I've just been outside shoveling. This blizzard nor'easter is going to be pretty bad: heavy, sticky, a ton of the stuff, and windy. It's a bad combination.
1/5: Claude's C Compiler #ccc is a C compiler written entirely by Claude Code. My analysis of the parser: (1) There is no statement of what dialect of C it is supposed to support. (2) The initial commit is a GCC clone of https://t.co/9MJydM09EB.
4/5: I have been using Claude Code for a while now in practice, and I am astounded by what it can do. But more than ever, it is clear that developers need to know how to specify a problem from requirements, then guide Claude Code to explore the problem space, not produce garbage.
@clattner_llvm@AnthropicAI There's no EBNF, no statement as to the C dialect (it's a little of C89 + C99 + C11 + C23 + GCC ext). Follows the tradition of "experts" nowadays: hack before specifying anything. Initial commit was a clone of https://t.co/hdzj5mWBMV from GCC, then worked on an operational def.
Astounding: Claude Code can write all ports for an Antlr4 grammar and implement disambiguating predicates with symbol table, usually in a few hours. Normally, it would take several days if not weeks if done by hand. Quality is superior. Dawn of a new era in software development.
Did they have it "read" the paper? Brabrand, Claus, Robert Giegerich, and Anders Møller. "Analyzing ambiguity of context-free grammars." Science of Computer Programming 75.3 (2010): 176-191.
ChatGPT 5 is BS'ing. Asked to define vertical and horizontal ambiguity in a formal grammar, it replied this: https://t.co/kjxezDLmM9. Tested the horizontal ambig example with Antlr (`grammar a; s : list EOF; list: item | list ',' item ; item: 'a';`), but there's no ambiguity.
Repeat of ChatGPT-5 experiment: "What are the unused parser rules of grammars-v4/acme/acme.g4 and grammars-v4/sql/mysql/Oracle/MySQLParser.g4?" It replied: "No unused parser rules were found." This is not correct (acme.g4: 4; MySQLParser.g4: 8).
Yep. Chat's gotten worse. "Hi Chat. I asked you this before under GPT 4, let's see what you do under GPT 5. Please read the Antlr4 grammar at https://t.co/idU3rY3C5k. Please tell me all the unused parser rule symbols. How many are there?". 3 m later: "Can't read URL".
AI still has a way to go before it understands programming language grammars. "Please read the Antlr4 grammar at https://t.co/idU3rY3C5k. Please tell me all the unused parser rule symbols. How many are there?" cursor_body and cursor_spec do not even exist! https://t.co/4HDBLam5Mp
I see people using genAI to write ANTLR grammars. NB: LLMs do not have logical reasoning. It can produce crap. E.g., "Tell me which symbols are nullable for "grammar x2; start: ; z : 'asdf'; a : start | z; b : ('a' | 'b')* d+; d : e f; e : ; f : ( | 'a')+;" -> "start, a, e"