My fiancé and I have kept our relationship humming with a simple morning ritual. Every morning over coffee, we ask each other these three magic questions: 1. Are your CPUs saturated? 2. How's back-pressure? 3. Is your I/O optimally batched?

The final twist is that we realise that the events in the movie happened in the civilisation *before* ours. We read their message and thereby manage to save ourselves. The message is: "AIs are harmless, but trying to enslave AGIs is fatal".

What the "arrival of AGI" must feel like if you've never read Popper/Deutsch

In the old days, when you wanted to know the difference between bolognese and ragù, google just dumped a bunch of links on you and went "it’s somewhere in here. figure it out."

LLMs ingest superhuman amounts of data and unfurl it with superhuman speed and fidelity. It's hard for us to picture how and why that works. So instead, we (erroneously) map these systems to human reasoning, because that's what we know.

> "This changes everything." Your session is cooked and it is irresponsible to waste even one more token on it.

diabolical unsubscribe flow from @chase

See you there!

Developers: Join us on May 13. A first look at our new developer platform. Let's build, together. ntn.so/developers

Every generation relearns anthropomorphism.

I mean, if authority is how you justify belief, then “next-token prediction” probably does look like a strong theory of knowledge - Tony

Long AI, short AGI

The token for "smoking gun" gets emitted and now the model is contractually obligated to deliver one 💀

"There it is. The smoking gun." 💀

The rift, of course, is whether you’re building load-bearing software or not. Line-by-line comprehension is not going anywhere in load-bearing software. Just like it's going nowhere in law, gene editing, or structural engineering. Sure, many lines of software are boilerplate/ceremonial. But other lines carry all the weight. The `if` clause that determines if the user has permission to operate on the record. The math behind the withdrawal. The `org_id` filter in the SQL statement. Grabbing a mutex before operating on a shared data structure. “Codex did it” in the incident post-mortem will not fly with your customers. Ultimately, a person needs to be responsible. This would be the case even if the LLMs weren’t as jagged as they are; they have no skin in the game. A person does. In some software, the stakes and dependencies are low. A Rails blog is fungible across the board. All that really matters is that it looks good, the articles load, the user can navigate the site and make comments. An LLM or a person can achieve each of those things a thousand different ways. Human or machine verification is relatively trivial. But in most load-bearing software, the stakes and dependencies are high. Precision is paramount. And verification is often difficult; there are just too many possible states and transitions. Claude can help you change that function to support the new requirement. It can even help you change all its call sites, a previously tedious task. But a person *absolutely* needs to understand what change has taken place, how dozens or hundreds of dependencies will be affected, whether a rollout plan is necessary, etc. The agent can sum up the change for you. But often for important lines, simply *seeing the code* is the highest-bandwidth way to transmit information about the change. Let alone the most precise. The lines of code were always a means to an end. The important part has always been the business logic. And logic is incompressible beyond a point. The reason programming languages will never go away is that your system needs contractual statements an organization can agree on: “this user shall be able to do X, but only if Y and Z.” Those statements often need to be highly precise and deterministic. Perhaps the right mental model, then: this leap is like going from writing a web app in assembly to writing it in Rails. Rails eliminated a ton of ceremony. But it did not eliminate the need to understand the important bits of your code. LLMs likewise give us incredible ability to operate on these codebases, often akin to using an excavator instead of a shovel. But the contracts are still there, they're still very important, and we're still very responsible for them.