Hacker News Reader

Hacker News

Comment by alwaysrunning | Hacker News Reader

Parent Full thread

alwaysrunning

[flagged]

View on HN

Reads like LLMs are slowly getting better at math.

Not much better. Look at the last line. (Pointed out by https://news.ycombinator.com/item?id=40204564)

Yeah, I did check it all before I commented, quietly preparing to come out with something a little more scathing.

It's nearly there though so it's genuinely an improvement on how they used to behave.

Leaving aside the general ridiculousness, that is, and without checking their actual daily revenue which I'm fairly confident is more likely to be wrong than the arithmetic.

> It's nearly there though so it's genuinely an improvement on how they used to behave.

It's writing the conclusion at the beginning, so none of that convincing-looking mathematical reasoning had any bearing on the provided answer. The sum in the fourth paragraph could have been implemented using a plugin: we have no reason to believe the actual language model is doing it (and lots of precedent to suggest it probably didn't).

Note: the original comment has been edited since these remarks were made; it originally said (warning: machine-generated nonsense):

---

It would take the combined daily revenue of T-Mobile, AT&T, and Verizon approximately 2.67 days to generate $196 million in revenue. To estimate the time it takes for T-Mobile, AT&T, and Verizon combined to generate $196 million in revenue, we need to calculate their combined daily revenue.

Let's assume the combined daily revenue of the three companies is:

T-Mobile: $45.5 million (as estimated earlier) AT&T: $125.6 million (as estimated earlier) Verizon: $349.3 million (as estimated earlier) Combined daily revenue: 45.5 million + 125.6 million + 349.3 million = 520.4 million

To generate $196 million in revenue, it would take the combined daily revenue of the three companies approximately:

520.4 million (combined daily revenue) ÷ 196 million (target revenue) ≈ 2.67 days

You're probably right, and I did wonder if it was just delegating the math. I remember they tried to get chatgpt to delegate to python for that sort of thing.

What interested me was that even though it got it backwards 520.4 / 196 isn't quite 2.67, without allowing for some weird floating point behaviour (it's 2.655)

The fact the final answer is close to the initial answer is surprising: the fact it's not the same is unsurprising. 2.67 was decided (plucked out of thin air) before any of the other numbers were invented.

LLM's are always running.

Why would you use revenue and not net profit for this calculation?

What would be the justification for using profit rather than revenue?

For years Amazon was one of the biggest companies in the world while never making a profit. If it engaged in wrongdoing should they have been given a fine of 0$ (or negative I guess) since they weren't profitable?

I don't think the parent is suggesting that they be fined based on their profit.

They're saying it's unusual to use revenue to talk about how easily a company could pay a fine, given that (cashflow cleverness aside) your revenue doesn't determine your ability to pay for arbitary extras.

In reality neither does net profit, as nonessentials and comp tend to consume just enough to leave a profitability which keeps shareholders happy.

Also the dimensional analysis is wrong.

X dollars/day ÷ Y dollars = 1/days (so, the answer is inverted)

Because profit is a poor indicator of financial flexibility. Amazon being the famous example of running at maximum growth for many years instead of trying to have fat profits. That said, revenue isn’t a great indicator either.

Because then you wouldn't be nearly as outraged by the "data" of course!

Or if you want to use revenue, revenue from the data sales.

Reply on news.ycombinator.com