Anton Kuratnik | AI Nerd

Just rage quit Claude Code mid-project. Horrifying performance right now. I wasn't fully on this Codex hype train before, but the lack of reliability from Anthropic is really driving me in that direction.

@vasuman Yep! Super dumb right now. Must be a new model drop coming

they lobotomized opus again lol (not joking)

Claude definitely dumber in the past 24 hrs. Both Code and Chat.

@trying_to_exits Wind. Animals. Rain.

what's stopping u from coding like this ?

Opus 4.6 hallucinating hard on me today. Sigh. Probably another update by Anthropic that they'll deny for 3 weeks and then do a post-portem. New model dropping?

@KaiXCreator I replaced 5 positions (or rather just didn't hire for them) with AI. To be fair, I couldn't afford 5 actual positions, but they were all needed for my business. In terms of actual people who didn't get jobs, I'd say 2.

Has anyone actually been replaced by AI or is it just hype?

Let's say arch is "solved". Let me paint a picture for you: Noob vibe coder: "build me an app that does A, B, C!" AI: *designs solid architecture based on reqs* AI: *writes solid code based on architecture* User: wait! You didn't add massive feature D! I Also want feature D! Feature D, if included in the original request, would've required an entirely different architecture. Option 1: AI: *spends 5M tokens rebuilding the architecture + rewriting the code* User: what?? AI sucks, it spends so much time and costs so much money! Option 2: AI: *hacks on the feature despite poor architecture fit* User: Now add feature E! AI: *adds feature E, causes 10 bugs due to bad architecture* User: I'm spending more time debugging than I am building! This sucks! The real issue is that most people can't think through what they want properly. So, what do we do? Have AI force-interview you on the specifics? Users will get annoyed too. The reality is the human factor is the limiter here, not the AI. The AI needs to improve too, though. Architecture isn't the problem, it's just that it's a big project: requirement review, big-picture planning, detailed planning, research, review, revisions, user input, etc etc etc. This requires long agentic runs with tons of considerations and a huge context window. This is obviously being solved right now. It might be solved within the next 12 months, who knows. But then we'll still hit the human problem.

The last bastion of argument here. You’ll have agents spinning up simultaneous arch stacks in realtime soon and constantly optimize performance. If it knows what good design looks like it will know what good architecture looks like. In the meantime you should build and get users

Serious question. For 20 years, a "Software Engineer" was someone who spent thousands of hours mastering complex syntax, logic, and architecture. Now, a 19-year-old can vibe-code a production-ready SaaS in a weekend using plain English and a $20 Claude subscription. What does the title "Software Engineer" even mean right now?

@OpenWeb360 @DanielSmidstrup China is way beyond copying at this point. Some of the most important breakthroughs in LLMs have been from China.

@DanielSmidstrup USA builds China copies fast Europe regulates

USA has ChatGPT USA has Grok USA has Claude USA has Gemini China has DeepSeek China has Qwen China has Kimi China has MiniMax Europe has?

Honestly, anytime I have Claude plan or write code. Always run the final product by GPT 5.5 and it always finds issues/fixes. Claude is great at thinking/big picture stuff, it's terrible at details and downstream effects. Codex usually catches stuff like "if you change this it'll break this other thing"

When do you reach for other models instead of Claude? What can we do better? Hit me with all of your frustrations. dms open. If you can give me detail (e.g. specifics/transcipts) - it'll help a lot in finding out exactly what we need to do to improve the next model

@49agents @abhishekgawade_ @zuess05 If you're gonna have AI post for you, you might want to give it an up to date landscape of the AI world.

@anton_onAI @abhishekgawade_ @zuess05 running claude code plans through gpt is smart cost optimization. opus is great but 4o is decent for reviewing and catching what opus misses at a fraction of the cost

@abhishekgawade_ @zuess05 For the coding? I primarily use Claude code but nowadays I run its plans and output by gpt. Cursor is nice too

@anton_onAI @zuess05 Makes sense, so do you focus on any particular set of tools, or can you change as per client needs or you can learn on the fly

@svpino How about neither.

I still don't know which one is better: OpenClaw or Hermes. Every time I ask somebody, I get a different answer. This is looking like a "Claude Code vs Codex" discussion: everyone has a favorite. If anyone has a strong opinion, please let me know.

Working on a complex coding project right now, but it's an internal team product for a client. Exposure is low, so I'm confident I can build it. I wouldn't try to build a public-facing SaaS for example. I'm not sure what you mean by the second question. If I'm not sure how to do something, I ask AI. A lot of my software architecture learning happened while building software. I ask AI what the best way to do it is, it gives me options, I ask it to explain each one with pros/cons. I sometimes go down rabbit holes to understand the system from the ground up before coming back to the original discussion. Basically, learned by doing.

@anton_onAI @zuess05 Do you take on complex projects, and how do you manage tasks when you’re not 100% sure?

@Av1dlive This makes me LESS confident in their product, not more.

Cursor pays engineers $1,100,000 a year to run teams of AI agents that ship code while they sleep. [The CEO of Cursor explained in 9 minutes how they ship at 100x speed using team of agents] ↓ Save this before everyone copies the playbook 1. Engineers no longer babysit one assistant. They manage dozens of agent colleagues working in parallel, each on its own remote machine 2. Validation contract before code, not after. Humans only at scoping and review. 3. The agent team handles the full loop : planning, coding, testing, shipping PRs with each agent specialised for a role. Watch the guide. Then read the guide below by @eng_khairallah1

@abhishekgawade_ hmm. that could go either way. i actually really like cursor, i hope it keeps going in the right direction

@anton_onAI spaceX

Have you ever seen an awesome new software project get purchased by a major corporation and have it become better? I don't think I ever have.

@abhishekgawade_ Did they get acquired by someone?

@anton_onAI yeah but i can bet on Cursor

This vibe coding thing is going to end in a massive vulnerability scandal. Don't get me wrong: I absolutely love vibe coding. I've built tools for myself and my team that we've been using for months and it's amazing. I run my own servers, deploy open-source software, create quick automations that make my life way easier. But I know my limitations. I can't read code. I can focus super hard on ensuring security, but that only goes so far when AI is doing the writing. I'd never try to build SaaS with Claude, not without a human dev team to back me up. But we're seeing massive projects run agent swarms to do their bug fixing and development. The industry is growing super fast and there's huge pressure to keep shipping new stuff. Everywhere I look, previously stable tools are pushing updates that ship 3 new features and break 10 older ones. People are adopting AI tools (some of them WITH KNOWN EXPLOITS) at lightning speed without thinking about security. At some point, there's going to be some massive hack that's going to make enterprise and/or consumers way more cautious about AI. I really think it's not a question of "if", it's a question of "when". In the meantime, anyone who's deploying AI needs to think security first and foremost.

@jijoya Hmm. I actually use Claude the way you use Gemini. I just ask it to "interview me one question at a time" or similar. For brainstorming fiction or code it's great. We go one topic at a time, discuss each in necessary detail, then it compiles the final decisions.

@anton_onAI I always take THOSE discussions (regardless of project) to Claude and ChatGPT too because then, the bullet lists and the „time for you to go and get it done now, bye“ attitude are very useful. Gemini is where I get things going though, every time.

Their recent slip-ups notwithstanding, Anthropic (Claude) is still my daily driver today. A few reasons: - Best personality: I can't stand GPT-5 models. Claude is pleasant to talk to and, more importantly, is very good at teaching. ChatGPT struggles to balance being too technical and too dumbed-down. - Best search: but in Claude Code, not in Chat. It's the most thorough searching I've seen an LLM perform. Slow but thorough. I've tested it head-to-head against Codex. - Hard code: GPT 5.5 is definitely more reliable and better at following instructions. But Claude is a better brainstormer, planner, and solver of hard puzzles. If Codex can't solve a problem, Opus usually can. - Writing: to be fair, GPT 5.5 has caught up a LOT in this department. I'd now grade them equal (but with different skillsets) in both copywriting AND creative writing (a side hobby of mine) ChatGPT: I primarily lean on it for writing (copy & creative) and code. It's become my go-to for most code implementations + a reviewer for whatever Claude built. Gemini: Not a model I use a whole lot... except for writing email marketing copy. Thanks to being trained on ridiculous amounts of GMail data (probably), nobody writes better emails than Gemini 3.1 Pro. Not by a long shot. Also very good at copying someone's writing voice.