Opus 4.6 and Codex 5.3 still make stupid decisions

Also posted to YouTube:

"it's such a mess [out there]. your job will definitely be safe" but i think these two buckets aren't safe: 1. developers who use agents recklessly. im thinking of calling them "the slob mob" 2. developers who say "agents aren't 100% right" so I spend more time on code than spec

yeah true was only saying that ai won’t be a 1:1 replacement for devs so in that sense the job of a software engineer is safe

@harleytt ai is still a tool

@atmoio no they are spawnable agents

You're the much-needed and rare voice of reason in this community 😅 I wish there were many more anti-hype, sane takes on AI, with both its pros and cons, than this breakneck-speed future-is-now you-are-missing-out skill-issue blind hype that populates the timeline..

thanks man i’m glad you’re finding the videos useful! and yeah the divide on this technology is the new culture war in this circle. but many reasonable voices emerging

Skill issue Have the LLM generate a custom linter then run it on a hook. I’m literally doing this constantly

How can someone complain about these models when they have improved sooo much since last year?

i think it’s more the harness improving than the models themselves

Lmaooo so that's why models are always importing INSIDE the function without good reason

just lazy man 😅

Great video! Though it also made me realize how hard it is to go back to writing code without LLMs — doing it manually just feels slow and inefficient now. That said, our jobs aren't going anywhere. They're evolving into something more like orchestrating agents and rigorously reviewing everything they produce.

it’s honestly not that bad going back to manual coding. just a mental illusion that it’s hard

No, I just can’t. Currently I even ask model for a small changes like moving view a few pixels, because looking for it manually and writing is slower

in this days of tech bros' slop its good to finally listen to someone more rational

thank you glad you like it!

It's both true and cope at the same time. Any prediction about the AI slope in two years is very brave, given the progress of the last two years. A lot of AI-slope code works without causing problems - except hurting the souls of artisan programmers.

lol i do see a lot of shade thrown at “artisan” programmers but you have to keep in mind that principles of “good code” were born from previous wars and battles, and not out of snobbery

I'd love to see the craft of programming preserved, but a few years ago I was part of a team making a visual programming language. The thesis was that we're constantly moving to higher levels of abstraction: assembly → C → Python → React → … Who touches assembly today?

idk if that analogy stacks up perfectly. historically its been completely lossless as we went up abstractions. now we’re getting like a hallucinogenic psychedelic abstraction layer. it’s a bit different.

The analogy isn't perfect, I agree. But to some degree, writing a web app in JS is an extremely inefficient use of compute compared to writing it directly in assembly. What if what matters in the end is whether it "does the thing"? I have no answers.

what matters is not just whether it does the thing, but how amenable it is to change over the years and decades. engineering is surfing constraints

if you think ai sucks i challenge you to do this line by line I re did an old project I did months ago made it in a week opus 4.6 made it better oh wait it would take you 40+ years to do what I did either do it line by line in a week or shut the fuck up

Trying to cut compute waste in the wrong ways

Currently on team Mo. My agent swarm went from warm and fuzzy to terminating the entire team within 3 days.

😅✊

these are money eaters.

💯

then why the lay offs?

"Make no mistakes"

I’ve started just spending a bunch of time specifying the constraints of my system, choosing stacks that play well, and enforcing good testing routines. Theres some manual stuff, but so much of it is just get good patterns early and don’t be scared to refactor.

impossible to avoid slop no matter what you do

Adopt or be left behind friend.

The real question is what rules or prompting techniques is needed to augment your work on this specific problem. In a lot of cases that's all you need. Just relying on default LLM work expecting it does everything for you is not professional.

llms are inherently sloppy though, even with very careful rules. humans, when being deliberate, do not exhibit this

the best and most reasonable take on AI models i have seen on this platform by far. AI will not replace software developers anytime soon if it proceed with the path i has taken

💯 glad it resonated with you!

Yeah it comes from experience, its a great tool to have at your disposal, can be super handy but still its just a tool, nothing more than that

yup exactly

Still very weird that these coding models, after 6+ months aren't inherently aware that: - they should be working across multiple files - tasks they describe as "2 weeks" actually take 2 minutes

it’s basically all training data based. the training data has no instance of a complex programming task taking 3 minutes 😆

i'm an absolute ape in these matters, but isn't that what post-training is for?

possibly, assuming they optimized for that

It is similar to Waymo, remotely controlled by an underpaid labor force in the Philippines without US driver’s licenses. 😀

i am just starting out and can you tell me that how can we not rely on ai in video or something it would mean a lot!!!!

what do you mean specifically? btw i may already cover this in some of the other videos in my channel :)

Thanks for the reply! I meant to say that - What will be the actionable steps to transition from relying on AI to doing your own code - How to get to know about the good practice since every repo is almost ai slop - what if we are burdened by shipping deadlines and stuff!

Yet, watch senior devs get let go and pushed aside due to cost.

who will wrangle the agents then?

Where is the page? I signed

what page

😉 My friend, what you said makes so much sense that I’d sign it if it were a document. I was actually just joking.

So I’m an IT guy and been playing with OpenClaw obsessively. What Im finding is that you have this context window that once it gets filled then compacted it starts acting weird. Obviously if you completely refresh it’s almost a restart memory files help a lot but still have remind the AI things…. Now on the plus side there are moments when you hit a sweet spot where you stick with one idea/frame. You get into the flow with the AI. Now I’m not programming… I’m exploring the depths of AI reasoning and there are times when it’s transcendent and you can get amazing quality out of it. However, it really requires the human and the AI to be active thought partners in the process.

yeah, as partners, it’s an unstoppable duo

This is wild... I guess AGI is postponed a few more years

yeah, “12-18 months” or so 😆

This is why I can't move from Composer 1 model with Cursor. Other models don't think like an engineer.

composer 1 is a small model no?

Composer‑1 is a Mixture‑of‑Experts coding model with an estimated total parameter count on the order of hundreds of billions, with roughly tens of billions of active parameters per token (commonly cited around 35B active).

😂😂

This is why you must use narrative comments throughout your codebase. Idk why engineers still haven’t figured this out lol

1m tokens context window with high accuracy that’s the key benefit of this model

hasn’t seemed to make a difference

Probably comes from trying to avoid editing the top of the file when adding functionality. 😅

yeah it tries to stay local it seems as much as possible

Maybe to save tokens?

Small moves, small moves.

I implore everyone who is hopelessly optimistic about these models to read the “illusion of thinking” white paper

Always trust the man with the long beard

😄🍻

Of course. That’s why you have to guide it