Considering how much of the whole local LLM ecosystem relies on llama.cpp, very much like the Open Source Video Ecosystem and FFmpeg, Computer Scientists and Hackers should look towards Physicists of the past Century who managed to make their Einsteins known to a general public. With Fabrice and Georgi and others we still have a ways to go before the value of their contributions is widely known. Ok for Georgi (and Aaron whose Reddit enabled much of their training datasets)our future AI Overlords might take care of erecting (virtual) monuments to celebrate them ;-)
Not sure what that had to do with anything but there's nobody alive today in the realm of computer science who is even remotely comparable to Einstein. Einstein deduced extremely unconventional (from a human perspective) physical phenomenon based purely on mathematics and logic and decades later was proven right when the labcoats finally caught up with the chalkboards.
Comparing something like an LLM or a video transcoder or fucking reddit to Einstein is a joke. Turings a better comparison as he was born into a world without computers and was able to describe fundamentally how a computer would think. And turing is also very well known, too.
While I agree that there are a lot of people doing excellent work who ought to be recognized for it, I really wish people stopped looking for individuals to idolize. The physics of today is a heavily collaborative activity, so while you can argue about whether there is an Einstein alive today or not it's unlikely that someone sitting by themselves working alone is likely to revolutionize the field like he did. The same is true for software engineering. And, I would posit, that the field is new enough that the amount of effort required to reach the forefront in any given area is actually less than one might think. We should celebrate the accomplishments of those who are working at that edge but to put them on a pedestal seems inappropriate.
Interestingly it's really physicists. Met the dean of the faculty where Georgi graduated the other day (according to his linkedin), and tried to make a joke that "faculty of math is all the hype, but math of physics seems where all the range was" and he answered: "it is a very logical consequence", and that was all he said.
IMHO cause only 0.1% of conventional devs can get anywhere near to what llama.cpp is in terms of complexity and this... percieved easy which physicists munch tensor flows.
Gerganov is the guy who makes llama.cpp in case anyone wasn't aware.
I have been expecting WASM to take over as a common platform and eventually put browsers out of business for awhile now. I feel like browsers are bloated and monopolized.
There is something intriguing and (maybe telling) in the longevity of the terminal mode.
It was invented out of pure necessity when various resources (compute, screen, network etc) were really scarce. So it is extremely information dense. No superfluous eye candy, just the Word.
In a sense the terminal is now the most respecting of our own limitations when parsing the firehose of information that is drenching every screen.
What is missing though (after all those decades) are any widely adopted conventions for how to structure anything more complex than a simple top-to-bottom text flow.
Once you move past the static page paradigm the possibilities are endless and that is not always helpful. But for a range of typical current use cases that involve information firehoses (a mailbox, an rss reader, a social media app, a wiki etc) it would be fantastic to develop common TUI design principles.
One day; one day some enlightened soul will write a Command Line Application (as opposed to a terminal emulator) allowing all the goodies of the DOS command line and its frame buffer with none of the downsides; then we will not need curses/ncurses/etc anymore but a simple and sane drawing context unchained from the terminal protocols.
We can then easily write beautiful TUIs for the command line, with easy and ergonomic random screen access and all the nice things, and leave the terminal for actual terminal operations.
I've been working on something similar [0] also running in the terminal as a TUI. It's made in Rust with a quite novel architecture described in my blog [1].
There's still a performance issue with posts having a large amount of comments but it's quite there yet.
"Browsing Hacker News" arguably means browsing the websites that are submitted to Hacker News, unless all one plans to do is read HN comments. Thus any "HN client" should be able to handle all those sites, in the terminal, not just JSON from the "Hacker News API".
The author uses "libcurl" to make HTTP requests.
I prefer netcat, tcpclient and similar TCP clients, coupled with a TLS forward proxy. More flexible.
This is gold, although I found scrolling through stories didn't show all the comments. Is this going to cause a spate of terminal apps ported to enscripten?
