Someone explain to me how this isn't AGI yet? My definition of AGI is a system that passes the touring test and is better than the majority of humans at most intellectual tasks. Break it down: artificial check, general check, intelligent (PHD level not enough for you?) check. We have that. What's really exciting is that we got something as powerful as GPT4 by using all the data on the internet, even though the internet is missing a lot of the step-by-step chain of thought internal reasoning that us humans engage in before we learn something for the first time. There was still enough of that data floating around to create the illusion of above average human level reasoning (most of the time, sometimes the illusion breaks). Now they have a model (strawberry) specifically designed to generate this custom data to train the next absolutely massive model which will make existing models look tiny and stupid. That model will be ASI in my opinion. You could argue that the watered down version of strawberry is ASI, just not extreme level ASI. How many humans do you know that have PHD knowledge and reasoning across as many domains as strawberry? I was on the fence earlier this year due to concerns about running out of quality data and needing to rely on synthetic data to train future models. I'm not an expert on AI, so that example of diffusion models degrading due to synthetic data concerned me for a bit. After that, I learned correctly curated synthetic data would work (I strongly suspected it would work better). Now we have proof of that, which is the last bit of evidence I needed to know that ASI is within reach. By the way I find it hilarious that people think strawberry is slow because it doesn't spit out an answer immediately lol.
It's not 100% accurate at anything yet, for if it was, it could solve many many MANY of the world's problems. We still need another MAJOR innovation in AI before AGI to be achieved. Shortly after, ASI will follow.
Great take imo. I think the main counter argument is the reliability but if you consider what they will be able to do with the training data that everyone is providing them with now (using o1) it might be less of an issue as this recursive loop seems to iron out the inconsistencies in LLM outputs. Time will tell if they can get it to 100 % but at this rate of progress I wouldn't be surprised.
The use case for the 0.1 preview is to raise venture investment in the upcoming round. It takes five minutes of exploration to see that the reason we aren’t getting much under-the-hood technical detail is that they haven’t resolved the underlying glass jaws; they’ve just trained around issues that are public. So it can count the letters in a word like "strawberry." Does that mean they have solved the underlying problem? The answer is clearly no, as demonstrated by this interaction: ChatGPT 0.1-preview “List 10 countries with only 2 occurrences of 'a' in their name” Thought for 23 seconds Here are 10 countries that have exactly two occurrences of the letter "a" in their names: 1. Algeria 2. Angola 3. Argentina 4. Armenia 5. Belarus 6. Colombia 7. Denmark 8. Iran 9. Iraq 10. Japan You can see 5/10 names are wrong in the response. It’s trivial to find other examples. This release has the feel of desperately bolting together a bunch of pieces in a Hail Mary attempt to take the “lead.” I’m still waiting for all the originally announced 4.0 features and Sora to be released. The hype machine is outdoing itself with this preview.
I'm super curious how you guys who keep crying about voice assistant get through life. This is a super intensive, incredibly complicated, breakneck industry at the moment. Fricking chill, learn what you can, use what you can, and enjoy the features as they come out. It's super weird that you're this stuck on thing thing.
sorry if this is a silly question, i use chat gpt daily but im not very technical. ive created a couple of gpts that I use for work, can I choose which model the gpt uses? ie 4o or 1o, how do i switch between the 2 when using my gpt
We need big advances in several disciplines before full dive VR: Neuralink (for brain/computer interfacing) + Fusion Energy (to power all of this advanced technology) + Quantum Computing (classical computing can't simulate real-time Quantum phenomena) + Nanotechnology (to monitor and negate any negative effects to our physical bodies, when our brains are fooled, e.g., so that falling off a cliff in full dive VR won't give your real body a heart attack). + AI (to power all of the NPCs and plot lines of whatever situation you request). This is probably the tech that we're actually closest to achieving.
I tried the one to generate one single file game html+javascript, i asked for a fps with movement and a detailed prompt, and one shot. I went to ask it some frontier research topic from quantum computing, and decide which promising research use for a novel paper, to explain the math and the steps to achieve the goal to the point to generate a paper and working code. I know that chatgpt2 did that, however. this was different, like the formulas made sense, the code was functional, the ideas or mixes between disciplines. I think this is the death of the pHD student as we know it.
I think this frameworks should use chat GPT-o1 for planning and designing the app and use claude 3.5 sonnet to do the actual code because it's just cheaper and better.
Thanks for the video... The model needs to leverage the thinking fast and slow system 1 vs system 2 frameworks to scale. Interestingly, came across a custom GPT that is already operating at this level of reasoning? Is this as a result of its custom instructions? If so, how can this be possible given the limitations inherent to GPT builder? Highly perplexing... these are certainly interesting times!😅
There is a surprising amount of things you can do with just custom instructions. The entire o1 model seems to be a set of custom instructions that were fine tuned into the model so it shouldn't be surprising that even within a GPT you can achieve interesting results even though they are still quite primitive.
These models from GPT 3.5 to o1 still stuggle with basic addition and subtraction that involves more than 20+ numbers... this is not limited to GPT, Claude struggles too.
Do you not hear yourself talking? Mumffs(months), furteemff(13th)???😂😂😂 I hear you say “things” correctly throughout the video but did catch a “fings” in there. How do you decide when to use “f” in place of “th”?😅 Apologies if I’m insulting you, but it’s not meant as an insult. I’m still enjoying the videos 👍
Sorry but o1 only is impressive when you test like it is some tool. The reasoning bs takes too long and it still sucks horribly in real world intelligence, not riddles and math and science. In fact it literally acted like an idiot and I could not even use it for more than 5 min, which was only good for like 5 prompts since it took so long just to say a stupid answer. They still got a LOOOONG ways to go... I guess I will just have to cerate the real one myself... stay tuned....
That UA-cam thumbnail felt like open ai official video
Omg I actually thought this too
The miniature monsters - I want the monsters! NOW, they are just SOOOO cute.
Someone explain to me how this isn't AGI yet? My definition of AGI is a system that passes the touring test and is better than the majority of humans at most intellectual tasks. Break it down: artificial check, general check, intelligent (PHD level not enough for you?) check. We have that. What's really exciting is that we got something as powerful as GPT4 by using all the data on the internet, even though the internet is missing a lot of the step-by-step chain of thought internal reasoning that us humans engage in before we learn something for the first time. There was still enough of that data floating around to create the illusion of above average human level reasoning (most of the time, sometimes the illusion breaks). Now they have a model (strawberry) specifically designed to generate this custom data to train the next absolutely massive model which will make existing models look tiny and stupid. That model will be ASI in my opinion. You could argue that the watered down version of strawberry is ASI, just not extreme level ASI. How many humans do you know that have PHD knowledge and reasoning across as many domains as strawberry? I was on the fence earlier this year due to concerns about running out of quality data and needing to rely on synthetic data to train future models. I'm not an expert on AI, so that example of diffusion models degrading due to synthetic data concerned me for a bit. After that, I learned correctly curated synthetic data would work (I strongly suspected it would work better). Now we have proof of that, which is the last bit of evidence I needed to know that ASI is within reach. By the way I find it hilarious that people think strawberry is slow because it doesn't spit out an answer immediately lol.
It's not 100% accurate at anything yet, for if it was, it could solve many many MANY of the world's problems. We still need another MAJOR innovation in AI before AGI to be achieved. Shortly after, ASI will follow.
Great take imo. I think the main counter argument is the reliability but if you consider what they will be able to do with the training data that everyone is providing them with now (using o1) it might be less of an issue as this recursive loop seems to iron out the inconsistencies in LLM outputs. Time will tell if they can get it to 100 % but at this rate of progress I wouldn't be surprised.
@@aiadvantage Do you think they need to get to 100% to be considered AGI?
@@henrismith7472 hmm no probably not. just like no human is 100 % either :D
The use case for the 0.1 preview is to raise venture investment in the upcoming round. It takes five minutes of exploration to see that the reason we aren’t getting much under-the-hood technical detail is that they haven’t resolved the underlying glass jaws; they’ve just trained around issues that are public. So it can count the letters in a word like "strawberry." Does that mean they have solved the underlying problem? The answer is clearly no, as demonstrated by this interaction:
ChatGPT 0.1-preview
“List 10 countries with only 2 occurrences of 'a' in their name”
Thought for 23 seconds
Here are 10 countries that have exactly two occurrences of the letter "a" in their names:
1. Algeria
2. Angola
3. Argentina
4. Armenia
5. Belarus
6. Colombia
7. Denmark
8. Iran
9. Iraq
10. Japan
You can see 5/10 names are wrong in the response. It’s trivial to find other examples. This release has the feel of desperately bolting together a bunch of pieces in a Hail Mary attempt to take the “lead.” I’m still waiting for all the originally announced 4.0 features and Sora to be released. The hype machine is outdoing itself with this preview.
Thanks!
Very nice, but I am still waiting for the 4o voice assistant :(
Who cares. The stupid voice assistant won‘t solve science problems.
I'm super curious how you guys who keep crying about voice assistant get through life. This is a super intensive, incredibly complicated, breakneck industry at the moment. Fricking chill, learn what you can, use what you can, and enjoy the features as they come out. It's super weird that you're this stuck on thing thing.
Same
sorry if this is a silly question, i use chat gpt daily but im not very technical. ive created a couple of gpts that I use for work, can I choose which model the gpt uses? ie 4o or 1o, how do i switch between the 2 when using my gpt
When will all the tech be implemented in virtual reality so we can put our minds in different dimensions
We need big advances in several disciplines before full dive VR:
Neuralink (for brain/computer interfacing)
+ Fusion Energy (to power all of this advanced technology)
+ Quantum Computing (classical computing can't simulate real-time Quantum phenomena)
+ Nanotechnology (to monitor and negate any negative effects to our physical bodies, when our brains are fooled, e.g., so that falling off a cliff in full dive VR won't give your real body a heart attack).
+ AI (to power all of the NPCs and plot lines of whatever situation you request). This is probably the tech that we're actually closest to achieving.
What's wrong with this dimension 😄
Great idea with switching model when you are not satisfied with 4o 🙏 thx
Great video, as always, I genuinely look forward to these.
Thanks for the kind words :)
Games getting REAL😮
I tried the one to generate one single file game html+javascript, i asked for a fps with movement and a detailed prompt, and one shot.
I went to ask it some frontier research topic from quantum computing, and decide which promising research use for a novel paper, to explain the math and the steps to achieve the goal to the point to generate a paper and working code.
I know that chatgpt2 did that, however. this was different, like the formulas made sense, the code was functional, the ideas or mixes between disciplines. I think this is the death of the pHD student as we know it.
Crushing the recon for us
At your service
I got 2 usages before the full reset, first day used all my stuff then next day got reset usage.
Nice!
I think this frameworks should use chat GPT-o1 for planning and designing the app and use claude 3.5 sonnet to do the actual code because it's just cheaper and better.
I agree. Seems like for code generation Sonnet 3.5 is still king.
What!? I had no idea that Google Notebook had a major upgrade!
Love that app
someone tell runway to let negative prompts in gen 3 they simply wont listel tell them pls
Thank you for the great video!
you are my favorite youtuber that covers this type of stuff . keep doing the good work man ❤
*digital hug*
Great information thanks
Thanks for the video... The model needs to leverage the thinking fast and slow system 1 vs system 2 frameworks to scale.
Interestingly, came across a custom GPT that is already operating at this level of reasoning? Is this as a result of its custom instructions?
If so, how can this be possible given the limitations inherent to GPT builder?
Highly perplexing... these are certainly interesting times!😅
There is a surprising amount of things you can do with just custom instructions. The entire o1 model seems to be a set of custom instructions that were fine tuned into the model so it shouldn't be surprising that even within a GPT you can achieve interesting results even though they are still quite primitive.
I still haven't seen a use case showcasing complex programming projects with over 300 LOC
Strange, on openrouter I have almost unlimited o1 and o1 mini. Almost because there is some requests per minute cap, 50 or something.
These models from GPT 3.5 to o1 still stuggle with basic addition and subtraction that involves more than 20+ numbers... this is not limited to GPT, Claude struggles too.
Well o1 seems to nail addition and subtraction every time now no?
@@aiadvantage it doesn’t give it 30 numbers and ask it to add them or subtract them. Watch it reason and confidently return the wrong answer
Děkuji za skvělou práci🔥
Velmi rad
We really got Strawberry before Advanced Voice Mode
Sad but true
when do yall think we will achive fdvr or really realistic vr
Nobestudy intelligence version one is breakthrough
What's the difference between thinking and slow response?
Didn't this channel change its name to "David Shapiro" for a few hours? What the hell was that?
What do you mean?
And o1 mini is cheaper than 4o.
1o is way too expensive for the use cases you talk about
Martin Deborah Gonzalez Steven Allen Kenneth
Naozaj vieš po slovensky?
Samozrejme :)
I♥️NYCU
Devin is a scam which has been proven to not be real, wish OpenAi did their research before letting them be added on their UA-cam
you said 48 times AI in this video
You are probably right lol
@@aiadvantage hahah AI...... Love your videos bro
Do you not hear yourself talking? Mumffs(months), furteemff(13th)???😂😂😂 I hear you say “things” correctly throughout the video but did catch a “fings” in there. How do you decide when to use “f” in place of “th”?😅 Apologies if I’m insulting you, but it’s not meant as an insult. I’m still enjoying the videos 👍
Nah I always appreciate feedback like this. Will watch out thanks (or fanks) :D
Part of o1 seems a bit of a dog and pony show
Sorry but o1 only is impressive when you test like it is some tool. The reasoning bs takes too long and it still sucks horribly in real world intelligence, not riddles and math and science. In fact it literally acted like an idiot and I could not even use it for more than 5 min, which was only good for like 5 prompts since it took so long just to say a stupid answer. They still got a LOOOONG ways to go... I guess I will just have to cerate the real one myself... stay tuned....