正在加载视频...
视频加载失败
this is so fun Below: an LLM loop is hooked into the Lisp REPL, and the LLM adds functions to the REPL as needed. These functions persist across application restarts. Instead of the entire read_file + edit_file + bash compile loop, it's now just prompt functionality in the application... show more
14,975 次观看 • 1 个月前 •via X (Twitter)
29 条评论
smalltalk had the persistent live image in the 70s, the twist is the llm writing its own functions into it.
lisp people waited 60 years for this
Image-based Lisp all over again, with an LLM as the hacker at the REPL. The catch Smalltalk folks hit decades ago: once functions persist in the image, the source of truth isn't a file anymore. I'd want every LLM-added defun auto-dumped to git so you can diff what it grew.
didn't know lisp has a repl Actually I just realized I don't know anything about lisp
to learn more
for code and to try it yourself
I've done something very similar with fsharp in
I thought my Lisp days were over now that I’m no longer authoring the code, but it might just be the other way around
love seeing the devenv update prompt there :D devenv update devenv ;) if you wanna go to latest
I'm about to fuck around with this for no less than 16 hours today
as you should! it’s addictive. essentially you can delete majority modern dev tooling companies with this one simple pattern….
I am going to replace your lisp implementation with LFE (lisp flavored erlang) and see if I can make it replicate
do it! i’m heavily using erl at work rn btw :)
linking in case you haven't come across it
catchup with @mikehostetler btw
@arthurcolle 👋
@GeoffreyHuntley 👋 hold on I gotta unfollow someone else to follow you, one moment
Hook this to USD, theatre set, generative Final Cut Pro, Logic Pro, and have AI validate details from the Planet of the Apes movie series. Start at horse riding analysis for learning to balance.
I did this with a vibe coded clj interpreter in march (bad idea too many bugs) and was planning to have the interpreter break on error (instead of failing) with edit tool to change and resume. Have you tried something like that?
A tiny rehearsal room for each new function would be neat 😄 Can the loop test a replacement in isolation before swapping it into the running app? Especially curious what happens when a failed experiment has already mutated live state.
LISP is soooo slow. Why is LISP so much slower than C, even though they’re roughly the same age?
stealing this for next weeks reset (already out of claude and codex and shortly muse :( )
and you agent can decide to put something like (fad:delete-directory-and-files "/"), he-he
fun fun
bash bash
do you have a repo anywhere for this
probably easy to make some kind of perfect memory data structure too
How delightful
functions persisting across restarts is the bit that matters. I landed on the same idea from the other side: the agent logs what it built to a sessions.md after each step and the next agent reads it first. how do you stop the REPL from piling up dead functions?
![this harness stack is f*cking gold TypeSafe AI + LangChain finally dropped the layer that makes agents affordable at company scale: jev running inside the langchain agent loop as middleware the problem every big team hits: each fork in the loop is another full LLM call. at 10 agents that's a bill. at 10,000 it's the budget > routing before: every task goes to the frontier model now: jev picks the least costly model that can finish it. lookups go fast, architecture goes powerful > guardrails before: the "is this command dangerous?" check sat in the closed parts of claude code, codex, cursor now: AutoModeMiddleware(tools=["bash"]). one line, any agent, blocked before the tool runs > decisions before: one LLM call per question now: every question about the same state in one request, answered in parallel > audit before: "the model decided" now: the answer, the probability and the confidence stay in agent state > adoption no rewrite. pip install langchain-typesafe and drop it into the agent you already run TypeSafe reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification. the LLM still reasons and writes. jev takes the thousands of small calls in between already live: browserbase agents for fractions of a cent, a live trading agent, email triage at scale full loop mapped below: 7 agents, 4 jev gates per step count the forks in your agent loop tonigh](https://image.24vids.com/tw-2104876056893104461/amplify_video_thumb/2104866254146203648/img/J8LnvuFvwbGYaJ7N.jpg)
