John Y 🔸 @yanjo115

Gutenberg. ex-anthropic, meta 🔸 10% Pledge with @givingwhatwecan gutenberg.ai SF Joined August 2015

Tweets

80
Followers

406
Following

2K
Likes

16K

John Y 🔸 @yanjo115

2 weeks ago

@fjzzq2002 @jiaxinwen22 Noice

0 0 1 295 0

View Details

John Y 🔸 @yanjo115

3 weeks ago

@LeoMcKeeReid @NeolithicAI Congrats Leo!!

0 0 3 189 0

View Details

Ah I see now. So there's hundreds of thousands of features, and we grabbed the most interesting ones we found after a cursory scan. None of these besides "delve", "it's not X, it's Y", and maybe some flowery vocab choice seems to jump out as good candidates. We may dig for more features later if we revisit this experiment

0 0 0 14 0

View Details

John Y 🔸 @yanjo115

2 months ago

What features constitute AI-generated text? We used our data-centric interp SDK on @pangram's public dataset and found some interesting features

6 7 65 3K 15

View Details

John Y 🔸 @yanjo115

2 months ago

@thkostolansky 🫡

1 0 2 56 0

View Details

John Y 🔸 @yanjo115

2 months ago

@NelsonEdmundb Features are diverse enough that the easiest soln is to have a judge review them. Maybe they could route the easily regex-able ones

1 0 0 56 0

View Details

John Y 🔸 @yanjo115

2 months ago

@QiaochuYuan @max_spero_ @pangram it'd be a quick and fun experiment to run! we're in talks

0 0 4 37 0

View Details

John Y 🔸 @yanjo115

2 months ago

This was a quick experiment that merits more investigation. We ran this with our product, which is now in public beta! Check out our quickstart guide here console.gutenberg.ai/d/docs

1 0 17 611 2

View Details

John Y 🔸 @yanjo115

2 months ago

We also had features that captured over-formatting, florid metaphorical language, and more

2 0 14 967 0

View Details

John Y 🔸 @yanjo115

2 months ago

@akrolsmir So cool!

0 0 2 93 0

View Details

John Y 🔸 @yanjo115

3 months ago

@MavorParker awesome work!

0 0 2 189 0

View Details

John Y 🔸 @yanjo115

5 months ago

@tomjiralerspong @TrentonBricken Very cool!

0 0 2 259 0

View Details

John Y 🔸 @yanjo115

5 months ago

@WorkshopLabs Very cool!

0 0 7 445 0

View Details

John Y 🔸 @yanjo115

5 months ago

@thesophiaxu aha yea maybe it's like how early waymo needed a human overseer. i have no idea what the laws are around this but it still seems like a good idea for gradual deployment

0 0 0 23 0

View Details

John Y 🔸 @yanjo115

6 months ago

@LLMPsychologist @browser_use Super cool!

0 0 4 903 0

View Details

John Y 🔸 @yanjo115

7 months ago

Thanks to @goodstartlabs and @alxai_ the dataset. And thanks to @nicetomeetyu2 , @yuqi_s_29 , and Matt Olson for their great work collaborating on this! arxiv.org/abs/2602.05183

0 2 9 705 3

View Details

John Y 🔸 @yanjo115

7 months ago

We also compared our successful training run to an unsuccessful one. If you just look at the reward, you find they diverge around step 9. However, when we compare the features between them, there’s this prominent feature that corresponds to “difficulty dealing with a minor bug in the environment”, which the goodrun overcame, and the badrun did not. Looking at this signal, we can see the runs diverging at step 6, giving earlier signal, and an explanation, as to why one run succeeded and the other did not.

2 1 6 722 0

View Details

John Y 🔸 @yanjo115

7 months ago

New paper! We used Sparse Autoencoder (SAE) embeddings to understand how agent behavior actually changes during post training. We analyzed thousands of rollouts in a Diplomacy environment and uncovered generalized reward hacking, weird roleplay, and the root cause of an unsuccessful training run. All of which our LLM baseline couldn't find.