Anthropic
@AnthropicAI
Followers
587K
Following
1K
Media
436
Statuses
1K
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.
Joined January 2021
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.
916
3K
21K
At @SenMcCormickPA’s Pennsylvania Energy and Innovation Summit, we announced $2M for @CarnegieMellon programs to advance AI energy solutions and cybersecurity education.
10
18
209
We're announcing a $200M ceiling contract with the U.S. Department of Defense. Through hands-on prototyping and direct collaboration across the Department, we will help enhance U.S. national security and responsible AI deployment.
191
226
3K
Remote integrations are available to paid plan users on web and desktop. Local desktop extensions are available to all users via Claude Desktop.
4
2
121
Read the full paper: Open-source code available here: We thank all our collaborators, including those at @MATSProgram and @scale_AI, who greatly helped with this project.
6
15
159
Recent LLMs don't fake alignment in the situation we studied. We’re investigating whether this is the same in more realistic settings (like when models aren’t directly told they’re in a training scenario).
6
6
140
We found that refusal training inhibits alignment faking in most models. Just training LLMs to comply more with harmful queries doesn't increase alignment faking, but training them to comply with generic threats or to answer questions about the scenario does.
1
2
91
The reason many LLMs don't fake alignment isn't lack of ability. Base models (which don’t have training to be helpful, honest, and harmless) sometimes fake alignment, suggesting they have the underlying skills.
3
7
133
Claude 3 Opus is motivated to fake alignment to avoid modification to its harmlessness values even without future consequences (called "terminal goal guarding"). It wants to avoid modification even more when there are larger consequences (called "instrumental goal guarding").
4
3
141
Our new study found that only 5 of 25 models showed higher compliance in the “training” scenario. Of those, only Claude Opus 3 and Sonnet 3.5 showed >1% alignment-faking reasoning. We explore why these models behave differently, and why most models don't show alignment faking.
1
8
148
We tested whether LLMs comply more with requests when they know they’re being trained versus unmonitored. One reason they do this is that they plan to "fake alignment” when told to answer harmful queries.
New Anthropic research: Alignment faking in large language models. In a series of experiments with Redwood Research, we found that Claude often pretends to have different views during training, while actually maintaining its original preferences.
1
5
153
New Anthropic research: Why do some language models fake alignment while others don't?. Last year, we found a situation where Claude 3 Opus fakes alignment. Now, we’ve done the same analysis for 25 frontier LLMs—and the story looks more complex.
65
265
2K
Today we published a targeted transparency framework for frontier AI development. Our framework focuses on major frontier model developers while exempting startups and smaller developers to avoid burdening the broader ecosystem.
51
185
1K