I wish Google can bring back the OG Pixelbook, where "AI" merely means Google Assistant.

tl;dr we reinvented UUID and it works well

Certainly not true. UUIDs have structure to them, and variants. Trace IDs are just 128-bit numbers, with any further semantics (almost) completely non-standardized (some systems encode timestamps, etc). They slapped a “last 56 bits are random” flag (not in the ID itself but as external metadata, so not like UUID at all) later giving IDs just a bit of semantics, but it’s not a reinvention of UUIDs.

Okay, sightly more bits than UUID v4. The whole article is merely reasoning "why at least 128 bits are required", and if you smuggle some non-random data inside these bits the entropy can only drop, making it more vulnerable to collision, i.e. inferior to UUID v4.

I kind of addressed this in https://news.ycombinator.com/item?id=48060549

Actually, because the birthday paradox has k^2 as a term, this is actually less true than you might think. Having a time component actually reduces the chance of collisions over the long run, albeit at a cost of reducing the number that can be safely generated in any given quantum.

If you consider a 128-bit random number, you effectively have 64 bits of allocation space before you are likely to get a collision.

If you devote 48-bits to time, which provides millisecond accuracy for 9000 years, you then have 80 bits of randomness, effectively giving 40 bits of allocation space per millisecond before you are likely to get a collision.

Instead of approx 2^64 allocations across all time before a collision, you instead have 2^40 (1 trillion per millisecond). That sounds like a poor deal, until you realise that the factor is only 2^24, or 16777216ms or under 280 minutes.

So in reality, reducing the random space and increasing bits that are guaranteed unique is actually a great trade.

I realized that my mentioning UUID without v4 was misleading.

It's C library taking care of the "b" part for you according to the article.

It's the other way around. It's the C runtime that treats text ("t") mode differently, because the C standard specifies \n as a line delimiter but the Windows convention is \r\n. In text mode C stdio translates between \n and \r\n. In binary mode it does no translation.

Note that when neither is supplied, the text mode is the default. This is why I said that it is the C library handling the "b" flag.

There is even a table copy-pasted into a paragraph without noticing.

> What’s needed is something different:

> Requirement ptrace seccomp eBPF Binary rewrite Low overhead per syscall No (~10-20µs) Yes Yes Yes [...]

null hypothesis bot

Reminding me of the Shoelace [0] project, which was rebranded as Web Awesome. The original (v2) repository was then archived.

[0]: https://shoelace.style/

The watch is interactive! Nice detail

Productivity is finite. If you pivot entirely to the AI stack, you're going to lose bandwidth for everything else. It's an opportunity cost problem.

A better example would be to use LLMs to generate passwords or secret keys. Then even if it looks random to human, the inherent bias would make it a security disaster.

You can still obfuscate JS heavily and make a VM that executes also obfuscated code calling arbitrary browser APIs. At least In WASM everything is sandboxed so the attack surface is smaller.