eob

Simon I hope from this day hence, your bio always includes:

"Simon Willison, among other things, is an advocate for the inclusion of pelican geometry in LLM training datasets."

I think because you don't know /which/ developer you're going to get.

One interesting aspect of LLMs is that each one, weights frozen, can be thought of as a single developer whose work you have already evaluated.

The cost of finding, evaluating, and negotiating with a new human is tremenous.

These comments are so filled with cool projects and visualizations I’ll throw my own out there:

Trying to generate non-slop, print-worthy recipe books for different diaspora communities.

www.robotbookclub.com

I’m pretty sure the recipes have surpassed slop. The photos are pretty close. The layouts, intros, and “chef photos” need a lot of work though.

This is how I write down recipes for myself! There’s got to be some cognitive archetype that works better with this style encoding.

Do you have any suspicion about what is different between the backends?

That's an absolutely bonkers statistic: it would mean spurious differences in hosting container overwhelm the performance differences between models.

My guess it's just the emergent behavior that results when a company doesn't provide developers time to fix bugs.

If their week is already booked full just trying to keep up with the roadmap deadlines, a bug ticket feels like being tossed a 25lb weight when you're drowning.

You could say: "but have pride in your work!"

But if your company only values shipping, not fixing, that attitude doesn't make it through the first performance review.

You think like a person who’s debugged large systems failures before :). That feels very plausible.

Some outlets reporting T-Mobile and ATT as well.

I assume state on state cyber attacks are commonplace but get minimized to avoid public fear.. perhaps this will be the first notable one.

I build coding agents for a living, and I'm struggling to map this onto the set of things I do at work.

In general, interoperability and user choice are really important for us to get right as the community of people building AI platforms...

Have others reading this document been able to map it onto their work?

As a specific example:

ai://bank/service/payments?amount=10&currency=USD

I'm not sure what this is representing here. Is it a way to encode a clickable link to chat with `bank` about `service/payments` with a few additional args attached?

Third party apps can’t use the network though. Iirc there’s an async message queue with eventual delivery that each app gets, which it can use to send messages back and forth with a paired phone app.

Bravo -- this is fantastic.

I've been waiting for this ever since reading some interview with Orson Scott Card ages ago. It turns out he thinks of his novels as radio theater, not books. Which is a very different way to experience the audio.

Or vice versa - perhaps some subset of the "thought chains" of Cyc's inference system could be useful training data for LLMs.

An aspect of these self-improvement thought experiments that I’m willing to tentatively believe.. but want more resolution on, is the exact work involved in “improvement”.

Eg today there’s billions of dollars being spent just to create and label more data, which is a global act of recruiting, training, organization, etc.

When we imagine these models self improving, are we imagining them “just” inventing better math, or conducting global-scale multi-company coordination operations? I can believe AI is capable of the latter, but that’s an awful lot of extra friction.

Are you comfortable sharing the video & lip-sync stack you use? I don't know anything about the space but am curious to check out what's possible these days.

This looks fantastic — congrats on the launch.

As an armchair observer, the agents + browser space feels like it’s waiting for someone to make the open source framework that everyone piles on to.

Proxy rotation sounds like a solid way to monetize for businesses.

There was an SF office of 18F -- IIRC it was in the building to the right of the Civic Center park as you looked at it from the Bart stop. They were great folks from every encounter I had with them.

Let's presume this gets developed to ShadCN/Tailwind level quality.

In that world, what would be the tradeoffs between:

- NextJs + Tailwind + ShadCN

- NextJS + ?? + Shoelace

I don't have a good sense of how web components compare to practice of "copy-by-value, then compile-into-binary" UI shipping that's common in the Next + Shad world these days.

Have you ever managed a complex, dynamic, changing system and found that the optimal size based on current conditions was 20% less than it was at some prior time?

I can think of all sorts examples.

So GCS customers will trust their codegen product. (Engineers aren’t the buyer; corp suite is)

I think Taiwan is worth adding to the list of healthcare systems that folks study. I believe it may be similar to the Swiss system.

Everyone has national health insurance, but you also get to choose where to go, and some doctors also offer non-insured services that you can pay for out of pocket.

The result is universal coverage combined with a competitive market that drives prices down and encourages innovation.

I know this is just anecdata, but having held an insurance card there for a while, our family was always able to see our family doctor the same day we called. And the one or two times a specialist or emergency room was needed, there was minimal hassle.

I'm sure there are problems with it, too -- I just don't know what they are. As a customer/patient, it seemed to work far better than the American system I'm used to.

A Project Jengo grant for using Agents/LLMs to identify prior art could be a fantastic experiment..

This is still mostly true. Kids books also have bopomofo rubies, like the kana rubies in Japanese. And occasionally you'll see bopomofo as a typographic choice to represent a sound that feels more natural in Taiwanese amidst an otherwise Mandarin sentence.

This is just my personal experience, but I think the big change in the past 15 years isn't Bopomofo -> Pinyin, but rather Wade Giles -> Pinyin. Bopomofo seems equally prevalent, but the Wade Giles romanizations on street signs have begin to get replaced with Pinyin for the sake of non-native speakers who are almost certainly more familiar with Pinyin than WG.

This is incredible --- @chearon thank you for open sourcing this!

I think most folks probably don't realize how difficult it is to go from HTML -> PNG programmatically. You get hit with a thousand papercuts related to either Node<>Browser differences or HTML<>Canvas differences.

Surprising to see another spacecraft at 1:30

This has some pretty great ideas within it. A few that stuck out to me in case others don't want to skim the full spec:

Start with YAML/JSON/TOML, but then add:

- Optionally strict typing and structs

- Inheritance

- Dynamically computed properties

- Functions & object methods, which are really more like computer properties

- If/Else conditionals, which are really more like a ternary

.. and managed to do it while still making it look like a simple config file.

The closest analogue from memory might actually be Terraform.

The one thing that had me squinting was the object extension syntax. It looks powerful.. and I understand its value.. but wow that's ends up being eye-full of () {} and [] all in one code bock.

The longitudinal studies of the of the kids who were 0-10 during the pandemic are going to be fascinating.

As a tiny example: I've anecdotally heard that kids in the 5-10yo bracket need their parents in the bed to fall asleep at a higher rate than would typically be normal. Probably because parents cuddled their kids to sleep during the lockdowns as a stress reducer for both of them, rather than turning out the lights and leaving the room.

MLE | training data

Just writing hand wavily :)

As for practical use cases, one is to find an approximate optimization to a function

- You want to find the min/max of some probability distribution P(x)

- P(x) is too complicated to find a closed-form min, but you can draw samples from it.

- So instead, you carefully construct some OTHER probability distribution Q(x|θ) that you claim is structurally similar "enough" to P(x), parameterized by θ.

- Now you find the theta which minimizes the KL divergence KL(P(x) || Q(x|θ)), which is equivalent to delivering you the parameters of θ to Q(x|θ) that make it [approximately] "most" similar to P(x) without ever having minimized P(x)

It was a trick that came up a lot when AI consisted of giant Bayesian plate models for each specific task that you had to hand-optimize.

There used to be a hack to get around password protected screen savers in which you’d do the following (IIRC):

- Drag the password window around the screen furiously

- Press Ctrl-Alt-Del as you did it

I think it would just dump you straight to the desktop.

I remember doing it in Circuit City as a kid to access computers.

ChatGPT is amazing but ultimately unwieldy for directed, long running relationships more nuanced than general themed chit chat.

We’re in the nascent stages but I think there will probably always be a community of folks who want to add more nuance to the communication, whether it’s reveries that enact a mood or goal, tie-ins to other services, etc.

Eg imagine wanting to have your ChatGPT DnD master also keep some kind of score. It may be ultimately easiest to put a wrapper around a themed GPT window that imposes a predictable way to do that rather than require everyone to figure out how to prompt it correctly.