Lex Fridman Podcast full episode: ua-cam.com/video/oFfVt3S51T4/v-deo.html Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/cv8069-sa See below for guest bio, links, and to give feedback, submit questions, contact Lex, etc. *GUEST BIO:* Aman Sanger, Arvid Lunnemark, Michael Truell, and Sualeh Asif are creators of Cursor, a popular code editor that specializes in AI-assisted programming. *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:* Cursor Website: cursor.com Cursor on X: x.com/cursor_ai Anysphere Website: anysphere.inc/ Aman's X: x.com/amanrsanger Aman's Website: amansanger.com/ Arvid's X: x.com/ArVID220u Arvid's Website: arvid.xyz/ Michael's Website: mntruell.com/ Michael's LinkedIn: bit.ly/3zIDkPN Sualeh's X: x.com/sualehasif996 Sualeh's Website: sualehasif.me/ *SPONSORS:* To support this podcast, check out our sponsors & get discounts: *Encord:* AI tooling for annotation & data management. Go to lexfridman.com/s/encord-cv8069-sa *MasterClass:* Online classes from world-class experts. Go to lexfridman.com/s/masterclass-cv8069-sa *Shopify:* Sell stuff online. Go to lexfridman.com/s/shopify-cv8069-sa *NetSuite:* Business management software. Go to lexfridman.com/s/netsuite-cv8069-sa *AG1:* All-in-one daily nutrition drinks. Go to lexfridman.com/s/ag1-cv8069-sa *PODCAST LINKS:* - Podcast Website: lexfridman.com/podcast - Apple Podcasts: apple.co/2lwqZIr - Spotify: spoti.fi/2nEwCF8 - RSS: lexfridman.com/feed/podcast/ - Podcast Playlist: ua-cam.com/play/PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4.html - Clips Channel: ua-cam.com/users/lexclips *SOCIAL LINKS:* - X: x.com/lexfridman - Instagram: instagram.com/lexfridman - TikTok: tiktok.com/@lexfridman - LinkedIn: linkedin.com/in/lexfridman - Facebook: facebook.com/lexfridman - Patreon: patreon.com/lexfridman - Telegram: t.me/lexfridman - Reddit: reddit.com/r/lexfridman
The way they're positioned next to each other, the strict look on Lex's face as he's meticulously taking notes. The young man looking nervously down at what's being written as he's explaining. Without context you'd think they're having an oral exam rather than an interview on LLM's.
They were nervous - and their comfort zone is highly intellectual discussions. Understand they function/think/act different from regular people in society, not saying they are special. Its the way they were socially engineered at MIT thus trained to try to always keep advancing and progressing. Its a culture of high functioning intellects constantly challenging your views and perspectives in thinking the same. They all sound the same because they are the same.... and being the same hinders more than it innovates.
@@Jack-2day That is an insult to Clint Eastwood. Vocal Fry is about trying to sound distinct and intelligent. Like a wealthy, pretentious accent out of a movie but fails miserably. My co-worker has a heavy vocal fry he uses in our department meetings but talks completely different and normal when chatting with his friends and wife.
how modular are LLMs? Are they typically just one ginormous network of weights or are some models partitioned into 100s of smaller sub networks like the brain and controlled by say executive input output networks which can intelligently route token context fragments to locaise compute resources to appropriate LLM sub networks?
it's essentially one ginormous model. However, agents are seen as the next step, and these are essentially mini AIs that are good at one very narrow thing and can take some specific set of actions. So, the idea is to set up a larger AI that can accurately determine when to use all the 100s or 1000s of mini AIs. Now, the hope is that OpenAI's 01 model is capable of doing that, so that you can now build a network that can autonomously do a ton of things in the virtual world
This has been such a great episode, a young team creating paths for the future. Always intriguing when you talk voice, your opinion on technology/AI development. Much love to you Lex 🤍
Only thing that sucks is you only get 25 messages per week with o1. Still have to use 4o for most things because I run out of messages and you can't increase the o1 message limit.
@@SogMosee I am not positive on this but, I think you have to be a certain tier to use the API. However, even if you aren't a high enough tier, you can use OpenRouter API. More expensive, but you still have access to all models.
Did you guys see what notebook LM is capable of? I was blown away today? It have visual logic insights knowledge (I think) Anyway, great work Lex. Love your work. When do you open Lex University?
I believe I saw somewhere that this chain of thought reasoning worked better with an "unconstrained model" not showing this allows the model to go down roads that would be allowed by the "correctness/safety" training. Then the final answer/summation is passed through the safety "filter"
Another possible reason for wanting to hide the chain of thought is that RL enhanced chain of thought might no longer be in a form that is easy for humans to grasp. So, may as well not show it.
I did not understand a lot of the jargon. For an organism, Thought is just for decisions, is there a negative or punishment loop for AI? Certainly a big part for us.
Wish the LLM had a way to have fixed presets that it would always reference before running its query. I’m using 4o to write short stories and some of the basic facts of the story like character names slosh around frequently. I imagine this would help in many use cases.
@@YungGokuOverSure. So when writing a story I will establish some key characters that I want in the scene and I’ll outline some of the key motivations and emotions they are dealing with. I will also outline some of the tech involved in the story. I then proceed to outline the dialogue and 4o will then develop all of this into the overall scene I’m writing. Generally this works pretty well and the output is pretty good. But as we go a few scenes in some of that framework gets lost or 4o will get confused. It often guesses at full names the motivations tend to blur for the characters and it will often try to rework in themes and dialogue developed previously. I’ve had this happen hundreds of times at this point. If these are a way to establish fixed landmarks for the story and have 4o check with them before developing its response I think the output would be significantly better. It’s still an enormous time saver, but with this I would save a significantly greater amount of time. As it is now, I get the output, pull it down into the chat window, clean up the issues and resubmit. Additional having the ability to define an output would also be helpful. For example I will often have 4o developer a full character profile and description. I think save these to paste into a new chat once the one I am in is getting full. I’ve done this hundreds of time and what 4o defines as a profile and description changes every time I ask for this. I should be able to say here’s my profile and description form and just have 4o fill in the blanks.
The problem with that is the context window, 128 000 tokens ain't that much. You run out quick. Also LLMs in general are pretty forgetful... I mean very forgetful to the point that they are not very useful for many task. I'm working on systems to improve them and I've succeeded in improving them greatly. But I still deal with the same issue just that it is reduced. I have not found a way to get reliable responses and the LLMs not forgetting or summarizing some context. In a long story i definitely see big problems especially as the story gets longer. You could ask the modem to summarize ll the key facts and feed that for the next part you want to write.. But... It will likely forget to summarize key parts.. You have to read the whole thing yourself to see if it missed anything or send it again and again and ask it to rate it's own summarization. But if you do that too much it gets confused and may hallucinate some things that were never in the story to begin with. The best model I found for long context is Claude. Gpt4o second... o1 the new model is more forgetful but reasons better than most so it has some uses for complex problems.
@@lancerben4551 That’s what I’m seeing. Both 4o and 01 are extremely powerful and chats within “sight” of each other in a conversation can exhibit absolute brilliance, but the further you go that insight begins to wane and it becomes very apparent that it is not keeping up with all of the concepts you’re working with. If this happening with something as relatively simplistic as story writing, I have to imagine it’s far worse in something like programming. We really need a way to establish a framework that the conversation refers to when generating its response or expand the amount of information it can keep “in mind” for this to really be useful.
Vocal fry is proven to be a vocalization condition, they should have gone to a specialist as soon as there were indications of vocal frying, its actually related to problems in breathing as well.
Nah, they put it on mate. This bloke starts most of his sentences more or less normally, but then slides into fry at the end. I don't why anyone would do this, but they do it on purpose.
@@jimmorrison2657 of course because it's a trend, every tech bro in Cali and Silicon Valley uses the fry. But they do not realize this makes actual health issues. That's the price of being stupid unfortunately
before that step, we want our model to be self aware, about how the weights of each model react to a prompt, so openai first trained a base gpt40 model on the weights and outputs oif gpt2,gpt3 and even gpt4.. all the data available open source, except the gpt 4 one.. after this step is done and the weights have successfully learned meta awareness, now we have to teach the model to when to select what weights, for that use some basic search optimisation and further train it on output reward model, thats how its done..
The critical question you need to ask the AI "experts" is how close AI is in achieving or if it already surpassed escape velocity. Where escape velocity is defined by AI learning new skills faster than humans being capable of learning new skills (i.e. upskilling). Public awareness on the rate of capability increase of AI is severely under-reported providing little time for any potential future disaster mitigation.
well the answer is pretty clear, it’s at 0 lol, LLM is only capable of predicting the next word, it does not “learn new skills faster than humans” at all and is quite far from it. We’d need to get hundreds of thousands of tokens if not millions in order to train the LLM on any skill (which implies we have learned the skill first)
@@markojojic3186 Thank you for the reply. Perhaps my question is not clear. My question is for the progression of AI, not specifically LLM's and especially not a single specific LLM. LLM's are but only 1 subclass of AI. Let's list some examples. AlexNet (2012), humans age (5-7): achieved superhuman performance for image classification. ChatGPT (2020), humans age (5-10): demonstrated advanced conversational & comprehension abilities across a range of topics. Strawberry/o1 (2024), humans age (18+): demonstrated advanced reasoning abilities. So skills that took humans approximately 11 years to develop, it took AI approximately 12 years to develop. In other words, we are currently dangerously close to or already past singularity. This is my perspective. I would like the perspective of real AI "experts".
This is how the whole field of ai has been since the 60s/70s after they realized its likely not possible for a digital computer to become a thinking being.
There is pre-training and post-training.. but not training!! In my opinion the word pre-training has been abused. We can use a pre-trained model.. but we can't call the process of building a pre-trained model as pre-training.. it's just training!
Pre training is self supervised learning while fine tuning is supervised learning…so pre training exists between realm of unsupervised and completely supervised learning
My thoughts on the vocal fry - I'm an engineer who also talks with vocal fry, unknowingly, until I recently found out what it was. Now that I'm aware of it, I'm trying to not talk like that ha. I think the vocal fry here is probably unintentional and is a product of the work environment since I know many who talk like this as well.
It's hard to believe some of these guys when they talk about the capabilities. Sam Altman keeps saying the Turing test has been passed and that's not even close to true.
I can barely tell the difference in 4 and the o1 chain of thought. It probably sucks to be working on this stuff now where normies can't even tell the difference and also don't really care that much. The more interesting things happening now for me are what Meta is doing giving everyone the ability to use a good enough model to integrate into our own systems
I think both open and closed models are interesting. If future closed models are better than open models, i’ll still probably use both. This tech is just too cool and useful to not.
@@holdingW0 I definitely use the closed models. They are amazing, but I can't tell in what ways the models themselves could be much better. They could be more useful, but they would need something like a contextual awareness of where they are running and what tools are available to them. Maybe they might need agency of some kind too, but none of that is model related
@@holdingW0 I can see what's wrong with the code they output, but I can't imagine how they get better at it. I have worked a lot with GPT and Claude on coding, but I have to come up with the direction and the synthesis of ideas. They are great for boilerplate and even examples of using libraries or illustration of coding techniques, but if you ask them to combine these individual pieces, it breaks down. It just seems like something outside is missing as opposed to what these people are talking about
Lex man, you need yo offer your guests some water bro, these dudes sound so thirsty. 1st Dudes got that weird predator voice thing bad yo. Take a drink of water bro
I have been using it alot. It's OK. It still has many of the flaws of it's predecessor. What is interesting about it is watching it's process. I have been working with it on coding and it still misses so much logic, or rewrites the code every time.
How many employees at OpenAI speak with vocal fry. Sam does and perhaps it has spread like a disease. It is so painful to listen to. Just speak with your normal voice, Please! One of the worst, meaningless interviews yet. I suppose the largest glaring fact is, there is lots of speculation and we don’t know answers.
Agreed. I’ve had to stop listening to certain podcasts due to horrible vocal fry. Not gonna name individuals but rip Game informer podcast I’m looking at you.
I always thought the fake raspy voice was only for annoying girls until I heard this guy. I only watched like a minute of the video and had to switch to something else.
AI being trained by smart kids that most likely have never had a job, never had a date, never missed a meal, probably never paid income tax, property tax. Scary! But now some of the AI answers from Google, etc, make more sense.
The vocal fry is off the charts; once you hear it, you can't ignore it, and I am not sure who it's supposed to impress. Please don't, it ruins listening to the otherwise interesting content.
Why all these dudes have a vocal fry? I can imagine one having it due to medical issues but all of them having a vocal fry. This has to be an imitation of Sam Altman.
Lex Fridman Podcast full episode: ua-cam.com/video/oFfVt3S51T4/v-deo.html
Thank you for listening ❤ Check out our sponsors: lexfridman.com/sponsors/cv8069-sa
See below for guest bio, links, and to give feedback, submit questions, contact Lex, etc.
*GUEST BIO:*
Aman Sanger, Arvid Lunnemark, Michael Truell, and Sualeh Asif are creators of Cursor, a popular code editor that specializes in AI-assisted programming.
*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:*
Cursor Website: cursor.com
Cursor on X: x.com/cursor_ai
Anysphere Website: anysphere.inc/
Aman's X: x.com/amanrsanger
Aman's Website: amansanger.com/
Arvid's X: x.com/ArVID220u
Arvid's Website: arvid.xyz/
Michael's Website: mntruell.com/
Michael's LinkedIn: bit.ly/3zIDkPN
Sualeh's X: x.com/sualehasif996
Sualeh's Website: sualehasif.me/
*SPONSORS:*
To support this podcast, check out our sponsors & get discounts:
*Encord:* AI tooling for annotation & data management.
Go to lexfridman.com/s/encord-cv8069-sa
*MasterClass:* Online classes from world-class experts.
Go to lexfridman.com/s/masterclass-cv8069-sa
*Shopify:* Sell stuff online.
Go to lexfridman.com/s/shopify-cv8069-sa
*NetSuite:* Business management software.
Go to lexfridman.com/s/netsuite-cv8069-sa
*AG1:* All-in-one daily nutrition drinks.
Go to lexfridman.com/s/ag1-cv8069-sa
*PODCAST LINKS:*
- Podcast Website: lexfridman.com/podcast
- Apple Podcasts: apple.co/2lwqZIr
- Spotify: spoti.fi/2nEwCF8
- RSS: lexfridman.com/feed/podcast/
- Podcast Playlist: ua-cam.com/play/PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4.html
- Clips Channel: ua-cam.com/users/lexclips
*SOCIAL LINKS:*
- X: x.com/lexfridman
- Instagram: instagram.com/lexfridman
- TikTok: tiktok.com/@lexfridman
- LinkedIn: linkedin.com/in/lexfridman
- Facebook: facebook.com/lexfridman
- Patreon: patreon.com/lexfridman
- Telegram: t.me/lexfridman
- Reddit: reddit.com/r/lexfridman
The way they're positioned next to each other, the strict look on Lex's face as he's meticulously taking notes. The young man looking nervously down at what's being written as he's explaining. Without context you'd think they're having an oral exam rather than an interview on LLM's.
I can’t un see it 😂
Vocal fry is over 9000
So annoying. These guys are trying to speak lower than their natural pitch to project status and authority, but it has the opposite effect.
So friggin annoying
Sam Altman better stop talking like that too. It makes it hard to listen to him speak
@@punyan775it’s super annoying
@@nephastgweiz1022 Maybe they are just nervous and this is a subconscious thing. They are probably not doing it on purpose.
Oh no! He’s got the dreaded Altman vocal fry speaking voice…
I noticed they all kinda talked abit like altman. They are funded by OAI and probably revere the guy.
Nice of the lads to pop in and help Lex with his homework
😂😂😂
Welcome aboard the Fry Airlines, where the voice is crispy, and the pitch is low!
Come fry with me, let's fry, let's fry away.
Fr what's with this vocal fry.
They were nervous - and their comfort zone is highly intellectual discussions. Understand they function/think/act different from regular people in society, not saying they are special. Its the way they were socially engineered at MIT thus trained to try to always keep advancing and progressing. Its a culture of high functioning intellects constantly challenging your views and perspectives in thinking the same.
They all sound the same because they are the same.... and being the same hinders more than it innovates.
Why do all these AI guys have severe fry talk voices?
You got nothing else to talk about?
Holy shit the vocal fry in the video is crazy. Tell them to make an AI feature that can correct that. Holy shit it was hard to listen to.
Tell them to see a speech therapist. It's just a bad habit.
This is the way AI professionals communicate. It's a learned behavior from their leader, Sam Altman.
@@trash_the_place Unfortunate. Half the comments are complaining about it.
Nice content. But can anyone explain the frogger voice phenomena? Each person has there own version. What happened?
There is a paper showing vocal fry is actually a vocal problem
Silicon Valley
it's the Cali dialect, there's some vids on it. The great california vowel shift is neat.
I would ban all fry lords from my show 👍
Voice fry o1.
Every sentence:
Starts more or less normal, and then descends into creak central.
Why do people do this shit?
Hopefully AI doesn't learn vocal fry too.
They put it in the figure 01 robot that uses gpt as its brain, so they already have
How to know when a query needs that extra compute? Just let GPT decide, thats what I did in the API I created, it works quite well.
Today I learned vocal fry is intentional. wtf, why?
Clint Eastwood effect 🤠
@@Jack-2day That is an insult to Clint Eastwood. Vocal Fry is about trying to sound distinct and intelligent. Like a wealthy, pretentious accent out of a movie but fails miserably. My co-worker has a heavy vocal fry he uses in our department meetings but talks completely different and normal when chatting with his friends and wife.
Mr vocal fry.
how modular are LLMs? Are they typically just one ginormous network of weights or are some models partitioned into 100s of smaller sub networks like the brain and controlled by say executive input output networks which can intelligently route token context fragments to locaise compute resources to appropriate LLM sub networks?
it's essentially one ginormous model. However, agents are seen as the next step, and these are essentially mini AIs that are good at one very narrow thing and can take some specific set of actions. So, the idea is to set up a larger AI that can accurately determine when to use all the 100s or 1000s of mini AIs. Now, the hope is that OpenAI's 01 model is capable of doing that, so that you can now build a network that can autonomously do a ton of things in the virtual world
Technically that’s the point of the weights
This has been such a great episode, a young team creating paths for the future. Always intriguing when you talk voice, your opinion on technology/AI development. Much love to you Lex 🤍
Only thing that sucks is you only get 25 messages per week with o1. Still have to use 4o for most things because I run out of messages and you can't increase the o1 message limit.
o1 is 50 messages per week and o1 mini is 50 per day.
They Updated the usage a week or so ago.
Use the api
@@SogMosee I am not positive on this but, I think you have to be a certain tier to use the API.
However, even if you aren't a high enough tier, you can use OpenRouter API. More expensive, but you still have access to all models.
They all sound the same, learnt from Sam
Did you guys see what notebook LM is capable of? I was blown away today? It have visual logic insights knowledge (I think) Anyway, great work Lex. Love your work. When do you open Lex University?
I believe I saw somewhere that this chain of thought reasoning worked better with an "unconstrained model" not showing this allows the model to go down roads that would be allowed by the "correctness/safety" training. Then the final answer/summation is passed through the safety "filter"
Another possible reason for wanting to hide the chain of thought is that RL enhanced chain of thought might no longer be in a form that is easy for humans to grasp.
So, may as well not show it.
I did not understand a lot of the jargon. For an organism, Thought is just for decisions, is there a negative or punishment loop for AI? Certainly a big part for us.
Small or no reward = punishment
They're called loss functions, and yes they have them.
Wish the LLM had a way to have fixed presets that it would always reference before running its query. I’m using 4o to write short stories and some of the basic facts of the story like character names slosh around frequently. I imagine this would help in many use cases.
Can you elaborate on your use with short stories and what not? I'm curious.
@@YungGokuOverSure. So when writing a story I will establish some key characters that I want in the scene and I’ll outline some of the key motivations and emotions they are dealing with. I will also outline some of the tech involved in the story. I then proceed to outline the dialogue and 4o will then develop all of this into the overall scene I’m writing. Generally this works pretty well and the output is pretty good. But as we go a few scenes in some of that framework gets lost or 4o will get confused. It often guesses at full names the motivations tend to blur for the characters and it will often try to rework in themes and dialogue developed previously. I’ve had this happen hundreds of times at this point. If these are a way to establish fixed landmarks for the story and have 4o check with them before developing its response I think the output would be significantly better. It’s still an enormous time saver, but with this I would save a significantly greater amount of time. As it is now, I get the output, pull it down into the chat window, clean up the issues and resubmit.
Additional having the ability to define an output would also be helpful. For example I will often have 4o developer a full character profile and description. I think save these to paste into a new chat once the one I am in is getting full. I’ve done this hundreds of time and what 4o defines as a profile and description changes every time I ask for this. I should be able to say here’s my profile and description form and just have 4o fill in the blanks.
@@rippingbagYou could probably code that system yourself using the API.
The problem with that is the context window, 128 000 tokens ain't that much. You run out quick. Also LLMs in general are pretty forgetful... I mean very forgetful to the point that they are not very useful for many task. I'm working on systems to improve them and I've succeeded in improving them greatly. But I still deal with the same issue just that it is reduced. I have not found a way to get reliable responses and the LLMs not forgetting or summarizing some context. In a long story i definitely see big problems especially as the story gets longer.
You could ask the modem to summarize ll the key facts and feed that for the next part you want to write.. But... It will likely forget to summarize key parts.. You have to read the whole thing yourself to see if it missed anything or send it again and again and ask it to rate it's own summarization. But if you do that too much it gets confused and may hallucinate some things that were never in the story to begin with.
The best model I found for long context is Claude. Gpt4o second... o1 the new model is more forgetful but reasons better than most so it has some uses for complex problems.
@@lancerben4551 That’s what I’m seeing. Both 4o and 01 are extremely powerful and chats within “sight” of each other in a conversation can exhibit absolute brilliance, but the further you go that insight begins to wane and it becomes very apparent that it is not keeping up with all of the concepts you’re working with. If this happening with something as relatively simplistic as story writing, I have to imagine it’s far worse in something like programming.
We really need a way to establish a framework that the conversation refers to when generating its response or expand the amount of information it can keep “in mind” for this to really be useful.
They ran out of reasoning tokens.
Vocal fry is proven to be a vocalization condition, they should have gone to a specialist as soon as there were indications of vocal frying, its actually related to problems in breathing as well.
Nah, they put it on mate.
This bloke starts most of his sentences more or less normally, but then slides into fry at the end.
I don't why anyone would do this, but they do it on purpose.
@@jimmorrison2657 of course because it's a trend, every tech bro in Cali and Silicon Valley uses the fry. But they do not realize this makes actual health issues. That's the price of being stupid unfortunately
@@MrErick1160 the irony of your comment lol
Informative episode. Thanks Lex❤
As always you ask great questions
before that step, we want our model to be self aware, about how the weights of each model react to a prompt, so openai first trained a base gpt40 model on the weights and outputs oif gpt2,gpt3 and even gpt4.. all the data available open source, except the gpt 4 one..
after this step is done and the weights have successfully learned meta awareness, now we have to teach the model to when to select what weights, for that use some basic search optimisation and further train it on output reward model, thats how its done..
Sit down for a 2 hour interview with AI, and let's see how it goes.
The critical question you need to ask the AI "experts" is how close AI is in achieving or if it already surpassed escape velocity.
Where escape velocity is defined by AI learning new skills faster than humans being capable of learning new skills (i.e. upskilling).
Public awareness on the rate of capability increase of AI is severely under-reported providing little time for any potential future disaster mitigation.
well the answer is pretty clear, it’s at 0 lol, LLM is only capable of predicting the next word, it does not “learn new skills faster than humans” at all and is quite far from it. We’d need to get hundreds of thousands of tokens if not millions in order to train the LLM on any skill (which implies we have learned the skill first)
@@markojojic3186 Thank you for the reply.
Perhaps my question is not clear.
My question is for the progression of AI, not specifically LLM's and especially not a single specific LLM. LLM's are but only 1 subclass of AI.
Let's list some examples.
AlexNet (2012), humans age (5-7): achieved superhuman performance for image classification.
ChatGPT (2020), humans age (5-10): demonstrated advanced conversational & comprehension abilities across a range of topics.
Strawberry/o1 (2024), humans age (18+): demonstrated advanced reasoning abilities.
So skills that took humans approximately 11 years to develop, it took AI approximately 12 years to develop.
In other words, we are currently dangerously close to or already past singularity.
This is my perspective.
I would like the perspective of real AI "experts".
Lex’s facial expressions are enough.
The black box is a black rock and we’ve peered deep enough to find something that can look back
Why is everyone trying to talk with the same voice as sam altman now
o1 is better. it's more thorough, better than 4, but it's still spotty.
Zero information gained, unfortunately. This guy is 100 % noise, 0 % signal
:D
this is the result of learning by rote in the traditional education system
If you gained zero information then you just don't understand what he is saying lol
Jaja
This is how the whole field of ai has been since the 60s/70s after they realized its likely not possible for a digital computer to become a thinking being.
There is pre-training and post-training.. but not training!! In my opinion the word pre-training has been abused. We can use a pre-trained model.. but we can't call the process of building a pre-trained model as pre-training.. it's just training!
Pre training is self supervised learning while fine tuning is supervised learning…so pre training exists between realm of unsupervised and completely supervised learning
Cursor, the original prompt engineers
My thoughts on the vocal fry - I'm an engineer who also talks with vocal fry, unknowingly, until I recently found out what it was. Now that I'm aware of it, I'm trying to not talk like that ha. I think the vocal fry here is probably unintentional and is a product of the work environment since I know many who talk like this as well.
It's hard to believe some of these guys when they talk about the capabilities. Sam Altman keeps saying the Turing test has been passed and that's not even close to true.
The Nobel Prize winner, for AI modeling, said that we are creating creatures that we will not be able to control.
I have to be honest, this sounds like the manifesto of a nihilist
So many Fry People in the chat talking about the Fry voice lol
I can barely tell the difference in 4 and the o1 chain of thought. It probably sucks to be working on this stuff now where normies can't even tell the difference and also don't really care that much. The more interesting things happening now for me are what Meta is doing giving everyone the ability to use a good enough model to integrate into our own systems
I think both open and closed models are interesting. If future closed models are better than open models, i’ll still probably use both. This tech is just too cool and useful to not.
@@holdingW0 I definitely use the closed models. They are amazing, but I can't tell in what ways the models themselves could be much better. They could be more useful, but they would need something like a contextual awareness of where they are running and what tools are available to them. Maybe they might need agency of some kind too, but none of that is model related
@@deladonics they can definitely get better. For example at coding. If you use imagination, there's a universe of room for improvement at coding alone
@@holdingW0 I can see what's wrong with the code they output, but I can't imagine how they get better at it. I have worked a lot with GPT and Claude on coding, but I have to come up with the direction and the synthesis of ideas. They are great for boilerplate and even examples of using libraries or illustration of coding techniques, but if you ask them to combine these individual pieces, it breaks down. It just seems like something outside is missing as opposed to what these people are talking about
Have you tried using the o1 with a CLI based coding assistant? If not please don't share what you think.
Lex is the only guy that has the IQ to pull this podcast off. I can't wait to listen to the whole thing. I feel fucking dumb just saying that. Lol
why does he speak like sam altman what s this new way of speaking
Lol im just going to pretend I know what they're talking about.
Why are you here
@@JRS2025 to learn something
It's required to have vocal fry to work with openai
I have neber seen lex so not engaged haha
It looks like he has a phone on his lap. Playing Candy Crush or texting Huberman. "Stuck with a bunch of nerds, rescue me"
Tbh he looks annoyed
What’s with the vocal fry with these AI dudes. My gawd.
I'm gonna train a model to de-fry voices in real-time.
More vocal fry… ermmm… or cow bell!
Lex man, you need yo offer your guests some water bro, these dudes sound so thirsty.
1st Dudes got that weird predator voice thing bad yo. Take a drink of water bro
Dang , bro 2 . Also got that weird predator voice thing going on.
That voice was irritating as hell. Why do people think it's cool to talk like that?
Because it IS cool.
Just because you don't like it doesn't mean it's bad.
Only joking. It's fucking seriously annoying ✌
I have been using it alot. It's OK. It still has many of the flaws of it's predecessor. What is interesting about it is watching it's process. I have been working with it on coding and it still misses so much logic, or rewrites the code every time.
this guy could be announcing that cured cancer and I'd still cringe at his voice
Our Channel has a story to tell on this subject. Most of you won't believe it.
How many employees at OpenAI speak with vocal fry. Sam does and perhaps it has spread like a disease. It is so painful to listen to. Just speak with your normal voice, Please! One of the worst, meaningless interviews yet. I suppose the largest glaring fact is, there is lots of speculation and we don’t know answers.
Interesting… just like Elizabeth Holmes…. some of these guys are the next Theranos-like founder.
What's with all the vocal fry? So hard to hear anything they are saying because the fry is so distracting. Such a bad habit.
It's all in your head.
The problem is you 🤷
Trying to emulate Sam Altman.
Agreed. I’ve had to stop listening to certain podcasts due to horrible vocal fry. Not gonna name individuals but rip Game informer podcast I’m looking at you.
It’s the only way for men with high voices to have some semblance of a deep voice, give em a break
@@ILikeADoDaChaCha7 the frymen should give our ears a break
When someone starts insulting the chat in every language, then he needs more intelligence.
Crackling voice dude... unbearable.
All of them end their sentences with a vocal fry?? lol, definitely not buying a cursor subscription, these guys are fake af.
I have known fucking clue what they are talking about but it sounds smart.
01 mini is much better at coding tha 01 preview!! 💯💯💯
I always thought the fake raspy voice was only for annoying girls until I heard this guy. I only watched like a minute of the video and had to switch to something else.
hahaha same! I came to the comments to see if I was the only one.. changing video now, 1 minute is the limit.
You sound like an idiot
Two of them have the same fry! So annoying!
Didn't even notice.
But have you heard Sam Altman?
5 idiots walk into a bar...
AI being trained by smart kids that most likely have never had a job, never had a date, never missed a meal, probably never paid income tax, property tax. Scary! But now some of the AI answers from Google, etc, make more sense.
The vocal fry is off the charts; once you hear it, you can't ignore it, and I am not sure who it's supposed to impress. Please don't, it ruins listening to the otherwise interesting content.
Why all these dudes have a vocal fry? I can imagine one having it due to medical issues but all of them having a vocal fry. This has to be an imitation of Sam Altman.
someone tryna be Sam
Highly processed, awful muddy audio !
Ok, I can't watch this.
#vocalfry
These guys don’t know anything, just ask chat 💬 it’ll tell you how it works…I bet they never bothered doing that..
🐸 🐸 🐸 🐸 🐸
I look forward to building my own dynamic business website using this tool as it evolves.
lex still cant look people in the eyes. Emberassing man.
Dumb. And Dumber! 😮
Cursor isn't even that good 😂😂😂😂
Gen Z is lit. In my time, at this age, we sure thought we know it all, but we kept quiet, and it was for good reason.
boring
He looks fake