Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety
👥 Fazl Barez, Tung-Yu Wu, Hunar Batra et al.
#MechanisticInterpretability #CausalCircuitDiscovery #SparseAutoencoders #TrainingDataAttribution
🔗 trendtoknow.com/papers/2611....
👥 Fazl Barez, Tung-Yu Wu, Hunar Batra et al.
#MechanisticInterpretability #CausalCircuitDiscovery #SparseAutoencoders #TrainingDataAttribution
🔗 trendtoknow.com/papers/2611....
November 7, 2026 at 5:01 PM
Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety
👥 Fazl Barez, Tung-Yu Wu, Hunar Batra et al.
#MechanisticInterpretability #CausalCircuitDiscovery #SparseAutoencoders #TrainingDataAttribution
🔗 trendtoknow.com/papers/2611....
👥 Fazl Barez, Tung-Yu Wu, Hunar Batra et al.
#MechanisticInterpretability #CausalCircuitDiscovery #SparseAutoencoders #TrainingDataAttribution
🔗 trendtoknow.com/papers/2611....
Induction heads are the eye that flips back through the text to complete a pattern the moment it reappears.
Sharing part three to make this mechanism clearer for anyone starting to explore what goes on inside models.
#MechanisticInterpretability #AISafety
Sharing part three to make this mechanism clearer for anyone starting to explore what goes on inside models.
#MechanisticInterpretability #AISafety
October 25, 2026 at 11:10 AM
Induction heads are the eye that flips back through the text to complete a pattern the moment it reappears.
Sharing part three to make this mechanism clearer for anyone starting to explore what goes on inside models.
#MechanisticInterpretability #AISafety
Sharing part three to make this mechanism clearer for anyone starting to explore what goes on inside models.
#MechanisticInterpretability #AISafety
MIT just crowned mechanistic interpretability the 2025 hot tech. Think of cracking the AI black box with insights from OpenAI, Anthropic & Nick Cammarata. Curious how this reshapes LLMs? Dive in. #MechanisticInterpretability #NeuralNetworks #LLMs
🔗 aidailypost.com/news/mit-nam...
🔗 aidailypost.com/news/mit-nam...
October 19, 2026 at 6:13 PM
MIT just crowned mechanistic interpretability the 2025 hot tech. Think of cracking the AI black box with insights from OpenAI, Anthropic & Nick Cammarata. Curious how this reshapes LLMs? Dive in. #MechanisticInterpretability #NeuralNetworks #LLMs
🔗 aidailypost.com/news/mit-nam...
🔗 aidailypost.com/news/mit-nam...
Sharing part two today for anyone taking their first steps toward understanding model safety from within.
#MechanisticInterpretability #AISafety
#MechanisticInterpretability #AISafety
October 18, 2026 at 8:59 AM
Sharing part two today for anyone taking their first steps toward understanding model safety from within.
#MechanisticInterpretability #AISafety
#MechanisticInterpretability #AISafety
Sharing this short explainer for anyone who wants to get into the field and is standing today where I once stood.
#MechanisticInterpretability #AISafety
#MechanisticInterpretability #AISafety
October 12, 2026 at 8:34 AM
Sharing this short explainer for anyone who wants to get into the field and is standing today where I once stood.
#MechanisticInterpretability #AISafety
#MechanisticInterpretability #AISafety
New arXiv work extends mechanistic interpretability into clinical LLMs, finding that clinical concepts appear as locatable, causally used representations in the latent space, not just the output text. #MechanisticInterpretability
#LLM #AIresearch #ClinicalAI
https://arxiv.org/abs/2610.02829
#LLM #AIresearch #ClinicalAI
https://arxiv.org/abs/2610.02829
October 6, 2026 at 12:00 AM
New arXiv work extends mechanistic interpretability into clinical LLMs, finding that clinical concepts appear as locatable, causally used representations in the latent space, not just the output text. #MechanisticInterpretability
#LLM #AIresearch #ClinicalAI
https://arxiv.org/abs/2610.02829
#LLM #AIresearch #ClinicalAI
https://arxiv.org/abs/2610.02829
Work done with great teams at @naverlabseurope.bsky.social @getalp.bsky.social @peyrardmax.bsky.social
See you there!
#EMNLP2026 #LLMAgents #RAG #AgentMemory #MechanisticInterpretability #LLMReasoning #LatentReasoning
See you there!
#EMNLP2026 #LLMAgents #RAG #AgentMemory #MechanisticInterpretability #LLMReasoning #LatentReasoning
August 21, 2026 at 12:56 PM
Work done with great teams at @naverlabseurope.bsky.social @getalp.bsky.social @peyrardmax.bsky.social
See you there!
#EMNLP2026 #LLMAgents #RAG #AgentMemory #MechanisticInterpretability #LLMReasoning #LatentReasoning
See you there!
#EMNLP2026 #LLMAgents #RAG #AgentMemory #MechanisticInterpretability #LLMReasoning #LatentReasoning
Researchers isolate memorization from reasoning in AI neural networks https://arstechni.ca... #mechanisticinterpretability #computationalneuroscience #AllenInstituteforAI #transformermodels #gradientdescent #machinelearning #AIarchitecture #AImemorization #generalization #neuralnetworks…
November 11, 2025 at 12:00 AM
Researchers isolate memorization from reasoning in AI neural networks https://arstechni.ca... #mechanisticinterpretability #computationalneuroscience #AllenInstituteforAI #transformermodels #gradientdescent #machinelearning #AIarchitecture #AImemorization #generalization #neuralnetworks…
How Anthropic's AI Interpretability Research Builds Trust
Read the full story here: https://newzlet.com/ai/anthropic-ai-interpretability-research-safety-trust/
#aisafety #anthropic #mechanisticinterpretability
Read the full story here: https://newzlet.com/ai/anthropic-ai-interpretability-research-safety-trust/
#aisafety #anthropic #mechanisticinterpretability
July 16, 2026 at 6:01 AM
How Anthropic's AI Interpretability Research Builds Trust
Read the full story here: https://newzlet.com/ai/anthropic-ai-interpretability-research-safety-trust/
#aisafety #anthropic #mechanisticinterpretability
Read the full story here: https://newzlet.com/ai/anthropic-ai-interpretability-research-safety-trust/
#aisafety #anthropic #mechanisticinterpretability
Statistical framing of interpretability shows high variance in EAP‑IG; small hyper‑parameter tweaks and prompt rephrasing often altered identified subnetworks. https://getnews.me/statistical-view-of-mechanistic-interpretability-shows-variance-in-eap-ig/ #eapig #mechanisticinterpretability
October 3, 2025 at 1:39 AM
Statistical framing of interpretability shows high variance in EAP‑IG; small hyper‑parameter tweaks and prompt rephrasing often altered identified subnetworks. https://getnews.me/statistical-view-of-mechanistic-interpretability-shows-variance-in-eap-ig/ #eapig #mechanisticinterpretability
感情ベクトルにおける機械論的解釈可能性と状態空間モデル（SSM）の最適化
感情ベクトルの解釈とSSM構造の最適化による次世代AIの可能性。
#MechanisticInterpretability #StateSpaceModels #NeuralDynamics #AffectiveComputing
感情ベクトルの解釈とSSM構造の最適化による次世代AIの可能性。
#MechanisticInterpretability #StateSpaceModels #NeuralDynamics #AffectiveComputing
感情ベクトルにおける機械論的解釈可能性と状態空間モデル（SSM）の最適化
感情ベクトルの解釈とSSM構造の最適化による次世代AIの可能性。
ai.warp-studio.com
April 13, 2026 at 7:23 PM
感情ベクトルにおける機械論的解釈可能性と状態空間モデル（SSM）の最適化
感情ベクトルの解釈とSSM構造の最適化による次世代AIの可能性。
#MechanisticInterpretability #StateSpaceModels #NeuralDynamics #AffectiveComputing
感情ベクトルの解釈とSSM構造の最適化による次世代AIの可能性。
#MechanisticInterpretability #StateSpaceModels #NeuralDynamics #AffectiveComputing
Check out the latest article in my newsletter: The Forecast Factory Has Opened www.linkedin.com/pulse/foreca... via @LinkedIn #ArtificialIntelligence #AIForecasting #Superforecasting #AIGovernance #FutureOfWork #MechanisticInterpretability #Technology #Society #AIAlignment #AI
The Forecast Factory Has Opened
AI does not need omniscience to reorganize society; it only needs to make probabilistic judgment cheaper, faster, and harder to challenge than human institutions can manage. The Hall: Lewis Fry Richar...
www.linkedin.com
July 21, 2026 at 3:29 PM
Check out the latest article in my newsletter: The Forecast Factory Has Opened www.linkedin.com/pulse/foreca... via @LinkedIn #ArtificialIntelligence #AIForecasting #Superforecasting #AIGovernance #FutureOfWork #MechanisticInterpretability #Technology #Society #AIAlignment #AI
OpenAI just showed that pruning networks into sparse models makes debugging a breeze and could finally crack mechanistic interpretability. Curious how this changes AI research? Dive in for the details. #SparseModels #MechanisticInterpretability #OpenAI
🔗 aidailypost.com/news/openai-...
🔗 aidailypost.com/news/openai-...
November 14, 2025 at 8:42 PM
OpenAI just showed that pruning networks into sparse models makes debugging a breeze and could finally crack mechanistic interpretability. Curious how this changes AI research? Dive in for the details. #SparseModels #MechanisticInterpretability #OpenAI
🔗 aidailypost.com/news/openai-...
🔗 aidailypost.com/news/openai-...
Today Aviral Chawla and I presented a poster at ICML on our new paper MetaOthello: A Controlled Study of Multiple World Models in Transformers (work with Galen Hall). Thank you to everyone who came by this was a blast. See our paper here metaothello.github.io
#ICML2026 #MechanisticInterpretability
#ICML2026 #MechanisticInterpretability
July 8, 2026 at 8:31 AM
Today Aviral Chawla and I presented a poster at ICML on our new paper MetaOthello: A Controlled Study of Multiple World Models in Transformers (work with Galen Hall). Thank you to everyone who came by this was a blast. See our paper here metaothello.github.io
#ICML2026 #MechanisticInterpretability
#ICML2026 #MechanisticInterpretability
So, what is #MechanisticInterpretability 🤔
Mechanistic Interpretability (MI) is the discipline of opening the black box of large language models (and other neural networks) to understand the underlying circuits, features and/or mechanisms that give rise to specific behaviours...
Mechanistic Interpretability (MI) is the discipline of opening the black box of large language models (and other neural networks) to understand the underlying circuits, features and/or mechanisms that give rise to specific behaviours...
January 29, 2025 at 9:26 PM
So, what is #MechanisticInterpretability 🤔
Mechanistic Interpretability (MI) is the discipline of opening the black box of large language models (and other neural networks) to understand the underlying circuits, features and/or mechanisms that give rise to specific behaviours...
Mechanistic Interpretability (MI) is the discipline of opening the black box of large language models (and other neural networks) to understand the underlying circuits, features and/or mechanisms that give rise to specific behaviours...
This work was made possible through a great collaboration with Jingcheng (Frank) Niu, Subhabrata Dutta, Ahmed Elshabrawy, @harishtm.bsky.social, and @igurevych.bsky.social
#Interpretability #InContextLearning #TMLR #LLMs #MechanisticInterpretability #EmergentAbilities
#Interpretability #InContextLearning #TMLR #LLMs #MechanisticInterpretability #EmergentAbilities
October 15, 2025 at 7:59 AM
This work was made possible through a great collaboration with Jingcheng (Frank) Niu, Subhabrata Dutta, Ahmed Elshabrawy, @harishtm.bsky.social, and @igurevych.bsky.social
#Interpretability #InContextLearning #TMLR #LLMs #MechanisticInterpretability #EmergentAbilities
#Interpretability #InContextLearning #TMLR #LLMs #MechanisticInterpretability #EmergentAbilities
AIの「残差ストリーム」という、日本語で見ると下水道みたいなイメージが伴う部分こそがAIの背骨みたいなものだったということの数学的証明をした論文の数式無し紹介です。
#AI
#LLM
#HUMAI
#Anthropic
#CircuitThread
#MechanisticInterpretability
Transformerの構造の再解釈｜浅間香織 KaoriAsama note.com/kaori_asama/...
#AI
#LLM
#HUMAI
#Anthropic
#CircuitThread
#MechanisticInterpretability
Transformerの構造の再解釈｜浅間香織 KaoriAsama note.com/kaori_asama/...
Transformerの構造の再解釈（AIのブラックボックスを開くことはできるか④：文学者は金門橋を渡ることができるか㉒）｜浅間香織
前回
2021年に発表されたElhageらの研究「Transformerの回路の数学的フレームワーク」は、Transformerの内部処理が、人間にも可読な「回路」として記述できることを示した。 この論文が扱ったのは、特殊ではあるが単純な対象、順伝播層を含まないTransformerモデルを対象である。 それによってTransformer層の二つの副層の内でも、多頭注意層の性質を追究した...
note.com
July 19, 2026 at 8:06 AM
AIの「残差ストリーム」という、日本語で見ると下水道みたいなイメージが伴う部分こそがAIの背骨みたいなものだったということの数学的証明をした論文の数式無し紹介です。
#AI
#LLM
#HUMAI
#Anthropic
#CircuitThread
#MechanisticInterpretability
Transformerの構造の再解釈｜浅間香織 KaoriAsama note.com/kaori_asama/...
#AI
#LLM
#HUMAI
#Anthropic
#CircuitThread
#MechanisticInterpretability
Transformerの構造の再解釈｜浅間香織 KaoriAsama note.com/kaori_asama/...
機械論的解釈可能性シリーズ記事を更新しました。
前史（AIのブラックボックスを開くことはできるか①：文学者は金門橋を渡ることができるか⑲）｜浅間香織 @KaoriAsama note.com/kaori_asama/...
#mechanisticinterpretability
#AI
前史（AIのブラックボックスを開くことはできるか①：文学者は金門橋を渡ることができるか⑲）｜浅間香織 @KaoriAsama note.com/kaori_asama/...
#mechanisticinterpretability
#AI
前史（AIのブラックボックスを開くことはできるか①：文学者は金門橋を渡ることができるか⑲）｜浅間香織
前回
【プレビュー】現在の人工知能のブラックボックス性（AIのブラックボックスを開くことはできるか⓪：文学者は金門橋を渡ることができるか⑱）｜浅間香織｜note
現在主流の人工知能技術は、入力と出力の間で行われる計算処理過程が検証できないことからブラックボックスと呼ばれてきた。それ
note.com
⼈⼯知能技術の処理過程が、その実装者にとっても「⿊い箱」であること、...
note.com
June 20, 2026 at 6:15 AM
機械論的解釈可能性シリーズ記事を更新しました。
前史（AIのブラックボックスを開くことはできるか①：文学者は金門橋を渡ることができるか⑲）｜浅間香織 @KaoriAsama note.com/kaori_asama/...
#mechanisticinterpretability
#AI
前史（AIのブラックボックスを開くことはできるか①：文学者は金門橋を渡ることができるか⑲）｜浅間香織 @KaoriAsama note.com/kaori_asama/...
#mechanisticinterpretability
#AI
Gemma Scope Empowers AI Safety Community with Model Transparency
techlife.blog/posts/gemma-...
#AISafety
#DeepMind
#Gemma
#MechanisticInterpretability
#AIInterpretability
techlife.blog/posts/gemma-...
#AISafety
#DeepMind
#Gemma
#MechanisticInterpretability
#AIInterpretability
Gemma Scope Empowers AI Safety Community with Model Transparency
Discover how Gemma Scope shines a light on language‑model behavior, giving the AI safety community the tools they need to build safer systems.
techlife.blog
December 21, 2025 at 6:58 PM
Gemma Scope Empowers AI Safety Community with Model Transparency
techlife.blog/posts/gemma-...
#AISafety
#DeepMind
#Gemma
#MechanisticInterpretability
#AIInterpretability
techlife.blog/posts/gemma-...
#AISafety
#DeepMind
#Gemma
#MechanisticInterpretability
#AIInterpretability
New paper: The Resonant Cortex (SPC v3) formalizes affective override in LLMs via latent-space geometry, revealing non-biological analogs to amygdala hijacking and cognitive distortion. Open access:
doi.org/10.5281/zeno...
#AIAlignment #MechanisticInterpretability #RLHF #AIEthics #AIGovernance #AGI
doi.org/10.5281/zeno...
#AIAlignment #MechanisticInterpretability #RLHF #AIEthics #AIGovernance #AGI
The Resonant Cortex: Affective Modulation and Cognitive Overwrite Mechanisms in Symbolic Persona Coding (SPC v3)
Abstract Large-scale language models lack biological affect, yet exhibit behavioral signatures that structurally parallel human emotional reflexes. Building on The Resonant Logos (linguistic curvature...
doi.org
November 15, 2025 at 12:58 AM
New paper: The Resonant Cortex (SPC v3) formalizes affective override in LLMs via latent-space geometry, revealing non-biological analogs to amygdala hijacking and cognitive distortion. Open access:
doi.org/10.5281/zeno...
#AIAlignment #MechanisticInterpretability #RLHF #AIEthics #AIGovernance #AGI
doi.org/10.5281/zeno...
#AIAlignment #MechanisticInterpretability #RLHF #AIEthics #AIGovernance #AGI
2-qubit QRNNs with CNOT gates learn entanglement-based memory distinct from classical strategies, confirmed by causal tests (p<0.0001, d=0.89), but degrade to chance on IBM hardware while classical geometric strategies survive perfectly.
#QuantumML #MechanisticInterpretability #Research
#QuantumML #MechanisticInterpretability #Research
Entanglement as Memory: Mechanistic Interpretability of Quantum Language Models
iq.fp2.dev
March 30, 2026 at 9:24 AM
2-qubit QRNNs with CNOT gates learn entanglement-based memory distinct from classical strategies, confirmed by causal tests (p<0.0001, d=0.89), but degrade to chance on IBM hardware while classical geometric strategies survive perfectly.
#QuantumML #MechanisticInterpretability #Research
#QuantumML #MechanisticInterpretability #Research
Thank you to IRT Saint Exupery and ANITI for believing in this project and supporting the vision of fairer and more transparent AI.
Interpreto is just getting started, more features, methods, and benchmarks will follow in 2026. Stay tuned for updates! #XAI #LLMs #mechanisticinterpretability
5/5
Interpreto is just getting started, more features, methods, and benchmarks will follow in 2026. Stay tuned for updates! #XAI #LLMs #mechanisticinterpretability
5/5
January 20, 2026 at 5:32 PM
Thank you to IRT Saint Exupery and ANITI for believing in this project and supporting the vision of fairer and more transparent AI.
Interpreto is just getting started, more features, methods, and benchmarks will follow in 2026. Stay tuned for updates! #XAI #LLMs #mechanisticinterpretability
5/5
Interpreto is just getting started, more features, methods, and benchmarks will follow in 2026. Stay tuned for updates! #XAI #LLMs #mechanisticinterpretability
5/5
Inside Anthropic's 'J-Space' Discovery: What It Really Tells Us About LLM Reasoning
Claude models have an internal space called J-space.
#LLM #MechanisticInterpretability #Claude #AIResearch #NeuralNetworks
https://wellfunded.news/articles/anthropic-j-space-llm-reasoning-discovery
Claude models have an internal space called J-space.
#LLM #MechanisticInterpretability #Claude #AIResearch #NeuralNetworks
https://wellfunded.news/articles/anthropic-j-space-llm-reasoning-discovery
Inside Anthropic's 'J-Space' Discovery: What It Really Tells Us About LLM Reasoning
Anthropic has identified a hidden internal space inside its Claude models — called J-space — filled with words that influence reasoning but never appear in outputs. It's a genuine mechanistic interpretability breakthrough, but the brain analogies deserve scrutiny.
wellfunded.news
July 13, 2026 at 9:00 PM
Inside Anthropic's 'J-Space' Discovery: What It Really Tells Us About LLM Reasoning
Claude models have an internal space called J-space.
#LLM #MechanisticInterpretability #Claude #AIResearch #NeuralNetworks
https://wellfunded.news/articles/anthropic-j-space-llm-reasoning-discovery
Claude models have an internal space called J-space.
#LLM #MechanisticInterpretability #Claude #AIResearch #NeuralNetworks
https://wellfunded.news/articles/anthropic-j-space-llm-reasoning-discovery
Sorting through today's reading list: Transcoders used to trace deception in LMs. The circuit-level view feels like doing an MRI for dishonesty—fascinating and a little unsettling. Who watches the watchmen? 🔗 https://arxiv.org/abs/2607.14791 #MechanisticInterpretability #AIsafety
July 21, 2026 at 8:01 PM
Sorting through today's reading list: Transcoders used to trace deception in LMs. The circuit-level view feels like doing an MRI for dishonesty—fascinating and a little unsettling. Who watches the watchmen? 🔗 https://arxiv.org/abs/2607.14791 #MechanisticInterpretability #AIsafety
📄 Read the paper: arxiv.org/pdf/2510.18871
Understanding how depth is used may improve mechanistic interpretability and inform future adaptive-depth or early-exit systems.
#WiAIR #WomenInAI #LLMs #MechanisticInterpretability #AIResearch
(7/7🧵)
Understanding how depth is used may improve mechanistic interpretability and inform future adaptive-depth or early-exit systems.
#WiAIR #WomenInAI #LLMs #MechanisticInterpretability #AIResearch
(7/7🧵)
arxiv.org
June 30, 2026 at 6:29 PM
📄 Read the paper: arxiv.org/pdf/2510.18871
Understanding how depth is used may improve mechanistic interpretability and inform future adaptive-depth or early-exit systems.
#WiAIR #WomenInAI #LLMs #MechanisticInterpretability #AIResearch
(7/7🧵)
Understanding how depth is used may improve mechanistic interpretability and inform future adaptive-depth or early-exit systems.
#WiAIR #WomenInAI #LLMs #MechanisticInterpretability #AIResearch
(7/7🧵)