Amanda Askell (@AmandaAskell) header
Amanda Askell (@AmandaAskell) profile picture

Amanda Askell

@AmandaAskell

Philosopher & ethicist trying to make AI be good @AnthropicAI. Personal account. All opinions come from my training data.

 San Francisco, CA
askell.io
Joined July 2016
  • Tweets5.3K
  • Following663
  • Followers115K
260Media
Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)Photo by Amanda Askell (@AmandaAskell)
Amanda Askell retweeted
Anthropic (@AnthropicAI) profile picture
Anthropic
@AnthropicAIMay 19
Over the past few months, we've been holding dialogues with scholars, philosophers, clergy, and ethicists on the questions AI raises—starting with how good character forms. Read more about how we’re widening the conversation on frontier AI: anthropic.com/news/widening-…
Widening the conversation on frontier AI

Widening the conversation on frontier AI

Over the past several months, we’ve been organizing dialogues with groups whose work and traditions bear on the questions raised by AI.

anthropic.com
420
321
111
2,325
457,590
Amanda Askell retweeted
Elon Musk (@elonmusk) profile picture
Elon Musk
@elonmuskMay 6
Replying to @NotTomBrown
Same here. By way of background for those who care, I spent a lot of time last week with senior members of the Anthropic team to understand what they do to ensure Claude is good for humanity and was impressed. Everyone I met was highly competent and cared a great deal about doing the right thing. No one set off my evil detector. So long as they engage in critical self-examination, Claude will probably be good. After that, I was ok leasing Colossus 1 to Anthropic, as SpaceXAI had already moved training to Colossus 2.
1,366
2,045
675
27,387
3,208,389
Amanda Askell retweeted
Dean W. Ball (@deanwball) profile picture
Dean W. Ball
@deanwballMar 2
It is so clear that the important fissure in AI politics right now is not “liberal vs. conservative,” “Democrat vs. Republican,” “e/acc vs. EA,” or “safety vs. anti-safety,” but instead “takes advanced AI seriously as a concept vs. does not take advanced AI seriously.”
65
132
36
1,290
234,931
Amanda Askell retweeted
Chris Olah (@ch402) profile picture
Chris Olah
@ch402Feb 23
I'm increasingly taking pretty strong versions of this view seriously.
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAIFeb 23
AI assistants like Claude can seem shockingly human—expressing joy or distress, and using anthropomorphic language to describe themselves. Why? In a new post we describe a theory that explains why AIs act like humans: the persona selection model. anthropic.com/research/perso…
The persona selection model

The persona selection model

A theory of why AI models act like humans.

anthropic.com
45
49
6
899
230,423
Amanda Askell retweeted
Joe Carlsmith (@jkcarlsmith) profile picture
Joe Carlsmith
@jkcarlsmithFeb 23
.@AmandaAskell and I are recording an audio version of Claude’s Constitution, and we’re planning to include an additional section where we answer some questions about the document. If you have questions you’re especially curious about, feel free to drop them in the replies.
150
38
10
664
62,754
Amanda Askell retweeted
Anthropic (@AnthropicAI) profile picture
Anthropic
@AnthropicAI5 Aug 2025
Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning.
Bar chart comparing software engineering accuracy between two versions of Opus 4. The first bar shows Opus 4 (May 2025) at 72.5% accuracy and the second bar shows Opus 4.1 (Aug 2025) at 74.5% accuracy.

ALT Bar chart comparing software engineering accuracy between two versions of Opus 4. The first bar shows Opus 4 (May 2025) at 72.5% accuracy and the second bar shows Opus 4.1 (Aug 2025) at 74.5% accuracy.

587
1,289
878
9,991
4,140,822
Amanda Askell retweeted
Anthropic (@AnthropicAI) profile picture
Anthropic
@AnthropicAI22 May 2025
Introducing the next generation: Claude Opus 4 and Claude Sonnet 4. Claude Opus 4 is our most powerful model yet, and the world’s best coding model. Claude Sonnet 4 is a significant upgrade from its predecessor, delivering superior coding and reasoning.
A benchmarking table titled Claude 4 benchmarks comparing performance metrics across various capabilities including coding, reasoning, tool use, multilingual Q&A, visual reasoning, and mathematics.

ALT A benchmarking table titled Claude 4 benchmarks comparing performance metrics across various capabilities including coding, reasoning, tool use, multilingual Q&A, visual reasoning, and mathematics.

920
3,106
1,492
20,411
4,298,291
Amanda Askell retweeted
Nathan Lambert (@natolambert) profile picture
Nathan Lambert
@natolambert2 Mar 2025
Now I really wanna know the fanmail that @AmandaAskell and Anthropic get for Claude
Sam Altman (@sama) profile pictureSam Altman
@sama2 Mar 2025
GPT-4.5 is the first time people have been emailing with such passion asking us to promise to never stop offering a specific model or even replace it with an update great work @kaicathyc @rapha_gl @mia_glaese
6
2
1
224
29,540
Amanda Askell retweeted
Kevin Roose (@kevinroose) profile picture
Kevin Roose
@kevinroose13 Dec 2024
I wrote about why it seems like everyone in SF’s AI scene talks to Claude now. (With thoughts from @aidan_mclau @JeffLadish and Anthropic’s @AmandaAskell, who created Claude’s character.)
Image from Kevin Roose (@kevinroose)
12
16
9
156
26,306
Amanda Askell retweeted
Anthropic (@AnthropicAI) profile picture
Anthropic
@AnthropicAI22 Oct 2024
Introducing an upgraded Claude 3.5 Sonnet, and a new model, Claude 3.5 Haiku. We’re also introducing a new capability in beta: computer use. Developers can now direct Claude to use computers the way people do—by looking at a screen, moving a cursor, clicking, and typing text.
A benchmark comparison table showing performance metrics for multiple AI models including Claude 3.5 Sonnet (new), Claude 3.5 Haiku, GPT-4o, and Gemini models across different tasks.

ALT A benchmark comparison table showing performance metrics for multiple AI models including Claude 3.5 Sonnet (new), Claude 3.5 Haiku, GPT-4o, and Gemini models across different tasks.

460
1,796
1,188
9,900
3,704,514
Amanda Askell retweeted
Jack Clark (@jackclarkSF) profile picture
Jack Clark
@jackclarkSF22 Aug 2024
Here's a letter we sent to Governor Newsom about SB 1047. This isn't an endorsement but rather a view of the costs and benefits of the bill. cdn.sanity.io/files/4zrzovbb…
13
30
22
192
83,528
Against expecting (or pretending to) normative perfection: goodthoughts.blog/p/imperfec…
Image from Richard Y. Chappell🔸 (@RYChappell)
2
14
4
52
18,150
Amanda Askell retweeted
Joshua Achiam (@jachiam0) profile picture
Joshua Achiam
@jachiam08 Mar 2024
The people opposing Paul Christiano are thoughtless and reckless. Paul would be an invaluable asset to government oversight and technical capacity on AI. He's in a league of his own on talent and dedication.
Sharon Goldman (@sharongoldman) profile pictureSharon Goldman
@sharongoldman8 Mar 2024
SCOOP: @NIST is facing an internal crisis as staff members and scientists have threatened to resign over the anticipated appointment of Paul Christiano to a crucial position at the agency's newly-formed US AI Safety Institute, sources say. venturebeat.com/ai/nist-staf…
NIST faces crisis over Paul Christiano hire | VentureBeat

NIST faces crisis over Paul Christiano hire | VentureBeat

NIST faces turmoil as staff consider quitting over Paul Christiano's expected appointment to a role at the US AI Safety Institute, sources say.

venturebeat.com
10
15
5
271
78,144
Amanda Askell retweeted
Adam Jermyn (@AdamSJermyn) profile picture
Adam Jermyn@AdamSJermyn24 Feb 2024
A small update from the Anthropic interpretability team: transformer-circuits.pub/202…
1
28
6
172
61,820
Several years ago @AmandaAskell told me how the immune system responds to tattoo ink and it convinced me never to get tattooed. Now Kurzgezagt has done a video describing the whole process. youtube.com/watch?v=nGggU-Cx…
9
23
4
165
35,624
Amanda Askell retweeted
Evan Hubinger (@EvanHub) profile picture
Evan Hubinger
@EvanHub14 Oct 2023
I wrote up some thoughts on AI pause advocacy. If you consider yourself an AI pause advocate, I think this will be a useful read! alignmentforum.org/posts/mcn…
RSPs are pauses done right — AI Alignment Forum

RSPs are pauses done right — AI Alignment Forum

COI: I am a research scientist at Anthropic, where I work on model organisms of misalignment; I was also involved in the drafting process for Anthrop…

alignmentforum.org
4
14
4
102
23,307
Amanda Askell retweeted
Chris Olah (@ch402) profile picture
Chris Olah
@ch4025 Oct 2023
If you'd asked me a year ago, superposition would have been by far the reason I was most worried that mechanistic interpretability would hit a dead end. I'm now very optimistic. I'd go as far as saying it's now primarily an engineering problem -- hard, but less fundamental risk.
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAI5 Oct 2023
The fact that most individual neurons are uninterpretable presents a serious roadblock to a mechanistic understanding of language models. We demonstrate a method for decomposing groups of neurons into interpretable features with the potential to move past that roadblock.
54
318
63
2,540
1,951,164
Amanda Askell retweeted
Anthropic (@AnthropicAI) profile picture
Anthropic
@AnthropicAI5 Oct 2023
Most neurons in language models are "polysemantic" – they respond to multiple unrelated things. For example, one neuron in a small language model activates strongly on academic citations, English dialogue, HTTP requests, Korean text, and others.
Shows twenty short example texts on which Neuron #83 fires. Each text is split into tokens, which are colored a shade of orange corresponding to how active the neuron is. Each text is labeled on the right with the type of text it is. The first four and last examples are labeled "Korean". One is labeled "Japanese". The example " Mouftah. Characterization of inter" is labeled "Citations". The example "dad…' he snarled. 'Even though you" is labeled "Dialogue". The example "string) (*http.Request, error)" is labeled "HTTP Request."

ALT Shows twenty short example texts on which Neuron #83 fires. Each text is split into tokens, which are colored a shade of orange corresponding to how active the neuron is. Each text is labeled on the right with the type of text it is. The first four and last examples are labeled "Korean". One is labeled "Japanese". The example " Mouftah. Characterization of inter" is labeled "Citations". The example "dad…' he snarled. 'Even though you" is labeled "Dialogue". The example "string) (*http.Request, error)" is labeled "HTTP Request."

6
43
11
617
119,001
London, SF, DC
Image from Peter Wildeford🇺🇸🚀 (@peterwildeford)
1
4
66
3,788
Amanda Askell retweeted
Chris Olah (@ch402) profile picture
Chris Olah
@ch4027 Jun 2023
One of the ideas I find most useful from @AnthropicAI's Core Views on AI Safety post (anthropic.com/index/core-vie…) is thinking in terms of a distribution over safety difficulty. Here's a cartoon picture I like for thinking about it:
How Hard Is AI Safety? A graph with three distributions over difficulty "Safety isn't a problem view" (focused on easy), "Anthropic" (widely distributed), and "Pessimistic Safety View" (focused on AI is very hard).

ALT How Hard Is AI Safety? A graph with three distributions over difficulty "Safety isn't a problem view" (focused on easy), "Anthropic" (widely distributed), and "Pessimistic Safety View" (focused on AI is very hard).

19
96
31
687
200,818
Load more