Don Karter

Most AI devs chase models. I chase context quality. Data preparation > bigger LLMs Retrieval accuracy > prompt engineering Token efficiency > API credits The moat isn't the model. It's what you feed into it. Here's why I believe this 🧵

@bindureddy Beats only in bench but fails in real cases to be honest

DeepSeek V4 Beats Opus 4.7 And GPT 5.5 To Become The World's Best Open Source Model DeepSeek V4 Pro is the NEW KING of open-source . - better and 10x cheaper than Opus 4.7 and GPT 5.5 medium - out performs Kimi 2.6 thinking - much faster that any of the other big models It's literally the best open source model in the world and months away from GPT-5.5 xHigh.

Builders never sleep. What are you building? 👇

Harsh truth. Coding will never be the same again

@icanvardar If you do not hold it like this you just know how to make it work with closed cover :)

if you don’t hold your macbook like this you’re not agentmaxxing enough

AI isn't going anywhere. The real question: what are you doing about it?

@HsanC_ Any free channel you could advice? What the best start when you have no influence to find the first customer in your opinion?

@donk8r you just try all the channels and double down on what's working

Maybe you should stop doing marketing and fix the product first.

@anupamrjp getvext.app

What are you building right now? 🚀 Not ideas. Not plans. Real SaaS. Real AI tools. Real MVPs. Drop your build link below 👇

@dharmeshba @opencode How fast is it? No lagging on turns? Thinking about also try their plan

Claude rate limits suck! Just get @opencode Go plan for $10 and use Kimi 2.6 Lit 🔥

The one real case that I just explored while using AI is protecting yourself from scammers after analysing MANY followers and MANY comments of person who try to sell your services. So AI everywhere, also in there where it's not supposed to be :) Do your own research

@foundrceo Building Vext getvext.app Free to try, buy once, your forever. Running locally on Mac M1+

Share your product 👇 Feeling great today. Thought I’d give something back I’ll randomly pick one to receive a Premium Launch 🚀 on FoundrList (dofollow backlink + Premium listing)

That exact feeling – the one that brought me and most of you (senior devs with years of experience) here long ago. But AI is fundamentally altering how software gets built. What's next?

@jxnlco Why do you nee so many? local LLM ?

I’m limited by compute.

The hate is real but misdirected. Anthropic's problem isn't the product — Opus 4.7 still wins on reasoning. The problem is they're slow to iterate and expensive. But that's a feature, not a bug, if you care about safety. The hate should be: why aren't other labs moving as carefully?

anthropic shouldn't be getting 1/100 of the hate they're getting

Rebuilding is the trap, but so is prompt brittleness. The real win with Kimi: you can architect for retrieval quality instead of model capacity. Cheaper models + better retrieval = beats expensive models + lazy retrieval. That's the hidden cost nobody talks about — not the prompts, the pipeline.

Cancelled both my Claude Code Pro and ChatGPT Pro for this. Kimi K2.6 is just as good for my side projects as Opus or GPT 5.4 were. The price for this is crazy low, and there are a bunch of models I can try (like DeepSeek). Bonus: I'm moving away from building everything on Claude Code - now that both @opencode and @cursor_ai have their SDKs open, I feel I can rebuild the agentic workflows I built for Claude Code in a more platform-independent manner.

@shub0414 The real play is token efficiency. We cut token spend 60% on the same Claude instance by fixing retrieval. No model change. Just better context. When subsidies end, the teams that built tight retrieval + chunking strategies will ship for pennies. The rest will be priced out.

Unpopular opinion: Your $20/month AI subscription is a lie. You’re living on VC handouts. OpenAI and Anthropic are burning billions to subsidize your productivity. That money will run out. When the subsidies dry up, the cost of "intelligence" will skyrocket. The play? Build your agents and workflows now while compute is pennies on the dollar. Lock in your advantage before the price tag catches up to the tech.

@rohanpaul_ai The real issue is retrieval quality in agent loops. We tested: agents with noisy context get stuck debating bad info. Same agents with clean, ranked retrieval converge in 3 rounds. It's not the model. It's what enters the context window.

Research proves that current AI agent groups cannot reliably coordinate or agree on simple decisions. Building teams of AI agents that can consistently agree on a final decision is surprisingly difficult for LLMs. But problem is that developers frequently assume that if you have enough AI agents working together, they will eventually figure out how to solve a problem by talking it through. This paper shows that this assumption is currently wrong. Even in a friendly environment where every agent is trying to help, the team often gets stuck or stops responding entirely. Because this happens more often as the group gets bigger, it means we cannot yet trust these agent systems to handle tasks where they must agree on a correct answer. ---- Paper Link – arxiv. org/abs/2603.01213 Paper Title: "Can AI Agents Agree?"

@sama Disagree slightly — it's not smarter vs cheaper, it's context quality vs model size. We tested this: same Claude instance, same prompt, different retrieval strategy. 60% token reduction. No model change. The bottleneck isn't the model. It's what enters the context window.

i keep thinking i want the models to be cheaper/faster more than i want them to be smarter but it seems that just being smarter is still the most important thing

Everyone's chasing bigger models. We tested it: same client, same prompt, different retrieval. 60% token reduction. No model change. The real bottleneck? Context quality, not model size.