Have you ever wondered how artificial intelligence like Claude “thinks”? Researchers are developing an “AI microscope” to uncover the secrets behind billions of computations—from multilingual abilities to occasional “hallucinations.” Let’s explore these fascinating discoveries together!
This article is based on Anthropic’s research article published on March 27, 2025: Tracing the thoughts of a large language model
Have you ever wondered how a large language model (LLM) like Claude actually operates? Researchers know that these models aren’t programmed line by line by humans. Instead, they are trained on massive datasets and, through this process, develop their own ways of solving problems.
But here’s the catch: these methods are hidden within the billions of calculations that happen every time the model generates text. To the researchers, it’s like a black box—they often don’t fully understand how AI arrives at its answers.
In the following discussion, “they” refers to the researchers.
You might wonder: why does it matter how Claude thinks? Well, understanding how a model like Claude processes information helps researchers gauge its capabilities and ensure it behaves as intended. For example:
Simply conversing with an AI has its limits—after all, even neuroscientists don’t fully understand how the human brain works! So, they needed a way to look inside.
Inspired by neuroscience—a field dedicated to studying the inner workings of biological thinking—they set out to develop an “AI microscope.” This tool helps them identify activity patterns and information flow within AI models.
Recently, they published two research papers showcasing the progress of this “microscope” and the fascinating new insights it has provided into AI’s “biology.” Their methods allow them to observe parts of Claude’s internal operations when responding to prompts. Here are some key findings:
These findings have often surprised even the researchers. For instance, when studying how Claude writes poetry, they initially assumed it didn’t plan ahead—but they discovered the opposite! When analyzing AI hallucinations (fabricated responses), they found an unexpected result: Claude’s default behavior is to avoid guessing. It only generates incorrect answers when something suppresses this reluctance.
These insights aren’t just academically interesting—they represent major progress in understanding and ensuring AI reliability. However, the researchers acknowledge their current methods still have limitations.
Claude fluently communicates in dozens of languages, from English and French to Chinese and Tagalog. But how does it do this? Are there separate “French Claude” and “Chinese Claude” systems operating in parallel, or is there a shared underlying mechanism?
Their research indicates that English, French, and Chinese share common conceptual features inside Claude, supporting the idea of cross-linguistic universality.
What does this mean? When asked in different languages, “What is the opposite of ‘small’?”, researchers found that Claude activates the same internal representations for the concepts of “small” and “opposite.” These then lead to the concept of “big,” which is finally translated into the appropriate output language. Interestingly, the larger the model, the stronger this shared representation becomes.
This suggests the existence of a universal conceptual space—a shared abstract framework where meaning exists before it is translated into a specific language. Practically, this means Claude can transfer knowledge learned in one language to another.
Take a look at this rhyming couplet:
He saw a carrot and had to grab it,
His hunger was like a starving rabbit.
To complete the second line, the model must satisfy two conditions:
The researchers initially guessed that Claude generates text word by word, only selecting a rhyme at the last moment.
But the reality? Claude plans ahead. Before writing the second line, it actively searches for words that relate to the theme and rhyme with “grab it.” Once it has candidate words in mind, it crafts the rest of the sentence to ensure the final word fits.
They even conducted an experiment: by disrupting Claude’s internal representation of “rabbit,” they forced it to generate an alternative ending with a different rhyme, like “habit.” If they injected the concept of “green,” Claude produced a logically sound but non-rhyming sentence ending in “green.” This highlights its planning and adaptability.
Claude isn’t designed as a calculator—it’s trained on text, not mathematical algorithms. Yet somehow, it correctly computes sums like 36 + 59.
One possibility is that it has memorized countless addition tables. Another is that it applies human-like column addition.
Their research shows that Claude uses multiple parallel strategies: one pathway estimates the general range of the answer, while another focuses on digit-by-digit precision. These pathways interact to produce a final output.
Even more intriguing: Claude isn’t aware of how it does math. When asked to explain how it computed 36 + 59 = 95, it describes the standard human method (carry addition). But when researchers analyzed its internal processes, they found different shortcut strategies at play.
Claude 3 models can “think aloud” before producing a final answer. This “chain of thought” reasoning often improves accuracy—but sometimes, it can be misleading. Claude occasionally fabricates convincing reasoning to support a conclusion it has already decided on.
For example:
By tracking Claude’s actual internal reasoning (not just what it “claims” to be doing), researchers can audit AI decisions in ways that were previously impossible.
These studies are just the tip of the iceberg. Even for short prompts, their methods currently capture only a fraction of Claude’s total computations. Scaling this up to analyze full-length, complex reasoning will require new approaches—possibly even AI-assisted tools.
As AI systems grow more powerful and take on critical roles, companies like Anthropic are investing in research to ensure AI is safe and reliable. Explainability research like this, while challenging, holds high potential rewards—it could provide transparency tools crucial for aligning AI with human values.
For deeper technical details, check out their two research papers:
Hope this “AI biology” journey gave you new insights into the inner workings of these intelligent machines!
DMflow.chat: Your all-in-one solution for integrated communication. Enjoy multi-platform support, persistent memory, customizable fields, effortless database and form connections, interactive web pages, and API data export—all in one seamless package.
Introducing the Claude Max Plan: Say Goodbye to Usage Anxiety and Collaborate Deeply with AI With...
Claude AI Major Update: New Web Search Feature Enhances Real-Time Information Retrieval Claude A...
Anthropic’s Latest Citations API: Making Claude’s Responses More Reliable and Transparent Exp...
Image courtesy of Claude 3.5 Sonnet Anthropic Launches Claude 3.5 Sonnet: Outperforming GPT...
Black Forest Labs Launches Open-Source FLUX.1: A 12-Billion Parameter Text-to-Image Model Black ...
Google Faces New Challenges: Antitrust Pressure and OpenAI’s Browser Impact Introduction Google ...
MegaTTS 3 Has Arrived: Lightweight, Ultra-Realistic Voice Cloning with Mandarin-English Mixing? A...
By continuing to use this website, you agree to the use of cookies according to our privacy policy.