I wrote my own harness in Emacs and it’s completely ridiculous how well it works. Auto-compact is the only missing feature on my list. Claude‘s approach, if I understand it correctly, invalidates a lot of cached context, and I‘m thinking about a more cache-friendly strategy.
Humans will not regurgitate longer segments of code verbatim. Even if we wanted to, we couldn’t do it because our memory doesn’t work that way. LLM on the other hand can totally do that, and there’s nothing you can do to prevent it.
If the censoring is at the DNS level, can the admin please replace the domain name in the url with the ip address to which it should resolve? Thank you.
It actually interesting that Carlsen (likely the best classical player of all time) hasn’t overfit classical chess to the point where it hurts his ability to play other variants.
Sure, the market chose Markdown, but this simply led me to the conclusion that the market isn’t worth following. Of course the mismatch creates some friction, but the benefits of org-mode, for me personally, easily outweigh that.
What problem is this trying to solve and does it actually succeed at solving it? I‘m struggling to see the appeal given that the JS still needs to model the internal structure of the template in order to fill the slots.
The pixels may be 7.5 microns but you’re forgetting that they are viewed through a lens. The point stands: 4K pixels for the full field of view, which is a lower density than 4K for a small screen.
> Data were collected from 229 participants working in a management position in
the United States through Amazon Mechanical Turk. Participants were compensated at a rate equivalent to $7 per hour.
I have a hard time imagining people in "management positions" working on Mechanical Turk for less than (US) minimum wage.
If I understand correctly, participants also didn't rate applications with photos, but just photos. Seems possible and perhaps likely that the pragmatics of this specific task would activate biases that may not be at play (or to a lesser degree) in an actual application process where the photo would be just one piece of information. The task really seems designed to maximize effects of beautification while sacrificing generalizability to real-world settings.
The thing that got me annoyed was the bold claim that it‘s the fastest terminal emulator even though many people reported that other terminal emulators were faster. For instance, urxvt was 3 to 5 times faster last time I checked (in addition to also being rock solid).
Great essay. Basically, they say that surprise doesn't tell us so much about the world but more about our model of the world. The degree of surprise can essentially be seen as a measure of our ignorance about some aspect of the world. As such surprise can serve as a guide for improving our understanding. However, it would be a mistake to see ignorance necessarily as a shortcoming. A simplifying model of the world is preferable if it gets the job done. In contrast to that, a perfectly accurate model of the world would render decision-making computationally intractable and is therefore not desirable.
> When context gets too long, maki compacts history automatically: strips images, thinking blocks, and summarizes older turns.
Don’t the summaries of older turns effectively invalidate the context cache, such that you consume less tokens but more expensive tokens?