Hacker News Front Page | buzzwords.technology

As oppose to what everybody thought 3 years ago, LLMs are not very good with finance. Makes sense when you think about it as AI in 2025 still rely heavily on language transformers, and transformers are more or less probability calculators.

Ai agents in accounting firms are way less effective than Ai agents in law firms for example. 1 wrong word in the bylaws can break your entire company. Accounting is more precise, it tolerates less mistakes, and the rules that govern it are much less option to intrepretation.

In accounting, the rules are more grounded. Not in precedent but in a generally accepted specific rulebook. Gross profits mean the same in the US as it does in Bangladesh, and it will mean the same 100 years in the future.

In accounting, probability is very bad. A single bad journal entry in a 100,000 ledger can distort the entire balance sheet. And we know that no LLM will ever reach 100% accuracy rate because its reasoning is probabilistic, not absolute. Campfire raised $65m and built an entire accounting module from scratch, they only managed to achieve ~95% accuracy rate.

It might feel like 95% is very close to 100%, but that is not how it works. A football player such as Arbeloa, who ranks in the 80th percentile, is not remotely comparable to someone in the 90th percentile. Now imagine a player who is better than 99% of all football players who ever lived; that superstar is on an entirely different level, and Arbeloa would lose 0–10 to him every single time.

All of which without mentioning the 99.99th percentile, where the gap becomes unimaginable to human minds. Accounting follows the same power law; 99% accuracy rate might sound close to perfect, but across 100k ledger entries, that means ~1k bad entires. You can hire someone to go through the bad ones, but how can we know which ones are going to be bad. Its definitly a cost saver, but it's not the same as coding or legal. Also, all the major labs are going after accounting after they finish with coding, so margins are going down fast.

Yes, most accountants have an accuracy rate below 90%, but humans are allowed to make mistakes because YOU CAN SUE A HUMAN. That might not be a big deal in other areas, but when it comes to finance, companies suddenly get very risk averse. Its not even the hullication part that scares them, but the part where the Ai missed something. Bookkeeping can be divided into smaller tasks and automated fine, but accounting is a different story. Its about accountibility more than anything else.

Humans have a MAJOR advantage over AI; we can be held legally accountable. If it weren't for that, most of us will be obsolete already. Under Sarbanes Oxley, any company no matter how small must have a "human" finance officer (CFO), and a "human" controller. Courts won't allow an Ai attorney to present a case no matter how good it was at the job.

Comments URL: https://news.ycombinator.com/item?id=50213770

Points: 4

# Comments: 3

Kimi K3.1 scores 55 on Artificial Analysis, placed on the Pareto line

Hacker News Front Page

4 hours 21 minutes ago

Article URL: https://artificialanalysis.ai/models/kimi-k3-1-pro

Comments URL: https://news.ycombinator.com/item?id=50213337

Points: 96

# Comments: 18

I turned my TV into a Game Boy Color with a Pokemon town as my home screen

Hacker News Front Page

5 hours 5 minutes ago

Article URL: https://twitter.com/robbietilton/status/2111936458923851923

Comments URL: https://news.ycombinator.com/item?id=50212668

Points: 51

# Comments: 15

Air Purifiers as White Noise Machines

Hacker News Front Page

5 hours 17 minutes ago

Article URL: https://www.jefftk.com/p/air-purifiers-as-white-noise-machines

Comments URL: https://news.ycombinator.com/item?id=50212507

Points: 32

# Comments: 32

Can we replace a $150k Looker with a bucket of HTML files?

Hacker News Front Page

5 hours 31 minutes ago

Article URL: https://david.lequin.dev/blog/bi-software-is-dead/

Comments URL: https://news.ycombinator.com/item?id=50212320

Points: 92

# Comments: 55

Show HN: Searching for optimal strategies in a puzzle game with chance (Drop7)

Hacker News Front Page

5 hours 32 minutes ago

Sixteen years ago, I wrote a Java program which tried to pick the ideal move in a puzzle game called Drop7. On each turn, the player only has seven choices for their next move, but there is also an element of chance which turns this simple game into a surprisingly difficult problem involving search, hidden information, long-term planning, and chance.

I renewed my investigation this weekend, into whether there is an autonomous strategy that averages more than one million points per game. An experienced human player can score a million points with long-term strategic thinking, but the best automated strategies I could come up with are currently averaging around 300k-400k points. One of the easiest ways to contribute to advancing this research is to compete on the global leaderboard and submit your human strategy - https://drop7.dev/compete.

This project has also been an enlightening experiment into using agents as research collaborators and evidence collectors.

Repository: https://github.com/keshavsaharia/drop7-bench

I’d especially appreciate feedback from people who are knowledgeable in game theory, stochastic planning, learned value functions, or heavy-tailed evaluation. How would you approach long-horizon action ranking here? I’d also love to hear whether the research makes sense to someone encountering the project for the first time.

Comments URL: https://news.ycombinator.com/item?id=50212292

Points: 10

# Comments: 0