[chairhairair]

cron, but completely unreliable. How nice.

LLM heads will say “it’s not completely unreliable, it works very often”. That is completely unreliable. You cannot rely on it to work.

Please product people, stop putting LLMs at the core of products that need reliability.

[kenjackson]

It's all a matter of degree. Even in deterministic systems, bit flipping happens. Rarely, but it does. You don't throw out computers as a whole because of this phenomena, do you? You just assess the risk and determine if the scenario you care about sits above or below the threshold.

[dkjaudyeqooe]

A bit flip is a rare occurrence in an array typically tens of billions large.

The chance that the flipped bit changes a bit that results in a new valid state and one that does something actually damaging is astronomically small.

Meanwhile LLM errors are common and directly effect the result.

[kenjackson]

My point is that your confidence level depends on your task. There are many tasks for which I'll require ECC. There are other tasks where an LLM is sufficient. Just like there are some tasks where dropped packets aren't a big deal and others where it is absolutely unacceptable.

If you don't understand the tolerance of your scenario, then all this talk about LLM unreliability is wasted. You need to spend time understanding your requirements first.

[great_psy]

When’s the last time you personally had a bit flip on you?

[mhitza]

You generally cannot know because we don't measure for it? Especially not on personal computers, maybe ECC ram reports this information in some way?

In practice I think it happens often enough, and I remember a blackhat conference talk from around a decade ago where the hacker squatted typoed variants of the domain of a popular facebook game, and caught requests from real end users. Basing his attack on the random chance of bitflips during dns lookups.

Related, but not the video I was referring to

https://news.ycombinator.com/item?id=5446854