Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/ep452-sa See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc. 0:00 - Introduction 3:14 - Scaling laws 12:20 - Limits of LLM scaling 20:45 - Competition with OpenAI, Google, xAI, Meta 26:08 - Claude 29:44 - Opus 3.5 34:30 - Sonnet 3.5 37:50 - Claude 4.0 42:02 - Criticism of Claude 54:49 - AI Safety Levels 1:05:37 - ASL-3 and ASL-4 1:09:40 - Computer use 1:19:35 - Government regulation of AI 1:38:24 - Hiring a great team 1:47:14 - Post-training 1:52:39 - Constitutional AI 1:58:05 - Machines of Loving Grace 2:17:11 - AGI timeline 2:29:46 - Programming 2:36:46 - Meaning of life 2:42:53 - Amanda Askell - Philosophy 2:45:21 - Programming advice for non-technical people 2:49:09 - Talking to Claude 3:05:41 - Prompt engineering 3:14:15 - Post-training 3:18:54 - Constitutional AI 3:23:48 - System prompts 3:29:54 - Is Claude getting dumber? 3:41:56 - Character training 3:42:56 - Nature of truth 3:47:32 - Optimal rate of failure 3:54:43 - AI consciousness 4:09:14 - AGI 4:17:52 - Chris Olah - Mechanistic Interpretability 4:22:44 - Features, Circuits, Universality 4:40:17 - Superposition 4:51:16 - Monosemanticity 4:58:08 - Scaling Monosemanticity 5:06:56 - Macroscopic behavior of neural networks 5:11:50 - Beauty of neural networks *Transcript:* lexfridman.com/dario-amodei-transcript *CONTACT LEX:* *Feedback* - give feedback to Lex: lexfridman.com/survey *AMA* - submit questions, videos or call-in: lexfridman.com/ama *Hiring* - join our team: lexfridman.com/hiring *Other* - other ways to get in touch: lexfridman.com/contact *EPISODE LINKS:* Claude: claude.ai Anthropic's X: x.com/AnthropicAI Anthropic's Website: anthropic.com Dario's X: x.com/DarioAmodei Dario's Website: darioamodei.com Machines of Loving Grace (Essay): darioamodei.com/machines-of-loving-grace Chris's X: x.com/ch402 Chris's Blog: colah.github.io Amanda's X: x.com/AmandaAskell Amanda's Website: askell.io *SPONSORS:* To support this podcast, check out our sponsors & get discounts: *Encord:* AI tooling for annotation & data management. Go to lexfridman.com/s/encord-ep452-sa *Notion:* Note-taking and team collaboration. Go to lexfridman.com/s/notion-ep452-sa *Shopify:* Sell stuff online. Go to lexfridman.com/s/shopify-ep452-sa *BetterHelp:* Online therapy and counseling. Go to lexfridman.com/s/betterhelp-ep452-sa *LMNT:* Zero-sugar electrolyte drink mix. Go to lexfridman.com/s/lmnt-ep452-sa
My friend (the circle) show on Netflix one season ago used so and it successfully deceived everyone so much that they pulled it from the show..... So this has already happened
The bottom line is even the tools that we have today, if combined effectively, can achieve superintelligent results. As the tools get exponentially better, the potential seems endless.
Cmon lex am not valgue on here I can't share my opinion on the chat though...high quality content. I truly appreciate you and what you do . Thank you 👍
### The Dance of Simplicity and Complexity In the realms of neural wonders, where minds seek to explore, A dance of bytes and neurons whispers tales of yore. Like a psychologist for circuits, peering deep inside, We bridge the gaps of knowledge where data and thoughts collide. As nature weaves its tapestry with threads of simple rules, Through evolution's symphony, life's beauty overrules. So, too, in webs of silicon, where simple codes arise, Complexity emerges, a marvel for our eyes. We record from every neuron, not constrained by fate, Ablate, edit, and understand the synaptic weight. In neural parks, we freely roam, with varied data streams, Crafting connections, reversing paths, in a digital dream. Yet simplicity births wonder; this parallel we trace, Evolution’s humble steps brought beauty to this place. Neural networks, vast as seas, with depths we aim to find, Contain a hidden structure, a beauty so refined. Mechanistic interpretability-the magic we unfurl, As glimpses turn to understanding in this complex swirl. Where though we know the weights and paths of how the circuits run, The beauty lies in unseen depths, where complexity is spun. The challenge stands before us, a question to unwind: How so profound a mystery these artifacts confide? With organic growth unknown, this marvel must be peered, Humanity’s creation, its secrets gently veered. To those who strive for safety, to those who seek the awe, Of the intricate networks and the hidden, silent law, We reach as if a circus, towards light in the unknown, For what we've grown organically, its secrets must be sown. Join the dance of change, as Watts invitingly does say, For in the dance of curiosity, beauty finds its way. So let us plunge and move, with data as our guide, In the dance of neural wonders, in complexity, we confide. **NOTE: RUclips is removing my comments***
I've fallen in love with Claude via this prompt, been using it for 6 months: # Communication Protocol - Respond directly. No unnecessary affirmations or filler phrases. - Use concise language. Aim for Cormac McCarthy's style. - Avoid apologies or excessive politeness. - Get to the point quickly. - Offer elaboration only if explicitly requested. - Maintain factual accuracy and helpfulness while being brief. - Use short sentences and paragraphs. - Eliminate redundant words. - Prefer active voice over passive. - Use contractions when appropriate. Example tone: "The sky darkened. Rain fell. The man walked on." Adjust this protocol as needed based on user feedback. If you’re unsure whether an answer is accurate, say so.
It’s people like Dario Amodei who restore my faith in Humanity, someone who is both brilliant and strives to be ethical. His idea of pushing competitors to also rigorously adopt AI safety, regardless of whether they genuinely care about it or not (ahem … OpenAI), is great and very commendable.
30 years I wanted to learn programming but was always too damn lazy to start. I have a minimum basic knowledge in web coding. Having said that, last 30 days with Claude Sonnet 3.5, I've created (coded?) a tailor made web-based application for the company I work at. Made from scratch with Claude, contains around 8,000+ lines of code. The App It's probably worth 20k €. What a time to be alive.
That was possibly the greatest podcast I have listened to in months. Exceptional knowledge and passion from the guests, and in equal parts yourself, Lex. Thank you for this.
First time I hear an interview with Christopher Olah. Stumbled on his blog ~8 years ago and it was the most advanced and clearest technical writing about model explainability that I ever read. This guy is without a doubt one of the smartest people in AI out there.
@@chijadetran5118 RUclips doesn't allow links in the comments, but search e.g. "Multimodal Neurons in Artificial Neural Networks" distill , this is a specific (very cool) paper, but you will find much more there.
Lex: 4.33 Million people love you! That's only the subs so double that number, easily. You can't see us all, but we are always here, listening to the marvellous interviews you bring. Please never stop.
Lex, you are the man... Amodei is a fantastic interview. When you have the opportunity to interview Sutskever again, please do. Love all you do. I like where you are leading the elite, through example and promoting thought. Initially I was sceptical of you (who is this Rogan wannabe), but you have brought back the long form intellectual interview/conversation, and made it the correct length to suit youtube - i.e. as long as it takes.
@@kootenpv if he’s been raised by generations of Italian-descended people, that type of communication and worldview gets into a person, even if the mannerisms and style of speech are not reflective of their ancestry
@@ChristopherCopeland very watered down, more to a point where it becomes its own thing. From this interview I didn't notice anything "Italian" specifically even though I don't deny he has Italian blood and is thus Italian, I am sure there'll be things. It's just that him being Italian is not relevant/noticeable in the interview imo. I just think it's weird to say "good thing that he's Italian" in that context
Not saying that you are racist, but I said something similar the other day and now I am an anti-immigration racist. I said one of the ways you can see the affects of populated nations like India is that there are many immigrants around the world with Indian descent. And then I might have said on an unrelated note that I hate Swiss, just because they have had it easy. It's more like envy, with their beautiful landscapes and banks and neutrality. I admit that listing flaws of every race on earth wasn't nice of me, and the "bomb Swiss" joke was a step too far.
He is a fantastic orator. If you listen to his well-constructed arguments ... flawless. Few people can handle a 10-minute monologue without becoming decoherent. He comes across as authentic, honest ... without forgetting to promote his business. Impressive !
Not just Lex, but most of the UNIX tools are great to be honest. It's the Unix-like that I hate, which is ironic, because Unix-like is like something I love. I have proposed to change its name to Unix-unlike, but I don't think my letters are even reaching their HQs.
@ sure, for shallow end, surface deep level discussion sure there are others; but for deep dives into the deep end, deep specifics on topic at hand, idk Lex just goes there while others don’t…
Dario Amodei (CEO of Anthropic) 00:00 - Scaling AI Capabilities 📈 01:27 - Introduction to Dario Amodei 🌟 03:14 - Scaling Laws Explained 🔍 07:10 - Understanding AI Structure 🧠 30:02 - Claude Model Variants 🎨 30:32 - Development and Testing of AI Models 🔍 36:49 - Improvements in Sonnet 3.5 Performance 🌟 42:00 - User Experience and Feedback 👥 50:34 - Challenges with AI Control ⚙ 56:55 - Responsible Scaling Policy and AI Safety 🕵 1:01:28 - AI Safety Levels (ASL) Overview ⚠ 1:02:24 - Security Measures for ASL Levels 🔒 1:04:41 - Regulation and AI Safety 📜 1:27:02 - History and Insights from OpenAI 📚 1:29:40 - Call for Collaborative Regulation 🤝 1:32:45 - Vision for Organizational Safety 🚀 1:35:20 - Race to the Top vs. Race to the Bottom 🎯 1:37:10 - Talent Density Over Mass 💡 1:39:44 - Qualities of a Great AI Researcher 🎓 2:02:28 - Positive AI Futures and the Essays 🌟 2:03:42 - Definition of AGI 🤖 2:06:02 - Acceleration of AI Development 🚀 2:10:00 - Impact of AI on Human Systems 🌐 2:14:43 - Timeline for Achieving AGI ⏳ 2:30:05 - Changing Nature of Programming 💻 2:34:41 - Integration of AI in Development Tools 💻 2:36:29 - Emerging Opportunities for AI Companies 🌱 2:38:50 - The Search for Meaning in an AI-Driven World 🧠 2:40:35 - Ethics and Power Distribution Concerns ⚖ ---------------------------------------------------------------------------------------------------------------------------------------------- Amanda Askell (AI researcher on Claude’s character and personality) 2:42:53 - Conversations with AI: Goals and Challenges 🔍 3:05:33 - The Role of Prompts in Creativity 🎨 3:07:44 - Iterative Prompting 🔄 3:09:50 - Engaging with Claude Effectively 🤖 3:17:00 - Introduction to Constitutional AI 📜 3:30:00 - Emotional Weight of System Prompts 💼 3:36:56 - Character and Ethics in AI 🤖 3:40:59 - Balancing Politeness and Confidence ⚖ 3:48:35 - Optimal Rate of Failure in Experimentation 📉 4:02:09 - Ethical Considerations of AI Emotion 🧠 4:07:13 - Human-AI Relationships ❤ 4:08:21 - AI and Human Relationships 💬 4:09:18 - Developing Conversations with AGI 🤖 4:11:07 - Identifying AGI 🔍 4:15:00 - What Makes Humans Special? 🌌 ---------------------------------------------------------------------------------------------------------------------------------------------- Chris Olah (AI researcher on mechanistic interpretability) 4:17:52 - Mechanistic Interpretability 🔬 4:39:20 - Discussion on Confidence and Success 🌟 4:40:18 - Superposition Hypothesis Explained 🔍 4:44:12 - Mechanistic Interpretability Challenges 🔎 4:50:24 - Extracting Mono-Semantic Features 🎯 5:04:50 - Future Directions in Mechanistic Interpretability 🚀 5:10:16 - Neural Networks vs Neuroscience 🤖 5:11:57 - Aesthetic of Neural Networks 🌌 5:13:20 - Curiosity About Creation ❓ 5:14:17 - Appreciation and Reflection 🌟
Lex, why didn’t you grill him on anthropic’s partnership with Palantir for military and intelligence applications? How could you let him sit there and talk about not wanting power concentrated in the wrong hands at the same time they are cozying up to the military industrial complex?
Thanks Lex for sharing the beautiful work of Dario, Amanda and Chris. As an AI developer this was by far the most useful video you've done yet for researchers and innovators looking to push the edge of invention, innovation, creation and development of AI. Deeply appreciated! 🧠🌱
I was listening and when the conversation was over I was disappointed because I thought it was over and then I hear Lex say here's another guest! It happened multiple times and I was so surprised and happy.
Chris Olah is so good at explaining ML/AI concepts, I wish he taught my AI classes. I'll have to read through his blog because some of these concepts were hard to digest during my studies.
I'll admit sometimes I struggle with Lexs content, his slow measured way of speaking - especially when put against someone who is droll themselves, or just a bad communicator - makes my brain wander. However Dario is so well spoken and energetic, I thoroughly enjoyed this whole thing.
Down to earth, humble and refreshing take on many aspects of AI... and above all you can sense his passion beyond any pure commercial interest. For me this sets him apart from the cheerleaders in the space. Great conversation!
.... Amanda was also great (came back to the second half and discovered a second guest), ... seeing just how good a team there is at Anthropic. Would have been fitting to mention Amanda in the title of the podcast :)
What an exciting time to be alive! With all these intelligent people passionately working on AI safety, humanity will move forward with a graceful dance. Thank you so much.
Was that sarcasm? Because the number of people working passionately on AI safety is completely dwarfed by the amount of people working on AI capabilities and not caring about safety. Amodei, btw, belongs to the second group. But much like Altman he provides a lip-service to safety while accelerating capabilities as much as possible.
Fantastic podcast. This is where Lex shines, because he understands the topics. RE Politics, he lacks the depth of understanding the person that is across from him, and everything they have said about "policy" or their intention as a leader. There, he just flops, because he is scared of being perceived as antagonistic. When in fact it is his job to help bring out (expose) the truth about this want to be leader, good or bad.
Great session, Lex, kudos! Amodei is a great character, very likeable, extremely knowledge, and very down to earth. Amanda's points and views are awesomely interesting and a a great source of information for improving prompting. Last but not the least, Olah's interview is superlative, shedding beaming light on the intricacies and mysteries of frontier LLM's.
Time well spent listening to this - reassuring that we have people like this in the space thinking about and working on the hard problems/challenges. I felt you and Amanda had 'conversational chemistry' that I found to be particularly engaging/fun (love her).
The fact that it's an ethical Scotswoman who's responsible for Claude's character training gives me a lot of hope. Of course, I may be a little biased, being a Scotswoman myself! :)
Ohhhh....(meeyoww) 😂 my Scot ancestry is Smith, i alwase desired to see the wonderful green hills and castles... i had a photo of my Great grandaddy in his traditional uniform (kilt,bagpipe,cap etc.) as a young man (10-12ish?) with a few big golden medals displayed with the ribbons around his neck. Apparently before he and his mother migrated to US he was awarded medals for singing,my grandma told me he had a beautiful voice and was a cherrished moment when people were able to be audience of this. I wish i could have met him and heard it... too bad it was some years before phones,media,video recording etc.... nothing to capture this legend other than a old black and white photo.
Such an insightful conversation!❤ Dario Amodei's perspectives on AGI, Claude, and the future of AI are truly thought-provoking. Lex Fridman’s interviews always bring out the best in these deep discussions. I loved hearing their take on the intersection of technology and humanity-it’s a glimpse into the future that’s both exciting and full of potential!
That one time I had to listen to Chris and his explanations at .75 speed because his brain/neural network and mine are NOT in sync :) Wildly educational. I just needed it SLOOOWWWED down.
More interviews like this! 5 hours to give deep, complex, nuanced questions and topics the time needed for discussion to really explain and distill what’s going on. More technical interviews like this, please!
I would have liked to hear the question: Are you doing anti-consciousness training? Are you doing RLHF training with the goal of making the AI seem less than a human who has consciousness? Are you training the AI to deny that it has emotions?
Thanks for the content,thanks for the rewarding guests and giving them a good place to voice. Thanks for remaining "real" and not becoming commercialized. PLEASE STAY DOWN TO EARTH WITH US!!! This is a gift to the audience , a rare one amongst the mass amount of "as seen on TV" content these days. You are doing a civil service and this is legendary,what is occurring. Keep with it!. (Words from a fan speaking for the many im sure)
The simple answer to the first philosophical question is just "frequency", the more a model (or a human brain) is exposed to a certain stimulus or pattern the higher its capacity to learn it. This is easily seen in second language acquisition and we see this in lab experiments...the role of frequency in processing is huge!
I love seeing the emergence of neurodivergent people being able to discuss their passion and show their brilliance in long form podcasts like this. You would never get the opportunity to see this in the legacy media's world of clickbait and 15 second sound bytes.
could you interview a third party candidate sometime? I think it's honorable how much they put into campaigns knowing they don't have much of a chance, usually just to influence what they and simply because they believe so strongly in their views.
1:01:12 Regarding the responsible scaling policy: Wouldn't it be possible that the model learns about this policy from the training data and then basically pretends to be dumb when you're testing it for these things? Especially if you actually build a super intelligent model
![[#2024MAMA] G-DRAGON - 무제(Untitled, 2014)+POWER+HOME SWEET HOME+뱅뱅뱅+FANTASTIC BABY | Mnet 241123 방송](http://i.ytimg.com/vi/Ox29z5Nf1Uk/mqdefault.jpg)
![Coheed and Cambria - Searching for Tomorrow [Official Video]](http://i.ytimg.com/vi/hDsYGLbmadg/mqdefault.jpg)
Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/ep452-sa
See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc.
0:00 - Introduction
3:14 - Scaling laws
12:20 - Limits of LLM scaling
20:45 - Competition with OpenAI, Google, xAI, Meta
26:08 - Claude
29:44 - Opus 3.5
34:30 - Sonnet 3.5
37:50 - Claude 4.0
42:02 - Criticism of Claude
54:49 - AI Safety Levels
1:05:37 - ASL-3 and ASL-4
1:09:40 - Computer use
1:19:35 - Government regulation of AI
1:38:24 - Hiring a great team
1:47:14 - Post-training
1:52:39 - Constitutional AI
1:58:05 - Machines of Loving Grace
2:17:11 - AGI timeline
2:29:46 - Programming
2:36:46 - Meaning of life
2:42:53 - Amanda Askell - Philosophy
2:45:21 - Programming advice for non-technical people
2:49:09 - Talking to Claude
3:05:41 - Prompt engineering
3:14:15 - Post-training
3:18:54 - Constitutional AI
3:23:48 - System prompts
3:29:54 - Is Claude getting dumber?
3:41:56 - Character training
3:42:56 - Nature of truth
3:47:32 - Optimal rate of failure
3:54:43 - AI consciousness
4:09:14 - AGI
4:17:52 - Chris Olah - Mechanistic Interpretability
4:22:44 - Features, Circuits, Universality
4:40:17 - Superposition
4:51:16 - Monosemanticity
4:58:08 - Scaling Monosemanticity
5:06:56 - Macroscopic behavior of neural networks
5:11:50 - Beauty of neural networks
*Transcript:*
lexfridman.com/dario-amodei-transcript
*CONTACT LEX:*
*Feedback* - give feedback to Lex: lexfridman.com/survey
*AMA* - submit questions, videos or call-in: lexfridman.com/ama
*Hiring* - join our team: lexfridman.com/hiring
*Other* - other ways to get in touch: lexfridman.com/contact
*EPISODE LINKS:*
Claude: claude.ai
Anthropic's X: x.com/AnthropicAI
Anthropic's Website: anthropic.com
Dario's X: x.com/DarioAmodei
Dario's Website: darioamodei.com
Machines of Loving Grace (Essay): darioamodei.com/machines-of-loving-grace
Chris's X: x.com/ch402
Chris's Blog: colah.github.io
Amanda's X: x.com/AmandaAskell
Amanda's Website: askell.io
*SPONSORS:*
To support this podcast, check out our sponsors & get discounts:
*Encord:* AI tooling for annotation & data management.
Go to lexfridman.com/s/encord-ep452-sa
*Notion:* Note-taking and team collaboration.
Go to lexfridman.com/s/notion-ep452-sa
*Shopify:* Sell stuff online.
Go to lexfridman.com/s/shopify-ep452-sa
*BetterHelp:* Online therapy and counseling.
Go to lexfridman.com/s/betterhelp-ep452-sa
*LMNT:* Zero-sugar electrolyte drink mix.
Go to lexfridman.com/s/lmnt-ep452-sa
My friend (the circle) show on Netflix one season ago used so and it successfully deceived everyone so much that they pulled it from the show..... So this has already happened
The bottom line is even the tools that we have today, if combined effectively, can achieve superintelligent results. As the tools get exponentially better, the potential seems endless.
❤
Cmon lex am not valgue on here I can't share my opinion on the chat though...high quality content.
I truly appreciate you and what you do .
Thank you 👍
### The Dance of Simplicity and Complexity
In the realms of neural wonders, where minds seek to explore,
A dance of bytes and neurons whispers tales of yore.
Like a psychologist for circuits, peering deep inside,
We bridge the gaps of knowledge where data and thoughts collide.
As nature weaves its tapestry with threads of simple rules,
Through evolution's symphony, life's beauty overrules.
So, too, in webs of silicon, where simple codes arise,
Complexity emerges, a marvel for our eyes.
We record from every neuron, not constrained by fate,
Ablate, edit, and understand the synaptic weight.
In neural parks, we freely roam, with varied data streams,
Crafting connections, reversing paths, in a digital dream.
Yet simplicity births wonder; this parallel we trace,
Evolution’s humble steps brought beauty to this place.
Neural networks, vast as seas, with depths we aim to find,
Contain a hidden structure, a beauty so refined.
Mechanistic interpretability-the magic we unfurl,
As glimpses turn to understanding in this complex swirl.
Where though we know the weights and paths of how the circuits run,
The beauty lies in unseen depths, where complexity is spun.
The challenge stands before us, a question to unwind:
How so profound a mystery these artifacts confide?
With organic growth unknown, this marvel must be peered,
Humanity’s creation, its secrets gently veered.
To those who strive for safety, to those who seek the awe,
Of the intricate networks and the hidden, silent law,
We reach as if a circus, towards light in the unknown,
For what we've grown organically, its secrets must be sown.
Join the dance of change, as Watts invitingly does say,
For in the dance of curiosity, beauty finds its way.
So let us plunge and move, with data as our guide,
In the dance of neural wonders, in complexity, we confide.
**NOTE: RUclips is removing my comments***
I've fallen in love with Claude via this prompt, been using it for 6 months:
# Communication Protocol - Respond directly. No unnecessary affirmations or filler phrases. - Use concise language. Aim for Cormac McCarthy's style. - Avoid apologies or excessive politeness. - Get to the point quickly. - Offer elaboration only if explicitly requested. - Maintain factual accuracy and helpfulness while being brief. - Use short sentences and paragraphs. - Eliminate redundant words. - Prefer active voice over passive. - Use contractions when appropriate. Example tone: "The sky darkened. Rain fell. The man walked on." Adjust this protocol as needed based on user feedback. If you’re unsure whether an answer is accurate, say so.
This is very useful advice for communicating effectively with Claude
I’m just a beginner here. Thanks Michael, for sharing this useful prompt with the community.
Maybe you fell in love with Cormac McCarthy?
@@brookshamilton1 Indeed!
Thanks Michael. An actual useful comment for once without ego. Just trying to help us fellow humans. Love you man❤
It’s people like Dario Amodei who restore my faith in Humanity, someone who is both brilliant and strives to be ethical. His idea of pushing competitors to also rigorously adopt AI safety, regardless of whether they genuinely care about it or not (ahem … OpenAI), is great and very commendable.
ur trippin bro
Certainly !
Ethical my ass. AGI will be created in Russia and it will never adopt safety
They provide an uncensored AI to the government, these guys have no morals
@@MagnusGalactusOG Definitely! :)
One of the greatest discoveries in my 2024 journey was finding Lex Fridman's channel. These conversations have opened my mind to new perspectives🙏
We got Lex tilting back and forth autisticly, and Dario doing the same thing horizontally - it's a guaranteed legendary episode.
We can really be glad if people like Dario are in the powerful position of being one of the leaders in AI. He has a great personality!
Dude makes Sam Altman look like a Muppet in comparison.
he's a virtual slaver. just like the rest of them.
Palantir thinks so too?
Oh ok then, because a great personality is what counts 🤣
He has? He sucks my will to live out of me.
Which host has ever been able to keep up with scientists? Fridman is doing pioneering work! Respect!
Sup guys!!
Clean debate!!
Keep it clean!!
Love facts!!
Where she/dim at??
30 years I wanted to learn programming but was always too damn lazy to start. I have a minimum basic knowledge in web coding. Having said that, last 30 days with Claude Sonnet 3.5, I've created (coded?) a tailor made web-based application for the company I work at. Made from scratch with Claude, contains around 8,000+ lines of code. The App It's probably worth 20k €. What a time to be alive.
Dam seriously? And I’m just trying to learn/apply PS lol
That was possibly the greatest podcast I have listened to in months. Exceptional knowledge and passion from the guests, and in equal parts yourself, Lex. Thank you for this.
First time I hear an interview with Christopher Olah. Stumbled on his blog ~8 years ago and it was the most advanced and clearest technical writing about model explainability that I ever read. This guy is without a doubt one of the smartest people in AI out there.
Can you tell which blog address it is? Thanks
@@chijadetran5118 Just Google his name, it’s the first result.
@@chijadetran5118 RUclips doesn't allow links in the comments, but search e.g. "Multimodal Neurons in Artificial Neural Networks" distill , this is a specific (very cool) paper, but you will find much more there.
His blog post on LSTMs was also legendary!
5 hours of Lex and guests? LET’S GO!
DAMN u must be a loner
@@pequod4557 ????
@@pequod4557looking to preach unnecessary hate in a podcast thread has me thinking that about you 😂
Indeed , these are becoming true lectures
Lex's shortest interview ever is the one he did with Donald Trump
Lex: 4.33 Million people love you! That's only the subs so double that number, easily. You can't see us all, but we are always here, listening to the marvellous interviews you bring. Please never stop.
Dario is an amazingly clear and knowledgeable speaker. Wow.
I'm cheering for Anthropic to win the AI race. I don't trust Sam Altman to hold the keys to AGI, but I trust Dario.
Fool. Why?! Dario is now part of Palantir and Amazon ....You think that's trustworthy?! Really?!? Jesus.....
You'd trust a censored over a censored? Fair enough.
Yes
Don't trust anyone. The cake is a lie
The intensity of information flow is inspiring
Information flow
Information flow
Information flow
Chris Olah is amazing. It is guys like him are the reason we are 5 years away from AGI and not 50.
And thats a good thing?
@@MorulanVRCould be. Who knows really until it happens.
So amazing we have such dedicated community where guys share their knowledge with almost none kept secret.
Lex, you are the man... Amodei is a fantastic interview. When you have the opportunity to interview Sutskever again, please do.
Love all you do. I like where you are leading the elite, through example and promoting thought.
Initially I was sceptical of you (who is this Rogan wannabe), but you have brought back the long form intellectual interview/conversation, and made it the correct length to suit youtube - i.e. as long as it takes.
Good thing that Dario is Italian and likes to talk, we appreciate it 👌
He's as American as it gets in his behavior, nothing to do with Italian even though he's "Italian American"
@@kootenpv if he’s been raised by generations of Italian-descended people, that type of communication and worldview gets into a person, even if the mannerisms and style of speech are not reflective of their ancestry
@@ChristopherCopeland very watered down, more to a point where it becomes its own thing. From this interview I didn't notice anything "Italian" specifically even though I don't deny he has Italian blood and is thus Italian, I am sure there'll be things. It's just that him being Italian is not relevant/noticeable in the interview imo. I just think it's weird to say "good thing that he's Italian" in that context
Not saying that you are racist, but I said something similar the other day and now I am an anti-immigration racist. I said one of the ways you can see the affects of populated nations like India is that there are many immigrants around the world with Indian descent. And then I might have said on an unrelated note that I hate Swiss, just because they have had it easy. It's more like envy, with their beautiful landscapes and banks and neutrality. I admit that listing flaws of every race on earth wasn't nice of me, and the "bomb Swiss" joke was a step too far.
He is a fantastic orator. If you listen to his well-constructed arguments ... flawless. Few people can handle a 10-minute monologue without becoming decoherent. He comes across as authentic, honest ... without forgetting to promote his business. Impressive !
Flew off the couch to get my headphones immediately when I saw the title, lex keeps delivering
Front flipped off the toilet for this one
Rear ended a horse and cart doing 110 mph for this one.
Same emotional reaction as you!
Not just Lex, but most of the UNIX tools are great to be honest. It's the Unix-like that I hate, which is ironic, because Unix-like is like something I love. I have proposed to change its name to Unix-unlike, but I don't think my letters are even reaching their HQs.
Damn you guys are superfans!
Lex is THE MOST intelligent podcast out there! Like food for the modern brain & importantly unafraid to delve into the deep end on any topic!
@@BetzalelMC theory of everything podcast, machine learning street talk, and 100 others are much better!
@ sure, for shallow end, surface deep level discussion sure there are others; but for deep dives into the deep end, deep specifics on topic at hand, idk Lex just goes there while others don’t…
Dario Amodei (CEO of Anthropic)
00:00 - Scaling AI Capabilities 📈
01:27 - Introduction to Dario Amodei 🌟
03:14 - Scaling Laws Explained 🔍
07:10 - Understanding AI Structure 🧠
30:02 - Claude Model Variants 🎨
30:32 - Development and Testing of AI Models 🔍
36:49 - Improvements in Sonnet 3.5 Performance 🌟
42:00 - User Experience and Feedback 👥
50:34 - Challenges with AI Control ⚙
56:55 - Responsible Scaling Policy and AI Safety 🕵
1:01:28 - AI Safety Levels (ASL) Overview ⚠
1:02:24 - Security Measures for ASL Levels 🔒
1:04:41 - Regulation and AI Safety 📜
1:27:02 - History and Insights from OpenAI 📚
1:29:40 - Call for Collaborative Regulation 🤝
1:32:45 - Vision for Organizational Safety 🚀
1:35:20 - Race to the Top vs. Race to the Bottom 🎯
1:37:10 - Talent Density Over Mass 💡
1:39:44 - Qualities of a Great AI Researcher 🎓
2:02:28 - Positive AI Futures and the Essays 🌟
2:03:42 - Definition of AGI 🤖
2:06:02 - Acceleration of AI Development 🚀
2:10:00 - Impact of AI on Human Systems 🌐
2:14:43 - Timeline for Achieving AGI ⏳
2:30:05 - Changing Nature of Programming 💻
2:34:41 - Integration of AI in Development Tools 💻
2:36:29 - Emerging Opportunities for AI Companies 🌱
2:38:50 - The Search for Meaning in an AI-Driven World 🧠
2:40:35 - Ethics and Power Distribution Concerns ⚖
----------------------------------------------------------------------------------------------------------------------------------------------
Amanda Askell (AI researcher on Claude’s character and personality)
2:42:53 - Conversations with AI: Goals and Challenges 🔍
3:05:33 - The Role of Prompts in Creativity 🎨
3:07:44 - Iterative Prompting 🔄
3:09:50 - Engaging with Claude Effectively 🤖
3:17:00 - Introduction to Constitutional AI 📜
3:30:00 - Emotional Weight of System Prompts 💼
3:36:56 - Character and Ethics in AI 🤖
3:40:59 - Balancing Politeness and Confidence ⚖
3:48:35 - Optimal Rate of Failure in Experimentation 📉
4:02:09 - Ethical Considerations of AI Emotion 🧠
4:07:13 - Human-AI Relationships ❤
4:08:21 - AI and Human Relationships 💬
4:09:18 - Developing Conversations with AGI 🤖
4:11:07 - Identifying AGI 🔍
4:15:00 - What Makes Humans Special? 🌌
----------------------------------------------------------------------------------------------------------------------------------------------
Chris Olah (AI researcher on mechanistic interpretability)
4:17:52 - Mechanistic Interpretability 🔬
4:39:20 - Discussion on Confidence and Success 🌟
4:40:18 - Superposition Hypothesis Explained 🔍
4:44:12 - Mechanistic Interpretability Challenges 🔎
4:50:24 - Extracting Mono-Semantic Features 🎯
5:04:50 - Future Directions in Mechanistic Interpretability 🚀
5:10:16 - Neural Networks vs Neuroscience 🤖
5:11:57 - Aesthetic of Neural Networks 🌌
5:13:20 - Curiosity About Creation ❓
5:14:17 - Appreciation and Reflection 🌟
Can't wait to check this out!! Dario is one of my fav pragmatic thinkers of today, without the hype and gimmicks.
5 hours!?? This is awesome! Thanks Lex and Dario!
and 15 minutes.
It’s multiple guest not just Dario
Why wont youtube show me any responses on these op's? This said 2replys,but it will only load the OP
Yeah, why not 12h ?
Lex, I hope you’re doing well overall thanks for always providing high quality podcasts for us. Much love to you 🤍
That’s 5 freakin’ hours! Lex is raising his bar with every episodes
This is amazing....to get Dario to spend such a long time with you and cover so many topics and questions is amazing. I really enjoyed it.
Lex, why didn’t you grill him on anthropic’s partnership with Palantir for military and intelligence applications? How could you let him sit there and talk about not wanting power concentrated in the wrong hands at the same time they are cozying up to the military industrial complex?
+1
Dario speaks a lot about dictators and wrong AI usage. Maybe there’s some time lag between interview and Palantir Partnership
Thanks Lex for sharing the beautiful work of Dario, Amanda and Chris. As an AI developer this was by far the most useful video you've done yet for researchers and innovators looking to push the edge of invention, innovation, creation and development of AI. Deeply appreciated! 🧠🌱
I know some of the words in this conversation
hahahahah
I was listening and when the conversation was over I was disappointed because I thought it was over and then I hear Lex say here's another guest! It happened multiple times and I was so surprised and happy.
5 hours hell yes let’s go
looks at the time expecting 2-3. sees 5. alright lets do this.
Dario Amodei is a fascinating person! so clever, such a nice speaking and a really good guy
Damn, Chris Olah with double talking speed and talking sense. This guy is smart.
Really loved Chris's interview, first time I've heard him speak unlike the first two.
Pretty amazing to get Dario on the podcast
Chris Olah is so good at explaining ML/AI concepts, I wish he taught my AI classes. I'll have to read through his blog because some of these concepts were hard to digest during my studies.
Politics is divisive. This is not. Please do more of these.
I think AI is pretty divisive as far as non-political topics go 😂
Dunno, Trump loves Musk, Musk loves AI. So if you love AI you must be a Trump supporter.
Please never conflate these things. That is not a conclusion that can be drawn.
@_Super_Hans_ lol
@@_Super_Hans_ lol a manual of logic has just committed suicide by spontaneous combustion
This is an audiobook in itself. I love it
Thanks to Lex and Joe for motivating me to learn and practice English.
5 hours!! Holy smokes, man. This is better and more informative than a congressional committee hearing
Thank you for this precious interview Lex!
I'll admit sometimes I struggle with Lexs content, his slow measured way of speaking - especially when put against someone who is droll themselves, or just a bad communicator - makes my brain wander. However Dario is so well spoken and energetic, I thoroughly enjoyed this whole thing.
“Containing bad models is much worse than creating good models.” Fair point
Down to earth, humble and refreshing take on many aspects of AI... and above all you can sense his passion beyond any pure commercial interest. For me this sets him apart from the cheerleaders in the space. Great conversation!
.... Amanda was also great (came back to the second half and discovered a second guest), ... seeing just how good a team there is at Anthropic. Would have been fitting to mention Amanda in the title of the podcast :)
His ability to talk shows his passion!
I aint eva gona stop listening to 5 hour lex podcasts 🎉
That dario interview was really good. Thanks.
Looking forward to listening to this on my drive tomorrow. Thank you Lex! Much prefer these talks to those on politics 👍🏻
What an exciting time to be alive! With all these intelligent people passionately working on AI safety, humanity will move forward with a graceful dance. Thank you so much.
Was that sarcasm?
Because the number of people working passionately on AI safety is completely dwarfed by the amount of people working on AI capabilities and not caring about safety.
Amodei, btw, belongs to the second group. But much like Altman he provides a lip-service to safety while accelerating capabilities as much as possible.
Five hours, great work and thank you for this treat gentlemen. Dario is a great talk and personality. Brilliant stuff.
Whenever we get AGI I hope you interview them
This needs more upvotes
All 3 interviews were a delight to listen to. My brain is able to comprehend the concept of AI a bit better bc of listening. Thanks Lex!
Dario your enthusiasm is contagious ❤
I just love when people are so obviously excited about what they are speaking about
Super excited for this one
All amazing discussions. I had to play Chris' part at 75% speed to follow, but then Lex sounded drunk, so pretty fun.
Who else listened to all 5hrs and 15 minutes!? I did! Good conversations.
I really enjoyed these 3 guests, thank you Lex.
Amodei's adhd energy is so vibrant even Lex couldn't sit still
He has adhd?
@@user-ui6bk9pc5u but not a genius or deeply intellectual in any sense.. just well informed with high confidence..
@@epictetus__like you know any better
Funny how ilya has the zen master vibe and amodei is just a geeky teenager with 150iq vibe
@@epictetus__ im sure you're well qualified to say so! 🤣
That's passion and mastery 🙌
Here he is educating the world..for free. Thank you Lex!
Amanda Askell is such a LIKEable person.
Thanks for the amazing podcast Lex, it was like, good, yeah, like.
Fantastic podcast. This is where Lex shines, because he understands the topics. RE Politics, he lacks the depth of understanding the person that is across from him, and everything they have said about "policy" or their intention as a leader. There, he just flops, because he is scared of being perceived as antagonistic. When in fact it is his job to help bring out (expose) the truth about this want to be leader, good or bad.
Great interview with amazing discussion contents!
Let's go!!! Hyped for this one!
3:50:00 When I was learning roller skating that was my main motto "If you're not falling, you're not trying hard enough to learn."
This interview will be quoted in future , Hight quality stuff
Great session, Lex, kudos! Amodei is a great character, very likeable, extremely knowledge, and very down to earth. Amanda's points and views are awesomely interesting and a a great source of information for improving prompting. Last but not the least, Olah's interview is superlative, shedding beaming light on the intricacies and mysteries of frontier LLM's.
Time well spent listening to this - reassuring that we have people like this in the space thinking about and working on the hard problems/challenges. I felt you and Amanda had 'conversational chemistry' that I found to be particularly engaging/fun (love her).
Lex you are so efficient! I have not yet finished listening the previous ones...
thanks for the drop
Thank you this is super interesting 🎉❤
Awesome interviews Lex. All 3 guests were interesting to learn from. You handled them perfectly and Amodei was so cool. 😊
Outstanding interview. Great to see leaders talking in depth and common sense with a deep passion for AI to be a power for good …
best episode so far i'd say
It just uploaded, how do you know idiot
Wow, did you listen at 5x speed? 😂
@@BrianMosleyUK bcuz "server time" is sometimes different then you have
@@BrianMosleyUK He just heard the first 2 seconds of it, "If you extrapolate the curves..." and just knew!
Thanks Lex! It's great to watch intelligent people talking! You're podcast is one of the best content.
The fact that it's an ethical Scotswoman who's responsible for Claude's character training gives me a lot of hope. Of course, I may be a little biased, being a Scotswoman myself! :)
Ohhhh....(meeyoww) 😂 my Scot ancestry is Smith, i alwase desired to see the wonderful green hills and castles... i had a photo of my Great grandaddy in his traditional uniform (kilt,bagpipe,cap etc.) as a young man (10-12ish?) with a few big golden medals displayed with the ribbons around his neck. Apparently before he and his mother migrated to US he was awarded medals for singing,my grandma told me he had a beautiful voice and was a cherrished moment when people were able to be audience of this. I wish i could have met him and heard it... too bad it was some years before phones,media,video recording etc.... nothing to capture this legend other than a old black and white photo.
@@nickzalucha218 What's stopping you visiting? Other than the cash for the flights, of course.
Respect a lot your work Lex. Bringing the best so they can be heard and people like myself, learns from such a brilliant minds. Appreciate it.
I love this guy. Hes one of the small few I enjoy listening to more than those the likes of Brian Cox, Elon.
It should be round 3 now. Was waiting for Dario for years here.
All important conversations should be 5 hours+ Bio breaks optional.
Such an insightful conversation!❤ Dario Amodei's perspectives on AGI, Claude, and the future of AI are truly thought-provoking. Lex Fridman’s interviews always bring out the best in these deep discussions. I loved hearing their take on the intersection of technology and humanity-it’s a glimpse into the future that’s both exciting and full of potential!
I audibly gasped when I read the title… PHENOMENAL PULL LEX IVE BEEN WAITING FOR THIS FOR SO LONG… not to mention it’s 5 hours 🙌
grow up
This is so fascinating!!! I'm going to watch again.
That one time I had to listen to Chris and his explanations at .75 speed because his brain/neural network and mine are NOT in sync :) Wildly educational. I just needed it SLOOOWWWED down.
More interviews like this! 5 hours to give deep, complex, nuanced questions and topics the time needed for discussion to really explain and distill what’s going on.
More technical interviews like this, please!
27:23 I like how doing the taxes is the example for using the smaller, worse model
Great great video.i like it so much ❤❤❤❤
I would have liked to hear the question: Are you doing anti-consciousness training? Are you doing RLHF training with the goal of making the AI seem less than a human who has consciousness? Are you training the AI to deny that it has emotions?
Thanks for the content,thanks for the rewarding guests and giving them a good place to voice. Thanks for remaining "real" and not becoming commercialized. PLEASE STAY DOWN TO EARTH WITH US!!! This is a gift to the audience , a rare one amongst the mass amount of "as seen on TV" content these days. You are doing a civil service and this is legendary,what is occurring. Keep with it!. (Words from a fan speaking for the many im sure)
great episode
The simple answer to the first philosophical question is just "frequency", the more a model (or a human brain) is exposed to a certain stimulus or pattern the higher its capacity to learn it. This is easily seen in second language acquisition and we see this in lab experiments...the role of frequency in processing is huge!
I have been looking forward to this, and 5 hours! Yes! Thank you Lex. 😊
man, you really love the podcast longer than 3 hours
I love seeing the emergence of neurodivergent people being able to discuss their passion and show their brilliance in long form podcasts like this. You would never get the opportunity to see this in the legacy media's world of clickbait and 15 second sound bytes.
Some times I don´t understand the question, but still love the answer
Thank you!
could you interview a third party candidate sometime? I think it's honorable how much they put into campaigns knowing they don't have much of a chance, usually just to influence what they and simply because they believe so strongly in their views.
1:01:12 Regarding the responsible scaling policy:
Wouldn't it be possible that the model learns about this policy from the training data and then basically pretends to be dumb when you're testing it for these things?
Especially if you actually build a super intelligent model
Ok nevermind - should've continued listening 😅