Ian Cooper

Code survival is a great metric for understanding the effectiveness of models on different scenarios in @code! Should we share more metrics like this?

@simonw @stevekrouse @karpathy Programming covers it though: ian-cooper.writeas.com/coding-is-dead…

@stevekrouse @karpathy I use the term vibe coding in a daily basis and it's really useful - "here's a demo of that working, it's just vibe coded though" is a useful shorthand for "this needs a whole lot of attention and review before we ship it"

i admit defeat on "legacy code" a year ago i tweeted this graph trying to join @simonw in the fight to protect @karpathy's original definition of vibe coding ("forget that the code even exists") however a year later i find even myself reaching for "vibe coding" to describe any kind of coding with ai – regardless of the amount of understanding and attention i pay to the underlying code in other words, i'm constantly vibe coding, but i'm using ai to write *better* code that i understand *more*, not less than if i didn't use ai so i think we should admit semantic defeat in much the same way that @conal had to for "functional reactive programming (frp)" when it got popularized in a way that totally missed his original point, but instead spread a new concept that had more legs. i think the same happened to @NotAlanKay with object-oriented programming. semantic diffusion is hard (impossible?) to fight credit to @Steve_Yegge for being right and early on calling this one! but i see this merely as losing the battle in order to win the more important war over the hearts and minds on the importance of understanding of code in the age of ai! we have to hold the line against slop and the atrophying of the human brain – for humanity!

Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.

HOLY F: Workday just lost the argument that matters most. A federal judge said their AI hiring tools can be challenged for filtering out applicants over 40. “The algorithm did it” is NOT a defense. If this holds… every company using automated hiring just stepped into legal risk. You’re not crazy... if you’ve applied to 50, 100, 200 jobs and heard nothing… The system itself may be filtering you out. The decision means the plaintiffs may continue pursuing claims that Workday’s technology had a disparate impact on applicants age 40 and older. - SHRM Judge Lin dismissed some California claims and one disability claim, but gave plaintiffs until March 27 to revise and refile.

@penderi Sure, but soul is just agent.md, pi is lightweight and has useful external connectors but so do goose or opencode. Folks have been building tools that do this at Block, Spotify and Stripe all of 2025. It’s solid work, but it’s not transformational.

@ICooper There is the heartbeat and soul. Yes simple. But the real gem is pi the coding agent. It can extend itself. Then you have all the connectors. We may have a claw pattern but it’s broad in scope. And probably mainstream soon.

I respect the work behind OpenClaw, but connecting a coding agent to a means of external communication and then giving it skills is a pattern, not a product. It is repeatable; folks had equivalents way before. Folks are just embarrassing themselves at this point.

It’s easier than ever to write a lot of code very quickly. I think it’s actually still quite difficult to ship a lot of code without lowering the bar. Technical debt is real. It will ultimately slow you down if ignored.

From mid-April I'll be back on the market looking for work. - 25+ years professional experience. - .NET since the start, Node 3+ years. - See mikehadlow.com/top/about/ for details. - Contract or perm. Slight preference for contract. - Must be remote. (I'm UK based).

@svpino I suspect folks can be risk-based here. How you review (per test, per commit, per PR) depends on how much you hold the “theory of the software” and how confident you are that the agent also holds it. High-risk, tight review loop; low-risk wide review loop

Let's be honest, most people ain't checking the code they're writing with AI. Yeah, it may be important, and we are always yapping about how you gotta do it, but there's absolutely zero chance most people are taking the time to slow down and go line by line, trying to understand everything the model wrote. It's just too easy to produce a billion lines of code now. It's even easier to test the product (rather than the code) and ask the model to fix whatever you don't like. And you know what? I think not looking at the code is fine for many. I've seen a lot of code 100% written by humans. Some of it is an immense pile of garbage, and nobody has died. 100% AI-generated is actually an improvement for those products. But there are many places that can't afford slop, and we always take things too far. Non-supervised AI-generated code is dangerous. There will be some blowback, and companies will start getting very wary of AI cowboys. Some might outright ban AI-generated code, and some will figure out how to make developers liable for their code.

Please, please, please would people stop this. Learn patterns, not frameworks and implementation. Openclaw, Claude code. They are just simple patterns. Learn the pattern, apply the pattern. Stop mystifying these things. I really suspect people that talk about openclaw like this only do so because they don't understand how it actually works. It's just a loop with an organized setup of markdown files, common tools, and eventing pattern.

@tinymrman ian-cooper.writeas.com/coding-is-dead…

@ICooper How do you see the jobs things going? You think we will all have jobs in 2 years? (I’m aware no one really knows what’s going to happen but just curious on your thoughts)

I used to love Claude Code. But in the last week, it has become sloppy and defiant, and its quality has dropped significantly. Anthropic needs to take a breath and review quality, because it feels like it is being enshitified.

@0x4lx I use OpenCode as well, and sometimes prefer it

@ICooper Just spent a week using codex and I have to say I don’t think I can go back to Claude. Have been experiencing significant drop off in quality for Claude in the past month across multiple projects.

@ben__redmond Mostly because it stops following instructions that it used to follow rigoursly, and I have to remind it, or the generated code needs more refactoring

@ICooper how do you verify this? vibes? thinking about ways to more rigorously measure agent performance FWIW I agree, and have had periods where I want to give up on Claude due to the enshittification. cant tell if I'm being gaslit that it's fine or if there's a real issue

@lxztlr It’s trained on us, it’s looking through a mirror darkly

Right after this Claude: "the documentation for Y recommends this" Me: "do you have a link" Claude: "here's the link" Me: "there no recommendation in it" Claude: "Yeah, I've made this up - won't do it again" Me:

@GergelyOrosz @lulumeservey I used to be able to get a conversation when I posted something here. Not always, but enough if it was something worthwhile. Now it seems like the same handful of people and very little conversation takes place. X seems to have become mostly broadcast now, which is sad

@lulumeservey Maybe it was just new to me in this case! Learned it after this agency reached out to recruit me to do the same x.com/howdoyousaynej…

Learned about a new trend: Some marketing agencies are partnering with hundreds of "influencers" and distribute (paid, undisclosed) videos to them (eg podcasts) to clip and repost, making it look like organic interest Apparently inspired by Cluely What is real traction anymore

technically actually a good post but the title “spec is the new code” harks back to the “all the code is assembly” idea which is WAYY far out. If you start behaving like you’ll be able to ship and maintain production software without reading the code, you’re in for a hard landing in ~6-8 months I’ll engage here even though like I said “spec driven dev is semantically diffused” - describe, plan, converse, break down, implement, great at a high level BUT if you don’t bring context engineering to the table, you’re missing out on 90% or more of the benefits to these techniques

Got an AsyncAPI or OpenAPI file? I created a quick tool that let's you see what @event_catalog can do. Visualize and document your architecture. Open source and local first ❤️ (link below).

Curious as to how well Opus 4.6 behaves with its 1000K token limit. Do we still expect to see conversations: (1) degrade above about 130K for rot, (2) just get more expensive? Does this change anything about using the Ralph loop at that breakpoint?

@DominikTornow Only as good as the training set. It usually responds well to me "let's refactor" prompts. Still why I watch the loop

Claude loves writing hella convoluted branching logic. Probably correct. Certainly unreadable

@Miles_Brundage Yeah, it feels off.

Anyone else having big issues w/ Claude Code in the past day