Hacker News Reader

Comment by crosen99 | Hacker News Reader

Hallucinations are a real problem right now, and that needs to be solved. But we don’t need the tech to be perfect in order for it to have a massive impact.

One question is how far off we are from a model that is reliable enough to autonomously handle hard business transactions. But an entirely different question is how much better could many user experiences be that don’t involve an automated culmination of a hard transaction.

I think of you differentiate this way between transactions with a hard impact (ie charging a customer or shipping a product) vs. informational transactions, you can allow for a lot more to happen on a much faster timeline.

View on HN

> Hallucinations are a real problem right now, and that needs to be solved. But we don’t need the tech to be perfect in order for it to have a massive impact.

As long as the use cases are considered properly.

Consider I hit a pet emergency website that is powered by ChatGPT. I ask it what to do because my dog has just had an encounter with a specific poisonous animal. It gives me advice based on a similar, but slightly different situation and my dog dies.

Alternatively say a telecommunications vendor (or for that matter, many SaaS vendors) implement it instead of detailed product docs. It gives me wrong advice and I misconfigure a router. As a result, an emergency call is mis-routed and someone dies.

AI chat isn't even remotely ready for these scenarios yet and there are still plenty of critical infrastructure services that rely on things like email. Therefore even your basic webmail SaaS isn't a suitable target for this type of tech.

Where could you genuinely deploy it safely?

Well, you need one level of readiness for life and death apps and a whole other level for entertainment apps and then there’s everything in between.

One can debate just how much of the market falls into the safer range of the spectrum, but the key question may be how far we are in time from much more reliable performance and not how far the current performance falls below some reliability cutoff.

For the sake of this discussion, I think how far we are in time is the more compelling question. And, I don’t know that we are nearly so “remote” in this sense as suggested by the phrase “isn’t even remotely ready.”

Unless there is specific reason to believe otherwise, I’d expect such improvements could appear in very short order given the recent pace of model improvements in other ways.

Reply on news.ycombinator.com