Hacker News Reader

jensb1on One month coding with GLM 5.3 Flash

Yeah agree, misunderstood the article intent as well.

jensb1on How GLM built its own inference infrastructure

What is "illegal" about it?

jensb1on Ask HN: Is it just me or is Claude Code getting worse?

Not more stupid IMO, but significnatly less token efficient or they have decreased tokens for subscription users, difficult to know.

jensb1on Toy Lisp compiler using LLVM done with AI in 8hrs

Hi, always wanted to create a toy lisp compiler in LLVM for some reason, but never had the time to get it working.

Took a new stab at it with Claude 3.7. Got a working version up in less than 8hrs in total with passing tests! Never wrote a single line of code. I would estimate the number of cursor iterations to about 500.

Process was: 1. Ask Claude to create a boiler plate JIT in LLVM (see below shared chat) 2. Asked Claude on the architecture and what classes should be created (did this in cursor): - lexer - parser - AST - Code gen 3. Implemented each part separately, lexer, parser and AST took 1hr, codegen took ~7hrs.

It can define variables, functions and do some basic stuff: try: (defn factorial [n] (if (= n 0) 1 (* n (factorial (- n 1))))) (factorial 5)

Hope you like it!

Here are some chats if you are interested in the first part of the coding (the rest was done in cursor):

LLVM - https://claude.ai/share/3c5a1f92-389f-4274-8348-dcceaee8f140

Lexer - https://claude.ai/share/fa32b8bd-0b8e-4faa-bc09-fc3dd9bb5ad3

jensb1on Exllamav2: Inference library for running LLMs locally on consumer-class GPUs

typo

jensb1on Exllamav2: Inference library for running LLMs locally on consumer-class GPUs

Could someone explain how token generation speed relates to latency for the first token to be outputted?

And if anyone have any metrics on latency on a 4090 for the 70B model, that would be very helpful.