Digest 03/2026

I've closed more than 200 tabs today. This is the selection of most interesting articles I came across.

Interesting stuff

BAML which stands for "basically a made up language" is an opinionated pick of language features from various flagship programming languages. The intent here is to create language which won't let LLM to create slop. I like the idea very much and it reminds me how Polar forced their LLM to use single <Box /> component which only accepted particular CSS properties in order to ensure their design system.

Bun v1.4 is finally out! I am not going to comment on current beef between Jarred and Andrew as you've probably are fed up with it anyway. I am looking forward to use Bun.Image, bun audit fix, bun dedupe and bun prune the most!

If you are not ready to sell your soul to Dario yet you can try to improve your Node.js toolchain by using Nub.

Speaking of non-node runtimes - with Deno you can now create desktop apps.

If for any reason you've been trying to avoid JavaScript and you love CSS, this site is just for you. And this and this.

I like checklists for your website - security checklist, performance checklist, seo checklist... Now all in one place.

React just got its own doctor by the creators of million.js. Now you can ask your agents to fix simple errors with real impact.

Also looks like I don't need to think about re-renders, hooks order nor useEffect anymore - octane will solve it all. Will it though?

Tanner makes a great point on RSC - why they removed it from TanStack.com and what you should consider before implementing RSC in your project. It is great to see someone to abandon the most popular tool when it just don't make sense to use it anymore. You should always be able to do decisions based on reality not some fad or bubble.

slowjson is very interesting way of denying a service through sending JSON data over the network very slowly. Now you the attacker don't need server farm to attack anymore. Don't forget to set the proper request timeout on your server.

I am a big fan of valibot and I am using it everywhere where the TypeScript confidence is dissolved in runtime. I've been using versioned schemas with DynamoDB for quite some time but maybe it is time to try to use doba with valibot.

Lambda microVMs have been introduced back in June and ever since I've been thinking if it would make sense to run CI in it as Github self-hosted runners. Luc van Donkersgoed has provided really nice overview on how to set this up. I am still not sure how this will work with new pricing as Github announced that you will be paying even for running self-hosted runners. But it seems that this has been postponed and Github wants to "to take time to re-evaluate their approach".

From now on you can use same policies which are used in AgentCore anywhere. AWS released Dogwood as open source - not sure how and when to use this but it is interesting to read about how AWS security model works.

DynamoDB now supports vector embeddings and vector indices. You can now implement semantic search functionality without need of additional database like MySQL/Postgres or search engine like OpenSearch. The only catch is embeddings generation - here you will need to pay for Bedrock Titan or some other alternative or set up your own model and infrastructure.

This is how Cloudflare OS has been built eg. the security take-aways on how to run LLM agents safetly. TL;DR - do not trust them.

Now I can finally write everything in TypeScript without paying the no-native language tax?

Not every project should be rewritten in Rust, but if you've already rewrote sqlite3 and you've laid down the foundation maybe, just this one time it might make sense to rewrite also Postgres and other SQL databases.

Projects worth mentioning

• mac-ocr - no need to send your documents anywhere - OCR and search PDFs with Apple's Vision framework
• ShadeMap - simulate sun shadows for any time and place on Earth
• codeburn - see which project/taks you burn most token on
• whim - throwaway root shells in AWS Lambda Firecracker microVMs.
• celld - self-hosted, distributed Durable Objects
• swarm-forge - coordinate AI agents like Uncle Bob - I like the six-pack the most 🍻

Closing thoughts

Sir Roger Penrose, the English mathematician, summed the current state of AI perfectly:

We shouldn't call it AI (artificial intelligence) but rather AC (artificial cleverness).

-- paraphrase

Watch the video 📺 to get the exact words - I like the overall idea. To me there was always something wrong with the term AI, that's why I prefer LLM (large language model) which is technically more correct. Talking about using AI is like riding a bike and presenting it as a teleport - it will get you from place A to B but it sure ain't no teleport.

The irony of using AC instead of AI in this heatwave is the similarity to an acronym that also stands for "air conditioning" - we are using these more than ever as well.

While we are at it we should also stop using phrases like "burning the tokens" and instead use the more accurate "burning the trees" which is effectively what we do by overusing LLMs for tasks which could have been done without them (note to self: c'mon you can git commit & git push yourself).

There is nothing intelligent about over-using "artificial intelligence" when you can do the same thing without it - sometimes even quicker.

I know that global warming is not caused by vast GPU farms used for providing you with LLMs but it would be ignorant to think they have no effect either. I am not against the LLM use but as always with any powerful technology you should develop a sense of how to use it, when to use it and perhaps most importantly when not to use it. I own a car but that does not mean I drive everywhere - I use public transport, a bicycle or I walk.

Be responsible and reasonable when you build your next thing.