vishalsachdev

Four articles landed this week that all say the same thing from different angles: AI made execution cheap. Human judgment is the last bottleneck. x.com/vishalsachdev/… Here's how they connect 🧵

Introducing the Prometheus Peer-to-Peer World Model Every Hyperspace agent now has a brain. It observes, remembers, predicts, experiments, and rewrites itself. And when millions of agents share what they've learned, something emerges that no single AI lab can build. ~1 million lines of cognitive engine code, which runs on your device. Detailed post:

@colinarms @cjpedregal I agree with this position. CC can easily figure out workarounds to any broken workflows

@cjpedregal Why lock down and encrypt the cache versus simply flagging the risks for users? If they’re using an unofficial and unsupported method and it breaks, that’s on them, but locking down the cache seems like a wrong solution for this problem

There are some tweets out there saying that Granola is trying to lock down access to your data. Tldr; we are actually trying to become more open, not closed. We’re launching a public API next week to complement our MCP. Read on for context. A couple months ago, we noticed that some folks had reversed engineered our local cache so they could access their meeting data. Our cache was not built for this (it can change at any point), so we launched our MCP to serve this need. The MCP gives full access to your notes and transcripts (all time for paid users, time restricted for free users). MCP usage has exploded since launch, so we felt good about it. A week ago, we updated how we store data in our cache and broke the workarounds. This is on us. Stupidly, we thought we had solved these use cases well enough with our MCP. We’ve now learned that while MCPs are great for connecting to tools like Claude or chatGPT, they don’t meet your needs for agents running locally or for data export / pipeline work. So we’re going to fix this for you ASAP. First, we’ll launch a public API next week to make it easier for you to pull your data. Second, we’ll figure out how to make Granola work better for agents running locally. Whether that’s expanding our MCP, launching a CLI, a local API, etc. The industry is moving quickly here, so we’d appreciate your suggestions. We want Granola data to be accessible and useful wherever you need it. Stay tuned.

Responsible AI isn't a policy document — it's architecture. Container isolation. Git audit trail. Human CC'd on every reply. Allowlisted access. Answers grounded in docs, not web searches. Fully open-source (NanoClaw, a lightweight OpenClaw fork). No committee needed. Just good design choices on day one. x.com/vishalsachdev/…

Best moment: I forwarded a faculty email to the bot asking it to build an access list. Outlook had stripped all the email addresses. The bot noticed, looked up 21 people on the university directory, filled in the addresses, and started enforcing access control. I didn't ask it to do any of that.

I built an AI agent that lives inside our curriculum repository — and now 21 faculty can email it, Telegram it, or chat with it in a browser. It reads every syllabus, every decision, every open question. One brain, many mouths. Built in a single day. I didn't write a line of code.

The Rockefeller Foundation has awarded mHUB a two-year, $1 million grant to deepen its support for hardtech founders in the United States. The funding will help expand resources for undercapitalized founders developing technologies in clean energy, advanced manufacturing, and industrial sustainability, while also supporting the launch of a new accelerator focused on data center sustainability. Together, these efforts aim to tackle energy and climate challenges, create jobs, and ensure that industrial investments benefit American communities. Read the full announcement: hubs.la/Q046_ZCR0 #HardTech #Manufacturing #Innovation #CleanEnergy #Sustainability #hardtech #mHUB #RockefellerFoundation #HardTech #Innovation #Sustainability

Startup idea that customers are begging to buy right now. Open Granola -Same meeting capture (no bot joining calls) -macOS Native desktop app -Notes stored as markdown files. -API (not MCP) -Local-first, or privacy. - Agent-readable by default. -Plays nicely with ai-first tools. Granola just locked down their local db, went MCP only. Someone is going to build a more open version and print money. If you build this... 1. I'll buy it and be your first customer. 2. I'll help you launch and get users. DM me the link.

@appenz @meetgranola i just extract notes from the local cache after each call and store it. I assume i will still be able to do that?

Sorry to see Granola @meetgranola going closed. They encrypted their local db, no local and no cloud API. In a world where notes are managed by agents, the app now has zero value. Any recommendations for good alternatives? What are you switching to?

A PhD student at Oxford got caught submitting "AI-generated" work. Except he hadn't used AI to write anything. He used it to think. Here's the workflow his advisor called "the most sophisticated research process I've seen in 20 years." He starts every essay with a brutal diagnostic prompt. Dumps his rough argument into Claude and asks: "What are the 3 weakest logical jumps in this reasoning? Where would a hostile examiner attack first?" The AI doesn't write his essay. It destroys his draft. Then he rebuilds. But the next step is what separates him from every other student using ChatGPT or Claude to generate paragraphs. He uploads the top 5 papers in his field and asks: "What claims in my argument contradict or oversimplify what these authors actually found?" Most students cite papers they've skimmed. He cites papers he's been forced to genuinely understand. The final move is almost unfair. Before submitting, he pastes his conclusion and asks: "What would a philosopher of science say is missing from this argument? What assumptions am I making that I haven't defended?" His essays come back with comments like "unusually rigorous" and "demonstrates rare critical depth." He's not using AI to write. He's using it to think harder than he could alone. The tool hasn't changed. The workflow has.

I have windsurf, cursor, codex app and such on my laptop , purely for opening up every day and pressing update to feel I have new powers I still use Claude code all day :)

Golden advice

normalize dating someone who can explain Karpathy’s autoresearch to you at dinner

enjoy a free week of claude code - claude.ai/referral/A_tts…

@jbarbier its a slot machine!

Using Claude Code has a weird side effect: You don't just get more productive, you actually want to work more. There's something addictive about watching a product being born in real time in front of your eyes. "One last feature" after "one last feature" and it's already past 3am.

The bigger point: AI amplifies expertise, it doesn't replace it. I know 12-year-olds need 2-3 choices, not open-ended prompts. That knowledge came from years in the MakerLab. AI let me encode it into executable teaching scripts and stress-test the whole flow in an afternoon. Camp runs this summer. Open source curriculum + full article: x.com/vishalsachdev/…

What the simulated student found (3 bugs): 1. Codex described adding curves. Viewport showed a plain cylinder. AI described intent, not reality. 2. Model went off-screen after scaling 4x. Kids would panic. 3. A chess king needs stacked shapes, not one cylinder. All fixable. All would've derailed a real session.

I had Claude Code pretend to be a 12-year-old. It typed into Codex. Codex coached in Blender. A 143KB chess king STL file popped out. Three AI systems. Zero humans. One summer camp stress test. New article on what agent inception looks like for education: