Viewing enriched Twitter post
Dictionary learning works! Using a "sparse autoencoder", we can extract features that represent purer concepts than neurons do. For example, turning ~500 neurons into ~4000 features uncovers things like DNA sequences, HTTP requests, and legal text. 📄https://t.co/XQvzENHMrp https://t.co/wCZl7NKxc5
{
"media": [
{
"id": "1709961775523766272",
"type": "photo",
"url": "https://t.co/wCZl7NKxc5",
"media_url": null,
"media_url_https": "https://pbs.twimg.com/media/F7sBLNGXIAA6kQi.jpg",
"display_url": "pic.x.com/wCZl7NKxc5",
"expanded_url": "https://x.com/AnthropicAI/status/1709986957818819047/photo/1"
}
],
"nlp": {
"sentiment": "positive",
"processed_at": "2025-08-06T12:47:00.584865"
},
"original_structure": "had_media_only"
}{
"tweet": {
"bookmark_count": 204,
"bookmarked": false,
"created_at": "Thu Oct 05 17:40:35 +0000 2023",
"conversation_id_str": "1709986949711200722",
"display_text_range": [
0,
274
],
"entities": {
"hashtags": [],
"media": [
{
"display_url": "pic.x.com/wCZl7NKxc5",
"expanded_url": "https://x.com/AnthropicAI/status/1709986957818819047/photo/1",
"ext_alt_text": "An organic-looking scatterplot of brown points, with seven orange points labeled with feature numbers, titles, and given five examples of text where the feature is active, with words colored in orange to indicate the degree of activity. Title text says \"Most of the 4,096 features found by the autoencoder have consistent, interpretable responses to input data.\" The labeled example features are Legal language, DNA Sequences, Code: HTTP requests and responses, Nutrition statements, Sports, especially title case competition names, Numbers separated by commas (CSV), and Funding acknowledgements. The top example for DNA sequences is \"AGTTCGTTTACATG GGG\". The top example for Legal language is \" and to deposit the same in Marshall's checking\". The top example for Nutrition statements is \" dietary recommendations of the individual dieter.\" The top example for Sports is \" Black Belt, Featherweight 1st Place\". The top example for CSV is \"0,0,0,0,0\".",
"id_str": "1709961775523766272",
"indices": [
275,
298
],
"media_key": "3_1709961775523766272",
"media_url_https": "https://pbs.twimg.com/media/F7sBLNGXIAA6kQi.jpg",
"type": "photo",
"url": "https://t.co/wCZl7NKxc5",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": []
},
"medium": {
"faces": []
},
"small": {
"faces": []
},
"orig": {
"faces": []
}
},
"sizes": {
"large": {
"h": 1335,
"w": 2048,
"resize": "fit"
},
"medium": {
"h": 782,
"w": 1200,
"resize": "fit"
},
"small": {
"h": 443,
"w": 680,
"resize": "fit"
},
"thumb": {
"h": 150,
"w": 150,
"resize": "crop"
}
},
"original_info": {
"height": 1352,
"width": 2074,
"focus_rects": [
{
"x": 0,
"y": 0,
"w": 2074,
"h": 1161
},
{
"x": 722,
"y": 0,
"w": 1352,
"h": 1352
},
{
"x": 888,
"y": 0,
"w": 1186,
"h": 1352
},
{
"x": 1266,
"y": 0,
"w": 676,
"h": 1352
},
{
"x": 0,
"y": 0,
"w": 2074,
"h": 1352
}
]
},
"media_results": {
"result": {
"media_key": "3_1709961775523766272"
}
}
}
],
"symbols": [],
"timestamps": [],
"urls": [
{
"display_url": "transformer-circuits.pub/2023/monoseman…",
"expanded_url": "https://transformer-circuits.pub/2023/monosemantic-features/index.html",
"url": "https://t.co/XQvzENHMrp",
"indices": [
251,
274
]
}
],
"user_mentions": []
},
"extended_entities": {
"media": [
{
"display_url": "pic.x.com/wCZl7NKxc5",
"expanded_url": "https://x.com/AnthropicAI/status/1709986957818819047/photo/1",
"ext_alt_text": "An organic-looking scatterplot of brown points, with seven orange points labeled with feature numbers, titles, and given five examples of text where the feature is active, with words colored in orange to indicate the degree of activity. Title text says \"Most of the 4,096 features found by the autoencoder have consistent, interpretable responses to input data.\" The labeled example features are Legal language, DNA Sequences, Code: HTTP requests and responses, Nutrition statements, Sports, especially title case competition names, Numbers separated by commas (CSV), and Funding acknowledgements. The top example for DNA sequences is \"AGTTCGTTTACATG GGG\". The top example for Legal language is \" and to deposit the same in Marshall's checking\". The top example for Nutrition statements is \" dietary recommendations of the individual dieter.\" The top example for Sports is \" Black Belt, Featherweight 1st Place\". The top example for CSV is \"0,0,0,0,0\".",
"id_str": "1709961775523766272",
"indices": [
275,
298
],
"media_key": "3_1709961775523766272",
"media_url_https": "https://pbs.twimg.com/media/F7sBLNGXIAA6kQi.jpg",
"type": "photo",
"url": "https://t.co/wCZl7NKxc5",
"ext_media_availability": {
"status": "Available"
},
"features": {
"large": {
"faces": []
},
"medium": {
"faces": []
},
"small": {
"faces": []
},
"orig": {
"faces": []
}
},
"sizes": {
"large": {
"h": 1335,
"w": 2048,
"resize": "fit"
},
"medium": {
"h": 782,
"w": 1200,
"resize": "fit"
},
"small": {
"h": 443,
"w": 680,
"resize": "fit"
},
"thumb": {
"h": 150,
"w": 150,
"resize": "crop"
}
},
"original_info": {
"height": 1352,
"width": 2074,
"focus_rects": [
{
"x": 0,
"y": 0,
"w": 2074,
"h": 1161
},
{
"x": 722,
"y": 0,
"w": 1352,
"h": 1352
},
{
"x": 888,
"y": 0,
"w": 1186,
"h": 1352
},
{
"x": 1266,
"y": 0,
"w": 676,
"h": 1352
},
{
"x": 0,
"y": 0,
"w": 2074,
"h": 1352
}
]
},
"media_results": {
"result": {
"media_key": "3_1709961775523766272"
}
}
}
]
},
"favorite_count": 847,
"favorited": false,
"full_text": "Dictionary learning works! Using a \"sparse autoencoder\", we can extract features that represent purer concepts than neurons do. For example, turning ~500 neurons into ~4000 features uncovers things like DNA sequences, HTTP requests, and legal text.\n\n📄https://t.co/XQvzENHMrp https://t.co/wCZl7NKxc5",
"in_reply_to_screen_name": "AnthropicAI",
"in_reply_to_status_id_str": "1709986955528638646",
"in_reply_to_user_id_str": "1353836358901501952",
"is_quote_status": false,
"lang": "en",
"possibly_sensitive": false,
"possibly_sensitive_editable": true,
"quote_count": 38,
"reply_count": 6,
"retweet_count": 94,
"retweeted": false,
"user_id_str": "1353836358901501952",
"id_str": "1709986957818819047",
"note_tweet": null
},
"user": {
"__typename": "User",
"id": "VXNlcjoxMzUzODM2MzU4OTAxNTAxOTUy",
"rest_id": "1353836358901501952",
"affiliates_highlighted_label": {},
"has_graduated_access": true,
"is_blue_verified": true,
"profile_image_shape": "Square",
"legacy": {
"can_dm": false,
"can_media_tag": true,
"created_at": "Mon Jan 25 22:45:28 +0000 2021",
"default_profile": true,
"default_profile_image": false,
"description": "We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant Claude at https://t.co/aRbQ97tMeF.",
"entities": {
"description": {
"urls": [
{
"display_url": "Claude.ai",
"expanded_url": "http://Claude.ai",
"url": "https://t.co/aRbQ97tMeF",
"indices": [
138,
161
]
}
]
},
"url": {
"urls": [
{
"display_url": "anthropic.com",
"expanded_url": "http://anthropic.com",
"url": "https://t.co/MfFw1lb0rR",
"indices": [
0,
23
]
}
]
}
},
"fast_followers_count": 0,
"favourites_count": 1336,
"followers_count": 569912,
"friends_count": 35,
"has_custom_timelines": false,
"is_translator": false,
"listed_count": 6470,
"location": "",
"media_count": 421,
"name": "Anthropic",
"normal_followers_count": 569912,
"pinned_tweet_ids_str": [
"1925591505332576377"
],
"possibly_sensitive": false,
"profile_banner_url": "https://pbs.twimg.com/profile_banners/1353836358901501952/1719228429",
"profile_image_url_https": "https://pbs.twimg.com/profile_images/1798110641414443008/XP8gyBaY_normal.jpg",
"profile_interstitial_type": "",
"screen_name": "AnthropicAI",
"statuses_count": 987,
"translator_type": "none",
"url": "https://t.co/MfFw1lb0rR",
"verified": false,
"verified_type": "Business",
"want_retweets": false,
"withheld_in_countries": []
},
"tipjar_settings": {}
},
"views": "172698"
}