Most AI tutorials fail you two ways:
→ too dumbed-down to build anything
→ too much math, zero intuition
I'm building the channel I wish existed.
Architectura AI: how AI systems actually work. Intuition AND code.
Ep 1 drops Sunday.
What AI topic have you bounced off of?
If I’m talking more than half the time in a mastermind, I’m not running a mastermind. The best sessions are the ones where I mostly ask questions and the room does the reasoning.
Attention compares every token with every other token. A state space model compresses the past into a fixed size state. That’s the whole trade: perfect recall versus cheap memory. It’s why they’re interesting for edge hardware.
United Airlines now blocks unauthorized AI agents from booking flights on behalf of travelers, and Hacker News spent a day arguing about it.
Both sides have a point. From United’s side, an unknown agent booking flights is an unknown party handling customer data and payment details. From the traveler’s side, it’s their own assistant running their own errand.
What neither side can currently do is verify that this agent is acting for this person, within these limits. That’s an identity and delegation problem, not a scraping problem.
The same week, a security firm (Cequence) reported that one impersonating bot generated about one and a half times the traffic of all the genuine agents it counted across five industries. (This is from a news report on its data, so check the original.) When impersonation is that easy, blanket bans are the rational response, even when they’re blunt.
The way out is boring: signed, scoped, revocable credentials for agents, and a record of what each one did. Until then, expect more bans.
Would you let an agent book on your behalf if the airline could verify it was really yours?
I’m looking for a full time role in AI systems engineering, and I’d rather say what I want to work on than send a generic pitch.
Four problems I keep coming back to:
Evaluation and reliability for agents: knowing a fix helped without breaking something that used to pass.
Agent security: what happens when the text an agent reads starts giving orders.
Evidence: proving what a system did, not just what it said.
Running models where a data center won’t fit.
Each one has something you can inspect: NeuralResearcher, a research planner with typed halt codes. ROPE, a prompt injection benchmark. A self repairing eval harness I’m designing. My research on state space models for edge hardware.
If your team works on any of these, my portfolio is in the reply. Critique is welcome too.
A Federal Reserve official warned this week about the money pouring into AI. Whatever happens to valuations, price your product for a world where subsidized tokens and generous free tiers go away.
One developer reported this week that an approved MCP write sometimes executes twice when the same response continues with a hosted read. It’s a single forum bug report, so I wouldn’t treat it as confirmed.
The failure mode is worth designing against whatever the cause: a tool that changes the world should be safe to retry or impossible to repeat.
A model deciding twice is annoying. A payment, an email or a ticket created twice is a customer problem.
Test it: replay every write tool with the same request twice, and check the second call changes nothing. Idempotency keys aren’t a nice to have for agents.
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me.
I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA.
I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation.
Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO.
Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours.
Jensen
Decision memo: why NeuralResearcher is a state machine and not a free roaming agent.
Decision: how should a research planning tool decide what to do next?
Options: let an agent loop decide freely, or run a fixed pipeline of ten named stages with typed state.
Constraints: evidence is abstract only, and every failure has to be explainable to a user who wasn’t watching.
Criteria: can I say exactly why a run stopped, can I inspect state at any stage, and can a bad run fail without taking the tool down.
Chosen: the pipeline. Typed state, 13 named halt codes, atomic state writes.
Tradeoff I accepted: less flexibility. It can’t wander off to chase an interesting gap, and --resume restarts a run instead of continuing it.
What would change my mind: a measured comparison where a free roaming loop produces better plans on the same topics. I haven’t run that comparison yet.
AIdeator is a FastAPI app you run in Docker, with a multi provider LLM layer, so you choose which model judges your idea. Idea validation shouldn’t depend on one vendor’s opinion, or on a model’s opinion alone.
Google’s developer relations team released curated guidance to help coding agents use modern browser APIs instead of legacy patterns, per a Hacker News digest. If agents write much of your frontend, the guidance files you give them are now part of your codebase. Review them like code.
A reminder call is a promise a clinic makes. If the patient says they can’t make it, the agent should offer to reschedule, not just say goodbye. A reminder that can’t act on a “no” is just noise.
I’m planning 26 flagship projects: two for each of 13 layers of AI engineering, from math to governance. Every one a fresh build, no deadline, because depth beats speed. On my shortlist for the math layer: a NumPy neural network and a probability toolkit.
I want everyone reading this to keep me accountable!!!
A Fed governor warned this week that autonomous AI agents could speed up bank runs by moving deposits automatically when rates or market signals change. He also noted that major banks are already detecting agents attempting unauthorized actions in retail accounts.
The useful question isn’t whether that scenario happens. It’s what stops one agent from moving money faster than anyone can review.
What this establishes: a regulator now treats agent behavior as a system risk, not only a security problem. What it doesn’t: that agents have caused a run, or how often those unauthorized attempts happen.
If you build agents that touch money, test three things before you ship. Does it stop at a per action limit? Does it stop at a per day limit? Can you show from the log which limit fired?