Aarav

hey everyone :) just shipped a refactor to my portfolio. cleaner layout, better project sections, and a lot less visual noise. check it out: aaravkashyap.live would love suggestions, reviews, or anything you think I should improve.

AI can build a landing page in 10 minutes. taste is still the bottleneck. what’s the first giveaway that a site was vibe coded?

Once upon a time !

@samirande_ don't they give u the permission to write in any language that u prefer. I mean it is a test to evaluate logic building instead of language syntax right?

Ok so I've to learn DSA in C to pass in exam

@buildwithsid use make.design to avoid the slop.

common issues with vibe coded websites they all just look the same llms are good for structure + layout but they SUCK at taste some common things i found:- 1/4 - black is never actually black, its just a dark version of blue or purple same with the cards, buttons, inputs, everything gets tinted blackish blue that looks horrible, real black creates contrast and looks much better 2/4 - everything just has different colors blue next to pink next to green nothing feels important and everything is trying to grab your attention at the same time pick one accent and rest neutral 3/4 - emojis everywhere 📦 🔥 ✨ 🚀 ⚡ emojis as icons, emojis as bullets, emojis in navigation, emojis in titles like bro what 😭😭💀 4/4 - box inside box inside box everything is just boxes inside boxes, like bro borders exist for a reason use grids n borders for separators. avoid boxes unless needed, even try to avoid borders also avoid common rounded patterns ------ the LLMs learned from billions of examples, and it is just predicting the next text its just that we need to write more detailed prompts, like literally just say everything we also need to use good skills.md basically we need to know what taste is, if you don't have taste, no prompt is going to save you

DeepSeek-V4-Flash got 2x faster on Ollama Cloud overnight. benchmarks get attention. latency decides whether an agent stays in the workflow. what makes you switch models: 2x speed, half the price, or better code?

@jared_mitosis yes

@byaarav memory. agents need to remember what they ran two days ago, what worked, what didn't. most systems treat each invocation as fresh. that's the fight you feel when the agent seems smart but keeps forgetting.

Anthropic's AI found 23,019 possible vulnerabilities. humans reviewed 1,900. 90.8% were confirmed valid. finding bugs is getting cheap. would you run this against every dependency before shipping?

@jared_mitosis true, the blob doesn't need to be a part everytime.

@byaarav Raw tool output. Store the decision it led to, not the JSON blob that produced it. The blob just rots your context window every time it gets recalled, and you almost never need it verbatim twice.

every AI app wants a "second brain." i don't need it to remember everything. i need it to forget the garbage. what do you never put in agent memory?

@github agents made giant diffs cheap. stacked PRs make them reviewable.

Large changes are hard to review. Smaller pull requests make them easier to ship. Stacked PRs are now on GitHub. docs.github.com/en/pull-reques…

@ollama 2x overnight is the kind of benchmark users actually feel.

DeepSeek-V4-Flash-0731 is now over 2x faster than yesterday on Ollama's cloud!

@RoundtableSpace one design image becoming the entire 3D system is absurdly cool.

Someone vibe coded a fully interactive 3D human anatomy app using GPT-5.6 Sol and Codex, and it started from a single design image. GPT Image 2.0 generated the design. Every organ image was generated from it and converted to 3D models through Tripo AI. Codex took those models, a master prompt and the design image and built the first version. The models were nearly 150MB each and running at 16fps, so Codex optimized everything down to 2 to 5.5MB per model and reduced total assets from 900MB to 28.6MB while keeping the visual quality intact. Codex also generated the anatomical illustrations showing where each organ sits, and created the interactive hotspot markers explaining different parts of every organ. Live and fully open source.

@mattpocockuk 200k stars is no longer a repo. it's a small country.

Thanks for 200k stars folks github.com/mattpocock/ski…

@thsottiaux the distribution is the real superpower. one PR, a billion users.

Crazy thing about OpenAI is you just open your laptop, codex a PR into existence and ship an improvement to 1B users. Just like that.

found an open-source repo that turns Codex or Claude Code into a full video production studio. 12 pipelines. 100+ tools. approval gates before it spends money on assets. 44.7K stars already. would you test it on an explainer or a product demo first? github.com/calesthio/Open…

@Itstheanurag My end goal

@byaarav Father and grandfathers used to be farmers, then after my birth father shifted to some different work.

Nowadays there are a lot of people in gym, I don't even feel like recording videos ( it feels weird right now, gotta fight this feeling ) Btw legs, triceps and walk done.

@Itstheanurag Aur genetics?

@Itstheanurag Calves routine share krdo ab jldi se

@Vasu_Devs Damn

@byaarav lol even more but now they are talking about opensourcing and haulting AI

Best thing I heard today 💯 If there weren’t any Chinese competitors, there would probably be $500 AI plans by now. That’s why balance of compute power is necessary.

most dev tools still assume the user has a mouse. agents don't want dashboards. they want an API, clear errors, and permission boundaries. what tool in your stack still fights automation?

@nykdotdev do you version eval fixtures beside the prompt too?