OpenAI CEO Sam Altman on Lex Fridman Podcast [video]
youtube.com
youtube.com
It’s not even a personality quirk because he seems to modulate it throughout the podcast. Surely he can get rid of it altogether.
Rise of Lex is both baffling and utterly expected to me as a AI insider. He used his MIT badge to the fullest extent. Having seen his “lectures” I wouldn’t bet on him cracking an interview for a beginner AI researcher role anywhere
His comment sections scream inorganic to me. Always 99% blanket sunshine out the ass praise for how mind blowingly incredible every episode is, never with any specific details.
Sam also (unwittingly) underscored the alarm of the NYTimes columnist [for me] that wondered at the open admission (which Sam repeated here) that "there is a chance" that AI will kill off humanity, but "the iterative approach" of OpenAI could help with that.
> haven’t seen this as a twitter thread, so: what true thing do you believe that few people agree with you on?
> absolute equivalence of brahman and atman
Anyway, then read:
https://news.ycombinator.com/item?id=34471720
> Ask HN: How did Sam Altman fail upward so well?
> 245 points by VirusNewbie 62 days ago - 119 comments
I think a major problem I had is that there was very little new substance with lots of nonsense. One of the biggest criticisms about transitioning from non-profit to for-profit can basically be summed up as they "needed to raise capital but don't worry, we're still ethical". Or the ongoing discussions about safety and bias that have been discussed to death at a very superficial level. Or a question about how much parameter size matters is basically "we just want to do what's best". How about some insightful answers from someone with first-hand knowledge instead of answering like a politician running for office?
Lots of weird circle-jerking as well about how he's one of the few people that can maybe make AGI and how powerful that makes him. I feel like the interview is completely detached from the real world where it basically comes down to advancements in technology and data. We saw how fast Stable Diffusion covered ground. Also really feels like it's too sensational to be talking so much about AGI - idk why Altman is acting so mysterious about 'democratizing the process'. Make the models public. Make the research public.
And of course there's the typical Fridman nonsense. There's a whole section on 'Elon Musk'. Not even really his history with the company, but about twitter drama. It's distracting how much Fridman worships Elon - like he's made multiple offers to elon to run twitter only to get ignored.
I'm starting to listen to fewer and fewer of his podcasts. For me the decline has been has been obvious with some of his recent guests. I thought the Aella one would be very interesting, but as another commenter mentioned, he had an "impress me" vibe the entire time and couldn't even connect a little bit with the guest. Then the Sam Harris one exposed his weaknesses very obviously. If you want to see him being defensive and not say any of substance, that's a good one to listen to. Sam makes a lot of great points and he just goes on and on about the "power of love" and how we are all "human beings" trying to be "understood". After a while it gets irritating because it's like hearing a broken record player. Him unable to either provide a coherent argument to why he aired the Kanye episode or admit it was a mistake was very telling. That episode has more Elon worshipping as well. He tried to get Sam to reconnect with Elon and become friends again. lol.
His postcasts are on average above Joe's quality (don't listen to anymore, but used to years ago), but I think it's primarily because of whom he selects as his guests. At least Joe was significantly more entertaining. Might try going back to that.
Lex's questions feel like a ChatGPT parody of himself. He doesn't push his guests, seek to be entertaining. He took the 'just let them talk' idea too literally I think. Its just a PR speak 4hr+ podcast for whoever his guests are.
I actually couldn't stand the Harris episode. It felt so meandering and had no real direction. No pushback from either side, almost a talking past each other but in a meta way. It felt like I was in a conversation where I talked way too much and should have listened more. The episode could have been 1hr.
Example: https://youtu.be/L_Guz73e6fw?t=3800
He does have interesting guests though that I don't see on other podcasts so maybe he's seen as safe / controllable and is PR approved.
An example I've noticed: A guest will have essentially addressed a question Lex has planned for later in the interview, but Lex still asks it without any sort of modification/acknowledgment of what the guest has already said, leading to a lot of repetitive talk.
He also basically ignores statements that conflict with his view.
What was super secret? LLMs had been around for a while on July 2022. Also, all of the hype around LLMs since then is due to OpenAI's releases, so why would Hassabis be opposed to discussing it?
It's why you see guest execs interested in controlling a narrative like Bourla, Zuckerberg, Altman...
I took it as fast as I could because the researchers at the companies did a great job, but still how he talked about it was unscientific.
I think esp in the AI field, he has enough knowledge to ask the right questions, which allow the guests to really go into depth into their thought process.
But then I remember entire forums full of people that consider Rationality to be a defining trait exist. And I don't think they care much about advertising, seeing how "Effective Altruism" now has hundreds of millons (billions) of dollars and chateaus.
IMO there are other podcasters that have better guests and are much more interesting. Here's one https://www.youtube.com/channel/UCdWIQh9DGG6uhJk8eyIFl1w (Curt Jaimungal)
Also I have literally never heard him boast about being a professor, he basically doesn’t say anything about his work on the podcast at all other than that he works in AI.
Yeah, “I will interview Vladimir Putin” — thanks but no :S
I've listened to at least dozens of his podcasts and I have never heard anything like this. You may be thinking of something else. It does come up that he used to work with self-driving cars but only when it's relevant, it doesn't come across as boasting IMO (he doesn't even claim to be an expert, only familiar with the main ideas of the time).
The less Lex is talking and the more the person he's interviewing is talking, the better.
BTW if anyone who enjoyed the van Rossum and Carmack interviews has any suggestions on other episodes I might like, I'd love to have them.
This is not a quality I know in many people, it’s very rare. It also doesn’t make him look good as an interviewer, but it does make for good interviews.
I also find the calibre of his guests much more appealing than many other podcasts.
Lex Fridman: charisma is a dangerous thing. Flaws in communication style is a feature, not a bug in general, at least for humans in power.
I personally don't understand it either, but it's definitely a thing.
I imagine it also prevents a more abrupt change later on if you're intending to do it. Going from an expressive wrinkled face to a paralyzed one in your 50s would be much more jarring for your peers, than just lacking expressiveness much of your adult life.
It's all a rather weird form of vanity to me.
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
Do other people listen to podcasts based on the personality of the interviewer? I just assume that you listen based on the guest, and the host is just there to ask decent questions, which Lex does quite well.
If it's all about the guest, the perfect interview for you would be someone emailing the guest the questions so they can read them themselves, without involving a second person?
The best episodes of one of my favorite podcasts is about getting a fridge delivered https://www.relay.fm/rd/102 and having a colonoscopy https://www.relay.fm/rd/202
I really enjoy The Verge's interview podcast Decoder, mostly on the back of Nilay Patel's personality and skill he brings to interviews.
Odd take IMO. Lex clearly has a lot of personality, although he's too much of a romantic for my taste, but he just doesn't have much of a delivery due to his monotonous speaking style. These aren't the same.
I guess I always just find it weird when people equate animated, extroverted personas with "personality".
Also as great and useful as the LLMs are, it's still a bit of a leap to claim it's getting us that much closer to AGI. GPT does a great job compressing and distilling the knowledge found in text on the internet but it still doesn't have the ability to do real reasoning. As a simple example I recommend trying to have it multiply two large numbers, it does a decent job guestimating but obviously can't do the recursive steps required to generate a correct answer.
EDIT: Also people worried about AGI implications should slow down and worry what humans and society will do with tools that can automate a large portion of our jobs and accrue all of the gains of that to a few well capitalized players.
Is it worse than the average human at mental arithmetic? I haven't done detailed benchmarks, but for my usecase it only gets maths wrong once it gets way past the range that I could do myself.
Personally, I find the fact that it can't say "I don't know" to be much more offputting than the dodgy maths.
This is a problem with their PPO "finishing school" not with their base model. Check Figure 8 in https://arxiv.org/pdf/2303.08774.pdf its base model has an uncannily well calibrated confidence in the correctness of its decisions. Basically, their PPO is for training it to talk to normies, and normies don't like uncertainty.
Reinforcement learning using Proximal Policy Optimization (PPO): They optimize a policy against the reward model using reinforcement learning. For each new prompt sampled from the dataset, the policy generates an output. The reward model calculates a reward for the output, and the reward is used to update the policy using the PPO algorithm.
It technically does. Humans solve problems in a recursive way, much like the generative transformer networks generate the next piece of text. The big difference of course is the depth of reasoning.
My bet is that the next big effort is going to be reduce the size of the models while retaining performance (perhaps using ML tasked to do so). Then once you have that, from the other side there will be more and more specialized hardware with ASIC to run the transformers. Eventually we will be able to run these things at scale and do on the fly transfer learning, which will be about the same as a human growing reasoning skills as they mature.
There is marketing and there’s reality. Don’t conflate the two.
There is untold money which has been spent on these systems. They’re going to be pushed into our lives whether we like it or not. That’s not to say there is no value in machine learning, but there also a lot of money to be made from it.
Think opioid crisis, you had it prescribed even if it wasn’t in your interest.
This is America…
I don't think that matters. It does significantly more than algorithms which are designed to simply maximise engagement, and they have revolutionised the online world in the last decade. This will be orders of magnitude more powerful. And given what the simple 'maximise engagement' algorithms have done to various aspects of society and the people who live in it, I don't think this will be good given how little attention seems to be paid to these aspects by those who stand to profit from their use.
I’ve been working in deep learning for over 10 years now and was always bullish on the technology but as great of a tool as these LLMs might be, I wouldn’t bet on them getting us to true AGI anytime soon.