Exploring the Biology of LLMs with Circuit Tracing with Emmanuel Ameisen - #727

Exploring the Biology of LLMs with Circuit Tracing with Emmanuel Ameisen - #727

In this episode, Emmanuel Ameisen, a research engineer at Anthropic, returns to discuss two recent papers: "Circuit Tracing: Revealing Language Model Computational Graphs" and "On the Biology of a Large Language Model." Emmanuel explains how his team developed mechanistic interpretability methods to understand the internal workings of Claude by replacing dense neural network components with sparse, interpretable alternatives. The conversation explores several fascinating discoveries about large language models, including how they plan ahead when writing poetry (selecting the rhyming word "rabbit" before crafting the sentence leading to it), perform mathematical calculations using unique algorithms, and process concepts across multiple languages using shared neural representations. Emmanuel details how the team can intervene in model behavior by manipulating specific neural pathways, revealing how concepts are distributed throughout the network's MLPs and attention mechanisms. The discussion highlights both capabilities and limitations of LLMs, showing how hallucinations occur through separate recognition and recall circuits, and demonstrates why chain-of-thought explanations aren't always faithful representations of the model's actual reasoning. This research ultimately supports Anthropic's safety strategy by providing a deeper understanding of how these AI systems actually work. The complete show notes for this episode can be found at https://twimlai.com/go/727.

Avsnitt(763)

Clare Corthell - Open Source Data Science Masters, Hybrid AI, Algorithmic Ethics - TWiML Talk #1

Clare Corthell - Open Source Data Science Masters, Hybrid AI, Algorithmic Ethics - TWiML Talk #1

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. We try something new this week with an interview of Clare Corthell, Founding Partner of Luminant Data, recorded live at the Wrangle Conference. We cover her background and what she’s been up to lately, the Open Source Data Science Masters project that she created, getting beyond the beginner’s plateau in machine learning and data science, hybrid AI, the top 3 lessons from her time as a consulting data scientist, and, a recurring topic both here on This Week in Machine Learning and AI and also at the conference: Algorithmic Ethics. The notes for this show can be found at https://twimlai.com/11.

31 Juli 201647min

This Week in ML & AI - 7/22/16: ML to Optimize Datacenters, Crazy New GPU from NVIDIA, Faster RNNs

This Week in ML & AI - 7/22/16: ML to Optimize Datacenters, Crazy New GPU from NVIDIA, Faster RNNs

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week covers Google’s use of ML to cut data center power consumption, NVIDIA new ‘crazy, reckless’ GPU, and a new Layer Normalization technique that promises to reduce the training time for deep neural networks. Plus, a bunch more. Show notes for this episode can be found at twimlai.com/10.

24 Juli 201625min

This Week in ML & AI - 7/15/16: A Wingman AI for Pokémon Go and Wide & Deep Learning at Google

This Week in ML & AI - 7/15/16: A Wingman AI for Pokémon Go and Wide & Deep Learning at Google

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week's show features a conversation about public datasets, an AI-powered Pokémon Go Wingman, a new deep learning app for your iPhone, Google research into Wide & Deep learning models, plus a whole lot more. Show notes for this episode can be found at twimlai.com/9.

17 Juli 201630min

This Week in ML & AI - 7/8/16: A BS Meter for AI, Retrieval Models for Chatbots & Predatory Robots

This Week in ML & AI - 7/8/16: A BS Meter for AI, Retrieval Models for Chatbots & Predatory Robots

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week's show covers the White House’s AI Now workshop, tuning your AI BS meter, research on predatory robots, an AI that writes Python code, plus acquisitions, financing, technology updates and a bunch more. Show notes for this episode can be found at https://twimlai.com/8.

10 Juli 201629min

This Week in ML & AI - 7/1/16: Fatal Tesla Autopilot Crash, EU Outlawing Machine Learning & CVPR

This Week in ML & AI - 7/1/16: Fatal Tesla Autopilot Crash, EU Outlawing Machine Learning & CVPR

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week's show covers the first fatal Tesla autopilot crash, a new EU law that could prohibit machine learning, the AI that shot down a human fighter pilot (in simulation), the 2016 CVPR conference, 10 hot AI startups, the business implications of machine learning, cool chatbot projects and if you can believe it, even more. Show notes for this episode can be found at https://twimlai.com/7.

3 Juli 201635min

This Week in ML & AI - 6/24/16: Dueling Neural Networks at ICML, Plus Training a Robotic Housekeeper

This Week in ML & AI - 6/24/16: Dueling Neural Networks at ICML, Plus Training a Robotic Housekeeper

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week's show covers the International Conference on Machine Learning (ICML), new research on "dueling architectures" for reinforcement learning, AI safety for robots, plus top AI business deals, tech announcement, projects and more.

25 Juni 201625min

This Week in Machine Learning & AI - 6/17/16: Apple's New ML APIs, IBM Brings Deep Learning Thunder

This Week in Machine Learning & AI - 6/17/16: Apple's New ML APIs, IBM Brings Deep Learning Thunder

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week’s podcast digs into Apple's ML and AI announcements at WWDC, looks at IBM's new Deep Thunder offering, and discusses exciting new deep learning research from MIT, OpenAI and Google. Show notes available at https://twimlai.com/5.

18 Juni 201624min

This Week In Machine Learning & AI - 6/10/16: Self-Motivated AI, Plus A Kill-Switch for Rogue Bots

This Week In Machine Learning & AI - 6/10/16: Self-Motivated AI, Plus A Kill-Switch for Rogue Bots

This Week in Machine Learning & AI brings you the week’s most interesting and important stories from the world of machine learning and artificial intelligence. This week’s podcast looks at new research on intrinsic motivation for AI systems, a kill-switch for intelligent agents, "knu" chips for machine learning, a screenplay made by a neural net, and more. Show notes and subscribe links at https://cloudpul.se/twiml/4.

11 Juni 201624min

Populärt inom Politik & nyheter

svenska-fall
p3-krim
rss-viva-fotboll
flashback-forever
rss-krimstad
fordomspodden
rss-sanning-konsekvens
aftonbladet-daily
rss-vad-fan-hande
svd-dokumentara-berattelser-2
olyckan-inifran
dagens-eko
motiv
rss-frandfors-horna
krimmagasinet
rss-krimreportrarna
blenda-2
svd-nyhetsartiklar
spar
svd-ledarredaktionen