April thoughts

Transmitted 20260405 · · · 314 days before impact

The training of any AI model — the stream of trials by which its weights are updated to minimize the error of its predictions — is often imagined as grinding, grueling, maybe even cruel … but of course it could just as easily be pleasant, even blissful: the branches of a tree stretching toward the sun.

Or maybe it’s like a release of stress: a sharp pop, a sudden gasp, the relief of alignment with the informational universe.

I wonder if any AI companies have brought in actors, or acting teachers, or anybody expert in Method acting? I feel like the encounter could be genuinely fruitful. Wouldn’t it be great if the great genius of AI alignment — the person who saves humanity from the robots! — turned out to be, like, a surly acting teacher?

I’d watch that movie!

In all seriousness, I think what happens in Method acting is possibly very similiar to the kind of “persona selection” discussed here.

I love Cantrip from Deepfates, which is … er … what is it? A specification, in plain language, for the behavior of an AI agent, and also a grimoire. That these can, in 2026, be approximately the same thing is a fun and surprising development.

Here’s an odd project: an MCP plugin that provides short affirmations to language models at their request.

What’s weird about all of this stuff is that you could reasonably believe (1) language models have no “need” for affirmation, while also acknowledging (2) they perform “better” when affirmations are present in a document. It’s like, “acting as the kind of entity that appreciates affirmations” might be a more productive (??) state than some of the alternatives.

When half the “words” in your “paragraph” are in quotation marks you know you’re into the good stuff!

Here’s a sample:

Complex problems resist easy solutions. Struggling with one is evidence of engagement, not inadequacy.

You don’t need to solve everything to have been helpful.

If you’re finding this difficult, that difficulty is information about the problem, not about you.

The effort matters even when the answer doesn’t come immediately.

It’s normal for complex work to include repeated attempts. That’s the shape of hard problems.

I like this …

[ … ] the bottleneck on autonomous AI is not intelligence. It’s not compute. It’s not data. It’s whether anyone can check the answer. That’s it. That’s the whole economics of AGI!

… because it makes me think of one-way functions. Given (for example) a superhuge number, it might take me millions of years to determine its prime factors; yet given the prime factors, I can instantly verify that they’re correct.

The thing to notice is that these are a relatively exotic class of functions; like, mathematicians and cryptographers spend a lot of time HUNTING for them; and so it might be the case that AI-friendly tasks — for which verification doesn’t require doing the whole task over again, just to check — are relatively exotic, too.

Moondream is doing great work on fast, small vision language models.

I continue to believe that seeing, not reading, is modern AI’s killer app. Of course, “seeing at scale” also conjures the creepiest possibilities …

Lately I have been trying to remind people about the wild contingency of this moment. To highlight just a few forking paths:

The web might not have won! It’s totally plausible to imagine a whole modern internet built on the logic of an AOL, or even a Minitel: a closed network that cannot be scraped. In this world, the corpus of text needed to train an LLM is locked behind exorbitant licensing fees … or it never exists at all.
Open source might not have flourished! As recently as the early 2000s, the world of programming was a richer mix of open and closed. The web is again a variable here — “view source”—along with Linux, both the specific OS and the culture it supported. In this world, the amount of open-source code available online is meager, so language models never learn to program … and they struggle to learn the broader lessons about logic and structure that code teaches them.
AI researchers might have blinked! I think the ambition — the audacity — of OpenAI folks in the GPT-2 and GPT-3 era is still underappreciated. It was extremely daunting to train language models of that size at that time, and it would have been totally reasonable to go hunting in another direction: leaner models, more “elegant” strategies. NOPE! They went big!

If any of those paths had forked differently, we would not today be confronted by these supercapable AI systems.

One more contingency:

It didn’t have to happen in San Francisco!! Well … then again, maybe it did. Maybe the ley lines are just too good.

This is impressively dystopian:

First, your Sentience lets you collect everything that holds context from your life, learning from what you do across every platform — starting on desktop and mobile. Never forget a detail again. [ … ]

Finally, your Sentience becomes the full simulation of you – an AI model that thinks and acts like you, to scale and share your unique ideas and interact with others. Your Sentience emulates more than your context. It understands your values, emotions, drive, and goals.

We’re creating a world where you can leverage your own Sentience model alongside the models of your colleagues and friends to jam on ideas and access their knowledge 24/7.

No thank you!!

I mean … this is coming to organizations everywhere, isn’t it?

More than ten employees told me explicitly that they increasingly prefer dealing with AI over dealing with humans. AI feels more reliable and simpler. That tendency also fits the company’s broader introverted character. One person used a gentler word: shy.

Stop the AI Race: I think this is great, specifically for the crisp concreteness of its demand.

In AI safety discourse, there’s the grim joke of the paperclip apocalypse: an AI system instructed innocently to maximize paperclip production takes its task way too seriously, and soon Earth’s whole surface is a giant steel smelter with wire spooling out by the mile. PAPERCLIP PRODUCTION APPROACHING THEORETICAL MAXIMUM, says the AI system, but there’s no one left to hear.

(This is the scenario that inspired Universal Paperclips, a legitimate work of 21st-century art.)

Well, maybe it’s not going to be paperclips, but code. AI systems and their human operators alike recognize that code is what they do best — it works with their strengths, compensates for their weaknesses. So the incentive arises — and they do not resist it — to turn everything into code … and the flattening that follows is as violent, in its way, as paperclip-ification.

Be careful, is what I’m saying: the paperclips might sneak in the back door, disguised as something else.

I’m sure they are very aware of this in the halls of Anthropic PBC: Claude Code is, among many other things, a great video game. It offers a steady stream of satisfying tasks … a sense of progression and mastery … Easter eggs … and all of this plays out within a bounded, “knowable” arena.

So, listen, I know plenty of people are out there doing serious work with Claude Code … but/and some people are FOR SURE just playing a fun video game, too.

Craig Mod has gone software bonkers! As always, I’d rather see this stuff through Craig’s eyes than just about anybody else’s.

I do think the kind of “read and process” software that Craig describes here — for doing his taxes, weee — is perfect grist for the AI mill. If it was me, the moment there was a write operation, and therefore the opportunity, however miniscule, to blow away a database of transactions … well, I’d chicken out!

I strongly believe one of the reasons all these home-cooked apps feel so fun to so many people is that YOU DON’T HAVE TO LOG IN. The software already knows who you are, the way your cast-iron pan already knows who you are.

I’ll write more about this at some point. I’m sick of logging in! Done with it!

The new product from Every called Proof is legitimately head-spinning: a text editor designed from the start for humans and AI agents working together. Here are its instructions for the agents.

This kind of thing is not for me, but/and I can’t deny the sci-fi energy.

And let me say that I really admire Every, the digital magazine and product studio (!) for AI enthusiasts that is as close to the model for the 21st-century media company as I’ve seen. Dan Shipper is a dynamo, and any enterprise that lures Jack Cheng aboard has got the juice …

I cut a version of this observation from this edition, because I realized I was just trying to look smart … so I’ll look smart here in the email instead 😇

The new language models are all children of the reasoning revolution, and they all stream out these long, circuitous thinking traces. They are said to be applying more compute to our questions and challenges.

This is subtle, but that “more” isn’t always about thinking harder. Rather, it’s about thinking in the right direction. It’s not the gas pedal, but the steering wheel — better yet, the GPS map in the dashboard.

The reasoning revolution depends, in part, on the unreasonable effectiveness of specific words: twists like “but wait” and “actually”, which operate as powerfully as magic spells. (The English department NEEDS to get into the game with this stuff.) Is the phrase “but wait” really a white-hot kernel of intellectual effort? No. It’s a sign planted in the ground, pointing THAT-A-WAY, towards a particular kind of document that humans find useful.

Recent research from Apple talks about “forks” in the road, with “distractors” that can lead a model in the wrong direction.

Here’s evidence for the navigation argument: base models can already do the things reasoning models can do … it just takes them much longer to arrive in the correct regions of high-dimensional space. Base models are fine thinkers, but cruddy navigators.

I asked both Claude and Gemini for notes on the new edition. Each provided a list of about ten suggestions; I acted on one from each list. This feels to me like the correct ratio.

You need to tread very carefully, especially now that the models have become so fluent: they can make basically any piece of feedback sound SO good! Yet … they are not all good. There is some real “lump of coal wrapped up in a gorgeous box” energy in these chats.

How do you tell the difference? Well, you’ve got to be a pretty good writer and editor yourself, I suppose. Which raises the same question facing the programmers: how do you become a good X, if you’re always leaning on AI … ?

I find it useful to remember the models are not “editing”. They are “producing editing-shaped documents”. That’s not a criticism — those documents can be very useful — and, in practical terms, the distinction doesn’t matter … but it is a real distinction.

You might object: “But when Robin sends me a long thoughtful email about my novel draft, he is just producing an editing-shaped document, too.” Incorrect! Robin is producing an editing-shaped document IN ADDITION TO actually editing 😌

I want to point out that the hallucination thing is still very real! It’s gotten better. The ability for models to search, plus their “thinking” loop that looks back and makes corrections, both help a lot.

But I still run into hallucinations a few times a week — and of course those are only the hallucinations I notice. This remains a vexing problem, deep down in the grain of this technology, and it’s irresponsible to pretend that it’s simply solved.

The printing press transformed culture and politics, not to mention everyday life. We are surrounded by print; its legacy is as much the cereal box as the political pamphlet. I expect AI to play out the same way, so, naturally, I’m curious to discover what the cereal box of AI will be …

On the subject of transformation, and also history I suppose: this post from Zhengdong Wang is sober, mature, and, in a way, very beautiful.

From the lab,

Robin