Tony Nguyen

Stanford @CS153Systems, Week 1 (Full Lecture) AI Scaling, Bottlenecks, and Why Compute Isn't a Commodity Yet 00:00 Compute Coachella 00:29 Simple Life Heuristic 01:08 Uncertainty Creates Opportunity 01:42 Four Bottlenecks Framework 01:51 Empirical Proof Matters 02:05 Cloud Costs Are Shifting 02:15 Verifiable vs Fuzzy Progress 02:48 Scaling Predictability Explained 03:43 CapEx Explosion in Big Tech 04:06 Chips Aren’t Commodities 04:45 Compute Scarcity Conclusion

Building a Personal Knowledge Agent I've been using a personal knowledge agent called Pal (Personal Agent that Learns). It runs locally, talks to me over Slack, and tries to get better over time (still tuning this part). I posted about it a few weeks ago and wanted to share some key design decisions. The goal is that I feed it raw data (URLs, papers, notes, meeting context, tidbits about people) and it organizes everything into two layers: a compiled wiki for text-heavy knowledge (concepts, summaries, research), and a SQL database for structured data (notes, people, projects, decisions). It should connect to my email, calendar, and slack. Here are some details: 1) Markdown + SQL Markdown is great until you need to query across dimensions. "Everything related to Project X from the last two weeks across all sources". "Prep me for my meeting with Sarah" (pull her notes, recent emails, project context, calendar history). This is relational data, not document retrieval. SQL handles this well. 2) Navigation over Search The key insight behind Pal is navigation over search. Each data source keeps its native query interface. Databases get SQL. Email gets queried by sender and date. Files get navigated by directory structure. The wiki gets navigated by its index. No flattening everything into one vector store. The agent picks the right source for the right question through a metadata routing layer, not through embedding similarity. 3) Structured data (SQL) When I say "save a note: met with sarah from acme, she's interested in a partnership". Pal creates a row in a notes table, tags it with ['sarah', 'acme', 'partnership'], and links it to sarah's entry in a people table. When I later ask "what do I know about sarah?" it queries across notes, people, projects, emails, and calendar. Tags are the cross-table connector. A note about a meeting with sarah about Project X gets tagged so it shows up in both contexts. The agent owns the schema. It creates tables on demand. Notes, people, projects, decisions all emerged from natural conversation. "Save a note" creates a note. "Track this project" creates a project. The schema grows with usage. 4) Knowledge base (Wiki) The other half is a compiled knowledge base for things that need depth. Research, technical concepts, reference material. 4.1) Ingest: I feed it URLs, papers, articles, meeting notes. It fetches the content, converts to clean markdown, and saves to a raw/ directory with YAML frontmatter (title, source, date, tags). A manifest tracks what's been ingested and what's been compiled. 4.2) Compile: A dedicated Compiler agent reads uncompiled raw files and produces structured wiki articles. It breaks each source into concept articles, writes summaries, cross-links related concepts, and maintains a master index. Compilation is incremental. Only new files get processed, never the whole wiki. New information enriches existing articles rather than replacing them, and every claim links back to the raw source. 4.3) Query: The wiki index is designed to fit in one LLM call (~100 articles). When I ask a knowledge question, the agent reads the index first, picks relevant articles, then falls back to raw sources and live tools. I expected to need vector search for this. Turns out an auto-maintained index with brief summaries works surprisingly well at this scale. The LLM navigates it like a table of contents. 5) Learnings I'm still working on this part. The pieces are there but it's not where I want it yet. Because Pal is a team of agents, they all share a common learning store. Every time a retrieval strategy works, it gets saved. Every time I correct the agent, that correction gets saved with highest priority. Over time, the agent should route to the right source faster and give better answers without me tuning anything. 6) Architecture: Pal is a team of five specialist agents. Navigator is the workhorse. Researcher gathers sources from the web. Compiler turns raw into wiki. Linter checks quality. Syncer pushes everything to GitHub. Pull the wiki locally, read it in your IDE of choice, push changes back. 7) Scheduled tasks: Eight scheduled tasks run themselves between conversations: daily briefings, wiki compilation, inbox digests, weekly reviews, wiki linting, context re-indexing, and git sync. Results post to Slack. TLDR: raw data from any source gets ingested and organized into two layers: a compiled wiki for knowledge depth and SQL tables for structured breadth. Five context systems get navigated (not searched) to answer questions. A learning loop compounds every interaction. The wiki is just markdown backed by git.

introducing deskmate a curated map of good coworking spots in sf and south bay you can add your favorite to the map. expanding to nyc, paris, shanghai, singapore soon! deskmate-app.vercel.app thread with some of my go-to spots 👇

@0xJuliechen Miiii

I’m creating a group of the most cracked devrels in SF if you are: - dev rel - dev tool's marketing/ecosystem - hosts the best hackathon/developer meetup comment below! i will add you to the group : )

@krupaad cfbr

bit late to the recruiting cycle, but looking for a summer internship in ML/hardware/inference!! i've been working on CUDA kernel writing, FPGA acceleration and RTL. would love to find a team doing similar work this summer dual US/Canada citizen, can relocate anywhere DMs open :)

If you are a computer science student struggling to get an internship, here is some offbeat advice: do a random non-computer science internship. Sit in a factory this summer, observe their processes and accounting, and identify AI use cases. Sit with a chartered accountant during their busy June-July season and help with accounting using Claude. Join a sales team to understand their workflow and see how their outdated SaaS CRM could be improved with AI. Work with a customer relations team at a startup and build an AI bot for them. Even better, embed with a complex domain-specific team like geological exploration, pure sciences, or research and tinker alongside them. My personal belief is that in the next decade, people who combine computer science skills with business context in some domain will be rockstars. Pure software engineering is losing its leverage.

My ideal workflow: 1. In the morning, I have a 15-min standup meeting with a master AI who gives me progress of all subagents & tasks. I provide feedback & give instructions in one focused meeting 2. During the day, we all go our separate ways. Agents focus on their work; I focus on my work 3. In the afternoon, repeat step 1

@lauradang0 @christyoverflow me me me last weekend in SF for me and please i really want to be there

Still trying to find community in SF? @christyoverflow and I are hosting a No Work No Pitch Picnic next Sunday for people in their 20’s to meet each other. DM or comment and I’ll send you the partiful, hope to see you there!

generalists are about to win big If you understand a little of tech, business, and people, and can connect everything fast. you're sitting on a goldmine right now.

Hey, you got a cool project that you are building? Link it I want to yap about cool projects

@katedeyneka I did realize... but since this is every where on X today so might as well give a shot cuz what if there is something truly legit that I can take one foot in loll 🤣😝

@tonybuildd Happy April Fools' Day🌼😀

co-hosting a small meetup where three Anthropic researchers will walk through some post-training details behind Opus and the upcoming Mythos models keeping it intimate, ~30 spots reply or dm if you want in and i’ll send the event link

@MilksandMatcha Im a part of ChatGPT Lab and we focus on learning different use case of ChatGPT of students. Exploring a lot of ChatGPT capability lately and its fascinating. From a developer perspective, a lot of things that we learn from ChatGPT can be used in Codex, and I really need to try!

Giving away 5 Codex Pro plans Each person will get 3 months of free Codex Pro (highest tier). Winners will be selected from comments in 48 hours, comment below why you want it.

@michal_gren im interested

I kept saving design AI agents/skills across 10 different places, so i built one place for all of them. Anyone interested?

Introducing the "Follow builders" skill: the best way to stay on top of the insane happenings in AI I carefully curated 25 X accounts & podcasts that share the highest-quality, first-hand insights on AI (by builders from OpenAI, Anthropic, Google, OpenClaw, Replit, Vercel, Cursor...) Your OpenClaw/agent can remix my central feed & send you a personalized daily newsletter in whatever channel you like Already widely used with 2k+ stars on GitHub; link below

@0xSero Meeeeeee was at Factory Mission event so gooddddd

Do you want to try Droid? I’m doing a giveaway 3 people will win 100M Factory credits each.Thats 5 months of their 20$ a month subscription. Winners selected randomly from comments in 48 hours.

here's how Claude Code actually handles memory : all 8 phases 🧵 Our team at @mem0ai use @claudeai a lot, we deeply care about memory. here is a summary of how it works 👇 User Input -> Context Assembly -> History System -> API / Query -> Response -> Summary Phase 1: session init registers hooks, warms the memory cache, and kicks off async directory walks before the first render Phase 2: memory is discovered in priority order — managed enterprise policy → user global → project VCS → local per-directory → auto-generated → team shared Phase 3: three parallel pipelines merge into every API call: system prompt + memory section + user context. relevance prefetch selects up to 5 memory files via sonnet side-call Phase 4: the model can directly read/write memory files using FileReadTool, FileWriteTool, FileEditTool. background extractor and model writes are mutually exclusive Phase 5: after EVERY response, three background agents fire — extractMemories, sessionMemory, and autoDream. extractMemories is a forked agent that runs in parallel, capped at 200 lines / 25kb Phase 6: when context fills up, compaction summarizes old messages using a skipped summarizer, preserving min 10k tokens / 5 text-block messages Phase 7: memory lives across ~/.claude/, project root, sessions/, and agent-memory/ — auto memory is git-ignored, team memory is VCS-tracked Phase 8: self-improving loop across sessions — within-turn writes + end-of-turn extracts + session memory + auto-dream consolidations every 24h+ every touchpoint: launch → query → response → background agents → shutdown → next session shoutout to @ChaithanyaK42 for the beautiful excalidraw!

Say hello to agentOS (beta) A portable open-source OS built just for agents. Powered by WASM & V8 isolates. 🔗 Embedded in your backend ⚡ ~6ms coldstarts, 32x cheaper than sbxs 📁 Mount anything as a file system (S3, SQLite, …) 🥧 Use Pi, Claude Code/Codex/Amp/OpenCode soon