Recent submissions

Comment · 1 month ago
and the DEC Alpha RISC CPU was on fire, and ran circles around the desktop CISC CPUs
Comment · 1 month ago
> You can try to say that I’m arguing whatever you like.

I did my honest best possible interpretation of what you really meant from what you wrote.

>> We do know that zero parts of human decision making are based on predicting the next most likely letter based on a giant internet-based database. We do know that’s what LLMs do.

I read this as "The decision making of LLMs are based on predicting the next most likely letter based on a giant internet-based database."

Is that wrong?

I understood that your meaning was something like "LLMs can't reason, they just output likely letters"?

> If you’re claiming that the underlying structure of digital so-called neural networks is comparable to biological neural networks

No, I don't claim that.

What do claim is this: Regardless of how the LLMs were trained, they show overwhelming signs of being able to reason, and not just recall memorized information.

This doesn't mean that they always reason perfectly about everything.

But if they only memorized things and output the next likely letter, you would see them answering very badly much more often.

Comment · 1 month ago
> ... are based on predicting the next most likely letter based on a giant internet-based database. We do know that’s what LLMs do.

If you're claiming that the training objective tells us what kind of internal mechanisms the training produced, then I think that's just plain wrong.

Next-token prediction describes the optimization target, not the internal mechanisms that the training produced.

In the same way for the natural evolution of humans, DNA replication is the evolutionary objective. It's not a description of the internal mechanisms that evolution has produced.

As an example, we know that neural networks can be trained to develop generalized algorithms for arithmetic.

They might first memorize the training examples, then with further training transition to a solution that generalizes correctly to unseen examples.

In some cases we've even reverse-engineered the evolved internal mechanisms and found structured arithmetic algorithms rather than rote memorization. Interestingly, for modular addition this can involve Fourier representations, which isn't an algorithm I would have guessed gradient descent training of neural networks would produce.

Comment · 4 months ago
"The battle royale answers one question cleanly" smells ChatGPT-generated.

But that was the only thing I tripped on. I enjoyed reading the article in general.

Comment · 5 months ago
To me, "autocompletion" means a pure lookup function input->output.

That's a lot different from "general purpose processor which can act based on program logic, stored data, and input data".

Comment · 5 months ago
Sure, we agree there.

But it seems to me that if the LLM can effectively "execute" the instruction of how to take an input IP packet and generate a response IP packet based on a set or rules, then that's effectively a general purpose processor. And not an "auto completer", right?

Comment · 5 months ago
What do you mean that the "harness ran the commands"?

It looks to me like the LLM "executed" the logic in pure output tokens, not by using any kind of external tool calls?

Comment · 5 months ago
Do some people still claim "LLMs are just dumb auto completers"?

Because this seems to disprove that claim pretty convincingly?

Comment · 5 months ago
Opus 4.6 is already very good at troubleshooting all kinds of network problems if it has access to the command line tshark tool and the pcap files.
Comment · 7 months ago
From the video: https://youtu.be/OoJsPvmFixU?t=1942

  32:22 It's time to be honest. As engineers, we are not known to be effective communicators. That's just not a strong suit of an engineer.

  32:29 It is time to lay out and systematize the communication so that every organization that's involved knows far more information than they need. We have actual targets and dates that are actionable and they don't depend on a miracle in technological innovation occurring at some point on the Gantt chart.

  32:48 We actually have actionable things that have to happen in certain time frames. And if something doesn't happen, a critical technology isn't in development, we communicate that there's a schedule slip to everyone.
From what I've seen of recent very large engineering projects, saying "we need Gantt charts" is a good way to get sidelined/fired?

It seems that modern organizations absolutely loathe up front systems engineering and planning?

Comment · 7 months ago
Comment · 7 months ago
> Could you elaborate? How is this illegal if you declare taxes?

As someone else mentioned, the taxes are different.

Namely: Salary is taxed lower than dividends. So the German tax authorities checks very carefully that you don't pay salary instead of dividends. If they determine that you paid out dividends as a salary, then you'll be charged with tax fraud.

Now you might say, "I don't care about paying a bit extra in taxes, so I'll pay it as dividends as they wish"

The problem is that you can only pay dividends the year after you earn the money.

If you can set a fixed salary which you can keep paying throughout, and then wait for the dividend payments next year, that's fine.

But what if you want to pay yourself wildly different amounts of money each month based on how much you managed to charge your customers? You can't just keep adjusting your salary up and down every month with a corporation.

So here's where something like a sole proprietorship may be simpler from that aspect?

Another thing you want to look at is "how easy will it be to dissolve the operations?" With a GmbH/UG it takes several years and potentially many thousands of euros in accounting fees. Not sure about the foreign corps. I think German sole proprietorships are simpler in either case?

Also, Germany has a "Moving away tax" where you get taxed on the fictious value of your company if you move away from Germany. This fictious value can be quite a lot more than what you'd actually get if selling the company.

Yet another thing: Depending on your setup, you may be covered by different rules regarding health insurance and pensions. If you don't make a lot of money in the beginning, it may be best to stay in the government insurance. But if you think you'll make a lot of money, it can be better to be able to do private insurance instead? There are rules on how you can move back and forth between government/private, so this is another area to consider carefully.

This is my understanding as a layman, please check this with a competent local tax expert before acting on any advice here.

Comment · 7 months ago
Sure, it's not malicious. But it is very eager to get things done, and surprisingly inventive and knowledgeable in all kinds of workarounds.
Comment · 7 months ago
> As somebody from Germany, establishing a company is a bit tedious and bureaucratic.

I'm fairly sure the German tax authority will claim that you have a local German branch office since you live and work there.

That might be OK tax wise?

But I'd recommend starting with the tax situation in Germany.

Having limited liability through some kind of corporation can be nice.

But on the other hand, it becomes harder in Germany to pay out a varying salary as profits fluctuates throughout the year since the German tax authorities will see that as an illegal dividend payment from your company.

From this perspective it can be easier to set up some kind of sole proprietorship. Easier accounting etc and can pay out profits easier. But you get the personal liability.

This is not hard advice, just some things to point out that it gets complicated fast. So I'd recommend spending a few hundred euros on getting advice from a tax professional to begin with.

Comment · 7 months ago
Thanks, I agree 100% with that.

Live from the official Hacker News API

A modern reader for Hacker News · SvelteKit + Skeleton