Why Claude Can't Build a 1996 Website (But You Should Try)

"Turns out AI can write React apps in its sleep but struggles with table layouts and marquee tags—maybe we peaked in '96."

Hello Builders! 👋

A developer tried recreating the iconic 1996 Space Jam website with Claude and spectacularly failed. The post sparked 389 comments on HN, hitting a nerve we all feel but don't always admit.

That Space Jam site is pure 90s chaos—nested tables, spacer GIFs, inline styles everywhere. It's objectively terrible code, but it works. Claude kept trying to "fix" it with modern best practices, completely missing the point.

Magic Memory Restores Photos - Upload old photos, GFPGAN restores them instantly, nothing stored server-side. Built with Next.js 16 and Replicate.

Paul Graham AI Agent - Open-source chatbot trained on 120+ PG essays, powered by Nia API for startup and tech advice.

Unlock ChatGPT's Full Reasoning Power - Custom settings to stop the hand-holding and get pure logic mode from ChatGPT without safety theater.

Build Semantic Search with FAISS - Practical guide to sentence matching systems using transformers and FAISS for search engines and chatbots.

Python Chatbot From Scratch - Complete Python tutorial for building a functional AI chatbot with step-by-step code walkthrough.

Gemini's Photo-to-3D Model - Upload building photos and watch Gemini convert them into three-dimensional models automatically.

LLM-Powered Security Agents - Collection of AI agents that automate the tedious parts of security work, from threat analysis to incident response.

AI Language Tutor Hits Revenue - Vibe-coded language learning app just landed its second paying customer, targeting ten users by year-end.

One-Hour Video Processor - Pat Walls built video-to-audio converter with Claude in 60 minutes, claims it would've taken a week manually.

Multi-AI Coding Strategy - Developer spotted running same prompt through ChatGPT, Gemini, Claude, Grok, and DeepSeek then comparing responses in five different files.

Ephemeral Software Economics - Justine Moore argues vibe coding unlocks apps you build for a weekend then throw away—economics finally make sense.

7,371 Lines Shipped - Pat Walls kicked off a lines-of-code measuring contest, because apparently we're doing this again in 2025.

OpenAI's Adult Mode Incoming - Soft launch spotted for less filtered responses, finally acknowledging that developers building real apps need models that don't clutch pearls.

OpenAI Kills Ad-Like Suggestions - Turns off app recommendations after users noticed ChatGPT pushing apps that looked suspiciously like sponsored content.

Apple Locks Down iOS - New privacy controls let users block AI features from accessing app data, because not everyone wants their notes fed to models.

OpenAI's Enterprise Growth Surge - ChatGPT Enterprise usage up 8x year-over-year with workers claiming they save an hour daily, though cost sustainability remains questionable.

Google's Titans Architecture Drops - New approach gives AI actual long-term memory instead of the goldfish-brain context windows we've been dealing with.

LLM Telepathy Paper - Wild research proposes skipping text entirely—let LLMs pass raw hidden states and KV caches directly between models like neural telepathy.

That girl with 5 AI tabs open? She's not cheating—she's just honest about what we're all doing. If Claude can't nail a 1996 table layout, we're definitely allowed to bounce between models until something works.

What are you shipping this week (and how many AI tabs did it take)?

Keep shipping,

P.S. If you're restoring old photos with Magic Memory, please don't let AI make your dad's 90s mustache look too good. Some things should stay pixelated.