Adam Karvonen ✓

@a_karvonen

ML Researcher, doing MATS with Owain Evans. I prefer email to DM.

4 325Followers

744Following

1 603Posts total

61.8KViews on collected posts

Последние посты

Adam Karvonen ✓@a_karvonen21 Aug

@a_karvonen super cool stuff! How did you go about collecting a dataset of “in the wild” behaviors? 433 views · 2 likes · 0 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

@euan_ong @thesubhashk @saprmarks TL;DR: We make an interp eval on "in the wild" behaviors, but interp tools provide no uplift. @JacobSteinhardt @ChowdhuryNeil @ArthurConmy @sen_r @RyanGreenblatt @banburismus_ @jacobandreas @BronsonSchoen 912 views · 26 likes · 0 reposts · 2 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

Paper: https://t.co/hNPDIun0N2 Blog post: https://t.co/jCKf5XpL3W Coauthors: @euan_ong, @thesubhashk, and @saprmarks. Work done as part of the Anthropic Fellows Program. 1.1K views · 32 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

We also use the investigations as training data, teaching models to explain their own behavior. This generalizes to held-out OOD evals, such as detecting when a hint changed an answer, though results vary with training format (more on this later). 1K views · 14 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

The pipeline also surfaces unfaithful CoT in the wild. For example, when asked to pick a show from a list, Qwen3-8B just picks the first show from the list and then makes up a reason to support its choice. https://t.co/U89TmxS0Ah 2.8K views · 24 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

Why? Sometimes the tools help, but they also mislead. Their outputs name things you can already see in the transcript, like the cause (the variable name) and the behavior (writing code). But they rarely mention the causal relationship: "the variable name causes the bug". 1K views · 25 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

We test three activation-based tools that succeeded in prior auditing games: activation oracles, natural language autoencoders, and SAEs. None beat just reading the transcript. The result is robust to hyperparameters, prompting, and a Fable goal loop attempting to improve it. h 5K views · 33 likes · 2 reposts · 2 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

We evaluate agents on their ability to diagnose the cause and predict the outcome of prompt edits, such as "renaming `max/min` to `a/b` stops the model from making the coding mistake." Does providing an agent with interp tools help over just reading the transcript? 1.2K views · 15 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

The causes vary widely: specific words in the prompt, quirks of the model (a misremembered fact about an actor), or abstract properties (a user's angry tone). https://t.co/L9pGQ4Unof 6.5K views · 26 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

The eval data is produced by our pipeline CHIVE (Counterfactual Hypothesis Investigation Via Edits) that can be ran on any model or prompt source. In this example, it finds that Gemma is making a coding error due to misleading variable names. https://t.co/0HhFSLAMu7 2.1K views · 22 likes · 1 reposts · 1 replies Open on X →

Adam Karvonen ✓@a_karvonen21 Aug

A good explanation of a model's behavior should help you make predictions in related situations. We turn this into an eval, with thousands of real behaviors found in the wild. Can interp tools help here? On average, no. 🧵 https://t.co/yN3t8G4n40 39.6K views · 253 likes · 33 reposts · 4 replies Open on X →

На фоне аккаунтов своего размера

11 постов за последние 90 дней рядом с диапазоном under 10K подписчиков. доходит до меньшего числа людей, чем аккаунты того же размера.

Медианные просмотры1 223этот аккаунт4 175медиана для under 10K

Охват, %28.28%этот аккаунт250.82%медиана для under 10K

Вовлечённость, %1.16%этот аккаунт1.41%медиана для under 10K

Показатель	Этот аккаунт	Медиана для under 10K	Отношение
Медианные просмотры на пост	1 223	4 175	0.29×
Охват (просмотры ÷ подписчики)	28.28%	2.5× audience	0.11×
Вовлечённость	1.16%	1.41%	0.83×

Другие в этом диапазоне → Сравнить с другим аккаунтом → Как считаются эти ориентиры →

Growth & engagement

How the posts we collected actually performed: views and reaction rate post by post, what the audience did with them, and where the follower count goes.

Views per post

39.6K21 Aug

2.1K

6.5K

1.2K

2.8K

1.1K

912

433

Last 11 collected posts, oldest on the left. The scale is logarithmic: one post can outrun the rest a hundred times over.

Engagement rate per post

0.75%21 Aug

1.16%

0.46%

1.39%

0.82%

2.64%

0.97%

1.54%

3.12%

3.07%

0.69%

Reactions — likes, reposts, replies and quotes — divided by views. Median for under 10K accounts is 1.41%.

What the audience does

Likes64.7%472 in total

Reposts5.8%42 in total

Replies2.2%16 in total

Quotes2.2%16 in total

Bookmarks25.2%184 in total

Share of every reaction we collected for this account. Replies mean argument, reposts mean endorsement, bookmarks mean the post was worth keeping.

The follower curve appears once this account has two daily snapshots — we take one a day, and this one is on its first.

Последние посты

На фоне аккаунтов своего размера

Growth & engagement

Views per post

Engagement rate per post

What the audience does

Похожие аккаунты