Comment by proc0 | Hacker News Reader

Comment by proc0 | Hacker News Reader

One recurring pattern I see with the current AI predictions is a kind of contradicting idea of AI being really intelligent agents, but also AI needs to be checked carefully. Either they're intelligent enough to not need humans or we are changing the definitions of intelligence.

For example, in the article:

> We're approaching a inflection point where the barrier to creating software will be primarily conceptual rather than technical.

And then...

> Developers will need to audit AI-generated code for vulnerabilities, implement security best practices, and ensure compliance with increasingly complex regulations.

Again, if AI is going to be so good at coding, why would it not be able to implement best practices, and generate perfectly compliant code with a few prompts? I think it's interesting that the promise of AI clearly implies that it will do everything humans can, yet I keep reading how engineers need to still check what the AI is doing. It's like self-driving cars that still need you to have your eyes on the road and hands on the wheel. Seems like the implied promise of the technology cannot quite reach its destination.

If we use the metaphor of the bird and the airplane, we're basically expecting airplanes to fly like birds, takeoff from the ground, flap its wings. Airplanes are much faster than birds, but needs a runway for takeoff and lots of fuel. Similarly current LLMs can synthesize huge amounts of text, summarize it, etc., but they have cognitive limitations that are crucial to solving problems in the way humans do.

I think there is something beyond this metaphor though. I think the brain is tapping into some algorithm from which mathematical reasoning emerges. This algorithm has side-effects that look like human reasoning, and it's also the missing ingredient to make machines properly communicate and collaborate with humans (and also allow them to be properly agentic).

> a kind of contradicting idea of AI being really intelligent agents, but also AI needs to be checked carefully. Either they're intelligent enough to not need humans or we are changing the definitions of intelligence.

Why is everyone so flummoxed by this? Your coworker is intelligent, but still needs code reviews.

Why is it than whenever people think of artificial intelligence the only options they see are dumb as a rock / pure parrot, or some omniscient god? There is no intelligence on earth that falls in either of those categories, but thats the only two options people can use to visualize AI and if it not one to them it must be the other.

Intelligence exists without being perfect gods. I feel like people have watched too much sci-fi.

That’s the thing though coworkers don’t actually need code reviews.

Years ago reviews of every code change before deployment weren’t a thing. Most of the time it was fine.

Today, it’s fairly rare for a code review to find a serious bug in a PR.

The majority of PRs from the majority of developers work fine. They do what they’re supposed to and they don’t bring production crashing down.

That’s not close to true for AI.

Rubber stamp PRs have the norm at every single place I have worked. No one has the time or mental energy for reading other’s code.

Also why would we not just get another AI to do code review? It would be significantly faster and cheaper than a human, if equivalent.

> Why is it than whenever people think of artificial intelligence the only options they see are dumb as a rock / pure parrot, or some omniscient god?

That's not a bad question. I think it's because of the implications of building an autonomous intelligent agent. If it's actually intelligent then it has to understand basic instructions and also follow simple logical reasoning to avoid pitfalls, consistently.

i.e., if a human artist is really good at generating art, if I ask them to please create a character with the left foot forward, it's an easy and obvious task they can do. Currently AI is really good at many things, which is great, but then it fails at simple things you would expect it to know if can do these other really hard things. This disconnect is what causes them to not be reliable, and therefore not as useful.

That said, who knows we could be months away from a leap that will give them internal logical structures that will allow them to actually reason from first principles consistently every time. In my opinion, this is the leap we need in order for all of the current hype to be real.

Are they tools? So we can apply the reliability measure, as to how often they help us do something more efficiently instead of hindering us?

Are they assistants, a step further than tools? So after training the can provide a more conxtetual help and we can offload menial tasks to them while whe do the more abstract thinking?

So far there’s no answer. They promises us assistants while delivering something that’s worse than any tooling. It’s an impressive tech, but not that useful on its own.

I think assistants is a good descriptions. There are several sci-fi universes where there exists these two kinds of AI, one is like what we have now (which is already incredible really) and the other is AI that has sentience and doesn't even consider itself to be artificial but just consciousness in a different substrate. I think the game Mass Effect has this, Warhammer 40K as well, and some others, but I think that's the distinction that we need. Assistant smart algorithms vs. human-like artificial minds.

I’m not a programmer so take my metaphor with a grain of salt.

I am, however, a writer, and so your comment made me think of the difference between writing and editing. It’s perfectly possible (and common, even) for someone to be an amazing writer but a terrible editor, and vice versa. The writer focuses on production, and typically even thinking about editing during that process is detrimental to the quality of the work.

The Ai-human coding situation seems relatively analogous to this.

In that case you would be using it as a tool, which of course has already made a huge impact and will continue to do so. There is however a promise being made by the leading AI companies, than in fact we will have the equivalent of human digital minds. The implication of this is that both the editor and the author can be AIs. It's just the direction Google, OpenAI, etc. are at least trying to go in.

Part of the discussion is whether this is hype or not, and I'm just saying that at the moment it seems like hype because they're clearly more like tools right now, and it's unclear if we will make the leap such that they do become digital minds without a few more breakthroughs.

Reply on news.ycombinator.com