> What exactly is the difference between lms and llmsterm?
With lms, LM Studio's frontend GUI/desktop application and its backend LLM API server (for OpenAI compatibility API endpoints) are tightly coupled: stopping LM Studio's GUI/desktop application will trigger stopping of LM Studio's backend LLM API server.
With llmsterm, they've been decoupled now; it (llmsterm) enables one, as LM Studio announcement says, to "deploy on servers, deploy in CI, deploy anywhere" (where having a GUI/desktop application doesn't make sense).
> I think I kind of have an idea what the author was doing, but not really.
Me neither; However, just like the rest I can only speculate (given the available information): I guess the following pieces provide a hint what's really going on here:
- "The quine is the quine" (one of the sub-headline of the article) and the meaning of the word "quine".
- Author's "scaffolding" tool which, once finished, had acquired the "knowledge"[1] how to add a CLAUDE.md baked instructions for a particular homemade framework (he's working on).
- Anthropic saying something like: no, stop; you cannot "copy"[1] Claude knowledge no matter how "non-serious" your scaffolding tool or your use-case is: as it might "shows", other Claude users, that there's a way to do similar things, maybe that time, for more "serious" tools.
---
[1]. Excerpt from the Author's blog post: "I would love to see the face of that AI (Claude AI system backend) when it saw its own 'system prompt' language being echoed back to it (from Author's scaffolding tool: assuming it's complete and fully-functional at that time)."
I haven't used any LLM deep research tools in the past; today, after reading this HN post, I gave Tongyi DeepResearch a try to see how it performs on a simple "research" task (in an area I've working experience in: healthcare and EHR) and I'm satisfied with its response (for the given tasks; I, obviously, can't say anything how it'll performs on other "research" tasks I'll ask it in the future). I think I'll keep using this model for tasks for which I was using other local LLM models before.
Besides I might give other large deep research models a try when needed.
I've known and worked with James Gosling for years before Java (Live Oak), on his earlier projects, Emacs at UniPress and NeWS at Sun, and fought along side him against Sun management trying to make NeWS free in 1990 (and I left Sun because they broke the promises they made us and spilled a lot of blood), so I didn't need to learn about Java's history from Wikipedia.
James's email that convinced me to go work with him at Sun on NeWS in 1990:
Here's a Stanford talk James Gosling gave about Java that I attended in 1995, where he talks about C++, his original tape copy program that turned into a satellite ground control system, how he holds the world record for writing the largest number of cheesy little extension languages to go, and his implementation of Emacs sold by UniPress (which RMS calls "Evil Software Hoarder Emacs"), and his design and implementation of NeWS (formerly SunDew), a PostScript based network extensible window system.
James Gosling - Sun Microsystems - Bringing Behavior to the Internet - 1995-12-1:
>Video of James Gosling's historic talk about Java, "Bringing Behavior to the Internet", presented to Terry Winograd's user interface class at Stanford University, December 1, 1995.
In that talk I asked him a couple questions about security and the "optical illusion attack" that he hedged on (44:53, 1:00:35). (The optical illusion attack is when the attacker simply draws a picture of a "secure" pop up dialog from your bank asking for your password.)
He mentioned off hand how a lot of the command and control systems for Operation Desert Storm was written in PostScript. That was his NeWS dialect of PostScript, and was written primarily by Josh Siegel at LANL called "LGATE", who later came to work at Sun in 1990 and rewrote the NeWS PostScript interpreter himself, then went on to write an X11 window manager in PostScript, again proving James's point that people always did a lot more with his cheesy little extension languages than he ever expected (which also held true with Java).
Josh's work on simulating Desert Storm and WWIII with NeWS at LANL:
I also saw Bill Joy's much earlier talk at the 1986 Sun Users Group in Washington DC, where he announced a hypothetical language he wanted to build called "C++++-=", and that he talked about in subsequent presentations.
I think that was the same talk when Bill said "You can't prove anything about a program written in C or FORTRAN. It's really just Peek and Poke with some syntactic sugar". More Bill Joy quotes:
James eventually realized that concept as Java, showing that the kernel inspiration of writing a "fuck you to C++" language existed long before James invented "Live Oak", even soon after C++ was invented. But "Java" was a much better name than "Live Oak" or "C++++-=" fortunately -- thanks to Kim Polese -- though not as succinct and musically inspired as "C#"!
>The peak computer speed doubles each year and thus is given by a simple function of time. Specifically, S = 2^(Year-1984), in which S is the peak computer speed attained during each year, expressed in MIPS. -Wikipedia, Joy’s law (computing)
>“C++++-= is the new language that is a little more than C++ and a lot less.”
-Bill Joy
>In this talk from 1991, Bill Joy predicts a new hypothetical language that he calls “C++++-=”, which adds some things to C++, and takes away some other things.
>“Java is C++ without the guns, knives, and clubs.” -James Gosling
From technology point-of-view, Humanity have progressed a lot; however, psychologically speaking, we're still almost the same as we're then (in 1960s).
> OpenAI says it has evidence DeepSeek used its model to train competitor.
> The San Francisco-based ChatGPT maker told the Financial Times it had seen some evidence of “distillation”, which it suspects to be from DeepSeek.
> ...
> OpenAI declined to comment further or provide details of its evidence. Its terms of service state users cannot “copy” any of its services or “use output to develop models that compete with OpenAI”.
OAI share the evidence with the public; or, accept the possibility that your case is not as strong as you're claiming here.
Also, there are so many innovations in their papers (Deepseek math, Deepseek v2/v3, R1) that I honestly wouldn’t even care. They figured out a way to train on only 2048 H800s when big companies are buying them in the hundreds of thousands. They created a new RL algorithm. They improved MoE. They improved the KV cache. They built an super efficient training framework.
With lms, LM Studio's frontend GUI/desktop application and its backend LLM API server (for OpenAI compatibility API endpoints) are tightly coupled: stopping LM Studio's GUI/desktop application will trigger stopping of LM Studio's backend LLM API server.
With llmsterm, they've been decoupled now; it (llmsterm) enables one, as LM Studio announcement says, to "deploy on servers, deploy in CI, deploy anywhere" (where having a GUI/desktop application doesn't make sense).
reply