search service - Kagi. But Kagi is built using Google’s data so it’s clear evidence that Google could do better, but won’t.

107

I’m just baffled at this mindset by COLLEGE STUDENTS! My wife’s instructions are essentially, “Turn to page 137 and write down the 3d word in the 4th paragraph.” These numbnuts instead submit responses saying, “More information is needed to respond.” :roll_eyes:

108

A post was split to a new topic: Kagi, the paid search engine

109

Heh, reminds me of a math assignment my teacher once gave me (this was in the 90s, long before LLMs). He reiterated several times before handing it out, “Make sure you read the instructions to the end first. They are important.”

The instructions started with “There are ten questions in this assignment. Read through all of them before you start, to make sure you understand the assignment.” In the middle were a bunch of math questions starting with basic arithmetic and working up to algebra and calculus. At the very end, the final instruction was simply “For a perfect score, turn this paper in without answering any questions. If you answered any of them, you will get a 0 for not having followed instructions.”

More than 80% of the class failed that assignment :slight_smile:

PS @beowulff Sorry, I totally forgot about this (your original format conversion request). I will try it again today, but I’m not too hopeful about my ability to succeed.

110

Don’t knock yourself out.
If it’s fun for you, do it. If not - it’s not a big deal for me.
Mostly just a curiosity thing.
Thanks.

111

I was probably in 5th or 6th grade when our teacher gave us a similar assignment, though it was several pages long, and at least 20 questions. The teacher handed it out, and maybe said something about “read the instructions carefully,” but wasn’t as blatant about it as your professor was; however, the first instruction on the first page clearly said, “read through all the questions before answering any of the questions.” Some of the questions were things like, “say the first ten letters of the alphabet out loud,” or “draw three triangles.”

Out of 25 or so of us in the class, one girl got it right.

112

Re: the OP’s original issue (converting from a plaintext-readable but undocumented & proprietary printed circuit board format into a newer, common standard), I finally had a chance to sic Antigravity (Gemini) and Claude Code on the problem.

These are Google’s and Anthropic’s agentic coding tools, meaning they let the AI interact with various tools online and on your local computer in order to build an entire app, not just answer in a chat interface.

Given this prompt:

In the sources folder, there are a .pcb and .osm file from the old macOS app OsmondCocoa. I’d like to reverse engineer these formats as much as possible (they are plaintext ASCII) in order to write a converter to more modern formats like KiCad. But there isn’t much documentation about these proprietary formats. The closest we can get right now is a Gerber export in the GerberFiles subfolder. Can you please try to compare all of these and try to write a converter as best as you can? Maybe start by writing tests based on what the converted output should be, starting from the parts of the file format you have the most confidence in and working down to the least. Then try to write a converter that can pass as many of those tests as possible.

Both were actually able to create a converted .kicad_pcb file that looked superficially similar to the original input. But not being an electronics person, I could not verify whether the output was actually correct. I’ve sent them over to the OP for validation, but it’s a start.

I wasn’t sure if the agents could do the job at all, so being able to even produce output that looks similar is a pretty good first attempt… maybe…? OP can say for sure once they take a look :slight_smile:

What practical tasks are AIs already good at today?
113

That’s pretty cool, @Reply . Looking forward to the OP’s… er… reply.

Inna and I are going to Ukraine sometime between August 15th and September 15th. I hate looking for flights, so, yesterday, I decided to use Gemini as a travel agent and look for flights for me. Since we pay for Gemini (pro version), I decided to create a “Gem” (i.e., an agent) so it can perform this repetitive task on a daily basis without being prompted.

I will revise this prompt as time goes by, but I wanted to test it first to see if it worked:

Today I got my first email:

I verified that these are actual flights and yes they are.

I may change the prompt to reflect different tier levels - “below $600 for economy OR below $1,000 for Premium OR…”, but I’m happy with the results I have so far. As we get closer to August and we have a more complete grasp of what our schedule entails, I will definitely revise it to look for specific dates.

Any suggestions for improvement? Additional cities, flight options, airlines, other?

What practical tasks are AIs already good at today?
114

The files that @Reply sent me were a good start.
I’m pretty impressed with what the AI was able to do.
It correctly created footprints from the Osmond file and used them in the KiCad file.
It replicated many (most) of the traces, although it made a bunch of mistakes with the copper flood.
It got the netlist correct, so the KiCad file retained most (maybe all) of the connectivity data.

The AI did make a bunch of mistakes with some hole positions. I’m not sure what caused that.
It didn’t honor the footprint silkscreen identifier locations, so that’s a mess. It also turned on silkscreen IDs for every via, which is wrong (but maybe an understandable mistake).

All in all, a surprisingly good attempt.

115

@beowulff,

Cool, thanks for details!

I think it’s up to you how much further you want to try and take this. My hunch is that AI could probably eventually make a good-enough converter, but only with your ongoing help, e.g.:

  • You can describe to it, in super-specific terms, what exactly is wrong and how it should be fixed. They seem to have access to good KiCad docs already (so being able to output in that format probably isn’t the issue), but properly ingesting Osmond does seem to be a roadblock right now. Some things they are able to guess at, but there are some parts they don’t have “known-good” information to work with so they can’t ascertain whether the output is correct.
    • The more detailed you can be, the more accurate the output will be. e.g. “In layer X, hole Y is wrong because ________, hole Z is wrong because ______” or whatever.
  • Or you can maybe make a semi-complex sample PCB in both Osmond and KiCad yourself, as an identical “before and after”, with all the features you need in both apps. That would give the AI an exact before/after that it can work from and it might be able to figure out the conversion that way.

You can also try to read the Osmond file yourself and explain to the AI what it means, e.g.:

Part void { Name Pack Loc { 163.08900m 124.38m 180 }
NameData { 40 156.61200m 123.87200m 0 }
ValueData { 50 163.72400m 126.92m 180 }
Pad A { Circle 175 }
Pad B { Rectangle 175 175 Spacing 75 }
Pin N { A A A 153.56400m 124.38m 3m P }
Pin P { B B B 163.08900m 124.38m 3m P }
}
Part void { Name Q1 Loc { 173.50300m 96.82100m 90 }
NameData { 40 152.04m 110.41m 0 Center }
ValueData { 0 173.50300m 96.82100m 90 }
Pad A { Circle 125 }
Pad B { Rectangle 100 90 Spacing 15 }
Pin 0 { A A A 151.91300m 105.71100m 125 P }
Pin D { B B B 169.69300m 105.71100m 45 P }
Pin G { B B B 169.69300m 103.17100m 45 P }
Pin S { B B B 169.69300m 108.25100m 45 P }

But anyway, working together with the AI, you might eventually get there.

I could try a few more prompts for you too if you give me specific directions (a list of detailed errors), but it’s probably easier for you to just try it yourself by signing up for a Claude Code or Antigravity/Gemini account if you were so inclined. It takes a lot of back-and-forth (AI does the work, you check it, tell it what’s right and what’s wrong, and so on over a few minutes/hours).

I think you will have better luck than me because I have no idea how to read a PCB diagram. I don’t know what silks, pads, pins, floods, etc. are. The AI does (way more than me), but I don’t know how to tell it how or why something is wrong.

I generally found Antigravity a lot easier to use, but your email said that Claude Code gave better results, so it’s up to you. Claude Code requires a lot more babysitting whereas Antigravity can work for long stretches more autonomously. I think OpenAI also has a similar offering (“Codex”? not sure), but I stopped using them after their recent partnership with the DoD so I dunno.

Any one of those agentic workflows would be much better suited for this use case than a generic chatbot, since the agents can write and run their own code and tests, call sub-agents as necessary, keep a working memory file, etc.

And for what it’s worth, for the $20/mo, it’s a hell lot cheaper than trying to pay a human developer to try to reverse-engineer that format on their own. They’d probably just ask their AI to do it anyway.

116

First, that’s cool! I didn’t know it could do that.

My only other thought is I think flights can change much quicker than once a day… maybe give it the ability to purchase the flight outright if it’s within certain criteria (X dollars, < Y hours and Z connections) if the airline lets you refund it within 24 hours anyway? Risky though. Who knows what else it’ll go on a shopping spree for :sweat_smile:

117

Here it is…the teachable moment. Instructors around me are also actively pondering what to do here, and in this situation, many have suggested students do both. Follow the instructions, fill out the requested data. AND also do the AI query and enter the AI version of this very specifically requested data. See the difference? Let students experience this first hand.

118

As this thread is more appropriate, I’m going to quote something I posted about in the Pit AI thread regarding a couple of more use cases for which I found AI to be very helpful. As it’s long, I’m going to put this in a details tag.

I’ve been reading about how AI is changing how coding (programming) is being done, much of this amplified over the past few weeks because of modifications to the Claude AI model.

So I decided to test it myself, using Gemini (which we already pay for in my business - it comes with our paid Workspace subscription.)

A note about myself: I am not a programmer, though I have run programming teams. As I explained to them: “I know what the inputs need to be. I know what the outputs need to be. I know, logically, how to get the output because, until now, I’ve brute-forced the inputs to give me my desired output. What I don’t know is how to tell the computer to take these inputs to create that output.”

IOW, I don’t know shit about programming.

Bona fides (or lack of them) aside, this weekend I decided to just go ahead and… given what I was reading about plain English being the programming language of the future… to see if I could get Gemini to programmatically solve two problems my company was having.

The first was to create a general knowledge internal web page about classifying bookkeeping transactions based upon inputs like dollar amount, industry, entity-type (sole proprietor, S-Corp, etc), and a description of what was needed. Using Gemini I built this thing to only use primary sources (Finra, IRS, others) and to be robust enough so that if the team member forgot an important detail, they could add it in later.

It fucking works like a charm, team. It’s an absolute game-changer and will assist our weaker team members into becoming better accountants and our great accountants into being superstars. It not only tells you the accounting treatment, it tells you why you do it this way, citing IRS Regs and GAAP standards, and alternative ways to do it based upon the legal entity type.

For example, I have the following issue:

“My new client bought a F150 for $55,000 in October 21st, 2023. She paid $15,000 in cash and took a $42,000 loan on the vehicle at 6.25% interest over 60 months. The extra $2,000 in the loan was because the document fees were rolled in the loan. She has not provided us with a statement, but we see an automatic payment of $634.43 leave her account every month like clockwork.

She totaled the truck on December 21st, 2025. Her insurance covered the loss, paying off the loan in full. She is an S-Corp. Please provide me with the following:

  1. The book entry for the purchase.
  2. Annual adjusting entries for interest and principal expenses.
  3. Annual adjusting entries for depreciation.
  4. The book entry for the loss of the vehicle.
  5. The amortization table you used to calculate the interest/principal payments.”

It then does all the math, gives me the book entries (which can be downloaded to CSV) and the amortization table.

But wait! It assumed that, being an S-Corp, we used accelerated depreciation… which the client didn’t do, their CPA just used a standard 5yr straight-line depreciation schedule. So I modified the query by saying:

“Oh, yeah, when they bought the truck, they were a sole-prop LLC, not a Sub-S. So they used a straight-line depreciation schedule of 5 years. Please provide me with updated depreciation book entries for each year.”

And it does so, easy-peasy.

The other use case is client specific. He wants specific administrative expenses allocated among his divisions based upon the divisions monthly revenue percentage. Why does he want this? Fuck if I know, but despite our telling him that this isn’t the best method, it’s what he wants.

So, also on the same day I created the Accounting Intelligence app, I created another HTML application where we just upload the P&L (by division) csv and the AI then builds a journal entry based upon the logic, which we then put into Quickbooks. A problem which took 15 minutes is now solved in 2.

I’ve taken the bottom example and now have a suite of web apps where we can upload documents, create a journal entry, and then put that JE into QuickBooks online. Processes that once took an hour, hour and a half, now take 3, 4 minutes.

What practical tasks are AIs already good at today?
119

Those are great uses.

A couple of months ago we shared in concert tickets with a friend and received a PDF with all our tickets in one doc (the theater doesn’t use an app). I could have just sent him all three and told him which to use, but I asked ChatGPT if it could split a PDF, and it did it in a matter of seconds based on my instructions. Way less complicated than what you’re doing, but nice to just have a thought and see it immediately executed.

What practical tasks are AIs already good at today?
120

We have four tickets to go see Swan Lake, us and another couple, and I just did that. Appreciate the heads-up!