larsfaye · HackerTrans

قبل 11 يومًاdiscuss

Uh, he's saying it because the kid is overly confident and acting as if he knows Williams' life because he saw a painting and drew a bunch of conclusions, so he was providing a life lesson to realize that experience provides a level of understanding that book knowledge will never be able to capture. And then he even re-affirms this to Will and emphasizes that he can't possibly know what makes him tick because he read Oliver Twist. Your tangent seems entirely unrelated, IMO.

larsfaye

قبل 22 يومًاdiscuss

I actually agree with you; I definitely feel I'm able to use LLMs to learn and explore concepts, but I've always been a self-taught and highly motivated person in the first place. Everything technical that I know, I know because I put in the work to learn it on my own. That's why I wish these tools were being advertised as models that help you do BETTER work, not MORE work. They're using them as excuses to lay off swaths of workers, instead of allowing them to uplift people's skillsets. And, of course, it invited a whole mass of individuals who use them to artificially elevate their perceived skills.

larsfaye

قبل 22 يومًاdiscuss

Oh ho ho....looks like I'll need to update my essay! The evidence is mounting higher and higher.

https://larsfaye.com/articles/agentic-coding-is-a-trap

larsfaye

قبل شهرينdiscuss

This is exactly how I feel. I've begun referring to them as a "Delegation Layer", which is a new addition to the stack.

larsfaye

قبل شهرينdiscuss

Not only can it not describe its reasoning, it can't "remember" if you ask it later; it can only observe what is. Nor can it be consistent; I've had it shift reasoning numerous times as the questioning continues, only to come full circle to its original statement while it apologizes profusely for being misleading.

The model will always be completing the story you start with it. There's no opinion to uncover because there's no experience that occurred. It's impossible to know where your influence ends and the model's factual basis begins.

larsfaye

قبل شهرينdiscuss

Yes—this is the part that nobody is telling you. That's the real shift that is happening.

larsfaye

قبل شهرينdiscuss

The "don't look at the code" movement is a fad that we simply have to go through to prove that the real problems were never obvious, but lie in the subtle nuances and interplay between all the moving parts of what makes software function in the first place. We have to abuse the tool before we learn to properly use the tool.

larsfaye

قبل شهرينdiscuss

I do agree with this. I was able to get exactly what I needed even out of GPT3.5. If you put enough parameters and examples, along with a real solid system prompt and (if you can) proper temperature and topK/topP, there's no reason they can't basically function like "smart typing assistants". The issue is that its a sliding scale of ambiguity to chaos. The more ambiguity, the more the LLM fills that in. And it can be very difficult and time consuming to know when you're being ambiguous or not (you don't know what you don't know, OR, you can't track what you aren't tracking).

Depending on the task, it can sometimes be just as arduous to produce enough guidance and guardrails to get the LLM to output exactly what you need that you can trust without issue or extensive review than it is to write it yourself and use the LLM just for ad-hoc generation. It's a constant balance and an endless amount of micro-decisions, honestly, but it's pretty essential to stay engaged and not YOLO with agents the way so many are. Most of my interactions with models these days are done in pseudo-code.

larsfaye

قبل شهرينdiscuss

> The claim you should know everything about everything you work on is an intensely naive one.

Author here. Where did you find I was stating that? As other users said, that's not at all found in my writing. The rest of your post goes on a tangent about this notion, but seems like its more of a personal pontification, rather than a critique of anything I wrote.

larsfaye

قبل شهرينdiscuss

Author here. Thank you. I absolutely was not stating that, and I don't think its possible or necessary to "know everything", but there's certainly a movement in the industry (and not a small one) that is advocating to abandon even looking at code. It started with Karpathy's "vibe coding" tweet and has extended to now where entire teams are performing "LGTM" level code reviews, while simultaneously reporting that they are finding it difficult to remember how to read, nevertheless write, code properly. "Thinking in code" is a form of thinking that is distinctly different than staying at the higher engineering levels of planning and architecture, and yet they are completely intertwined and interdependent. I plan better when I am engaging with code, and I code better when I know how to plan properly.

larsfaye

قبل شهرينdiscuss

Yeah, I can't help but chuckle when people say "Well, humans make mistakes, too".

That's what we're spending 7+ TRILLION dollars, destroying ecosystems to build datacenters, and ruining society's social contract on truth and employment for? To build something that produces the average quality of a human, all while making the same types of mistakes along the way?

Sounds like a shit deal, really.

larsfaye

قبل شهرينdiscuss

Author here....

"An increase in the complexity of the surrounding systems to mitigate the increased ambiguity of AI's non-determinism"

I'm referring to the layers of review that are needed to be put in place to reign in the fact that the code that is generated is blurry and obfuscated, and in amounts that exceed what someone can review in one or two sittings. I'm seeing multi-phase AI review stages that will try and distill review finer each time. Then another agent layer to document. Then another to create a PR. Then the human reviews, but uses Coderabbit locally. Then sends it back into the pipeline and round and round we go, all because there's simply too much volume of code to review.

larsfaye

قبل شهرينdiscuss

Author of the article here (had no idea someone shared it! Just found out through analytics).

I actually mention this exact thing it the article under the section "LLMs accelerate the wrong parts", which seems to be saying exactly what you're saying:

https://larsfaye.com/articles/agentic-coding-is-a-trap#llms-...

I would find it hard to believe there is any developer in history ever uttered the words:

"I really wish I had a tool that could generate code I don't understand, and at a rate faster than I can review".

comments