We analyse thousands of lines from a csv using an LLM. The only thing that worked for us was to send each individual line and analyse it one by one.

I’m not sure if that would work in your use case, but you could classify each line into a value using an LLM then hard code the trends you are looking for.

For example if you’re analysing something like support tickets. Use an LLM to classify the sentiment, and you can plot the sentiment on a graph and see if it’s trending up or down.

I’m really not sure I follow this argument.

A lot of software has friction to get to the value. This is often because of constraints not choice.

To give a concrete example of this, in my company we had users upload files for analysis. To get the export for the file, it took many steps. Not hard, but a lot to get done.

We switched it to an integration and now it’s 3 clicks. We’ve gone from 10% of users onboarding to 100%.

It doesn’t mean we get people to stay, but the barrier to understanding if our tool provides value to them has completely disappeared.

I’m very curious though, what value did you strip away when trying to make your product easier to use?

This also goes the other way too, you don’t assume the original is incorrect.

I see this a lot with developers who come in and start to criticise before understanding.

There is always a reason for why something is as it is, and it’s unlikely that the people before you were just idiots.

How do you introduce any tool/change to a team of people?

You get buy in, start having conversations see what AI people have explored. Have they tried claude? Do they prefer other tools? If so why? What are the objections. Actually listen. I’d also showcase what you can do. I love to present what codex has found when debugging something, or a prototype I’ve put together.

If you have the budget pay for subscriptions so they can play around.

Also, you say that development velocity is a big problem, but I would dive into why that is. You may be disappointed when velocity remains the same with AI tools.

My time in the military has made me hate acronyms with a fury.

https://acronyms-suck.com

If you focus on the differences that’s all you’ll see.

The opposite applies too, if you try hard enough you’ll find plenty of similarities with an Eskimo. We’re all just people in the end.

I’m curious have you ever been burnt or seen anyone burnt by copyright infringements in code?

Sometimes it’s super obvious because a game company steals code from a previous employer, but I have never seen this play out in entreprise software.

How is that extortion? If you use a library/project with a licence you need to abide by its terms. If you don't want to do that, then either write it yourself or find an alternative. People asking for the source code is not extortion, they're fully within their rights to do so.

I actually found the anecdote I was referring to:

https://news.ycombinator.com/item?id=44605873 (see comment by user yalok).

""" Just anecdotally, but this aligns with my observations on the trend/growth of successful useful open source projects that go with permissive BSD-like license. ~20 years ago there were way less of those than now. And as a SW developer doing client side/apps as well, using GPL/LGPL is a total pain and basically cost prohibitive, unless I work on my personal small project where I don’t care about having to/risking to open source the rest of the code and getting sued/cloned… Real life example from ~2010 - we ended up including an LGPL library in our mobile app code, and published/upstreamed all the modifications we did to that code (mostly ARM optimizations). Once the app became popular, our competitors came to us demanding the source code of our app - just because iOS didn’t support dynamic libraries (so we had to statically link it), and giving them the object code to relink it wasn’t enough for them (which would satisfy the spirit of LGPL), because they really wanted to see how we hacked around iOS camera input APIs… """

Arguably the competitors didn't have a case as the object code was provided, so I would stand by calling this extortion. Maybe the legal burdens were too high, so the company complied.

I didn’t even realise that was a possibility with iOS. Thanks!

Curious to hear more about this, do you have any examples?

We might release it as a framework of gems.

We created an Abstract controller that handles all of the typical behavior for a resource, auth, filtering, pagination, tenancy, import/export, serialization etc.

Then we expanded rails generators to cover ALL typical behavior. And the markdown file calls the generators.

It was a bit complicated to model polymorphic behavior but we got it working thanks to Ruby/Rails.

But the basic premise that made this work is: Use only restful actions; don’t turn it into RPC. Recognize that most RPC/graphql functions are state changes that could have been a patch request. So instead of /clients/activate its /clients with a status attribute for “activate” or “archive”. Then most nested routes aren’t needed, use accepts nested attributes for and return child ids in the show action. There’s more to it that this but by strictly following conventions and modeling the data for rest, the api ends up Super simple.

Our standard controller only whitelists strong params. All other behavior is automatic.

What do the markdown files look like?

Defines the resources, attributes and relationships. That’s it.

Client first_name last_name organization:references

Organization name tax_number has_many:Clients

I understand your pain, but this is very country dependent.

For example, most of what you describe can be solved using Xero in Australia.

> But it’s a hell of a lot better than sifting through thousands of transactions manually which accountants do and get wrong all the time.

I still wonder why humans getting things wrong is a problem, but LLMs getting more things more wrong more often than humans never is. At the very least you'll need a human accountant around to verify the LLM. Or I guess you could just practice "vibe accountancy" and hope things work out but that seems like a worse idea than a trained human professional. But I'm probably just a Luddite.

Also, I am admittedly not an accountant, but I don't think they manually sift through every transaction to verify compliance issues in every single case. That probably isn't how that works.

To be clear, I agree with you.

Our target market is accountants. I want to help them not replace them.

In the same way security audit tools aren’t replacing professionals, but that can help on initial scan.

> I still wonder why humans getting things wrong is a problem, but LLMs getting more things more wrong more often than humans never is

Some people hate humanity so much that they cannot wait to replace us all with AI so they never have to interact with another human ever again

That's honestly the only reason I can think that they are so biased toward AI

I find this take so strange, do you find no value in AI?

I don’t even want AI to replace us, but it’s a great tool with many use cases.

I weigh the economic value against the lives I believe is going to ruin and the damage I believe is going to do to society and the future of the human race and I do not find value there. I find ruin

There might be a way for us to adopt AI as a tool without bringing ruin to many people, but I don't believe that is the goal of anyone building AI.

As it stands, I don't believe there is anything ethical about AI in it's current form. So from that perspective, I vehemently deny there is any value in it

At one point in history, people like you were asking why anyone could be anti-slavery. After all, it was impossible to deny the economic value of slaves.

Why is AI unethical in its current form and what would make it ethical for you?

At this point I don't think even Sam Altman believes AI to be ethical, as much as he would like to believe such thing.

I think we misunderstand each other.

Accounting is quite broad and there are many examples of where LLMs can help.

For example, what is tax deductible is subjective to local laws. This is essentially a classification problem that LLMs are particularly good at.