Ryan Williams

Software development, sports, startups, podcasts, cars. Building Car Curious. I enjoy a nice API and a shiny object. #goducks

[Ryan Williams @ryanwi]

Still loving Langfuse for one view of my agent spend and eval scores across every model and provider. Best new tooling find of my week.

[Ryan Williams @ryanwi]

Nate's Substack today cited Andon Labs' Vending-Bench: max-effort Opus 4.8 scored lower than high-effort Opus 4.8 on long-horizon tasks. Both scored lower than Opus 4.7.

[Simon Willison @simonw]

Uber reportedly now caps coding agents at $1,500/month per employee per tool - seems sensible to me, but it's also an interesting hint at the value Uber thinks these tools are providing
https://t.co/6YT0lCzPml

[Ryan Williams @ryanwi]

iTerm2: "update available." me, with too many agents running: no thank you, the little guys are busy

[Ryan Williams @ryanwi]

@nateberkopec thanks for these, I've done 1 and 5 (fnox adopted from your newsletter, mise from new team), going to work on the others.

[Acquired Podcast @AcquiredFM]

We're experimenting with a PDF companion for each episode — tables, charts, timelines, etc.

Here's our first one for Vanguard. Let us know what you think!

https://t.co/09HtFXc3iS

[Ryan Williams @ryanwi]

Sources: Addy Osmani, Gautam Koul (Thoughtworks), Colin Breck, Farhan Thawar (Shopify), Joe Schmidt IV (a16z), Mike Loukides (O'Reilly), Ivan Chepurin (Evil Martians), Pierre-Carl Langlais (Vintage Data)

[Ryan Williams @ryanwi]

Been reading a lot about what AI-assisted dev actually costs. Three things that stuck with me today...

[Ryan Williams @ryanwi]

AI collapsed the cost of generating engineering work. It did not collapse the cost of reviewing, owning, or understanding it. The engineers winning right now aren't the ones running the most agents. They're the ones who know exactly what they shipped and why.

[Ryan Williams @ryanwi]

Got into all of it in today's issue, plus Ford telling 4,653 Bronco Sport and Maverick owners to stop driving over a steering failure: https://t.co/KCZps3MobN

[Ryan Williams @ryanwi]

Toyota just gave the save-the-manuals crowd exactly what it wanted: a two-seat, manual-only GRMN Corolla. Rear seats gone, no automatic, suspension tuned at the Nürburgring. Then it killed its Lexus EV flagship the same week.

[Ryan Williams @ryanwi]

@nateberkopec I don't know why this is surprising. I'm neither an AI maximalist nor anti-AI. I use it in many places, writing included. The allure is too great to expect most people to not use it for writing. But I hope we can keep exercising/improving editorial control

[Ryan Williams @ryanwi]

Full issue here:
https://t.co/lcLulvh4X9

Free to read, subscribe if you want it daily.

[Ryan Williams @ryanwi]

GM built a smaller Hummer. Electric, about Wrangler-sized, and they're very careful to say it's just a concept.

More in today's Suspension.

[Ryan Williams @ryanwi]

Neat project for devs:

Which developer tools do AI models actually recommend?

https://t.co/MF0fzyGTM8

Via https://t.co/Sj9qCWf7HS

[Ryan Williams @ryanwi]

today's issue: plus 1M new-car buyers permanently priced out, Rivian R2 launches June 9, USMCA talks begin

https://t.co/Q1YWXl6UYz

[Ryan Williams @ryanwi]

Lamborghini CEO says Ferrari's Luce backlash proves he was right

[Ryan Williams @ryanwi]

@RaulJuncoV Within-digest dedup (same item matches multiple searches, show once) is clear. The harder case: an item matched last week's digest and is still available. Do you send it again? How do you track "user has already been notified about this specific item" across sends?