That was me. The engine is deterministic so I wrote a beam solver for it. My score should've been 208 (pretty sure it could play forever; I capped the solver at a max time limit) but I messed up one of my moves (I was manually moving the pieces instead of submitting the final move list with curl).
Assuming OP means they used beam search on the sub trees of possible moves, which yields some local minimum / maximum in the objective space. Especially useful here because of the sheer multitude of possible moves, and also RNG.
Beam search can be googled for useful results, beam solver can not.
Curious how this compares to https://sentienceapi.com/. My understanding is that Sentience uses deterministic "semantic snapshots" to try and give agents a more reliable browser interface.
> agent-browser click "#submit"
> agent-browser fill "#email" "test@example.com"
> agent-browser find role button click --name "Submit"
I appreciate that there’s innovation in the space, we will get closer to the interface that’s most appropriate for models to tool-call. I’m going to check your link out, sounds interesting.
it's interesting to see how things will play out, but I really believe that doing Claude Code (maybe with Opus 4.6) + click tool + move_mouse tool + snapshot page tool + another 114 more tools is definitely not the best approach
the main issue with this interface is that the commands are too low-level and that there is no way of controlling the context over time
once a snapshot is added to the context those tokens will take up very precious context window space, leading to context rot, higher cost, and higher latency
that's why agents need to use very large models for these kind of systems to work and, unfortunately, even then they're very slow, expensive, and less reliable than using a purpose-made system
I wonder if a standardized interface will organically emerge over time. At the moment SKILL.md + CLI seem to be the most broadly adopted interface - even more than MCP maybe
Why is it that Tailwind causes such emotional reactions in people? Functional CSS has been around a long time now.
When functional css ideas were introduced it presented a novel solution for scoping issues, naming issues and bloat issues. I remember discovering basscss and tachyons around 2015. The paradigm immediately felt like a huge improvement in my workflow. The iteration was so much faster than what I was used to with Bootstrap, BEM, or SCSS- I felt like I could design in code for the first time. My favorite workflow for web design became loading basscss in a CodePen, with almost no idea of the outcome I wanted and iterating towards something. What’s kind of amazing is that it allowed for these sorts of quick sketches and improved maintainability in large projects.
It also led me to my first inkling of a design system- through the use of composable and constrained styling primitives. A lot of the ideas introduced with those libraries now saturate the frontend ecosystem. And keep in mind, functional CSS, or at least the principles behind it have evolved a lot since 2015, particularly when it comes to components and CSS-in-JS. The original creator of basscss has gone on to create a number of equally impressive projects that advance the main ideas: styled-system, rebass, basically everything from @jxnblk.
I haven’t used Tailwind or any other functional CSS philosophy over the past few years since I find primitive and composable styling components, CSS-in-JS, themes, and CSS variables offer even better workflows for building the applications I work on. When Tailwind initially made a splash I was already pretty into what I consider this next phase of styling libraries- so I was a bit confused why people were so excited about it when basscss and tachyons had been around for years. Looking through the docs it seemed like they had taken basscss and tachyons and layered a bit of Bootstrap on top (in the form of nicer visual defaults and some component patterns).
Outside of the core functional principles, I don’t like the idea of Tailwind as much as I do basscss/tachyons. I think Tailwind tries to put too much of your styling in atomic classes. Part of what makes basscss great is the small API and the constraints it introduces on reusable patterns. I don’t want all of CSS to be represented as atomic classes- just the stuff I reuse 90% of the time. It also seems like a bit too much tooling when one of the main advantages I saw in the older libraries were that they were just plain CSS- you could use them on pretty much any project, regardless of build tools or architecture. I used to start with basscss on a project and then add my own layers of patterns on top- extending the system as needed.
That said it still offers a lot of the good stuff, combined with great documentation and more thought put into component patterns. It’s a great starting point, especially if you don’t want to or can’t take on runtime costs for CSS-in-JS (this is also evolving but still a concern). Its success isn’t a fluke or a result of misdirection as the author of this critique alleges. Reading through Adam Wathan’s original post on functional css (https://adamwathan.me/css-utility-classes-and-separation-of-... he describes the exact same trajectory in CSS thinking I went through as a frontend developer through the 2010s. It’s a great read and I think speaks to the level of consideration put into the library. If I were to try a new project today using functional CSS I’d definitely give Tailwind a shot.
Having spent the past couple years working on a large design system, I am most excited about @react-aria and @react-stately. There is huge potential for allowing resource constrained teams to build out a truly customized DS in a fraction of the time. Even for a well resourced team, these hooks can make extending their existing DS and borrowing complex behavior much simpler/faster. The biggest challenge for a DS at scale is in how you can adapt it to the infinite edge cases that a product eventually uncovers- while staying consistent and leveraging existing code. I think these libraries offer the best model yet for achieving this. Congrats to @devongovett and the team at Adobe for realizing these ambitious projects and successfully open-sourcing them.
Thank you. This is the reason I can’t start to write. I love brevity and clarity.
Prolonged blog posts where the crucial information can be made up in 2-3 lines is now the norm and I can’t for the life of me adapt to this style of writing.
I would love to have a tldr; or executive summary (I know it has a different meaning, but you get the point) up on top and then the rest of the article, for those who are now intrigued to read more, available to read.
I want to call out https://www.regular-expressions.info/ since it was, for a long time, one of the better/best resources on regexes that I was able to find. I learned a lot from this ... guy, essentially.
but, I always try to add a warning when recommending - should use them only for the flavors supported (PCRE, Python, etc) I've seen many using it for cli tools and wonder why things like non-greedy or lookarounds don't work.
