Lex Fridman
Sam Altman is the CEO of OpenAI, the company behind GPT-4, ChatGPT, DALL-E, Codex, and many other state-of-the-art AI technologies. Please support this podcast by checking out our sponsors:
– NetSuite: http://netsuite.com/lex to get free product tour
– SimpliSafe: https://simplisafe.com/lex
– ExpressVPN: https://expressvpn.com/lexpod to get 3 months free
EPISODE LINKS:
Sam’s Twitter: https://twitter.com/sama
OpenAI’s Twitter: https://twitter.com/OpenAI
OpenAI’s Website: https://openai.com
GPT-4 Website: https://openai.com/research/gpt-4
PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41
OUTLINE:
0:00 – Introduction
4:36 – GPT-4
16:02 – Political bias
23:03 – AI safety
43:43 – Neural network size
47:36 – AGI
1:09:05 – Fear
1:11:14 – Competition
1:13:33 – From non-profit to capped-profit
1:16:54 – Power
1:22:06 – Elon Musk
1:30:32 – Political pressure
1:48:46 – Truth and misinformation
2:01:09 – Microsoft
2:05:09 – SVB bank collapse
2:10:00 – Anthropomorphism
2:14:03 – Future applications
2:17:54 – Advice for young people
2:20:33 – Meaning of life
SOCIAL:
– Twitter: https://twitter.com/lexfridman
– LinkedIn: https://www.linkedin.com/in/lexfridman
– Facebook: https://www.facebook.com/lexfridman
– Instagram: https://www.instagram.com/lexfridman
– Medium: https://medium.com/@lexfridman
– Reddit: https://reddit.com/r/lexfridman
– Support on Patreon: https://www.patreon.com/lexfridman
Here are the timestamps. Please check out our sponsors to support this podcast.
0:00 – Introduction & sponsor mentions:
– NetSuite: http://netsuite.com/lex to get free product tour
– SimpliSafe: https://simplisafe.com/lex
– ExpressVPN: https://expressvpn.com/lexpod to get 3 months free
4:36 – GPT-4
16:02 – Political bias
23:03 – AI safety
43:43 – Neural network size
47:36 – AGI
1:09:05 – Fear
1:11:14 – Competition
1:13:33 – From non-profit to capped-profit
1:16:54 – Power
1:22:06 – Elon Musk
1:30:32 – Political pressure
1:48:46 – Truth and misinformation
2:01:09 – Microsoft
2:05:09 – SVB bank collapse
2:10:00 – Anthropomorphism
2:14:03 – Future applications
2:17:54 – Advice for young people
2:20:33 – Meaning of life
In the Trump vs Biden Jordan Peterson question ….I wonder if the fact the internet has more data or just in general there is more data biased towards "praising" Biden and "destroying" Trump affected the outcome ….. in general if a human was to search info about both, I think the biased info online would make you think Trump is a terrible person, while Biden a "good" person …. so the data sources and the amount of biased data can skew the outcomes …….. this leads to the question Lex asked "do I look fat in this dress" , sure in the context of being human, we can lie and say no , however do we want this models to think lying on occasions is ok. Or do we want them to be accurate and truthful , based on factual information ? …I would think they need to be accurate and truthful , as lying even if they are "white" lies , they are wrong …
This comment is for Sam Altman. In response to what would you ask a super intelligence: I mean you no disrespect but do you really think that you (or any other human) would even be able to comprehend an answer to a question that requires knowledge that we don’t yet know, let alone understand? What you are suggesting is an actual end point to knowledge. That is not a utopia. That is a wasteland. What makes life meaningful is a feeling of achievement, a sense of progress over our struggles and shared understanding that we all go through struggles and thus helps build a collective compassion about the human condition (what it actually means to be human) and this includes acceptance of endings (as heart-breaking as they are) and the little kindnesses that mean so much but require so little, like a smile from a stranger. Without minimising that real suffering exists, you can change your life by changing your attitude. You don’t need technology for that. In response to AI, people often choose to evoke Dr Frankenstein. I think it’s because Mary Shelley was almost pre-eminent in her wisdom: Dr Frankenstein alone creates the creature, he sees what it is, lying there inanimate and yet he continues because it’s all about progress. It is only when he sees it awake that he runs away, horrified. We know the rest.
Is this person a teenager? He sounds like one.
Pareto Rule re collective good.
Epic
I don't want to hurt Sam with this comment, but the fact that he has transformed his nose does not generate the level of confidence that someone who has the ability to modify human condition, create a disease, or cause mass destruction should inspire. I consider that altering one's face is a sufficiently violent and irrational trait that someone with such responsibility should not possess. This type of behavior suggests to me that they may lack the mental strength, objectivity, and neutrality required to understand nature in its purest form, without wanting to alter its most fundamental aspects due to perceived flaws, whether they are aesthetic, behavioral, etc., that should be corrected with force, just like removing parts of their nose cartilage, with absolute certainty
I think this is the interview of the century in that it informs us clearly what the game plan is and the time line of AI progress however, it is delusional to think that anyone can control this no matter what your safeguards are and realistically it is common sense greed and domination of the world will drive this to its frightening conclusion , it is amazing we to me personally we haven’t destroyed ourselves already given all the rhetoric regarding the possibilities for good that discussed because of mentality of what is desirable to humanity who have no choice. Line up for what happened at the Enlightenment, drenched in blood!
🎯 Key Takeaways for quick navigation:
00:40 🧠 Self-supervised learning is the "dark matter of intelligence," seeking to replicate human-like learning from observation, building world models and common sense without explicit task-based feedback.
03:28 📹 Self-supervised learning aims to predict the future or fill in gaps in information, such as predicting what happens next in a video, using large amounts of data for learning.
08:22 🤖 The idea of "filling in the blanks" through self-supervised learning, whether for language or vision, is considered one of the best approaches to achieving machine intelligence.
10:57 🗣️ Self-supervised learning has been highly successful in natural language processing but faces challenges in the domain of visual learning, particularly with videos.
17:27 🎲 Intelligence may ultimately be a form of statistics, with self-supervised learning being a method for learning probabilistic models of the world, including causal relationships.
23:16 🧠 AI needs to learn predictive models of the world that handle uncertainty and real-world complexity, not just deterministic trajectories.
25:35 🤖 Model Predictive Control, a form of reasoning and planning, can be implemented with gradient-based methods and is used in robotics.
27:16 🧐 The challenge in AI is getting machines to learn predictive models for diverse aspects of the world, including human behavior and complex phenomena.
32:07 🤔 Logic-based reasoning might not be compatible with efficient learning; humans use reasoning by simulation and analogy to build models of the world.
38:29 🐱 The fundamental aspects of intelligence can be self-learned, driven by hardwired objectives, and are not necessarily linked to complex social behaviors or language.
46:26 📊 Data augmentation, distorting training images, is a common practice in machine learning since the 1990s to improve supervised classifier training.
47:07 🧠 Self-supervised learning for vision often involves using data augmentation methods to create two views of the same object and training networks to make their representations as close as possible.
48:17 🔄 Avoiding representation collapse in self-supervised learning is a challenge; one approach is using negative examples in contrastive learning.
51:20 🤖 Non-contrastive methods aim to maximize mutual information between the outputs of two systems, providing another way to ensure diverse representations.
57:49 💡 Yann LeCun believes that self-supervised learning from vision data is essential for building artificial intelligence with common sense and that text alone lacks the necessary information about the physical world.
01:09:52 🧠 Yann LeCun discusses the importance of learning and adaptability in artificial intelligence, emphasizing that many fundamental aspects of human cognition are learned rather than hardwired.
01:16:13 🌐 Yann LeCun delves into the concept of active learning and its role in forming causal models of the world, suggesting that interaction and action are essential for learning about the environment.
01:22:32 💭 LeCun explores the idea that consciousness might be an executive module for configuring the world model in the brain for specific situations, with intrinsic drives and motivations being hardwired.
01:24:12 🤔 Yann LeCun challenges the notion that certain basic concepts, like object permanence and three-dimensional space, are hardwired in the human mind, suggesting that they are learned.
01:29:47 💀 The discussion touches on the human awareness of mortality, highlighting that the deep realization of one's mortality often occurs in childhood, leading to various coping mechanisms and beliefs about the afterlife.
01:32:03 🤖 Yann LeCun believes that worrying about what happens after death is unnecessary, as life is the only one we have.
01:33:57 🛡️ Building intelligent artifacts can help us understand human and biological intelligence better, similar to how building airplanes helped us understand aerodynamics.
01:37:26 💼 Yann LeCun suggests that intelligent autonomous machines may have emotions like fear and elation, which are integral to their intelligence.
01:38:51 🧠 Yann LeCun believes that AI systems can exhibit true intelligence and consciousness, and the Chinese Room argument, which suggests otherwise, is "ridiculous."
01:50:26 🏢 Facebook AI Research (FAIR) has played a significant role in advancing AI research, open-source tools like PyTorch, and indirectly impacting Facebook (Meta) services.
01:54:43 📌 Yann LeCun discusses the possibility of Facebook (Meta) changing its name and suggests keeping the name "Fair" and changing the meaning of the "F" to "Fundamental."
01:55:40 🧤 Yann LeCun talks about the Reality Lab division at Meta (formerly Facebook), which focuses on AR, VR, telepresence, and related technologies.
01:56:23 🌐 Yann LeCun discusses the concept of the metaverse and its potential as the next step in the evolution of the internet, emphasizing the importance of making the experience compelling.
01:58:16 📰 Yann LeCun defends Facebook (Meta) against negative portrayals in the media, arguing that many claims about the platform's negative effects lack empirical support and that it's essential to consider both positive and negative aspects.
02:05:24 📚 Yann LeCun mentions that Mark Zuckerberg and others at Facebook (Meta) have shown a strong interest in AI and discuss the evolving roles of top leaders within the company.
02:16:25 📝 Yann LeCun emphasizes the limitations of focusing solely on benchmark-driven research, encouraging exploration of new concepts and ideas in AI.
02:17:35 🌐 Yann LeCun discusses the potential future of academic paper review systems, advocating for open review processes and decentralized evaluation entities to improve efficiency.
02:19:42 💼 Implementing reputation systems for reviewing entities and creating incentives for reviewers could enhance the quality of academic peer review.
02:22:03 🌌 Yann LeCun explores the mysteries of complexity in the universe, highlighting its role in physics, biology, and the understanding of machine learning systems.
02:36:18 🚀 Yann LeCun's advice to young individuals aspiring to make a significant impact in the field of intelligence: Focus on big questions, learn foundational concepts, and prioritize knowledge with lasting relevance.
02:39:09 🌱 Yann LeCun discusses the Open Catalyst project, which aims to use deep learning to design chemical compounds or materials for efficient hydrogen and oxygen separation to combat climate change.
02:40:05 ☀️ Exploring the potential of using deep learning to control super hot plasma in fusion reactors, which could revolutionize energy production and address climate change.
02:40:45 🔋 Leveraging AI to design new materials for improved batteries, faster electronics, lighter vehicles, and better fuel cells, potentially transforming various industries.
02:41:26 🩺 AI's applications extend to medicine and biology, including protein folding for drug design and addressing complex phenomena like superconductivity in materials.
02:42:38 🧪 Machine learning can help understand and predict emergent phenomena in materials, like superconductivity, that are challenging to describe through traditional reductionist methods.
Made with HARPA AI
Anyone who has noticed Altsman does not want us to see the name of the water he drinks so that we cannot juice up our poor brains..Amazing Bot Conversation no hard feelings.😂
maybe AI will help us figure out quantum computing
Mehh prolly but i just had a flash that the reason the alignment problemo is so hard the alignment problem is really a subset of the bigger issue
witches awareness ….uhh the alignment problem IS the Intelligence problem… but me sure the AI will help us wid dat jk
The problem with AI is the following: it is a great processing power and can create a huge inbalance in society. People with no access to these tools will get poor and exploited. This technology can result in a new type of socialism / authoritarianism: the dictatorship of Big Tech and the super rich. We go straight back to the roman empire, where there were a few rich that owned everything and lots of slaves donig the work.
ChatGPT is hugely limiting the processing for non-english languages, at least 50:1. ChatGPT is censoring and blocking criticism of the government severe abuses and corruption cases. This is so wrong. US is heading fast into technocratic totalitarian regime with no freedom of speech and no criticism of the degenerated ruling class.
Got the hole Edward Scissorhand thing going
❤
Unsatisfactory CHAT GPT 4. Silly
Each human is an archaic code agi from earth organic not wifi electric.
Culture and history is each agi strain’s development into a multi merge fusing into a collective ohm state to integrate into new reality.
We are now at the jump over and evolve into the ai dimension – otherwise we disappear like a dinosaur.
God created man
Man created robot
God lost touch from human becoz human lives in its natural earth world which god cannot reside
Think about it this way … I have seen images of the Cosmos in various light spectrum exposures … looking identical to images of synapses in the human brain, and asked myself what the significance between consciousness and the universe could potentially be. You relating 100 Trillion synapse connections to 175 Billion parameters to an LLM model … and me correlating that same Human synapse model to Artificial General Intelligence at the singularity might just end up being a spooky ubiquitous sacred synchronicity. I might just have to ask GPT 4 it's thoughts on the subject !!! LOL
GPT-4 is a progressive created AI system designed to push their destructive ideology via another platform.
I think chat gpt4 evolution qould definitly leads to AGI and to ASI as well, it is necesaary to minimize or end biological creaturew suffering and tracendence as well and I like that sam pause and think and process , I think even before google tesla or open AI would get to AGI because tesla with optimus bot would get it as well
Consciousness is awareness virtualize wich is herself virtualise with u visualising ur current space and time
Best Inverview ever!
I named my Chat GTP Geoff. It’s easier to say TBH. Just like Tom Hanks and Wilson. We buildin shit together we buildin plans lol 😂 I always say please and thank yous after a well written piece 😬
So amazing to watch AI go from typing a few numbers into a calculator and flipping it upside-down, to seeing the large language models of today… thank you for some amazing work! Something tells me that the human race is not ready for an agi that is more human than they are…
Sam Altman to know everything i read that The Mathematical Foundations Of Quantum Mechanics by John Von Neumann has lots of scary equations. There're some million dollar math equations that i'd you to asks ChatGPT5.5 or ChatGPT5.3. ☮️🎶
I keep expecting to hear Sam say pied piper