Hacker News Reader

Comment by tsunamifury | Hacker News Reader

This is the error of the thinking. It would be like saying software doesn't have a moat because thats just clever talking to a processor.

But no one would say that now, thats ridiculous. There is a sufficient degree of prompt engineering that is already defensible, I'm already doing it myself IMO. You'll see very sophisticated hybrid programming/prompting systems being developed in the next year that will prove out the case.

For example 30 parallel prompts that then amalgamate into a decision and an audit, with 10 simulation level prompts running chained afterwards to clean the output. These types of atomic configurations will become sufficiently complex to not be just for 'anybody'.

View on HN

The output of that many chained prompts is probably so unreliable that it's useless.

Again, this is a misunderstanding of what LLMs are capable of. These aren't chained, you can run parallel prompts of 15 personas with X diversity of perspective, that reason on a singular request, string, input or variable, they provide output plus audit explanation. You then run an amalgamation or committee decision (sort of like mixture of experts) on it to output variable or string. Then run parallel simulation or reflection prompts based on X different context personas to double check their application to outside cases, unconsidered context, etc.

It's pretty effective on complex problems like Spam, Trust and Safety, etc. And the applications of these sort of reasoning atomic configurations I think are unlimited. It's not just 'talking fancy' to an AI, its building processes that systematically improve reasoning to different very hard applied problems.

Brave are those who have set out to make prompt engg an entire industry, with gpt-5 and gemini lurking on the horizon.

Sam has specifically stated it's unlikely there will be a GPT5, and likely GPT4 is just a deeply prompt optimized and multimode version of GPT3.

But overall, hasn't that theme been true for like... all tech ever? You have to set up and build your own innovation path at some point.

> the applications of these sort of reasoning atomic configurations I think are unlimited

They are limited to applications in which the latency slo is O(seconds), knowledge of 2021-present doesn’t matter, and you’re allowed to make things up when you don’t know the answer.

There are, to be fair, many such applications. But it’s not unlimited.

Bounded =/= not unlimited

Interesting. Were you inspired by the Delphi decision algorithm, or did you happen to rediscover it? I agree it could provide a moat, though not a very wide one. All it takes is for your competition to talk to non-engineers, and all sorts of clever usages become possible.

See https://en.wikipedia.org/wiki/Delphi_method

Accidentally discovered its effectiveness when configured with LLMs, just applied with new technology. And it’s just one configuration example.

But in general many configurations are possible, but also you need to refine the personas a great deal to get it to work well.

> These aren't chained, you can run parallel prompts of 15 personas with X diversity of perspective, that reason on a singular request, string, input or variable, they provide output plus audit explanation. You then run an amalgamation or committee decision (sort of like mixture of experts) on it to output variable or string. Then run parallel simulation or reflection prompts based on X different context personas to double check their application to outside cases, unconsidered context, etc.

Sure, so you ensemble some results. You're back to the classical "hyperparameter" problem though that's faced ML for a long time-- what those personas are, what those subsequent prompts are, etc. require a fair amount of manual verification and tuning. And the search space is extremely vast.

Not to mention that something like this is likely to be very unperformant.

We’ve found way that works for us. Nothing magical, but it works in a generalized sense. And doesn’t have this need for significant manual tuning. Search space is vast, but since they can write their own prompts or select strategies, you can have then tree nav a library of strategies and they do pretty well.

That’s a lot of compute power you’re expending for funny computer parrot to hallucinate some text for you. Seems not even machines get to escape “decision by committee” lol.

Decision by committee is actually a pretty good technique for improving the performance of ml models but there's probably cheaper ways of doing it than hitting OpenAI's api 10 times a request

Okay that actually sounds like ensemble learning which is a known ml technique

Been working on ML for 10 years at goog. The difference is the context, subtlety and complexity of the problems you can feed this vs older models. Delphi, ensemble, whatever before lacked audit ability and explain ability —- LLMs excel at this. It allows you to apply it to much softer problems.

Reply on news.ycombinator.com