One gallon of almond milk requires 1,600+ gallons of water (likely much more — given some estimates are each almond requires up to 3 gallons of water). US almonds are almost entirely grown in California and consume 1.5T to 2T gallons of the state’s water.

It’s pretty dumb to compare AI/data center usage to water usage when something that’s actually wasteful such as almond milk exists.

There’s better metrics to make one’s anti-AI point.

As a reference, one gallon of cow milk requires about 250 gallons of water.

I feel like we need some citations for these figures. Our World In Data [1] disagrees. I've not seen any analysis that puts any plant milk at a higher water use figure than dairy.

[1] https://ourworldindata.org/environmental-impact-milks

>Almonds provide a real utility for humans: food.

Let's do not act as if there wasnt other food, a lot of it actually.

Meanwhile LLMs aren't that easy to replace, they're biggest human achievement up to date.

Search e.g

Strongly disagree.

Now when I use an internet search, the top answer is an AI guess/approximation to the result and can be completely inaccurate. I have to ignore that and move onto the traditional results and hope that they aren't A.I. slop websites.

LLMs have degraded the internet and search.

I'm talking about searching via LLM, not about experience of using Google.

True, you need extra UBO filters nowadays to block LLM results in search.

"Meanwhile LLMs aren't that easy to replace"

A BBS chatbot from the 80s (LISA) was almost as good as modern LLMs and that was just based on key words.

So, uh, yea, chatbots are very, VERY easy to replace.

>A BBS chatbot from the 80s (LISA) was almost as good as modern LLMs and that was just based on key words.

Of course, because it's not like whole world went crazy after ChatGPT and hundreds of billions of dollars are being constantly invested into many of technology stacks - software, hardware, semiconductor engineering in order to push this technology even further, lol.

> A BBS chatbot from the 80s (LISA) was almost as good as modern LLMs and that was just based on key words.

Lol. This is what anti-AIs actually believe!

It's really weird to me that datacenter fans all have detailed knowledge of almond farming, but never seem to notice that the nutritional value of ChatGPT is not on par with your average box of almond milk.

Because nutritional value is not the only thing of value. We invented this clever way of uniformly valuing things called "money" that abstracts over all of those differences though.

Indeed! Especially relevant because almond milk is sold at a profit, which is something the LLM industry has yet to achieve.

Too late.

Big fan of Windmill here! Use it heavily. The way you guys integrate seamlessly with Bun/DuckDB/Postgres/Python is just fantastic. Keep up the great work, Ruben!

It’s a very good post — and I do agree with the main ideas. It’s pretty remarkable how good writers like Scott Alexander and others are able to consistently pump out good writing, especially when the key does seem to always comes down to clarity (and mostly revision for me). Maybe reps are able to give that over time, but even with getting older and now being able to bounce ideas off LLMs == it still takes me so many iterations before I feel like my prose / ideas / outlines are worth sharing.

Thanks for this! I had been meaning to mark a phantom notification as read — and this snippet came in handy.

We’ve similarly had a lot of success with self-hosting Windmill.

Super insightful. I hadn’t been able to articulate the same feelings.

    Even as Next seppukus itself,
    people will likely just fall back 
    to React on Vite…

This is my exact read on the situation, as well. I’m not sure if anything can meaningful affect React’s domination in the short-term or medium-term, even with the accumulation of poor choices.

I haven’t tried tanstack-start, but I wouldn’t be surprised it becomes the defacto react framework instead of next. Everything by Tanner Linsley is just so well thought out and the DX is amazing. If his framework is the same level of quality, without any major gaps compared to next, it will probably blow next out of the water.

And Tanner is already a huge name in the typescript/react world, so I think there is actually a chance.

Thank you for your kind words and support!

This is really well done!

I tried Deno for awhile — the ability to run Jupyter notebooks was a cool idea — but I’ve been Bun-only for around a year now. It’s just significantly faster and easier to use.

Congrats, Tom!

I’ve been primarily a Python developer since 2012 and recently switched to uv. The ability to manage dependencies, venv, and multiple Python versions makes it best-in-class now. It really is a fantastic tool.