Video wird geladen...

Video konnte nicht geladen werden

Beim Laden dieses Videos ist ein Problem aufgetreten. Dies könnte an einem vorübergehenden Netzwerkproblem liegen oder das Video ist möglicherweise nicht verfügbar.

✨ New AI Interfaces powered by Interpretability I'm excited to share LatentLit, the result of my applied AI research fellowship with Goodfire Mechanistic interpretability isn’t just important for AI safety, it also gives us new ways to steer and interact with LLMs.

Thariq

337,547 subscribers

68,258 Aufrufe • vor 1 Jahr •via X (Twitter)

Bildung Wissenschaft & Technologie

Anya Rossi• Live Now

Private livecam show

11 Kommentare

Profilbild von Thariq

Thariqvor 1 Jahr

In LatentLit, you write stories by like a DJ might make music, adjusting knobs and dials using steering and seeing what effect they have. You might call it Vibe Writing. Try it out here!

Profilbild von Thariq

Thariqvor 1 Jahr

Our neural map highlights features like "Supernatural Discovery" and connects them to specific text passages they influence. Each colored bubble in the neural map represents a feature the model activates. The bigger the bubble, the stronger the influence. Hovering over these features will highlight where in the text this feature was activated most strongly.

Profilbild von Thariq

Thariqvor 1 Jahr

There are thousands of features to explore! These features represent individuals concepts that the LLM understands, but they may surprise you in joyful ways when you use them to steer text generation.

Profilbild von FRANK E ELKINS

FRANK E ELKINSvor 1 Jahr

Confused about AI? Get clarity today with Book VI - "The Rational Being!" Understand the benefits & risks, empower your future now by learning what AI really is and how it really works! Also check out the Free Weekly Newsletter "How Things Work: A Brief History of Reality"

Profilbild von julius tarng cyber inspector

julius tarng cyber inspectorvor 1 Jahr

@GoodfireAI neat! i got into a crazy broken state fairly easily playing w knobs tho. also when you remove a feature the % activation doesn't move down the indices with it

Profilbild von Thariq

Thariqvor 1 Jahr

@GoodfireAI Ohh thanks for trying it out! That's definitely busted, let me see what I can do to fix it.

Profilbild von Shane Farrow

Shane Farrowvor 1 Jahr

@livgorton @GoodfireAI This is so well done. I wanted to try something similar to get user feedback less intrusively than making them engage with buttons, and this is masterfully done.

Profilbild von Thariq

Thariqvor 1 Jahr

@livgorton @GoodfireAI Thank you! A lot of love went into this, so glad that I had Goodfire's support in pushing it through

Profilbild von Theo

Theovor 1 Jahr

@GoodfireAI Very very cool!

Profilbild von Avi Siegel

Avi Siegelvor 1 Jahr

@GoodfireAI This is 🔥 Need more products that showcase how different aspects of the neural network are impacting the end result.

Profilbild von Thariq

Thariqvor 1 Jahr

@GoodfireAI Thank you! Honestly we have the capabilities, I'm hoping more people will be inspired to use them. You can even feed in the inference from another AI into Goodfire's SDK and read out the features that way

Ähnliche Videos

What does the former co-lead of interpretability at Deepmind and now co-founder of Goodfire think about how to keep AI from becoming (his words) "a generally evil guy"? I got to speak with Tom McGrath at South Park Commons the day after OpenAI announced the Hugging Face hack. Good timing! (00:30) Why AI Interpretability Matters Now (02:15) When AI Reward Hacking Becomes Real (05:07) The Core Problem of Interpretability (07:23) Intentional Design: Steering What Models Learn (12:43) Neural Geometry: The Shapes Inside AI Models (22:29) Turning Interpretability Research Into a Product (25:28) How Research Changes Inside a Startup (32:48) Science Fiction, AI Risk, and the Future

What does the former co-lead of interpretability at Deepmind and now co-founder of Goodfire think about how to keep AI from becoming (his words) "a generally evil guy"? I got to speak with Tom McGrath at South Park Commons the day after OpenAI announced the Hugging Face hack. Good timing! (00:30) Why AI Interpretability Matters Now (02:15) When AI Reward Hacking Becomes Real (05:07) The Core Problem of Interpretability (07:23) Intentional Design: Steering What Models Learn (12:43) Neural Geometry: The Shapes Inside AI Models (22:29) Turning Interpretability Research Into a Product (25:28) How Research Changes Inside a Startup (32:48) Science Fiction, AI Risk, and the Future

Jonathan Brebner

12,759 Aufrufe • vor 1 Monat

“The hope is that ... just optimizing something to be sparse—without optimizing it to be interpretable—will stumble across that interpretable decomposition.” — Neel Nanda on sparse autoencoders for mechanistic interpretability and AI safety at the Vienna Alignment Workshop.

“The hope is that ... just optimizing something to be sparse—without optimizing it to be interpretable—will stumble across that interpretable decomposition.” — Neel Nanda on sparse autoencoders for mechanistic interpretability and AI safety at the Vienna Alignment Workshop.

FAR.AI

1,431,043 Aufrufe • vor 2 Jahren

Can we map the mind of an LLM? Our first mechanistic interpretability episode on Training Data featuring Goodfire founder Eric Ho (and our first cameo from Roelof Botha!) Goodfire is building an independent mech interp lab, led by some heavyweight researchers from the field (e.g. Lee Sharkey who has led a lot of important work in sparse autoencoders to "unscramble" LLMs and resolve superposition, Nick who has been a key pioneer behind auto interpretability) On this episode, Eric gives us a flyover of the technical results so far from this nascent field (universality, superposition), what's ahead in the research (going from circuits to weights, going from understanding to increasingly surgical editing), a preview of the real-world work they're doing already with Arc Institute, and the impact he expects Goodfire and the broader field to have on steering, safety, editing and more.

Can we map the mind of an LLM? Our first mechanistic interpretability episode on Training Data featuring Goodfire founder Eric Ho (and our first cameo from Roelof Botha!) Goodfire is building an independent mech interp lab, led by some heavyweight researchers from the field (e.g. Lee Sharkey who has led a lot of important work in sparse autoencoders to "unscramble" LLMs and resolve superposition, Nick who has been a key pioneer behind auto interpretability) On this episode, Eric gives us a flyover of the technical results so far from this nascent field (universality, superposition), what's ahead in the research (going from circuits to weights, going from understanding to increasingly surgical editing), a preview of the real-world work they're doing already with Arc Institute, and the impact he expects Goodfire and the broader field to have on steering, safety, editing and more.

Sonya Huang 🐥

19,379 Aufrufe • vor 1 Jahr

A model’s chain of thought acts like a scratch pad, offering a window into its reasoning. 📝 On the latest episode of our podcast, host Hannah Fry sits down with Neel Nanda to explore interpretability – the science of reverse engineering how neural networks learn and think. Timecodes: 00:00 Introduction 02:41 Motivation for interpretability research 04:01 Mechanistic interpretability 08:14 Chain of thought monitoring 18:14 Interpretability techniques 35:00 Auditing models for safety 48:53 What comes next for interpretability

A model’s chain of thought acts like a scratch pad, offering a window into its reasoning. 📝 On the latest episode of our podcast, host Hannah Fry sits down with Neel Nanda to explore interpretability – the science of reverse engineering how neural networks learn and think. Timecodes: 00:00 Introduction 02:41 Motivation for interpretability research 04:01 Mechanistic interpretability 08:14 Chain of thought monitoring 18:14 Interpretability techniques 35:00 Auditing models for safety 48:53 What comes next for interpretability

Google DeepMind

214,561 Aufrufe • vor 2 Monaten

How do we know what machines know? How can we understand the new, emerging behavior of machines? In this conversation between Been Kim from Google DeepMind and the BuzzRobot community, we dug into the challenges of interpretability and explored how humans and machines can better understand each other. Watch the full discussion on our YouTube channel. Thank you, Been, for taking the time to talk to us! Timestamps: 0:00 Intro 0:18 AI teaches grandmasters chess 05:30 AI neologisms 10:17 Extracting knowledge from machines 13:43 Interpretability research 17:19 Are we keeping up with AI progress 18:34 New AI related terms 20:26 The right direction for interpretability 23:35 AI lying 27:49 Conseptual maps 30:30 AI researchers bias 33:33 Generalizing AI teaching humans 35:11 Is AI sentient 36:04 Does AI has concepts 41:53 Progress in machine understanding

How do we know what machines know? How can we understand the new, emerging behavior of machines? In this conversation between Been Kim from Google DeepMind and the BuzzRobot community, we dug into the challenges of interpretability and explored how humans and machines can better understand each other. Watch the full discussion on our YouTube channel. Thank you, Been, for taking the time to talk to us! Timestamps: 0:00 Intro 0:18 AI teaches grandmasters chess 05:30 AI neologisms 10:17 Extracting knowledge from machines 13:43 Interpretability research 17:19 Are we keeping up with AI progress 18:34 New AI related terms 20:26 The right direction for interpretability 23:35 AI lying 27:49 Conseptual maps 30:30 AI researchers bias 33:33 Generalizing AI teaching humans 35:11 Is AI sentient 36:04 Does AI has concepts 41:53 Progress in machine understanding

Sophia

13,431 Aufrufe • vor 11 Monaten

From Palantir and Two Sigma to turning mechanistic interpretability into a production-grade platform, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are building Goodfire on the belief that “understanding what’s inside the model” is the path to safer, more controllable AI and Goodfire just raised a $150M Series B at a $1.25B valuation to scale that bet. We sat down with Mark and Myra to unpack what actionable interpretability looks like beyond demos: why SAE features can lose to raw-activation probes on real detection tasks, how they deployed token-level PII monitoring with Rakuten across English + Japanese without training on real customer data, and what it takes to do real-time steering on a trillion-parameter model (including the infamous Gen-Z feature). We also dig into hallucination detection, surgical post-training edits for bias/unlearning, the emerging equivalence between activation steering and in-context learning, and their push to bring interpretability into training so model development isn’t just “data in, weights out.” Goodfire mark bissell Myra Deng

From Palantir and Two Sigma to turning mechanistic interpretability into a production-grade platform, Mark Bissell (Member of Technical Staff) and Myra Deng (Head of Product) are building Goodfire on the belief that “understanding what’s inside the model” is the path to safer, more controllable AI and Goodfire just raised a $150M Series B at a $1.25B valuation to scale that bet. We sat down with Mark and Myra to unpack what actionable interpretability looks like beyond demos: why SAE features can lose to raw-activation probes on real detection tasks, how they deployed token-level PII monitoring with Rakuten across English + Japanese without training on real customer data, and what it takes to do real-time steering on a trillion-parameter model (including the infamous Gen-Z feature). We also dig into hallucination detection, surgical post-training edits for bias/unlearning, the emerging equivalence between activation steering and in-context learning, and their push to bring interpretability into training so model development isn’t just “data in, weights out.” Goodfire mark bissell Myra Deng

Latent.Space

19,389 Aufrufe • vor 7 Monaten

Neel Nanda is leading a Google DeepMind research team at 26. He and I discuss: • How that happened • “If your safety work doesn't advance capabilities, it's probably bad safety work” • Should people work at the safest or most reckless AI company? • An AI PhD – with these timelines?! • How to best operate in a big frontier AI company • Neel's distinctive uses of LLMs and which cold emails he answers • A common reasoning error in AI alignment • Why he (Neel Nanda) refuses to share his p(doom) This is part 2 of our conversation, part 1 was a comprehensive update on his research area: mechanistic interpretability, which I'll link below. Links to this episode of the 80,000 Hours Podcast below — enjoy!

Neel Nanda is leading a Google DeepMind research team at 26. He and I discuss: • How that happened • “If your safety work doesn't advance capabilities, it's probably bad safety work” • Should people work at the safest or most reckless AI company? • An AI PhD – with these timelines?! • How to best operate in a big frontier AI company • Neel's distinctive uses of LLMs and which cold emails he answers • A common reasoning error in AI alignment • Why he (Neel Nanda) refuses to share his p(doom) This is part 2 of our conversation, part 1 was a comprehensive update on his research area: mechanistic interpretability, which I'll link below. Links to this episode of the 80,000 Hours Podcast below — enjoy!

Rob Wiblin

111,865 Aufrufe • vor 1 Jahr

🚨 New Research LLMs are trained only on text... Yet their internal representations progressively organize in ways that resemble human perceptual geometry across different domains (like color, pitch, emotion and taste), with the structures peaking in intermediate layers before attenuating in deeper representations. 🥳 Accepted at ICML Mechanistic Interpretability Workshop 2026

🚨 New Research LLMs are trained only on text... Yet their internal representations progressively organize in ways that resemble human perceptual geometry across different domains (like color, pitch, emotion and taste), with the structures peaking in intermediate layers before attenuating in deeper representations. 🥳 Accepted at ICML Mechanistic Interpretability Workshop 2026

Lossfunk

63,914 Aufrufe • vor 2 Monaten

Can LLMs invent better ways to train LLMs? At Sakana AI, we’re pioneering AI-driven methods to automate AI research and discovery. We’re excited to release DiscoPOP: a new SOTA preference optimization algorithm that was discovered and written by an LLM! Our method leverages LLMs to propose and implement new preference optimization algorithms. We then train models with those algorithms and evaluate their performance, providing feedback to the LLM. By repeating this process for multiple generations in an evolutionary loop, the LLM discovers many highly-performant and novel preference optimization objectives! Paper: GitHub: Model: We proudly collaborated with the University of Oxford (Foerster Lab for AI Research (now part of BOLD)) and Cambridge University (Mihaela van der Schaar) on this groundbreaking project. Looking ahead, we envision a future where AI-driven research reduces the need for extensive human intervention and computational resources. This will accelerate scientific discoveries and innovation, pushing the boundaries of what AI can achieve.

Can LLMs invent better ways to train LLMs? At Sakana AI, we’re pioneering AI-driven methods to automate AI research and discovery. We’re excited to release DiscoPOP: a new SOTA preference optimization algorithm that was discovered and written by an LLM! Our method leverages LLMs to propose and implement new preference optimization algorithms. We then train models with those algorithms and evaluate their performance, providing feedback to the LLM. By repeating this process for multiple generations in an evolutionary loop, the LLM discovers many highly-performant and novel preference optimization objectives! Paper: GitHub: Model: We proudly collaborated with the University of Oxford (Foerster Lab for AI Research (now part of BOLD)) and Cambridge University (Mihaela van der Schaar) on this groundbreaking project. Looking ahead, we envision a future where AI-driven research reduces the need for extensive human intervention and computational resources. This will accelerate scientific discoveries and innovation, pushing the boundaries of what AI can achieve.

Sakana AI

556,037 Aufrufe • vor 2 Jahren

💬 “We're trying to make the most open approach to AI development that the world has ever seen.” 💡 Openness isn’t optional — it’s essential for the future of AI. That’s why Nemotron was built: to set new standards for speed, transparency, and versatility in enterprise AI. Hear Bryan Catanzaro, VP of Applied Deep Learning Research at NVIDIA, share how Nemotron gives developers the tools to build trusted, customizable AI for modern organizations. Watch the conversation 👇

💬 “We're trying to make the most open approach to AI development that the world has ever seen.” 💡 Openness isn’t optional — it’s essential for the future of AI. That’s why Nemotron was built: to set new standards for speed, transparency, and versatility in enterprise AI. Hear Bryan Catanzaro, VP of Applied Deep Learning Research at NVIDIA, share how Nemotron gives developers the tools to build trusted, customizable AI for modern organizations. Watch the conversation 👇

NVIDIA

52,641 Aufrufe • vor 1 Jahr

AI isn’t just about models, it’s about agents. Mainfunc’s Genspark platform, powered by Microsoft Azure and eight leading LLMs, is now evolving with Agent 365 and bringing agentic workflows to life. The result? Smarter automation, faster decisions, and real impact for knowledge workers. #MSIgnite

AI isn’t just about models, it’s about agents. Mainfunc’s Genspark platform, powered by Microsoft Azure and eight leading LLMs, is now evolving with Agent 365 and bringing agentic workflows to life. The result? Smarter automation, faster decisions, and real impact for knowledge workers. #MSIgnite

Microsoft

85,219 Aufrufe • vor 10 Monaten

A new way to browse is about to come online. All your Web3 activity, powered by AI. ARC Rewards just got upgraded. Unlock your share of the new $1M pool ✨

A new way to browse is about to come online. All your Web3 activity, powered by AI. ARC Rewards just got upgraded. Unlock your share of the new $1M pool ✨

ARC Terminal

139,013 Aufrufe • vor 8 Monaten

New In Person Podcast with Mike McCormick of Halcyon Ventures 📽️ Mike left a $1.5B fund to start Halcyon and has given career grants to help kickstart leading AI companies like Goodfire and Artificial Intelligence Underwriting Company He has deep takes on China vs US, Safety vs Accelerationism and AI.

New In Person Podcast with Mike McCormick of Halcyon Ventures 📽️ Mike left a $1.5B fund to start Halcyon and has given career grants to help kickstart leading AI companies like Goodfire and Artificial Intelligence Underwriting Company He has deep takes on China vs US, Safety vs Accelerationism and AI.

Tommy

13,391 Aufrufe • vor 11 Monaten

“AI gives you superpowers, but it just also gives you new ways to be annoying on the internet.” Today, Substack launched a new AI-detection feature with Pangram that lets readers scan posts, notes, replies, and comments for an estimate of how much AI was used. Founder and CEO Chris Best says AI itself isn’t the problem. He loves seeing creators use it to translate their work, improve English-language prose, conduct deep research, and sharpen original analysis. But he doesn’t want Substack flooded with undifferentiated content anyone could generate from the same prompt. As Best wrote this morning: “We don’t want to wait until your Substack app turns into LinkedIn before we start to learn and make progress.”

“AI gives you superpowers, but it just also gives you new ways to be annoying on the internet.” Today, Substack launched a new AI-detection feature with Pangram that lets readers scan posts, notes, replies, and comments for an estimate of how much AI was used. Founder and CEO Chris Best says AI itself isn’t the problem. He loves seeing creators use it to translate their work, improve English-language prose, conduct deep research, and sharpen original analysis. But he doesn’t want Substack flooded with undifferentiated content anyone could generate from the same prompt. As Best wrote this morning: “We don’t want to wait until your Substack app turns into LinkedIn before we start to learn and make progress.”

TBPN

33,384 Aufrufe • vor 2 Monaten

We finally have a product that helps with frontier AI research. Frontier LLMs are a) nerfed at AI research b) lack immediate GPU availability and c) lack access to the right skills and research material. "Teach me how vision models see" Goodfire Silico takes a vision transformer (ViT), loads ImageNet images, and for each head, visualizes the 196x196 matrix by creating animated arcs across patches. Haven't found another model or product that can do this level of comprehensive AI research, much less oneshot it.

We finally have a product that helps with frontier AI research. Frontier LLMs are a) nerfed at AI research b) lack immediate GPU availability and c) lack access to the right skills and research material. "Teach me how vision models see" Goodfire Silico takes a vision transformer (ViT), loads ImageNet images, and for each head, visualizes the 196x196 matrix by creating animated arcs across patches. Haven't found another model or product that can do this level of comprehensive AI research, much less oneshot it.

Deedy

74,579 Aufrufe • vor 1 Monat

Thank you for your unwavering support of Kling AI! We're excited to unveil a transformative evolution — an elevated new experience in visuals and functionalities, guided by our philosophy: "Kling AI, From Vision to Screen."🎉 【Visuals】 With a minimalistic logo, natural typography, and a black & white design, we eliminate distractions — putting you and your creations in the spotlight. This refined brand identity embodies clarity and the limitless potential of AI-powered storytelling. 【Functionalities】 This update also comes with exciting new features you’ve been waiting for! AI Sounds bring your content to life like never before, while a new file management system keeps your workflow smooth and organized! This evolution in our brand identity and the new features bring us one step closer to the future of AI creativity. We’re excited for what’s ahead, and grateful to have you with us on this journey!🥳💪 #KlingAI

Kling AI

74,330 Aufrufe • vor 1 Jahr

Crawling websites with the Realtime API 🗣️ Steer an AI agent with your voice as it crawls and browses websites in real time. Powered by OpenAI's new Realtime API, function calling, and . Check it out:

Crawling websites with the Realtime API 🗣️ Steer an AI agent with your voice as it crawls and browses websites in real time. Powered by OpenAI's new Realtime API, function calling, and . Check it out:

Eric Ciarla (hiring)

49,090 Aufrufe • vor 1 Jahr

Ultiverse is now powered by the latest in AI technology, creating endless gaming possibilities and deeper worlds. 🤝🏻 Get ready to interact with game content in new ways—engage with multiple characters in a dynamic, multi-modal simulation that evolves and responds to YOU! 🚀🫡

Ultiverse is now powered by the latest in AI technology, creating endless gaming possibilities and deeper worlds. 🤝🏻 Get ready to interact with game content in new ways—engage with multiple characters in a dynamic, multi-modal simulation that evolves and responds to YOU! 🚀🫡

Ultiverse

27,774 Aufrufe • vor 1 Jahr