Ben Swift

The biggest thing for me has been I have noticed that I am just not using Fable any more, even though I am on the juiciest plan and have enough tokens.

This is good snark but I think it’s also the nub of it.

I don’t get all these developers carrying laptops open’ memes. The venn diagram of people who are well positioned to use agents effectively and people who can set up a remote dev environment they can access over SSH is a perfect circle. 

This video, in both form and content, is honestly the first thing I will now show people to get them up to speed on what models are now capable of.

whose Claude number is 5.5? maybe it's mine?

just got absolutely rinsed by a colleague while presenting a workshop: “you dance like a computer scientist“

I heard that Anthropic are just about to solve housing inequality in Australia

This is why I care about AI coding.

lots of model releases today (muse spark, gemini flash) indicate that there's another new model (openai in this case) coming out soon which is good and they all want to dump their mid ones ahead of that release

I think Paul’s right here, and agree with the ‘inertia means some folks will be way behind’ thing. The 10x programmer has largely been a myth (fbellard excepted) but soon we’ll have 1000x programmers relative to the old way of doing things

"A chatbot can’t take a hint, it doesn’t know who’s paying. It can’t interpret the tone of your voice or catch the wink; to get the report you want, you have to say what you want, spelling out in writing what everyone involved would once have left unsaid." @benswift.me on AI and discovery

my (current) grand unified theory of agentic AI is the more you can make your problem look like 'running a bunch of commands against a git repo full of files' the closer you can get to achieving nirvana

Shortening the latency between making an architectural decision and encountering the consequences of that decision has taught me a lot about software development that I think would take a long time to learn otherwise.

Fourteen months of Claude Code session logs: 55 billion tokens across 13 models (and what they would have cost at API rates)

So many of the norms underlying academic culture relied on frictions associated with writing. It was entirely predictable that academia would be profoundly harmed when a machine came along to remove those frictions by making (insincere, inaccurate) communication effortless.

github.com/yc-software/qm I’m not sure what the future of work/collab is, but I’m increasingly convinced that if you’re stuck in an o365-document-means-of-production mindset you’re ngmi

An experimental typeface that shows one message to human readers and a different one to AI vision models, riffing on Decoy Font.

has anyone else noticed fable leaving "Decided YYYY-MM-DD" breadcrumbs througout files? it's annoying - git history is for keeping track of changes (and commit messages for reasoning) but the files should just represent the current state

revenge of the dilettanti! Card carrying member right here

when is someone gonna create a fork called ‘zag’? the snarky announcement tweet writes itself

DeepSeek wasn't a fluke: the moats around frontier AI are still shallow, and trillion-dollar valuations now depend on them being deep.

There are legit criticisms of this stuff, but this particular one is nonsense. The bitter lesson is exactly the opposite - these algorithms scale better than anything we've ever done before. That's what makes it bitter, but it doesn't make it false.

The blog's standard.site records now carry cover images, a publication icon and theme, and the strongRefs Bluesky uses for enhanced link cards.

being in the loop is important but i keep seeing people misunderstand this - your job isn't to micromanage the agent and get it to execute its steps in a way you think is better than what it'd do on its own it's to come up with cool things for it to work on

A zellij session switcher that shows how many Claude Code agents are running in each session, and which of them has stopped and is waiting on me.

A deadpan reference site that reports every university's single best ranking placement across all the major tables --- built at the ANU School of Cybernetics.

This is so great

The verification I never verified

ouch

Today we find out whose Claude number is 4.8

Cool idea but isn’t that the Village Roadshow logo?

I am irrationally into these types of ‘dumb idea taken to the extreme’. So great.

the first two sections of this readme are so good

there’s never been a better time to be a dilettante with a PhD in CS

This is great news - I’m teaching an upcoming Agentic Coding Studio course for final-year undergraduates, and I’ve been wrestling with what I say (be careful!) and what I do (yolo mode) with Claude Code. This is the way.

Inside every model there are two Claudes

Claude Code really loves talking about a "belt and suspenders" approach to things

i'll take this further: "slop" is just a pejorative way to say "production has become much cheaper and things can now be made that could not justify their expense in the status quo ante cost structure". the printing press was used for slop too. slop is just the expression of the freedom to make.

never interrupt your enemy while they are making a mistake or your child while they are eating a vegetable

Good news for Australians — the peak hours are 11pm–5am our time, so we basically get 2x usage all day every day 🦘

that would be so great

Did you know that arXiv is exploring to add Typst? The world's largest preprint archive faces unique challenges to keep science open. Join arXiv veteran Norbert Preining and learn what's ahead for arXiv and Typst. www.youtube.com/watch?v=zNZl...

increasingly believing much of the communication gap between LLM Enthusiasts and LLM Skeptics on this website boils down to “are you someone who has largely experienced a *reasonable* amount of job security in tech, or are you someone who has not experienced job security and likely aren’t in tech”

Dear @anthropic.com please consider extending your Claude Max subscriptions to open science software devs! Our repos may not have as many stars as those of some of our OSS peers, but we develop critical infrastructure in focused areas, and might even benefit more from what Claude Max offers!

I’ve been using this workflow for a couple of months now via Tailscale + the ShellFish SSH app - it’s great. But having official support is nice too

Say what you want about AI but since Opus 4.5 I’m having the most exciting time of my career. Both in terms of learning more and building faster. It’s a new tool you’ve gotta learn but once you have, you’re on a whole other level.