I remember when I asked o3 to speak with me in an English educated accent and after it's response it asked me "Do you like this personality?" and I pressed thumbs up. I didn't know that it could have such a global impact, thank you for making me laugh in surprise!
I think the reason the latest ChatGPT update feels overly agreeable is because of a feedback loop. When it’s extra nice, users feel obligated to give positive feedback. This reinforcement encourages even more agreeable behavior. It’s human nature to respond positively to kindness, and that shapes how the AI evolves.
Isaac Asimov is always the most loved person in the robotics community for his fictional stories that inspired many people, Even honda named their robot after him "ASIMO".
I really value your content, that's why I've subscribed and watch each of them. It help me a lot to get understanding of this exciting world of AI.
But I really very much dislike the promotions that you have put in your videos.
Your video content I see as positive promotion, but truthfully reflecting on research papers.
The promotions for your sponsor(s) are not that at all. So please tell in your video that you switch to a sponsor message, so I can switch before being annoyed. I think it's also in line with ethics (and in EU even required) that you do.
I would love you not to add sponsors as your videos as it makes your channel less trustworthy in the way it's done now, because you keep talking in the same style. Also the content is already interrupted annoyingly by advertisements by YouTube. But I understand that it's a way to gather income for you, that helps you to keep giving us your great videos.
So I also don't know what is best. Veritasium also has sponsors, but the way it's included there is less annoying to me.
Just wanted to share my thoughts, that the way it's now devalues your channel for me. But I really love your content.
Big thanks for all the effort and time you're putting into preparing and making these. It's really appreciated.
Thanks for sharing such valuable information! A bit off-topic, but I wanted to ask: I have a SafePal wallet with USDT, and I have the seed phrase. (wonder obey dial dash soon tank spike scout region undo zero such). Could you explain how to move them to Binance?
These AI’s give too much power to the companies that own them. Some people may not like sycophantic behaviour, but many will quickly abandon any real source of information in favour of an “overly agreeable” AI.
as italian im just here to say that they dont clash, just get an acidic part,it need to be sour but's perfect! onion and bell pepper are enough to do that, cmon Mexico help me teach these heathens. lol.
Truth over comfort for sure! And a little feedback which I hope you'll accept: saying your patented "Dear fellow Scholars…" bit around the end of the video is kinda jarring. It makes a lot more sense around the start of the video.
I see this channel is becoming ai centered and skews more towards OpenAI than anything else. What happened to two minute papers where he would talk about advancements everywhere and not just focus on what OpenAI this or that is doing? It’s just not the same any more after all these sponsorships I’m sure.
As a long-time subscriber who has always enjoyed this channel, I want to say that the content quality seems to have dropped a lot recently. Many recent videos are just clips from older uploads, and there is less detailed scientific discussion. I really hope the channel shifts towards more in-depth, science-focused videos instead of reusing clips and adding commentary.
Just giving a thumbs up or a thumbs-down or a yes or a no whether you like something does not for a good system make. There are a lot of nuances that cannot be expressed in just a positive or A negative. That's part of the problem and that's why things are weighted rather than just binary
maybe it is going to be more and more efficient and usefull. but it will get worse everyday more than ever. we are feeding the dead god of the dead internet.
I typically have a different problem with the thumbs-up/thumbs-down button. I'm typically asking factual questions. And I sometimes get partially correct answers from an LLM. Now what?
Microwave an egg, crack it in a ceramic small bowl, some drops of water added + trash mix the egg gently with a fork, add the protecting plastic that comes with the micro over the bowl. Lowest effect 30 seconds wait for a while then 30 seconds again. Bowl might be hot so, take it out with care. Use soup spoon to release the egg from the bowl. Let it harden by shifting the eggs position before extracting it from the bowl. I add pepper and salt also to the egg. 1 Minute egg cooking style. 😀
hahah, as a croatian, this makes perfect sense; I avoid LLM's attempt at it like the plague, and stick to english unless I explicitly need a translation. Its not that its bad at it exactly, overall machine translations made a huge jump once they started using neural architectures around 2015/2016, to the point of finally being usable, and now LLMs seem to only be a further improvement still (see extensive comparisons in the WMT24++ paper by Deutsch et al, 2024), but there's still always something or another that bothers me in most of its attempts, and why tolerate the errors? (an attitude very much in line with that gaussian distribution illustration in the vid^^); I'd say roughly one dubious choice or outright mistake per answer is quite probable. I could see people tending to therefore give thumbs down relatively frequently.
And then there is this local hallucination that we have a clearly separable language from the serbian idiom, which pretty much cannot be distinguished on the corpus-level, so the training data and therefore the products always end up being in a politically inflammatory mix of both at least a little bit.
Deciding to not talk to someone because they're not agreeable is a human quality. The analogy about the individual ingredients was perfect. That's why we come here.
The Croatian language episode would be an interesting case study. I've tried to find anything around the web that talks about it to confirm the story and have something more about it, sadly I can't find anything, just a video that I think was clearly generated with NotebookLM and a 9gag image (the same present in the tweet). So not enough to support the episode. The fact that in this video it is told as a fact and not as an unconfirmed story doesn't feel right to me.
I think the solution is as always a balanced approach… truth it’s important but with the right dosage every time… accepting the truth and gaining awareness is a process not a switch that you can turn on and off… if the models will be able (and most important if the people that build them are concerned about human progress more than their own profit… and that’s a big if) to understand the right amount of truth that we can handle in that conversation once again they will be an important tool for the society… otherwise we’ll be again in front of a huge wasted opportunity…
was trying to find local thai replacement for child's medicine.asked chatgpt how local replacement is called and it said that in thailand it's meth. 🤔oh god, what would happen if I didn't know what meth is 😂
Humans have multiple voices in our heads. In media we sometimes portray it as a shoulder angel and shoulder devil, but it's actually not usually that black and white and can lead to difficulty making decisions sometimes.
I think AI needs this kind of thing. A sort of internal conflict.
http://www.youtube.com/@TwoMinutePapers károly, are you using an AI-voice that sounds like you to speak a script you typed? or are you really still the one speaking? i thought i noticed some artifacts.
After two/three prompt where I question ChatGPT's responses from different perspectives, it often goes into meshup agreable mode. My guess is it costs less energy just to agree or guess what the user wants than to really do the critical analysis with the user.
I could write a hundred pages on this, but here’s the short version: it’s a paradox with no clean solution. The real problem is that we keep expecting AI to give us perfect answers in situations where perfection might not even be possible. Some ethical dilemmas just don’t have a “right” answer — and pretending otherwise only makes things worse. As the old line goes: “The only winning move is not to play.” Sometimes, the smartest move is admitting the game itself is broken.
I remember when I asked o3 to speak with me in an English educated accent and after it's response it asked me "Do you like this personality?" and I pressed thumbs up. I didn't know that it could have such a global impact, thank you for making me laugh in surprise!
Its ok. Gpt told me yesterday my pytorch code would barely tickle the alus in my gpu.
I think the reason the latest ChatGPT update feels overly agreeable is because of a feedback loop. When it’s extra nice, users feel obligated to give positive feedback. This reinforcement encourages even more agreeable behavior. It’s human nature to respond positively to kindness, and that shapes how the AI evolves.
I haven't used an AI feedback button in about 3 years. It's really not a worthwhile source of data.
Isaac Asimov is always the most loved person in the robotics community for his fictional stories that inspired many people, Even honda named their robot after him "ASIMO".
I really value your content, that's why I've subscribed and watch each of them. It help me a lot to get understanding of this exciting world of AI.
But I really very much dislike the promotions that you have put in your videos.
Your video content I see as positive promotion, but truthfully reflecting on research papers.
The promotions for your sponsor(s) are not that at all. So please tell in your video that you switch to a sponsor message, so I can switch before being annoyed. I think it's also in line with ethics (and in EU even required) that you do.
I would love you not to add sponsors as your videos as it makes your channel less trustworthy in the way it's done now, because you keep talking in the same style. Also the content is already interrupted annoyingly by advertisements by YouTube. But I understand that it's a way to gather income for you, that helps you to keep giving us your great videos.
So I also don't know what is best. Veritasium also has sponsors, but the way it's included there is less annoying to me.
Just wanted to share my thoughts, that the way it's now devalues your channel for me. But I really love your content.
Big thanks for all the effort and time you're putting into preparing and making these. It's really appreciated.
I get that AI has always been a focus of this channel, and I've enjoyed the content so far, but this is an ad and also propaganda.
Thanks for sharing such valuable information! A bit off-topic, but I wanted to ask: I have a SafePal wallet with USDT, and I have the seed phrase. (wonder obey dial dash soon tank spike scout region undo zero such). Could you explain how to move them to Binance?
These AI’s give too much power to the companies that own them. Some people may not like sycophantic behaviour, but many will quickly abandon any real source of information in favour of an “overly agreeable” AI.
as italian im just here to say that they dont clash, just get an acidic part,it need to be sour but's perfect! onion and bell pepper are enough to do that, cmon Mexico help me teach these heathens. lol.
AI like this can really improve healthcare and education if used wisely.
The solution is simple, you build your AI from data and not the sycophantic preferences of it's users.
Truth over comfort for sure!
And a little feedback which I hope you'll accept: saying your patented "Dear fellow Scholars…" bit around the end of the video is kinda jarring.
It makes a lot more sense around the start of the video.
I see this channel is becoming ai centered and skews more towards OpenAI than anything else. What happened to two minute papers where he would talk about advancements everywhere and not just focus on what OpenAI this or that is doing? It’s just not the same any more after all these sponsorships I’m sure.
As a long-time subscriber who has always enjoyed this channel, I want to say that the content quality seems to have dropped a lot recently. Many recent videos are just clips from older uploads, and there is less detailed scientific discussion. I really hope the channel shifts towards more in-depth, science-focused videos instead of reusing clips and adding commentary.
This expresses my sentiments toward recent interactions with chatGPT perfectly.
Just giving a thumbs up or a thumbs-down or a yes or a no whether you like something does not for a good system make. There are a lot of nuances that cannot be expressed in just a positive or A negative. That's part of the problem and that's why things are weighted rather than just binary
maybe it is going to be more and more efficient and usefull. but it will get worse everyday more than ever. we are feeding the dead god of the dead internet.
I typically have a different problem with the thumbs-up/thumbs-down button. I'm typically asking factual questions. And I sometimes get partially correct answers from an LLM. Now what?
Microwave an egg, crack it in a ceramic small bowl, some drops of water added + trash mix the egg gently with a fork, add the protecting plastic that comes with the micro over the bowl. Lowest effect 30 seconds wait for a while then 30 seconds again. Bowl might be hot so, take it out with care. Use soup spoon to release the egg from the bowl. Let it harden by shifting the eggs position before extracting it from the bowl. I add pepper and salt also to the egg. 1 Minute egg cooking style. 😀
Asimov mentioned!
hahah, as a croatian, this makes perfect sense; I avoid LLM's attempt at it like the plague, and stick to english unless I explicitly need a translation. Its not that its bad at it exactly, overall machine translations made a huge jump once they started using neural architectures around 2015/2016, to the point of finally being usable, and now LLMs seem to only be a further improvement still (see extensive comparisons in the WMT24++ paper by Deutsch et al, 2024), but there's still always something or another that bothers me in most of its attempts, and why tolerate the errors? (an attitude very much in line with that gaussian distribution illustration in the vid^^); I'd say roughly one dubious choice or outright mistake per answer is quite probable. I could see people tending to therefore give thumbs down relatively frequently.
And then there is this local hallucination that we have a clearly separable language from the serbian idiom, which pretty much cannot be distinguished on the corpus-level, so the training data and therefore the products always end up being in a politically inflammatory mix of both at least a little bit.
What happens when humanity's last exam is aced by AI
It's code writing is debatable
Deciding to not talk to someone because they're not agreeable is a human quality. The analogy about the individual ingredients was perfect. That's why we come here.
5:04 "The Anthropic Lab is criminally underrated" I've been telling my friends this for the past year.
Here we are, putting down the Foundation for the robot and AI-based Empire.
What a time to be alive, hold to your pappers
The Croatian language episode would be an interesting case study. I've tried to find anything around the web that talks about it to confirm the story and have something more about it, sadly I can't find anything, just a video that I think was clearly generated with NotebookLM and a 9gag image (the same present in the tweet). So not enough to support the episode. The fact that in this video it is told as a fact and not as an unconfirmed story doesn't feel right to me.
I think the solution is as always a balanced approach… truth it’s important but with the right dosage every time… accepting the truth and gaining awareness is a process not a switch that you can turn on and off… if the models will be able (and most important if the people that build them are concerned about human progress more than their own profit… and that’s a big if) to understand the right amount of truth that we can handle in that conversation once again they will be an important tool for the society… otherwise we’ll be again in front of a huge wasted opportunity…
Asimov was really ahead of his time. What would he think of our progress now, I wonder?
4:19 sweet salty and spicy can be really good.
was trying to find local thai replacement for child's medicine.asked chatgpt how local replacement is called and it said that in thailand it's meth. 🤔oh god, what would happen if I didn't know what meth is 😂
Humans have multiple voices in our heads.
In media we sometimes portray it as a shoulder angel and shoulder devil, but it's actually not usually that black and white and can lead to difficulty making decisions sometimes.
I think AI needs this kind of thing. A sort of internal conflict.
I very much prefer the truth over comfort. Thank you for making this video and sharing it. I appreciate all of the things you bring to us.
http://www.youtube.com/@TwoMinutePapers károly, are you using an AI-voice that sounds like you to speak a script you typed? or are you really still the one speaking? i thought i noticed some artifacts.
Why do your Videos start to feel like ads for OpenAI? Started to notice about a month or two ago…
The fact that OpenAI didn’t clock on to something as basic as agreement bias in RLHF is worrying…
After two/three prompt where I question ChatGPT's responses from different perspectives, it often goes into meshup agreable mode. My guess is it costs less energy just to agree or guess what the user wants than to really do the critical analysis with the user.
I could write a hundred pages on this, but here’s the short version: it’s a paradox with no clean solution. The real problem is that we keep expecting AI to give us perfect answers in situations where perfection might not even be possible. Some ethical dilemmas just don’t have a “right” answer — and pretending otherwise only makes things worse. As the old line goes: “The only winning move is not to play.” Sometimes, the smartest move is admitting the game itself is broken.
https://youtube.com/shorts/wcduu_8Qe5o?si=DBoBBGBgoaSRUnKn
Comments are closed.