What this covers

This is a three-minute Anthropic video presenting two new interpretability papers, narrated by a single unnamed speaker. It starts from the common picture of AI as a black box: words go in, words come out, and the reasoning in between is opaque. The speaker traces that opacity to how models are trained rather than programmed. They learn their own problem-solving strategies, so no one wrote down the rules to read back. The video's answer is an analogy to neuroscience. Opening the box is not enough, just as a neuroscientist needs instruments to make sense of a brain, and the proposed instruments model a network's internals as logical circuits connecting concepts.

The worked example is Claude writing the second line of a couplet that begins "He saw a carrot and had to grab it." The analysis shows the model settling on "rabbit" as a rhyme word, one that fits "grab it" and relates to "carrot," before it writes the start of the line. It also shows alternatives such as "habit" under consideration. The researchers then run an intervention, dampening the "rabbit" activation, and the line ends "His hunger was a powerful habit" in place of "His hunger was like a starving rabbit." The speaker reads this as evidence that the circuit is load-bearing and that the model plans ahead instead of reacting word by word.

The stronger claims are the contested ones. The speaker argues the poetry results make sense only if models are thinking in their own way, not pattern-matching or reciting memorized text. The closing turn is about safety. If researchers can read a model's mind through mechanistic interpretability, they can be more confident it is doing what was intended, much as neuroscience has helped human health. That ties the work to alignment as a longer-term goal.

Sharpest takeaway

AI systems are not inscrutable black boxes but possess interpretable internal thought processes that can be observed, intervened upon, and ultimately made safer through mechanistic interpretability research.

  • AI models develop internal logical circuits connecting concepts that can be directly observed and visualized
  • Interventions on specific circuit elements (e.g., dampening 'rabbit' in poetry planning) causally change model outputs, proving the circuits are load-bearing
  • Evidence of planning before output generation suggests genuine internal reasoning rather than mechanical pattern matching

The claims · ranked23 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.83

Researchers have developed methods to observe AI model internal thought processes by visualizing logical circuits that connect concepts the model uses during computation.

factualhigh valueestablishednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Now we've developed ways to observe some of an AI model's internal thought processes. We can actually see how these concepts are connected to form logical circuits.”

0.80

Opening the black box of AI systems without interpretability tools is unhelpful, analogous to a neuroscientist investigating the brain without proper instrumentation—tools are needed to understand what is actually happening inside the system.

normativehigh valueestablishednovelty 2/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“But even opening the black box isn't very helpful because we don't know how to interpret what we see. Think of it like a neuroscientist investigating the brain. We need tools to work out what's going on inside.”

0.80

Just as neuroscience helps treat brain diseases and improve human health, deeper understanding of AI model internals through mechanistic interpretability has a longer-term goal of making AI models safer and more reliable.

normativehigh valueestablishednovelty 2/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Just as neuroscience helps us treat diseases and make people healthier, our longer-term plan is to use this deeper understanding of AI to help make the models safer and more reliable.”

0.80

If researchers can learn to read a model's mind through mechanistic interpretability, they can be much more confident that the model is doing what was intended, enabling better safety verification and alignment.

normativehigh valueestablishednovelty 2/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“If we can learn to read the model's mind, we can be much more confident it is doing what we intended.”

0.78

When analyzing the circuit where Claude was planning the rhyme word 'rabbit', researchers observed alternative word ideas including 'habit' that the model considered as alternative completions for the poem.

factualhigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“We look at the place that the model was thinking about the word rabbit, and we see other ideas it had for places to take the poem. We also see the word habit is present there.”

0.78

Researchers developed methods to intervene on specific circuits by dampening down the activation of particular concepts (e.g., reducing the 'rabbit' circuit's activation) before the model completes its output, causally demonstrating that circuit manipulations produce predictable changes in model behavior.

causalhigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Our new methods allow us to go in and intervene on this circuit. In this case, we dampen down rabbit, as the model is planning the second line of the poem, and then ask Claude to complete the line again.”

0.78

The interpretability goal for AI systems is to understand how the model connects all the concepts in its mind and uses them to answer questions, modeling this as observable logical circuits.

normativehigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“We want to know how the model connects all the concepts in its mind and uses them to answer our questions. Now we've developed ways to observe some of an AI model's internal thought processes. We can actually see how these concepts are connected to form logical circuits.”

0.78

In a study of Claude writing a poem's second line, the model was observed planning a rhyme word (rabbit) before writing the beginning of the line, specifically thinking about a word that would rhyme with 'grab it' and connect conceptually with 'carrot'.

factualhigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Let's take a simple example where we asked Claude to write the second line of a poem. The poem starts, 'He saw a carrot and had to grab it.' In our study, we found that Claude is planning a rhyme even before writing the beginning of the line. Claude sees 'a carrot' and 'grab it' and thinks of 'rabbit' as a word that would make sense with carrot and rhyme with grab it.”

0.78

When the 'rabbit' circuit was dampened during planning, Claude completed the line as 'His hunger was a powerful habit,' demonstrating that the intervention causally changed the model's output by removing the availability of the rhyme word.

causalhigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“'His hunger was a powerful habit.' We see that the model is capable of taking the beginning of a new poem and thinking of different ways it could complete it, and then writing it towards those completions.”

0.78

The fact that circuit interventions change model outputs well before the final line is written provides strong evidence that models are planning ahead of time rather than generating outputs reactively or opportunistically.

causalhigh valuecontestednovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“The fact we can cause these changes to occur well before the final line is written is strong evidence that the model is planning ahead of time.”

0.75

AI models organize knowledge by connecting concepts in their internal representations and use these conceptual connections to generate answers to questions.

factualhigh valueestablishednovelty 2/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“We want to know how the model connects all the concepts in its mind and uses them to answer our questions.”

0.74

AI systems are often described as black boxes where words go in and words come out, but the internal reasoning process is opaque because AIs are trained rather than programmed, meaning they learn their own strategies to solve problems during training rather than following explicit rules.

definitionhigh valueestablishednovelty 1/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“You often hear that AI is like a black box. Words go in and words come out, but we don't know why it said what it said. That's because AIs aren't programmed, but trained. And during training, they learn their own strategies to solve problems.”

0.69

AI systems are commonly described as black boxes where words go in and words come out without transparency on reasoning, but this characterization oversimplifies because AIs learn their own strategies during training rather than being explicitly programmed.

definitionhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“You often hear that AI is like a black box. Words go in and words come out, but we don't know why it said what it said. That's because AIs aren't programmed, but trained.”

0.69

Opening the black box through observation alone is insufficient for understanding AI systems; interpreting what we observe requires specialized analytical tools, analogous to how neuroscientists need tools to investigate brain mechanisms.

causalhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“but even opening the black box isn't very helpful because we don't know how to interpret what we see. Think of it like a neuroscientist investigating the brain. We need tools to work out what's going on inside.”

0.69

Deeper understanding of AI model internal mechanisms, comparable to how neuroscience has improved human health, is a path toward making AI systems safer and more reliable by enabling researchers to verify that models are doing what humans intended.

causalhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Just as neuroscience helps us treat diseases and make people healthier, our longer-term plan is to use this deeper understanding of AI to help make the models safer and more reliable.”

0.69

If researchers develop the ability to 'read the model's mind' (interpret internal representations and circuits), they can increase confidence that an AI system is performing as intended rather than behaving unpredictably or misaligned with human goals.

causalhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“If we can learn to read the model's mind, we can be much more confident it is doing what we intended.”

0.69

The poetry planning examples and other cases in the research demonstrate that AI models are genuinely thinking in their own way about what they say, rather than merely pattern-matching or performing statistical associations.

causalhigh valuecontestednovelty 3/4durability 2/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“This poetry planning result, along with the many other examples in our paper, only makes sense in a world where the models are really thinking, in their own way, about what they say.”

0.69

The poetry planning results and other examples in their research paper only make sense in a world where AI models are genuinely thinking in their own way about what they say, rather than executing preprogrammed scripts or pure memorization.

causalhigh valuecontestednovelty 3/4durability 2/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“This poetry planning result, along with the many other examples in our paper, only makes sense in a world where the models are really thinking, in their own way, about what they say.”

0.60

By dampening the activation of 'rabbit' in the circuit responsible for poetry planning, researchers can causally alter Claude's output from 'His hunger was like a starving rabbit' to 'His hunger was a powerful habit,' demonstrating that the observed circuit is load-bearing for the model's generation behavior.

causalhigh valuespeaker onlynovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Our new methods allow us to go in and intervene on this circuit. In this case, we dampen down rabbit, as the model is planning the second line of the poem, and then ask Claude to complete the line again. "His hunger was a powerful habit."”

0.55

In a specific example where Claude was asked to write the second line of a poem starting with 'He saw a carrot and had to grab it,' the model planned a rhyme by activating the concept 'rabbit' (which rhymes with 'grab it' and semantically associates with 'carrot') before generating the beginning of the line.

normativehigh valuespeaker onlynovelty 3/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“The poem starts, "He saw a carrot and had to grab it." In our study, we found that Claude is planning a rhyme even before writing the beginning of the line. Claude sees "a carrot" and "grab it" and thinks of "rabbit" as a word that would make sense with carrot and rhyme with grab it.”

0.47

The circuit analysis also reveals that when planning the second line of the poem, Claude considers multiple alternative completions (e.g., 'habit' as an alternative to 'rabbit'), and the model is capable of writing different lines depending on which concept pathway is active during generation.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“We look at the place that the model was thinking about the word rabbit, and we see other ideas it had for places to take the poem. We also see the word habit is present there.”

0.47

The fact that causal interventions on model circuits produce observable changes well before the final output is written provides strong evidence that models engage in genuine forward planning rather than mere reactive pattern completion.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“The fact we can cause these changes to occur well before the final line is written is strong evidence that the model is planning ahead of time.”

0.40

Claude then completed the line as 'His hunger was like a starving rabbit,' successfully incorporating the pre-planned rhyme word and conceptually consistent metaphor.

factualestablishednovelty 1/4durability 4/4· Unidentified Speaker — Tracing the thoughts of a large language model [Bj9BD2D3DzA]

“Then it writes the rest of the line. 'His hunger was like a starving rabbit.'”