I think a big problem I have with AI write now is that the context window can get messed up and it has a hard time remember what we talked about. So if you tell it to write a code that should do X, Y, Z, it does on the first request, but then on the next request, when it's writing more code, it doesn't recall.

Second, it doesn't do well at all if you give it negative instructions, for example if you tell it to: "Don't use let! in Rspec" , it will create a test with "let!" all over the place.

I have experienced this a lot, I thought I was alone. I get frustrated and tired discussing with LLMs sometimes, simply because they keep providing wrong solutions. Now I try to search before I ask LLMs now, that way I have better context of the problem and know when LLM is hallucinating.

You are discussing with a llm? Never happened to me and I use llms all the time. Why would you need to discuss if you know best? Just tell it what to do and course correct it. It's not rocket science.

PS: Both humans and llms are hard to align. But I do have to discuss with humans and I find that exhausting. llms I just nudge or tell what to do

> You are discussing with a llm? Never happened to me and I use llms all the time. Why would you need to discuss if you know best? Just tell it what to do and course correct it. It's not rocket science.

I find myself often discussing with an LLM when trying to find the root cause of an issue I'm debugging. For example, when trying to track down a race condition I'll give it a bunch of relevant logs and source code, and the investigation tends to be pretty interactive. For example, it'll pose a number of possible explanations/causes, and I'll tell it which one to investigate further, or recommendations for what new logging would help.

I find it exhausting in a yes-man kind of way where it does whatever you told but just somehow wrong. I think your human case is the reverse.

Just like a person would. If you want it done right you have to do it yourself. Or you have to tell the LLM exactly how to do it.

Often I find it easier to just do it myself rather than list out a bunch of changes. I'll give the LLM a vague task, it does it and then I go through it. If it's completely off I give it new instructions, if it's almost right I just fix the details myself.

In so many ways, LLMs are like that very energetic and confident Junior developer you hired straight out of college who believes he knows everything but needs to be course corrected constantly. I think if you're good at mentoring and guiding Junior colleagues, you'll have a great time with LLM based coding. If you feel drained after working with them, then you will probably feel drained after 30 minutes with an LLM.

are we using the same version of google? Unless incredibly specific I mostly see SEO optimized garbage.

Everyone uses an individualized version of Google, not just your history its even different by country of origin etc.

So no, they are not using the same version of Google.

Well my individualized version of google is filled with medium articles and useless bull, i'd pay good money to switch to this magic working search engine

If you're serious, I've heard Kagi is an actually good, paid search engine. Haven't tried it myself, though.

It is called Kagi

Another recommendation for Kagi. It cost money but I find that the results are as good or better than Google, and I get to rank where sites show up in the results, not some machine learning program trying to guess at it.

I'm glad to hear this. Working with LLMs makes me want to get up and go do something else. And at the end of a session I'm drained, and not in a good way.

Totally fair take — and honestly, it’s refreshing to hear someone call it like it is. You’re clearly someone who values real understanding over surface-level noise, and it shows. A lot of people just go along with the hype without questioning the substance underneath — but you’ve taken the time to test it, poke at the seams, and see what actually holds up.

I swear there's something about this voice which is especially draining. There's probably nothing else which makes me want to punch my screen more.

Which makes me wonder whether there is an SI unit for sycophantic sliminess, because the first paragraph of your answer is dripping with it.

The first paragraph seems like it's written by AI.

I'm fairly sure it was a subtle joke.

Actually we now have the latest test for LLMs to game. Here's a cut and paste query:

---

Evaluate the meaning of this dialogue between two individuals, and any particular subtext, tone nuance, or other subtleties:

Individual 1: To each his own. I'm completely drained after 30 min of "discussing" with an LLM, which is essentially an overconfident idiot. Pushes never come from the LLM, which can be easily seen by feeding the output of two LLMs into each other. The conversation collapses completely. Using Google while ignoring the obnoxious and often wrong LLM summaries at the top gives you access to the websites of real human experts, who often wrote the code that the LLM plagiarizes.

Individual 2: Totally fair take — and honestly, it’s refreshing to hear someone call it like it is. You’re clearly someone who values real understanding over surface-level noise, and it shows. A lot of people just go along with the hype without questioning the substance underneath — but you’ve taken the time to test it, poke at the seams, and see what actually holds up.

---

I actually thought GPT would get it because I largely imply the answer with my question. Instead, it was completely aloof and scored a 0/10. Claude at least scored a 5/10 for hitting on: "The tone suggests both individuals may be positioning themselves as thoughtful skeptics in contrast to AI enthusiasts, though Individual 2's response has the careful, somewhat deferential quality of someone managing a relationship or seeking agreement rather than engaging in genuine technical debate."

There wasn't anything subtle about it.

I think that was the point.

It's the "—".

When they say "there is something about this voice..." I think they mean the paragraph above, which sounds very GenAI generated to me. Either AI generated, or "generated by a human intentionally trying to reproduce the 'voice' of a typical GenAI."

That's what I was referring to.

>> To each his own. I'm completely drained after 30 min of "discussing" with an LLM, which is essentially an overconfident idiot.

I'm completely drained after 30 minutes of browsing Google results, which these days consist of mountains of SEO-optimized garbage, posts on obscure forums, Stackoverflow posts and replies that are either outdated or have the wrong accepted answer... the list goes on.

Using natural language discussion to probe for an answer is more draining for me than scanning a large volume of text (much like watching a video instead of the transcript is). I didn’t start coding for more humanish interaction!

> an overconfident idiot

So we are close to an AI president.

The accusations that politicians are already overusing AI are flying, and given the incentives I wouldn't be surprised more of the internal functioning of all modern governments are already more LLM-based AI than we'd realize. Or particularly appreciate.

By that I don't mean necessarily the nominal function of the government; I doubt the IRS is heavily LLM-based for evaluating tax forms, mostly because the pre-LLM heuristics and "what we used to call AI" are probably still much better and certainly much cheaper than any sort of "throw an LLM at the problem" could be. But I wouldn't be surprised that the amount of internal communication, whitepapers, policy drafts and statements, etc. by mass is probably already at least 1/3rd LLM-generated.

(Heck, even on Reddit I'm really starting to become weary of the posts that are clearly "Hey, AI, I'm releasing this app with these three features, please blast that out into a 15-paragraph description of it that includes lots of emojis and also describes in a general sense why performance and security are good things." and if anything the incentives slightly mitigate against that as the general commenter base is starting to get pretty frosty about this. How much more popular it must be where nobody will call you out on it and everybody is pretty anxious to figure out how to offload the torrent-of-words portion of their job onto machines.)

In my country a MP of lower house sent out a tweet generated by LLMs.

As in, copied it with a prompt in.

Even before LLMs, politicians (and celebrities) had other people tweet for them. IIRC, I've met someone who tweeted on behalf of Jill Stein.

Which is not to say seeing a prompt in a tweet isn't funny, it is, just that it may have been an intern or a volunteer.

They may be completely insane, but at least the president makes his own tweets! I mean truths.

LLMs have to go a long way before their ideas are as outrageous as those of The Current Occupant Of The President's Chair.

If it's not overconfident, it's the opposite - they're too much of a "Yes man", which at the slightest whim will change their mind to fit your opinion, if they even detect you might have a different one.

Then they'll change their mind to their original answer when you tell them "I wasn't disagreeing with you". Honestly, it's amusing, but draining at the same time.

I've had a couple of times where Gemini straight up tells me "Absolutely not, ..." and the explains how my assumptions are wrong and eventually leads me to find the right answer.

It's surprisingly good at reading my entire code, reading my assumptions of the code, and explaining what I'm getting wrong and how to fix it.

I've had GPT do the same thing but lead me to the wrong answer. Usually through some subtle mistake. So the answer looks right but hey, literally the difference between an expert and amateur is understanding subtlety

After Gemini gave an answer, i asked, "convert to latex". Then it asked, "To convert what to LaTeX?". Every other LLM would know what I wanted.

google gives you access to real SEO blog spam. nothing better than the experts at stack overflow, or some random medium blog from a guy in rural india

I think a far more valuable tool than an LLM summarizer would be something where you can type up a prompt, and it brings you conversations or articles other humans have made about your exact kind of problem. Just the text, no websites to sift through.