o1/o3 could already out-think Claude in tough coding problems moons ago, but made frequent surface/syntax bugs and had no cli. it's unsurprising that gpt5 is now skewering Claude. the models were already smart enough. I'm curious how much of the diff was training, vs 'live

We can use this practice as a canary in the coal mine to measure whether Germany is serious about startups. As long as it persists, they're not.

fixed that, added a color url param to override the built-in color scheme on a per-link basis

colors are auto-assigned, which is (painfully) obvious from the green roses in the example.

it's Markdown-friendly. highlights are described by ordinary links, which can refer to nearby `pre` elements. (my own use-case is for Jupyter notebooks which are published using MkDocs or quarto) https://t.co/RBzqWwUdmD

made a little code-highlighting tool, for explanations that sit next to code blocks

The new "artifact" UI is extraordinarily useful - you can ask it to generate html/js, and these are immediately executed in a second column. Instead of the typical ChatGPT infinite chats, artifacts are updated "in place" (sort-of) and you can flip back and forth between versions.

Finally, I had it write a short guide explaining its work, asking it to focus on how the curve is generated, and how particles are placed and moved:

It worked so well I kept going and made the thing more complex & added visualizations to see how the particle placement algorithm was working, and to see how far particles moved from their origin over time. runnable demo:

I saw a neat wiggling gif on NYTimes and thought I'd take the latest Anthropic Claude model (claude-3.5-sonnet) for a spin to recreate it. it made a working-but-incorrect thing immediately, and subsequently corrected all obvious mistakes quickly, reasonably, & almost error-free.

Kinda like what @unkai, @mhuebert and @jackrusher built here, upgrading emacs' which-key and command palettes into a graphical command bar https://t.co/iTDEMLHDlh Easy to see how each button can be its own React-like component with arbitrary graphics

Embark: Dynamic documents for making plans. https://t.co/soobH5ge6L

I made Lude. It lets you build video games quickly. Here's a demo where I build Breakout - https://t.co/mWsTEPs5CT

Even by the high standards of Cathal Kelly's columns, this is a superb column. Gambling ads are changing how we view hockey /via @globeandmail

This was not cherry picked, first time trying the Edge sidebar

Bing's autocomplete is something else. jfc

https://t.co/FJoEfKqwN7 this is 🎉

Werner Herzog on the importance of prompt engineering:

how much tweet would a tweetbot tweet if a tweetbot could tweet tweets

🤔 so they've squeezed the "promoted" label into the size and shape of a blue checkmark. Impressive