Hacker News Reader

Comment by atonse | Hacker News Reader

I have a similar claude story (much less money though), with the IRS R&D tax credit. The auditing firm initially said we qualify for $0. But then I had claude analyze past R&D reports and our expenses and it found the problem. The auditor had miscategorized our company.

So claude drafted an email even pointing to the right Internal Revenue Code (IRS Law), and specify why we fall under a specific category. The auditor got back to me two days later admitting their mistake and said our company now qualifies for $8k in tax credits. And a few months ago, it identified items in our AWS that saved us $250 a month (paying for itself).

So now I joke that even if I have a claude max plan, I've still come out ahead financially.

View on HN

> So now I joke that even if I have a claude max plan, I've still come out ahead financially.

This is no joke; for better or worse, I see a day when I'm paying a lot more for this and it will be a bargain.

By my estimation (guess) you won't actually need to spend that much because the models are already getting a point where they don't need to get a whole lot better to be extremely helpful across many domains.

And it looks like those very helpful capabilities will continue to transfer to smaller models as well, as architectures and training regimes continue to refine.

I can fairly easily imagine a world where the only people needing to spend a lot of money on models are those that are using them to solve truly novel problems. The rest of us will get plenty of use at reasonable costs for the typical day-to-day helpful stuff.

All we need is something like Qwen3-coder-next but at Kimi K2.6 ability so it runs on laptop workstation hardware and we are set...soon?

In 2023 GPT-4 was allegedly 1.8T parameters. In 2026 we have ~100x smaller models (10-20B) that handily outperform it, and can indeed run on a laptop.

How does "outperform" translate to the propensity of an LLM to hallucinate?

There seems to be a mass delusion about how capable SOTA models actually are. That's my only explanation for how poorly I find them performing in basic knowledge tasks compared to how others describe their prowess.

I understand you to be implying that I shouldn't trust my perception that there's a meaningful difference in how much different models hallucinate. I will take that under advisement, but I am still interested in the answer to my original question.

>I understand you to be implying that I shouldn't trust my perception that there's a meaningful difference in how much different models hallucinate.

Nope. Also I'm not GP.

It highly depends on the task. For math and coding, sure. But for knowledge tasks GPT-4 is wayy better than even SOTA ~100B models. For my knowledge test cases the lines get blurry at >400B

I am eagerly awaiting being able to run a strong local model. I'd hand Apple $5k right now for a Claude in a box. I know the cost might not be there now, just saying that is around my ideal price point.

$10k might even be worth it - but i'm assuming that the more expensive it is the beefier it is too, which also means more electricity.. and i already run ~6 computers/servers in my house. If a power surge happens i'm going to go live in the woods lol.

You can run 6-12 month old state of the art models for that type of money,

like, yesterday.

Yea, but i don't consider them good enough. I barely consider SOTA good enough.

I'm hoping that by the time the rugpull happens with SOTA (claude/etc) that at-home will be in the 4.7-5.5 range? We'll see.

They were good enough 6-12 months ago.

Maybe your tooling is what’s keeping you from your dream.

I would do the same but my issue is that the models are changing so fast, so I don't want to be left out of the next model cuz it only runs on an even newer GPU or something like that.

But maybe my limited understanding is thinking of this wrong.

I wouldn't worry about hardware.

I've run the latest local models over the last year, including the recent Qwen 3.6 30B A3B, on a 9yo GTX 1080 and 32G RAM I have lying around[0]. If I can do that I don't think hardware will be a problem for you in the near term. The only updates I've needed were to Llama.cpp when a new class of model was released.

[0]: In my case, I want to see how local models perform on limited hardware, sacrificing context size and intelligence compared to SOTA models, so I have to really limit my expectations.

> I would do the same but my issue is that the models are changing so fast, so I don't want to be left out of the next model cuz it only runs on an even newer GPU or something like that.

I think the same, and it's why i stopped caring about running llama/etc at home last year. That coupled with the models being dumb by comparison to SOTA really make me fine with waiting.

But in a year or two it's going to be difficult to resist at home, assuming the pace of improvement holds.

Focus on what’s actually required for your workflows.

Anything beyond that is just hobby, or continued education.

Uh... get a UPS?

I do, though they're not as bullet proof as you'd hope to my understanding. Hell i have one at the house level too - since i have an EV sitting behind that as well.

Are you in a region that doesn’t mandate grounded electrical systems?

(UPS is still a great idea for your expensive gear.)

[sci-fi “AGI” scenario] What if those with elite model access philosophize in a way us mere mortals can’t understand, so the elites have to prechew the ideas for us to bring them to our level, and they control the narrative?

In reality now, curious about social implications generally. Does this go beyond problem solving? Maybe the intelligence per token you get via your free library card/membership is insufficient to compete with peers in dating/employment/etc. markets, thus puts you at disadvantage.

That isn't really philosophy, but rather doom and gloom theories. Control the narrative on what exactly, how I write a bootstrap script for my servers? Or what type of flower is in this photo. Not everything is politics luckily.

Real world AGI scenario:

that’s already how world financial markets and governance work,

and yes, the best of the best models

and $ for tons of compute

will, for now, remain at the top.

In my anecdotal experience there is a huge gap between GPT-5-mini which hallucinates relentlessly and Claude Opus or the latest GPTs which are fairly reliable. I'm hoping that gap can be closed with improved approaches for small models and that good reliability is achievable for LLMs without requiring absolutely mammoth computing resources.

For what it's worth, I also used GPT-5.2 (via duck.ai) this year for questions about taxes and it was helpful — which makes sense because there's an abundance of material about taxes out there to be synthesized, so a text predictor trained in that domain should do well.

I've seen this day sometime in December and not only with Claude. Wish I was joking on some days, feeling exhilarated on others.

To be fair regarding taxes it should be that you get all exemptions by default and the other side telling/justifying you why you don't qualify, instead of you not getting the things you should because you're not sure how to interpret the law or don't know this or that rule exists. Taxes shouldn't be that difficult, and the US version of it seems to be behind even the one I have here in europe (where my taxes are done "for me" for anything non business owner related). If the government is calculating your taxes anyway, they should just give the number to you instead of asking you a number and you better have the same as us or you're guilty of something.

Feels like a system that is deliberatly made to be more punishing for those who can't afford the help or the education to figure things out.

Or punishing to those that don’t pay for software and services to the companies that lobby for it to be this way.

I would agree with you on most situations (like 1040 personal income taxes especially).

But in this particular tax credit, there's no way for the gov to know automatically what percentage of payroll was spent in qualified R&D expenses, since it's day to day business operations. Which is why we are _forced_ to hire an outside firm and pay them thousands of dollars (when Claude did an even better job), just to analyze how much of our time qualified as R&D expenses.

The problem I have is that I am forced to have to find a firm to do this, and most firms won't even work with companies as small as ours. So then we're stuck and losing out on years of R&D tax credits at the moment, when I really don't need them anymore, to be honest.

i saw a meme once like:

IRS> Pay your taxes!

me> ok how much?

IRS> idk you have to figure it out

me> ...ok

IRS> if you get it wrong you goto jail

I remember a different ending

me> so you don’t know how much I owe?

IRS> no, we do…

me> ...ok

IRS> if you get it wrong you goto jail

> it identified items in our AWS that saved us $250 a month

I gave Claude Code access to the Azure CLI and had it review usage. It found a couple servers that were underpowered, but it also found a lot of things that weren't being used at all. The total reduction was $40k a year.

I use AI to help me do my accounting (how to categorize and account for things). It pays for itself because I need to spend money for far fewer hours from an accounting firm each month to make sure that captialized expenses, depreciation, tax credits (we have Historical Tax Credits from restoring an 1880s building) etc, is put in the books properly. The AI gets it right a high enough percentage of the time that I only the to have a real accountant look at things once a quarter to make sure it's all OK. I used to have a dozen questions every month.

Same... I had chatgpt go over my taxes (I do it myself) and it found a number of savings I qualified for

Can you explain the steps you followed? Did you just feed it the whole return?

Kind of. I first of all did the entire return with it. So we went step by step and yes I fed the forms one at a time. So I filled in 1040 as best I could. Then just asked it what to do next at each step. It helps I've done it before so most of the steps it returned were ones I've done before. However, it did mention several things that I had not heard of, and also some new taxes that I had to file due to some exceptional events last year. So all in all, a solid use case. This year I have an accountant, but it saved my butt this last year, and I will absolutely run through my accountants decisions with it. It has an encyclopedic knowledge and an immense capability to understand without getting tired.

very trustworthy of the system sharing your taxes with a third party

People's tax returns are essentially public (yes I know they're not allowed to disclose them). Didn't send the forms in with social security numbers.

This absurd concern for privacy is silly in my opinion. The moment something is submitted to the government it ought to be considered public. Even your social security number is essentially public for anyone who cares to find it.

I would not submit my bank account information to these services, or my passwords, obviously.

Honestly, tax returns should be public again. Would make everyone better behaved IMO. It was this way for most of American income tax history believe it or not.

To be clear, my information has already been part of several breaches anyway. What protects you ultimately is the law not information security. Of course this point is often lost on engineering / computer scientist types who don't understand how law works.

omg. this is assuming the gov is completely incompetent and leaky. it seems doge "mission accomplished" indeed.

What's the threat model here? OpenAI gets my social security number and Sam Altman steals my identity? OpenAI leaves an S3 bucket open to the public and my filled out 1040.pdf gets leaked to the world?

Oh no, OpenAI knows how much money I make and they're going to send me ads! Ads that are relevant to my interests. How connivingly evil of them!

What sorts of things did it find?

Well my tax situation changed a bit this year, so there were some new taxes I had to file. Additionally it helped with depreciation for real estate. Helped me identify the exact part on the appraisal where I would get the cost of the dwelling (vs the land), and then walked me through the various ways the IRS allows people to do this. Helped me do some tax planning for next year as well around mega backdoor roths. Not lifechanging, but useful.

Reply on news.ycombinator.com