davidpapermill

Thanks, that's really helpful.

What do you think happens in terms of testing/reviewing going forward? I'd really appreciate your thoughts on that.

I'm afraid law and morality are two separate concepts. You make a good point.

I've heard something along the lines of "Claude is like a compiler: source code is the new object code, you don't look at that anymore" many times.

And I don't really think this is true. Compilers are usually deterministic, and whilst we can find edge cases, it's nothing like an AI agent writing all the code for you.

I think you have two choices, given the Claude is a code generator and not a compiler: (a) you review most or all of code to make sure it makes sense, or (b) you trust but verify via a strong test suite, potentially also created by Claude.

The problem with (a) is that you lose a lot of the speed-up. The problem with (b) is that you have no human oversight and the code may be incomplete, badly designed, or plain wrong.

Currently we review all code because correctness is extremely important to what we do, but that comes at a cost.

I don't know what the answer is here, in general. Does trust build over time? Do the models just get so good we can trust them to make zero mistakes?

Keynote has probably the best UX of any software I’ve ever used.

Yes I agree, they’re an alien kind of intelligence.

But that doesn’t mean they don’t reason.

Can you give me one example that works on Claude right now?

I'm never sure whether this indicates "no reasoning present" or you've just hit an odd behaviour in the AI such that its reasoning fails. For example, you present a problem in a way that's dissimilar to the way problems are presented in its training set. That doesn't mean it's not reasoning, just it can only reason correctly in some circumstances.

It would've been nice to reserve "AI" for superior human-like intelligence capable of genuine common sense and reasoning.

What would a frontier API have to be able to do to satisfy you?

I think a big question is whether any of these labs can produce a model that is _ahead_ of Anthropic and OpenAI.

A related question is how much they're dependent on the APIs of Anthropic and OpenAI to achieve their results - whether through distillation or other uses.

If these models are derivative of Anthropic/OpenAI I would expect performance to be more narrow and progress to be limited.

We use US models for everything in practice, but we are looking at open-source models right now. So it may just be a turn of phrase hiding the reality. I can't imagine anywhere near 80% are relying on open-source as their primary models.

I don’t agree. I think businesses including SaaS are not primarily their tech, but are mostly composed of other functions: sales, marketing, customer success, product management, strategy etc. The value is actually more in these other components than in the code.

That AI number is a gross underestimate. It’s nearly that for this year alone. Wildly off.

Hadn't heard of it! Site here: https://rubyllm.com/

This looks really nice. Snippet from site:

chat = Chat.create! model: "claude-sonnet-4"

chat.ask "What's in this file?", with: "report.pdf"

Thanks, would really like to read the blog post.

Are you using any of the major agentic frameworks (Mastra, LangSmith etc)? Or is the AI harness etc entirely custom-built?

Upvote for you.

I think "Your Art" tagline is confusing - I thought it was for selling paintings.

Nice, which frameworks are you using? What did you learn from customers?

I'd consider a different name to avoid issues with supabase should you take off.

I wonder if AI agents could developed their own open-source ecosystem.

Some kind of knowledge-sharing seems inevitable, but the question is what shape and form will that take? We've seen wiki's, discussion forums, AI's posting to GitHub.

I feel like knowledge bases for AI will look somewhat different from our past experience.

How will you encourage sharing of solutions? I don't think "social proof for models" will be enough.

Papermill - the document engine for AI. Turn AI content into polished documents using a new document/templating language called Press [1]

It's essentially a high-quality alternative to the HTML-PDF route so many people take for document generation. It's designed as a tool for AI agents.

Something cool we've released - an MCP server makes it possible for models like Claude to design fully-featured documents. We use this internally and some of our customers now use it to quickly build new templates.

Just launched self-serve a few weeks ago. Continuing to develop the typesetter and language behind it.

Press is the language, docs are open [2].

Would love any feedback from folks that have worked on document generation, or people with experience doing HTML->PDF.

[1] https://papermill.io/

[2] https://docs.papermill.io/

This is well-written.

I always thought that the "answer" to programming would be that one day everyone would use Lisp and with awesome tooling and libraries things would be wonderful. In fact, my plan for retirement was to build high-quality libraries for a Lisp language to accelerate this process.

Does the rise of AI bring an end to this dream? Is that, once again, we have solved the problem by adding more cruft? Rather than a superintelligent AI writing in the best programming language available, we're going to just spam lots of Python code until it works?

Does this matter? I don't know, I just wanted a world of elegance.

Love this, thank you. Are you building the mobile app? If you don't, others will.

This is classic NYT puzzle territory.

I actually think we're in a strange situation with AI compute.

Right now, we have models that are statistical models of language, with a world model and reasoning "falling out" of a lot of effort.

It's like we've made something that's a little bit intelligent, and now we're trying to amplify that trick to create something that's quite intelligent. And - don't get me wrong - it works.

But it's also super, super inefficient. We're having machines "think out loud" to compensate for the quality of their thought processes. We elongate the path to make up for the progress made on a given step.

I tink there's probably a much smarter way of doing things that will require qualitative architectural (and quite possibly hardware) innovations. Right now we're on the path to a Dyson sphere: that's probably not going to be necessary once we figure out a smarter way to think.

I've repeatedly bounced on and off 20VC by Harry Stebbings, but this year I'm finally hooked.

The main draw is the episode released towards the end of the week with Jason Lemkin (SaaStr) and Rory O'Driscoll (Scale Venture Partners). With the pace of AI announcements, it's been a good place to recap and analyse the week's events. In particular, Rory's insights are usually spot-on.

Fantastic. How did you learn Clojure? I'm a bit of a fan.

I don't think so. I think Tesla merger with SpaceX, which has the Cursor team and reportedly working on foundation model there.

I imagine the EU would block any attempted takeover of Mistral given recent Anthropic and US govt actions.

Only people with billions of dollars can train foundation models, yes.

But a competitor to Anthropic at the product level? With open source models, very little barrier.

Are the trains located where the urban areas are, or are the urban areas built around the train network?

It's chicken and egg question, but in Manchester and London it's very clear that mass transit led to urban development, rather than the other way around.

It's very surprising that cities like Leeds have no mass transit at all, and sizeable cities like Liverpool and Birmingham don't have much.

Click and hold the back button.

A huge part of the job of Software Engineering is producing the right amount of code at the right time.

I'd go further and say that usually the goal is to use as little code as possible without sacrificing readability.

Brevity is compression, and compression surfaces the salient points of a problem.

Elegance often comes down to brevity.

It's a good question. We had a previous language, JDoc, based in JSON. It covered only part of the functionality of the XML-based language and was really only for machine-machine.

We researched a bunch of others: languages like LaTeX and Typst are obvious alternatives. We also considered a super-augmented version of Markdown. Even looked at YAML.

Last year we chose XML as the basis for our document language.

It's been a good choice for designing a new language, but we've been really surprised by the poor quality of the available parsers. We figured it would be a solved problem, but we'll be writing our own at some point.

I think with recent changes they still retain data on Enterprise, no? Or have I misread this?