Post #1753536

2026-04-19 05:16 UTC

If you have heard the buzzword "agentic AI" but avoided finding out what it meant until now: 1. Someone figured out an LLM can do JSON RPCs by typing out the JSON token by token. 2. The LLM is run in a harness that regexes out the JSON from its output and executes the RPC. 3. The response is catted into the LLM's context window, also in the form of JSON that the LLM just reads. 4. People connect these harnesses to system shells on their dev machines. 5. Fast forward, this is a trillion-dollar industry held together by markdown files asking the LLM to please not curlbash from the internet.

Replies (7)

@astraleureka@social.treehouse.systems 2026-04-19 06:03

@wren6991 this absurdity is on-par with the "claude re-emits vaguely json-shaped output repeatedly until the linter says it's valid" discovery from the recent source leak

Open ##1753537
@lritter@mastodon.gamedev.place 2026-04-19 07:21

@wren6991 yup that is precisely how it works. been examining this with gemma 4 the past week. i put the control target in a docker image where it has root; it's useful for user testing, but i feel silly trying to make it do anything else.

Open ##1753542
@hp@mastodon.tmm.cx 2026-04-19 09:37

@wren6991 the whole idea of please asking a piece of software to not perform its primary purpose is pretty wild. They consume tokens. You get billed for the tokens. What?!

Open ##1753545
@dickon@splodge.fluff.org 2026-04-19 10:44

@wren6991 6. It does so anyway. The whole mess is unsafe, broken in myriad, unfixable ways, and utterly disheartening. I hate the whole thing. We started to avoid blindly executing untrusted data as if it was code, shortly after the first buffer overflow was weaponised. This LLM nonsense entirely relies upon it, hence 'prompt injection' 'attacks', which aren't attacks, but a fundamental way the damned thing works in the first place, so unsafe by design. I just wish this craze would die.

Open ##1753546
@fabio@cosocial.ca 2026-04-19 19:21

@wren6991 I don’t know if you looked at the Claude leak but the way it builds a JSON response is: Make output -> does it look like JSON? -> burn tokens until JSON is parseable Peak engineering right here

Open ##1753547
@jafo@inuh.net 2026-04-19 20:27

@wren6991 Just to clarify: I get where you're going, but that is not, to my understanding (informed by ~3 years of heavy AI use) "agentic AI". What you are talking about is more called "tool use". Tool use is often used as part of agentic workflows, but are not what makes agentic agentic. You can do agentic workflows without tool use.

Open ##1753548
@Byte@rage.love 2026-04-19 23:21

@wren6991 infinite recursion is probably like, the least bad thing you could make an “AI” “agent” do

Open ##1753549