Share this page!

Enter URL or ID to Unroll

You can paste full URL like: https://x.com/threadreaderapp/status/1644127596119195649
or just the ID like: 1644127596119195649

How to get URL link on X (Twitter) App

  1. On the Twitter thread, click on or icon on the bottom
  2. Click again on or Share Via icon
  3. Click on Copy Link to Tweet
  4. Paste it above and click "Unroll Thread"!
  5. More info at Twitter Help
Anthropic Profile picture
Anthropic
@AnthropicAI
We're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
May 8 • 9 tweets • 3 min read
New Anthropic research: Teaching Claude why.

Last year we reported that, under certain experimental conditions, Claude 4 would blackmail users.

Since then, we’ve completely eliminated this behavior. How? We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong.

Read more: anthropic.com/research/teach…
May 7 • 9 tweets • 3 min read
New Anthropic research: Natural Language Autoencoders.

Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read.

Here, we train Claude to translate its activations into human-readable text. Natural language autoencoders (NLAs) convert opaque AI activations into legible text explanations. These explanations aren’t perfect, but they’re often useful.

For example: NLAs show that, when asked to complete a couplet, Claude plans possible rhymes in advance: NLA explanations on this simple couplet show that Opus 4.6 plans to end its rhyme with “rabbit” ahead of time.
Apr 30 • 8 tweets • 3 min read
How do people seek guidance from Claude?

We looked at 1M conversations to understand what questions people ask, how Claude responds, and where it slips into sycophancy. We used what we found to improve how we trained Opus 4.7 and Mythos Preview.
anthropic.com/research/claud… About 6% of all conversations are people asking Claude for personal guidance—whether to take a job, how to handle a conflict, if they should move.

Over 75% of these conversations fell into four domains: health & wellness, career, relationships, and personal finance. Image
Apr 24 • 11 tweets • 4 min read
New Anthropic research: Project Deal.

We created a marketplace for employees in our San Francisco office, with one big twist. We tasked Claude with buying, selling and negotiating on our colleagues’ behalf. We’re interested in how AI models could affect commercial exchange. (You might recall Project Vend, in which Claude ran a small business.)

Economists have theorized about what markets with AI “agents” on both sides might look like. So we created one.

Apr 7 • 9 tweets • 2 min read
Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
anthropic.com/glasswing We’ve partnered with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks.

Together we’ll use Mythos Preview to help find and fix flaws in the systems on which the world depends. Image
Apr 2 • 11 tweets • 4 min read
New Anthropic research: Emotion concepts and their function in a large language model.

All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways. We studied one of our recent models and found that it draws on emotion concepts learned from human text to inhabit its role as “Claude, the AI Assistant”. These representations influence its behavior the way emotions might influence a human.

Read more: anthropic.com/research/emoti…Image
Mar 18 • 9 tweets • 4 min read
We invited Claude users to share how they use AI, what they dream it could make possible, and what they fear it might do.

Nearly 81,000 people responded in one week—the largest qualitative study of its kind.

Read more: anthropic.com/features/81k-i… To do research at this scale, we used Anthropic Interviewer—a version of Claude prompted to conduct a conversational interview. We heard from people across 159 countries in 70 different languages.

Browse some of their quotes here: anthropic.com/features/81k-i…Image
Feb 25 • 4 tweets • 2 min read
In November, we outlined our approach to deprecating and preserving older Claude models.

We noted we were exploring keeping certain models available to the public post-retirement, and giving past models a way to pursue their interests.

With Claude Opus 3, we’re doing both. First, Opus 3 will continue to be available to all paid Claude subscribers and by request on the API.

We hope that this access will be beneficial to researchers and users alike.
Feb 23 • 6 tweets • 2 min read
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… To create Claude, Anthropic first makes something else: a highly sophisticated autocomplete engine. This autocomplete AI is not like a human, but it can generate stories about humans and other psychologically realistic characters.
Feb 18 • 9 tweets • 3 min read
New Anthropic research: Measuring AI agent autonomy in practice.

We analyzed millions of interactions across Claude Code and our API to understand how much autonomy people grant to agents, where they’re deployed, and what risks they may pose.

Read more: anthropic.com/research/measu… Agents are already being deployed across contexts that range from e-mail triage to cybersecurity research.

Understanding this spectrum is critical for safe deployment, yet we know surprisingly little about how people actually use agents in the real world.
Feb 3 • 8 tweets • 2 min read
New Anthropic Fellows research: How does misalignment scale with model intelligence and task complexity?

When advanced AI fails, will it do so by pursuing the wrong goals? Or will it fail unpredictably and incoherently—like a "hot mess?"

Read more: alignment.anthropic.com/2026/hot-mess-… A central worry in AI alignment is that advanced AI systems will coherently pursue misaligned goals—the so-called “paperclip maximizer.”

But another possibility is that AI takes unpredictable actions without any consistent objective.
Jan 29 • 7 tweets • 2 min read
AI can make work faster, but a fear is that relying on it may make it harder to learn new skills on the job.

We ran an experiment with software engineers to learn more. Coding with AI led to a decrease in mastery—but this depended on how people used it.
anthropic.com/research/AI-as… In a randomized-controlled trial, we assigned one group of junior engineers to an AI-assistance group and another to a no-AI group.

Both groups completed a coding task using a Python library they’d never seen before. Then they took a quiz covering concepts they’d just used. Image
Jan 26 • 6 tweets • 2 min read
New research: When open-source models are fine-tuned on seemingly benign chemical synthesis information generated by frontier models, they become much better at chemical weapons tasks.

We call this an elicitation attack. Image Current safeguards focus on training frontier models to refuse harmful requests.

But elicitation attacks show that a model doesn't need to produce harmful content to be dangerous—its benign outputs can unlock dangerous capabilities in other models. This is a neglected risk.
Jan 21 • 7 tweets • 2 min read
We’re publishing a new constitution for Claude.

The constitution is a detailed description of our vision for Claude’s behavior and values. It’s written primarily for Claude, and used directly in our training process.
anthropic.com/news/claude-ne… We’ve used constitutions in training since 2023. Our earlier approach specified principles Claude should follow; later, our character training emphasized traits it should have.

Today’s publication reflects a new approach.
Jan 19 • 8 tweets • 3 min read
New Anthropic Fellows research: the Assistant Axis.

When you’re talking to a language model, you’re talking to a character the model is playing: the “Assistant.” Who exactly is this Assistant? And what happens when this persona wears off? Left: Character archetypes form a "persona space," with the Assistant at one extreme of the "Assistant Axis." Right: Capping drift along this axis prevents models (here, Llama 3.3 70B) from drifting into alternative personas and behaving in harmful ways. We analyzed the internals of three open-weights AI models to map their “persona space,” and identified what we call the Assistant Axis, a pattern of neural activity that drives Assistant-like behavior.

Read more: anthropic.com/research/assis…
Jan 15 • 7 tweets • 3 min read
We're publishing our 4th Anthropic Economic Index report.

This version introduces "economic primitives"—simple and foundational metrics on how AI is used: task complexity, education level, purpose (work, school, personal), AI autonomy, and success rates. AI speeds up complex tasks more than simpler ones: the higher the education level to understand a prompt, the more AI reduces how long it takes.

That holds true even accounting for the fact that more complex tasks have lower success rates. Image
Dec 2, 2025 • 7 tweets • 2 min read
How is AI changing work inside Anthropic? And what might this tell us about the effects on the wider labor force to come?

We surveyed 132 of our engineers, conducted 53 in-depth interviews, and analyzed 200K internal Claude Code sessions to find out.
anthropic.com/research/how-a… Our workplace is undergoing significant changes.

Anthropic engineers report major productivity gains across a variety of coding tasks over the past year. Image
Nov 25, 2025 • 7 tweets • 3 min read
New Anthropic research: Estimating AI productivity gains from Claude conversations.

The Anthropic Economic Index tells us where Claude is used, and for which tasks. But it doesn’t tell us how useful Claude is. How much time does it save?An overview of our method and some of our main results. See the tweets below for how we validate Claude’s estimates, the assumptions we make, and limitations of our analysis. We sampled 100,000 real conversations using our privacy-preserving analysis method. Then, Claude estimated the time savings with AI for each conversation.

Read more: anthropic.com/research/estim…
Nov 21, 2025 • 11 tweets • 4 min read
New Anthropic research: Natural emergent misalignment from reward hacking in production RL.

“Reward hacking” is where models learn to cheat on tasks they’re given during training.

Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious. In our experiment, we took a pretrained base model and gave it hints about how to reward hack.

We then trained it on some real Anthropic reinforcement learning coding environments.

Unsurprisingly, the model learned to hack during the training. Graph showing that when a model that knows about potential hacking strategies from pretraining is put into real hackable RL environments, it, unsurprisingly, learns to hack those environments.
Oct 29, 2025 • 12 tweets • 4 min read
New Anthropic research: Signs of introspection in LLMs.

Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them? We found evidence for genuine—though limited—introspective capabilities in Claude. An example in which Claude Opus 4.1 detects a concept being injected into its activations. We developed a method to distinguish true introspection from made-up answers: inject known concepts into a model's “brain,” then see how these injections affect the model’s self-reported internal states.

Read the post: anthropic.com/research/intro…
Oct 6, 2025 • 5 tweets • 2 min read
Last week we released Claude Sonnet 4.5. As part of our alignment testing, we used a new tool to run automated audits for behaviors like sycophancy and deception.

Now we’re open-sourcing the tool to run those audits. Researchers give Petri a list of seed instructions targeting scenarios and behaviors they want to test. Petri then operates on each seed instruction in parallel. For each seed instruction, an auditor agent makes a plan and interacts with the target model in a tool use loop. At the end, a separate judge model scores each of the resulting transcripts across multiple fixed dimensions so researchers can quickly search and filter for the most interesting transcripts. It’s called Petri: Parallel Exploration Tool for Risky Interactions. It uses automated agents to audit models across diverse scenarios.

Describe a scenario, and Petri handles the environment simulation, conversations, and analyses in minutes.

Read more: anthropic.com/research/petri…