[rockskon]

And yet it's never "now". The promised results are never here.

I understand one of the chief innovations the AI industry produces is rhetoric and hype, but it's insufferable and repetitive.

A better AI isn't good enough. "Closer" to a stated goal isn't good enough.

Deliver results that have value to more than just enthusiasts and academics.

[margalabargala]

Sure it is. AI software development is here. It's not good enough for everything, but it's good enough for a majority of the changes made by most software engineers.

That's now. Right now, the tooling exists so that for >80% of software devs, 80% of the code they produce could be created by AI rather than by hand.

You can always find some person saying that it'll destroy all jobs in a year, or make us all rich in a year, or whatever, but your cynicism blinds you to the actual advances being made. There is an endless supply of new goalpost positions, they will never all be met, and an endless supply of chartalans claiming unrealistic futures. Don't confuse that with "and therefore results do not exist".

[rockskon]

No, it isn't. There is a gigantic chasm of difference between "80% of code they produce could be created by AI" and "80% of commits they produce could be created by AI".

Mixing the two up is how we get a massive company like Microsoft to continually produce such atrocious software updates that destroy hardware or cause BSODs for their flagship Operating System.

That's not replacing software development. That's dysfunction masquerading as capability.

And none of what I said is goalpost moving. They are the goalposts constantly made by the AI industry and their hype-men. The very premise of replacing a significant amount of human labor underlies the exorbitant valuation AI has been given in the market.

[margalabargala]

We'll just have to agree to disagree.

It appears that your understanding of AI code generation reflects the state of 1-2 years ago. In which case of course it seems like what people are describing as reality, feels 1-2 years away.

> There is a gigantic chasm of difference between "80% of code they produce could be created by AI" and "80% of commits they produce could be created by AI".

This is exactly the goalpost moving I am talking about. I said 80% of code could be AI-written, you agreed, and followed up with "oh but it doesn't matter because now we're measuring by % of commits".

[slopinthebag]

> That's now. Right now, the tooling exists so that for >80% of software devs, 80% of the code they produce could be created by AI rather than by hand.

Technically 100% of the code they could produce could be created by a ton of very specific AI prompts. At that level of control it would be slower than typing the code out though.

Just throwing out random numbers like this is complete nonsense since there's about a million factors which determine the effectiveness of an LLM at generating code for a specific use case. And it also depends on what you consider producing by hand versus LLM output. Etc.

[GeoAtreides]

>The promised results are never here.

Today I fed to Opus 4.6 five screenshots with annotations from the client and told it to implement the changes. Then told it to generate real specs, which it did. I never even looked at the screenshots, I just checked and tested against the generated specs. Client was happy.

I don't know what it means.