Is your US shipping still on hold? I ordered a bunch of Faction Paradox books which just arrived (thanks! looking forward to reading them) but I forgot to get the EDAs collection before the tariffs arrived.

There are countless papers I've reviewed where they say they used a 2GHz Intel processor when reporting performance numbers and I have to ask "which one of the 25 years worth of 2GHz Intel processors did you use?" System environment reporting standards are just terrible.

Many conferences also use openreview.net, so the reviews (for accepted papers after publication) tend to be open as well.

I've only ever published in CS so I can't compare it to elsewhere, but it's typically 20-25% at the top conferences. Not sure about the journals. So much stuff is on ArXiv now too, so lots of things become well known before they finish going through peer review.

The best journal in Machine Learning (jmlr.org) and most of the best conferences are entirely open access and free to publish in. It's been this way for at least 15 years and is working pretty well.

I've derived a bunch of amusement from asking different LLMs about various obscure Doctor Who spinoffs and watching each one straight up lie to me. The 90s novels must not have been in the books they scraped from the internet.

I watched that one last night. Things would have been much better for everyone over the next 6 years if they'd decided to drop her in the gamma quadrant.

Yeah, I have a rough idea what JVM startup looks like, and a similarly rough idea of how processes get launched & what libc does to bring up a C program, but a web browser is far beyond my expertise. Even node.js or CPython are systems I don't understand startup for.

Feels like the difficulty decrease from using a higher level language to implement fizz buzz is not commensurate with the difficulty increase in describing how the language runtime works. It's much easier to describe how a (*nix) C program executes than how Java, Python or JavaScript executes.

Thanks for the stickers, they are great. We were considering making a similar one with a pile of old Sun servers on, we're the office hardware horders.

Good job my new laptop still has a bunch of space for stickers because I need that one.

Is it a transformer decoder? If it's encoder or encoder/decider then an auxillary loss head like BERT has for NSP will train that into the representation. If it's a decoder then that's more of a pain as the place where you do the prediction matters and sticking it on the last token is bad.

Are you including things like Mission to the Unknown to get there, or people who are more Doctor adjacent?

Early Wikipedia? Before the tooling and culture built up it was a lot more vulnerable to trolling and misleading edits. I remember my teachers continually saying to check its sources rather than use Wikipedia text directly.

Given how they work getting to zero hallucinations requires incredibly sharp probability distributions over the next token, leading to no variability. So I suspect the hallucination problem isn't going away with generative systems.