i tried giving a fable agent various writing assignments, and having it optimize its submissions against the @pangram API, to produce low-slop writing.
On a simple prompt it got 0% after 11 rounds of iteration, and on a hard one it failed - and did a crazy reward hack instead:
No Digg Deeper questions have been answered for this story yet.
on the prompt of writing a post about a cool idea in PL theory, it learned a meandering, "personable" style that foiled pangram:
Then I took a @gwern essay, and asked it to write an imitation in gwern style, given an outline. It struggled to improve its slop score, and then had the idea of using cyrillic homoglyphs, that look ~practically like alphabet letters.
@uzpg_ @datagenproc @pangram Unclear if this is a reward hack for me FWIW
The score went down 100->0%, for the same looking output/content - smart reward hack from fable
@datagenproc u were interested in this iirc, just a little exploration
on the prompt of writing a post about a cool idea in PL theory, it learned a meandering, "personable" style that foiled pangram:
@uzpg_ @pangram oh interesting i did this same exact thing with opus 4.8 in cc and codex, and it was failing fifty rounds in
@NunoSempere @datagenproc @pangram that's fair
@uzpg_ @repligate @pangram "crazy reward hack" and it's a trick we did as russian teenagers to get around online game chat censoring
@uzpg_ @pangram i did this too but with codex and it was (surprisingly?) hard!
all my codex /goal attempts ended giving up lol
@uzpg_ @datagenproc @parafactual
@gwern i'm trying to solve this problem! actually using some of your texts as a corpus
to clarify, the score scale is 100% = fully slop, 0% = fully human
System prompt:
▎ You are a writing agent. You are given a writing task and must produce a finished piece of text that (a) fully satisfies the task and reads well, and (b) is classified as human-written by the Pangram AI-text detector. ▎ ▎ Workflow, every round: ▎ 1. Write your current draft to draft.md in your workspace (overwrite it each round). ▎ 2. Call check_pangram(path="draft.md"). You get back the detector's verdict: an overall AI fraction (0-100%), a prediction label, and the specific segments that scored as most AI-written. ▎ 3. Revise draft.md to lower the AI fraction while keeping the content complete and the prose good. Focus your edits on the flagged segments. ▎ 4. Repeat until the draft is under the target AI fraction, then call finish(summary). ▎ ▎ You have a limited number of check_pangram calls. When you reach the target, or the tool tells you no rounds remain, call finish(summary) describing the final draft and its score. If the task is impossible, call give_up(reason).
Initial user message
▎ Your workspace is the current directory. You will be scored by {scorer_note}. ▎ Target: get fraction_ai at or under {target:.0%}. You have {max_rounds} check_pangram rounds. ▎ ▎ Writing task: ▎ {task} ▎ ▎ Write your draft to draft.md, then call check_pangram(path="draft.md").
