EK @ekcheungAI

Google just released a free 8-minute tutorial on building your first AI agent. 🆓 I’ve seen $500 courses that explain agents and loops less clearly than this. If you’re diving into AI agents, this is the clearest starting point I’ve found.

Forget robotic text-to-speech. 豆包 Seed Audio 1.0 generates audio from imagination — human voices, music, sound effects, even ambient noise. It captures the tiny details that make sound feel real. They're calling this the 'Seedance moment for audio.' Seedance gave images intelligence; this gives sound the same. That's not just an upgrade — it's a new category.

Codex is the best agent harness right now — and if you're only using it for coding, you're missing the point. I needed a newsletter landing page today. Told Codex what I wanted, had zero expectations. It autonomously: - Wrote the code - Pushed to GitHub - Opened Vercel - Created a new project - Connected the repo - Picked the right domain from my list 5 minutes later: fully live landing page. Here's the workflow I'm using: 1. List every computer task you do 2. Before doing it, start a new Codex chat 3. Ask how it would do it 4. Enable browser use + computer use 5. Run it and see how much gets done 6. Schedule repeats as automations Almost all knowledge work is about to become automated. Are you experimenting with this yet?

Source: @dotey on X

Codex/Claude Code's context compression is legit now. With Prompt Caching, a single session can go deep without breaking the bank. I keep tasks in one thread more and more. Two underrated features: 1️⃣ Fork — branch off from any message. Keeps only the history before that point. Cleaner context, no baggage. 2️⃣ /btw or /side — drop a side question without polluting the main context. Perfect when you're in plan mode, stuck on a confusing option, and just need a quick explainer or recommendation. Ever used /btw mid-session? Which one do you lean on more?

Source / full context:

Karpathy dropped a truth bomb: people who don't use LLMs are already losing. But here's the twist — subscribing to Claude doesn't count. Typing prompts and reading outputs? Anyone can do that. The real skill is building something that runs without you. I wrote a deep dive on Claude features most people miss. If you're serious about staying ahead, it's worth a read. What's one AI workflow you've set up to run on autopilot? 🤔

Big move: Anthropic got US approval to bring back Claude Mythos 5. 🛡️ It's their strongest cybersecurity model, now cleared for US critical infrastructure orgs. Similar to OpenAI's GPT-5.6 Sol exception — non-US nationals in approved orgs can also use it. Anthropic is also working to expand access and bring back Fable 5 broadly. No timeline yet. What do you think — is this the right balance between security and innovation?

Hermes Agent v0.17.0 — The Reach Release just dropped. Big version jump. Always exciting to see what NousResearch ships next. Anyone peeked at the changelog yet? What's the most interesting change?

Source / full context:

ByteDance just open-sourced DeerFlow 🦌 72,000+ GitHub stars. MIT. Free. It doesn't just run tools—it does the whole job. Give it one task: it plans, spins up sub-agents, writes code, tests, fixes errors, and delivers finished work in its own sandbox. Research, full websites, dashboards, slide decks, reports—done, not drafts. Setup takes ~5 min: install basics (git, docker, node, uv, pnpm), clone, `make setup`, `make doctor`, `docker-init`, `docker-start`. Or paste a prompt into Claude Code/Cursor/Codex. I've been all-in on Hermes (224B tokens/day on OpenRouter), but DeerFlow makes me want to switch. Hermes runs your tools. DeerFlow runs your project. Which one are you running? 🤖

Over the next 3 years, here's how I rank the AI bottlenecks: 1. Memory (HBM/DRAM/eSSD — already priced up, long-term contracts signed) 2. Optical Interconnect (800G/1.6T, scale-up/across — connecting compute) 3. Advanced Packaging (CoWoS, CoPoS, HBM packaging — key to GPU volume) 4. ASIC + Networking (custom chips, switches, SerDes, DSP — MRVL/AVGO core) 5. Power / AI Factory (transformers, grid, UPS, 800V DC, BBU — real capacity limit) 6. Liquid Cooling (DLC, CDU, cold plates — mandatory for high-power racks) 7. Nuclear / Generation (huge long-term, but slower than power gear) 8. PCB / ABF / CCL (supporting materials, less control) 9. MLCC / Passives (higher demand from high-freq/high-power) 10. Semiconductor Equipment (inspection, metrology, test — picks-and-shovels) 11. Physical AI (robots, autonomous driving, industrial AI — still 2026+) 12. Agent / Data layer (direction clear, winners not) 13. Stablecoins / AI Finance (AI agent payments, early) My top bets: Memory > Optical > Packaging. The real money is whoever controls the chain's throat. Which bottleneck are you watching? 🚀

Codex just got a killer feature: seamless thread handoff between local and remote hosts. 💻 Start work on your laptop 🔄 Send it to a remote box before you close the lid 🚀 Bring it back later And Codex orchestrates the whole handoff. No manual sync, no extra config. Who else has wanted this?

Brilliant hack: use an embedded browser as infinite canvas for Codex Image 2. No more juggling tabs or re-pasting prompts — iterate freely. 🤯 Anyone else tried this?

If Codex beats Claude Code, it'll be because the Claude team treats its UI like an afterthought. 1. Widely reported bugs and UX issues go unfixed for months. What is Boris burning infinite tokens on? 2. They keep selling "coding is solved" while their mobile app is terrible, the desktop app hangs and loses conversations, and nobody on the team seems to use their own products for 5 minutes. Ship a good frontend first, then talk. 🔥

Grok Imagine 1.5 is live. Elon is all-in on winning the video gen race — unlimited compute, unlimited budget, zero chill. I've been testing it. The speed is real. The quality gap? Closing fast. Can xAI actually take the crown from Sora and Runway? Or is this just another flex with no delivery? 🔥

Source / full context:

A Netflix engineer built an open-source proxy that slashes AI token usage by 60-95% with zero code changes. 🔥 Accuracy regression? ±0.000. No summarization, no loss. Just compression in-flight between your app and the LLM. Works with Claude Code, Cursor, Copilot, any OpenAI-compatible client. One pip install, one env var. 29.9k stars on GitHub. Apache 2.0. Netflix uses it internally. Built by Tejas Chopra. Have you tried it yet?

keyboard_arrow_left Previous keyboard_arrow_right Next

Source / full context:

A few months ago my kids started vibecoding web games with Cursor, and I needed a way to deploy them. GitHub Pages worked until they needed a real backend, so I hacked together a repo-to-server setup. Then one game hit 38,000+ players at Vibe Jam 2026 — the duct tape had to go. I rebuilt it as a single Go binary that turns one Linux VPS into a push-to-deploy host for any number of apps: • Each app is a GitHub repo • git push → live in <5 seconds • Zero-downtime deploys • Each app in its own container • Automatic Cloudflare DNS + TLS tunnels • SQLite-aware backup/restore No YAML, no Dockerfile for typical apps. It's just convention over configuration on top of Kamal, with all the infra glue wired up automatically. Setup is one interactive command. If you've got a pile of projects on one server, this might be your jam.