Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Everything around LLMs is still magical and wishful thinking

dmitriid.com

Post reply on HN
Print view View original

161–170 of 377 posts

Re: Everything around LLMs is still magical and wishful thinking

#161

Post by dcre »

troupo wrote: post #71
dcre wrote: post #3

Similar argument to https://www.baldurbjarnason.com/2025/trusting-your-own-judge... , but I like this one better because at least it doesn’t try to pull the rhetorical trick of slipping from “we can’t know whether LLMs are helping because we haven’t studied the question systematically” to “actually we do know, and they’re shit”.

Wow. Quite a conclusion from an article that actually doesn't reach for that conclusion

It’s a bad and incoherent essay. It simultaneously wants to say “we can’t know”, “we do know and it doesn’t work”, and “we know it makes engineering faster and that’s bad because most engineering is bad.” This isn’t nuance, it’s confusion. Sorry!

Re: Everything around LLMs is still magical and wishful thinking

#162

Post by swader999 »

abletonlive wrote: post #125
tasty_freeze wrote: post #18

One thing I find frustrating is that management where I work has heard of 10x productivity gains. Some of those claims even come from early adopters at my work. But that sets expectation way too high. Partly it is due to Amdahl's law: I spend only a portion of my time coding, and far more time thinking and communicating with others that are customers of my code. Even if does make the coding 10x faster (and it doesn't…

I’m a tech lead and I have maybe 5X output now compared to everybody else under me. Quantified by scoring tickets at a team level. I also have more responsibilities outside of IC work compared to the people under me. At this point I’m asking my manager to fire people that still think llms are just toys because I’m tired of working with people with this poor mindset. A pragmatic engineer continually reevaluates what t…

"I’ve offered to share my workflows, prompts" That should all be checked in.

Re: Everything around LLMs is still magical and wishful thinking

#163

Post by leptons »

arendtio wrote: post #60

I think it is more like googling: when the search engine appeared, everybody had to learn how to write a good query, even though the expectation was that everybody could use them. With LLMs, it's quite similar: you have to learn how to use them. Yes, they are non-deterministic, but if you know how to use them, you can increase your chances of getting a good result dramatically. Often, this not only means articulating…

>Yes, they are non-deterministic

The LLM is more like a Ouija board than a reliable tool.

>I can ask it to write user stories or prototypes

By the time I write enough to explain thoroughly to an LLM what to write in "user stories" or "prototypes", I could have just written it myself, without the middleman(bot), and without the LLM hallucinating.

If half the time I spend with an LLM is telling it what to do, and then another half is correcting what it did, then I'm not really saving any time at all by using it.

Re: Everything around LLMs is still magical and wishful thinking

#164

Post by 0x696C6961 »

the__alchemist wrote: post #118
0x696C6961 wrote: post #50

Earlier quoted context omitted.

I find that the code quality LLMs output is pretty bad. I end up going through so many iterations that it ends up being faster to do it myself. What I find agents actually useful for is doing large scale mechanical refractors. Instead of trying to figure out the perfect vim macro or AST rewrite script, I'll throw an agent at it.

What sort of mechanical refactors?

"Find all places this API is used and rewrite it using these other APIs."

Re: Everything around LLMs is still magical and wishful thinking

#165

Post by awkwabear »

This seems like a really bad take.

I do PhD research for superconducting materials and right I've been adapting and scaling an existing segmentation model from a research paper for image processing to run multithreaded and took the training runtime per image from 55min to 2min. Yeah it was low hanging fruit but honestly its the type of thing that is just tedious and easy to make mistakes and spend forever debugging.

Like sure I could have done it myself but it would have taken me days to figure out and I would have had to test and read a ton of docs. Claude got it working in like half an hour and generated every data plot I could need. If I wanted to test out different strategies and optimizations, I could iterate through various strategies rapidly.

I don't really like to rely on AI a bunch but it indisputably is incredibly good at certain things. If I am just trying to get something done and don't need to worry about vulnerabilities as it is just data collection code that runs once, it saves a tremendous amount of time. I don't think it will outright replace developers but there is some room for it to expand the effectiveness of individual devs so long as they are actually providing oversight and not just letting it do stuff unchecked.

I think the larger issue is more how economically viable it is for businesses to spend a ton on electricity and compute for me to be able to use it like this for 20 bucks a month. There will be an inevitable enshittification of services once a lot of the spaces investors are dumping money are figured out to be dead ends and people start calling for returns on their investment.

Right now the cash is flowing cause business people don't fully understand what its good at or not but that's not gonna last forever.

Re: Everything around LLMs is still magical and wishful thinking

#166

Post by 0x696C6961 »

AnotherGoodName wrote: post #111
0x696C6961 wrote: post #50

Earlier quoted context omitted.

I find that the code quality LLMs output is pretty bad. I end up going through so many iterations that it ends up being faster to do it myself. What I find agents actually useful for is doing large scale mechanical refractors. Instead of trying to figure out the perfect vim macro or AST rewrite script, I'll throw an agent at it.

I disagree strongly at this point. The code is generally good if the prompt was reasonable at this point but also every test possible is now being written, every ui element has the all required traits, every function has the correct documentation attached, the million little refactors to improve the codebase are being done, etc. Someone told me ‘ai makes all the little things trivial to do’ and i agree strongly with…

> The code is generally good if the prompt was reasonable

The point is writing that prompt takes longer than writing the code.

> Someone told me ‘ai makes all the little things trivial to do’ and i agree strongly with that

Yeah, it's great for doing all of those little things. It's bad at doing the big things.

Re: Everything around LLMs is still magical and wishful thinking

#167

Post by selfhoster11 »

CharlesW wrote: post #46
louthy wrote: post #26

Earlier quoted context omitted.

> overall my productivity is 10-15% better. That is nothing to sneeze at, but it isn't 10x. It is something to sneeze at if you are 10-15% more expensive to employ due to the cost of the LLM tools. The total cost of production should always be considered, not just throughput.

> It is something to sneeze at if you are 10-15% more expensive to employ due to the cost of the LLM tools. Claude Max is $200/month, or ~2% of the salary of an average software engineer.

In the US, maybe. Several times that by percentage in other places around the world.

Re: Everything around LLMs is still magical and wishful thinking

#168

Post by Loughla »

thunky wrote: post #157
tasty_freeze wrote: post #18

One thing I find frustrating is that management where I work has heard of 10x productivity gains. Some of those claims even come from early adopters at my work. But that sets expectation way too high. Partly it is due to Amdahl's law: I spend only a portion of my time coding, and far more time thinking and communicating with others that are customers of my code. Even if does make the coding 10x faster (and it doesn't…

> One thing I find frustrating is that management where I work has heard of 10x productivity gains. That may also be in part because llms are not as big of an accelerant for junior devs as they are for seniors (juniors don't know what is good and bad as well). So if you give 1 senior dev a souped up llm workflow I wouldn't be too surprised if they are as productive as 10 pre-llm juniors. Maybe even more, because a ba…

The item lost is pipeline of talent in all of this though.

Precision machining is going through an absolute nightmare where the journeymen or master machinists are aging out of the work force. These were people who originally learned on manual machines, and upgraded to CNC over the years. The pipeline collapsed about 1997.

Now there are no apprentice machinists to replace the skills of the retiring workforce.

This will happen to software developers. Probably faster because they tend to be financially independent WAY sooner than machinists.

Re: Everything around LLMs is still magical and wishful thinking

#170

Post by selfhoster11 »

fleebee wrote: post #104
labrador wrote: post #19

I'm a retired programmer. I can't imagine trusting code generated by probablities for anything mission critical. If it were close and just needed minor tweaks I could understand that. But I don't have experience with it. My comment is mainly to say LLMs are amazing in areas that are not coding, like brainstorming, blue sky thinking, filling in research details, asking questions that make me reflect. I treat the LLM l…

I tried the "thinking partner" approach for a while and for a moment I thought it worked well, but at some point the cracks started to show and I called the bluff. LLMs are extremely good at creating an illusion that they know things and are capable of reasoning, but they really don't do a good job of cultivating intellectual conversation. I think it's dangerously easy to get misled when trying to prod LLMs for knowl…

This is very model-dependent. If you use something heavy on sycophancy and low on brain cells (like GPT-4o, the default paid ChatGPT model), you'll get lots and lots of cracks because these models are optimised for engagement.

That said, don't use model output directly. Use it to extract "shibboleth" keywords and acronyms in that domain, then search those up yourself with a classical search engine (or in a follow-up LLM query). You'll access a lot of new information that way, simply because you know how to surface it now.

Post reply on HN