It pains me to link to LinkedIn, but for a Rich Hickey take I’ll do it.

He calls out referring to AI as a “tool”:

As a lifelong tool builder and user, I know what tools are. LLMs are not tools [...] LLMs are automation.

Shots fired!

You don't direct the action of an LLM - you make requests and it acts on its own, without your energy or continuous direction

You have traded working with the world to working for a capricious robot.

What have we lost?

What do you miss by abdicating doing the reading? Pretty much everything. First and foremost, you miss learning how to write from those who can. You miss learning how to formulate an argument. You miss noticing things. You miss loading your background mind with fodder for novel ideas. You miss building trust in some authors and suspicion of others. You miss all the jokes. You miss all of the edges. You miss learning how to think.

(I love his point about how AI is not a “tool” for reading the same way a Kindle is a “tool” for reading or “glasses” are tools for reading. Just read the whole thing.)

Lastly:

Tools don't get "misaligned", they don't "go rogue", they don't "escape containment", "conspire" etc.

It just reeks of the offloading of responsibility. “I was mindlessly cutting an apple and the knife in my hand escaped containment and cut my finger. Not my fault!”

Anthropic sheds light on how they view the role of engineers in the AI future:

Very often, we found that the role of the engineers was to tell Claude “no.”

Sean Voisen (emphasis mine):

This means specifically telling the agent to actually use the intended design system, use these components in such and such a way, use these colors for these purposes and not other purposes, always lay out elements in this fashion using these pre-defined spacing tokens, avoid inventing new components, and so forth and so on. Constraining generation shrinks the disturbance variety of the whole system by reducing what’s possible.

This vibes with my experience (pun intended).

Being an editor and “constraining generation” for the LLM seems like the primary function.

Alex Wellerstein has a good post detailing the runaway improvements in speed and efficiency in war:

The first atomic bombs did with a single plane, in a single instant, what would have taken hundreds of bombers loaded with conventional munitions the work of many hours […]

In less than ten years, the bomber forces of the Strategic Air Command had so swollen that it was possible to imagine the complete destruction of the Soviet bloc in a matter of hours. By the 1960s, the deployment of intercontinental ballistic missiles (ICBMs) meant that thermonuclear payloads could be delivered from one side of the globe to the other in around half an hour.

The way nuclear weapons accelerate the process of human-led global destruction is craaaaaazy:

• Thousands of planes loaded with millions of pounds of traditional munitions can destroy a city in a few days.
• One plane loaded with one atom bomb can destroy a city in a few hours (time it takes plane to fly to location).
• Multiple planes with multiple atom bombs can destroy multiple cities in a few hours.
• Multiple planes with multiple thermonuclear bombs can destroy multiple nations in a few hours.
• Thousands of missiles with multiple thermonuclear bombs can destroy the world in minutes.

Fast! Efficient! Automated! I guess you call that — progress?

[we made systems to be] able to order up the deaths of half a billion people in less time than people typically get for a lunch break.

A glimpse at how war planners saw it all working:

An SLBM launches. About a minute later and a half later (95 seconds), NORAD gets the alert. If the missile was on a “depressed” trajectory, you now have between 135 seconds […] and 645 seconds […] before the detonation, depending on the SAC target. If it is on a “nominal” trajectory, you get between 245 seconds […] and 755 seconds

We’re talking about (seemingly) rationale human beings designing systems capable of hundreds of millions of deaths and thinking, “It’s probably best to communicate this in seconds rather than minutes because of how fast we’ve made it all.”

Paul Ford gives a metaphor for how people are amazed at what AI can do when they use it:

They’re giving birth to a baby in two hours and they’re like, “That’s a great baby. That’s a two-hour baby.”

We used to live in a world where people equivocated a large artifact, like a 250-page PDF, with hard work (consulting companies made a living off this equivocation).

But generative AI destroyed this illusion. Now we’re all readjusting and reevaluating how to judge good, thoughtful work.

Adam Silver:

It’s harder to include something and prove it’s unnecessary, than it is to leave something out and prove that it is.

Truth.

Let’s praise more people and organizations who say “No”.

Jason Goldman discussing how he thinks we can approach this moment of AI and all the worries around it:

Continually find ways to change your language so that you do not come at the question [of AI] with an inevitability and a powerlessness, of like “This is just happening to us and one way or another we just gotta go through it. One way or another this is going to happen because if we don’t do it China will do it.” [As if] it’s just this tidal wave happening to us. No, no, no. You need to remind people continuously that they have power, that they have agency, that we have the ability to change the trajectory of things. In the course of human history, that is what we have done consistently. We have met very hard, very seemingly-intractable challenges, and through a weird combination of human ingenuity and stick-to-it-ness and stubbornness we have managed to move the direction of the ship in a more positive direction. We need to tap into the spirit of “this is actually something that all of us get a say in and can do something about”. It is not an abstract iceberg dropping down on us from the post-singularity heavens.

Here’s the YouTube link if you just want to see this excerpt.

Dave Rupert complains about email:

Just because we participated in online commerce once doesn’t mean you get the liberty to email me for the rest of my life. […]

When did a ten drip email campaign for every online transaction become the standard? You bought an item → Item is getting ready to ship → Want another item? → Item shipped → Need more items? → Item out for delivery → The holidays are coming up → Package delivered → Let us know how we did → Mother’s day coming up → Father’s day coming up → Step-father’s day coming up → etc… It’s entirely too much.

I get it. I feel this same thing. But I still love email. Email is good.

If I had Dave’s email address, I’d send him one right now just to troll him. But since I don’t, I’ll write it here:

To: Dave Rupert
Subject: An Email You’ll Never Read
Body:
I know you’ll never respond to this, and you don’t need to. It’s not a TODO for you. It’s just a note saying I liked your blog post. That’s it. The end.

Mandy Brown:

there [is] something keenly human and joyful about putting words and type on the internet

For me, this was always the case.

And it still is.

And I hope it will always be.

Paul Ford, on his company’s blog:

Everything in tech is huge right now. Every system lives on someone else’s machine and tracks you everywhere. All the content you make with those systems is being gobbled by LLMs.

So he made a tool to make documents whose content lives entirely in the URL. Here’s how it works, as described by a document of the tool itself :

1. You clicked on this link. That loaded the document data into your browser—but it didn’t know what to do with it.
2. Your browser asked for the website at TinyLittleOffice. That’s the code.
3. The code ran in your browser, and it converted and displayed the data after the “#” and painted it onto your screen.
4. Now you see the document!

Pretty damn cool! The entire premise hinges on the fact that browsers don’t send anything after the # to the server. From the threat model docs:

A document is its link. The whole of it, compressed, sits after the # of https://tinylittleoffice.com/<app>/#/v1/<app>/<payload>, and a browser never sends the part after # to any server. The site is static files on Cloudflare; there is no account, no database of documents in the clear and no server code that receives one

This project is at the intersection of two of my loves: 1) URLs. and 2) stripping things down to their essence.

Here’s Ford on Bluesky describing some of his process and decision making on the product:

Just me and a bot doing basic dictionary and encoding stuff, it's all pretty basic. By eliminating images and locking in things like font choice and pre-canning templates you just end up with very small documents in general.

Cool!