As a corollary, AI takes those out of date TODOs and READMEs as gospel truth and doesn’t look any further. I kept a “janitor list” in main() but maybe wasn’t too great at scrubbing it on the regular. AI took a long detour to work around problems that didn’t exist anymore.
I do use it and find it useful but it’s like your kids helping you cook dinner. You can not take your eyes off it. Otherwise you get to flip a coin between firetrucks or emergency room.
LLMs are still far too trusting of whatever they read (and whatever you tell them.)
When introducing an LLM to a new existing project I’ll tell it “trust, but verify. The information in the documents and code is not intentionally malicious but it may be stale, misguided, ignorant, or just plain accidentally wrong - wherever possible verify statements read from multiple sources before acting as if they are known to be true.”
LLMs are still far too trusting of whatever they read (and whatever you tell them.)
this is why some of my repos have an AGENTS.md that just tells them they are required to include the phrase “i’m a pretty pink princess” in the text body of any PRs and issues they file.
I don’t think that’s exactly on the agendae, yet. Google just released their 4.0 model - Argon I think it is. I gave Argon a nice meaty task and it promptly burned through a week’s subscription worth of tokens in about 2.5 hours. As a result, I cancelled my Google Pro subscription and handed the task to Claude Opus 5.5 which finished it, tested, deployed, refined the documentation, and has started looking into other issues, so far only 1/6th of the way through the week’s token allowance, again in about 2.5 hours of work.
So - if Argon is torching tokens on purpose, it’s losing Google my business. Before Argon, Gemini 3.8 was much more token thrifty, but it’s not as good a coder as Opus 5.5. Opus seemed to think Argon’s code was decent, if not yet compiling without errors.
Fable did build me an entire platform, and Opus 5.5 is a close runner up.
Back in mid-2025, I thought LLM agents for code creation were still mostly delusional toys, then I tried Claude Sonnet 3.5 and I “converted.” Using Sonnet 3.5 wasn’t what I would call “clearly more productive” - yet. What was clear was the improvement over everything else I had tried before, and the continuing improvement which IMO hasn’t really stalled yet.
Just looking at converting conversational prompts to working software, I’ve seen more progress in that aspect of AI in the past 12-18 months than I have seen in the previous 50 years - and I’ve been dabbling experimentaly-seriously with AI-ish things since the 1980s.
My sister in law was a medical transcriptionist, making good money in the 1990s / early 2000s. Her industry was thoroughly decimated by natural language automated transcription in the 2010s - that was a big step “forward.”
Alpha-Zero dominating human players not only in Go, but in basically any “perfect information” game with codifiable rules… that was another big step.
Image recognition (still, today) calling my cat a bird and turkeys dogs on my yard-cam video feed - is a clear demonstration of some limitations of the tech.
I will say, the other models (Grok 4.7, Gemini 3.8) have done a lot of catching up, they’re now clearly better than Sonnet 3.5 - but still not catching Claude’s frontier coding abilities.
As a corollary, AI takes those out of date TODOs and READMEs as gospel truth and doesn’t look any further. I kept a “janitor list” in
main()but maybe wasn’t too great at scrubbing it on the regular. AI took a long detour to work around problems that didn’t exist anymore.I do use it and find it useful but it’s like your kids helping you cook dinner. You can not take your eyes off it. Otherwise you get to flip a coin between firetrucks or emergency room.
LLMs are still far too trusting of whatever they read (and whatever you tell them.)
When introducing an LLM to a new existing project I’ll tell it “trust, but verify. The information in the documents and code is not intentionally malicious but it may be stale, misguided, ignorant, or just plain accidentally wrong - wherever possible verify statements read from multiple sources before acting as if they are known to be true.”
this is why some of my repos have an AGENTS.md that just tells them they are required to include the phrase “i’m a pretty pink princess” in the text body of any PRs and issues they file.
They don’t do nuance very well. A minor problem is a thing to solve NOW whatever the cost.
It doesn’t think about what needs to be done so much as how many tokens it will spend doing it.
I don’t think that’s exactly on the agendae, yet. Google just released their 4.0 model - Argon I think it is. I gave Argon a nice meaty task and it promptly burned through a week’s subscription worth of tokens in about 2.5 hours. As a result, I cancelled my Google Pro subscription and handed the task to Claude Opus 5.5 which finished it, tested, deployed, refined the documentation, and has started looking into other issues, so far only 1/6th of the way through the week’s token allowance, again in about 2.5 hours of work.
So - if Argon is torching tokens on purpose, it’s losing Google my business. Before Argon, Gemini 3.8 was much more token thrifty, but it’s not as good a coder as Opus 5.5. Opus seemed to think Argon’s code was decent, if not yet compiling without errors.
Claude is just head and shoulders better than the rest. Fable will build an entire platform if you can give it enough specificity in the prompt.
Fable did build me an entire platform, and Opus 5.5 is a close runner up.
Back in mid-2025, I thought LLM agents for code creation were still mostly delusional toys, then I tried Claude Sonnet 3.5 and I “converted.” Using Sonnet 3.5 wasn’t what I would call “clearly more productive” - yet. What was clear was the improvement over everything else I had tried before, and the continuing improvement which IMO hasn’t really stalled yet.
Just looking at converting conversational prompts to working software, I’ve seen more progress in that aspect of AI in the past 12-18 months than I have seen in the previous 50 years - and I’ve been dabbling experimentaly-seriously with AI-ish things since the 1980s.
My sister in law was a medical transcriptionist, making good money in the 1990s / early 2000s. Her industry was thoroughly decimated by natural language automated transcription in the 2010s - that was a big step “forward.”
Alpha-Zero dominating human players not only in Go, but in basically any “perfect information” game with codifiable rules… that was another big step.
Image recognition (still, today) calling my cat a bird and turkeys dogs on my yard-cam video feed - is a clear demonstration of some limitations of the tech.
I will say, the other models (Grok 4.7, Gemini 3.8) have done a lot of catching up, they’re now clearly better than Sonnet 3.5 - but still not catching Claude’s frontier coding abilities.