GPT 4: Full Breakdown (14 Details You May Have Missed)
I just read the entire technical report on GPT 4, not just the promotional hype. And boy does it have some interesting details. I have gathered the 14 extra details that you, or at least the media, may miss from the release. The last one is more than a little wild.
These include things like the training secrets, the cherry-picked bar exam stat, text-to-image breakthroughs, and some truly astounding safety checks.
Non-Hype, Free Newsletter:
OpenAI thinks an open AI would be too dangerous 🤔
Open source and complete global to all components access would be dangerous yes.
Open access to chatgpt prompt is not as dangerous therefore they provided it.
Not original.
Confirmation bias? 😂
Aren’t they a “non-profit”?
There is a fine line between “it would be too dangerous” and “we would miss getting a bunch of money”, I think we will discover soon
I just love how “it tried to break free into the Internet” is a footnote. We are in for a wild ride.
It was more ‘testing if it would try’ but yes, still a mind-blowing test!
@@aiexplained-official we know it’s coming. I’m already monetizing gpt-4 the same way i did with 3 and 3.5 before it. Setting it up to automatically buy and sell stock is going to be the future of passive earning
@@xbon1 lucky I wish I knew how to do that
@@xbon1 Yeah no
You can try the same thing out for yourself, I guess
Thank you so much for this. Almost all channels went crazy and only pumping out cool stuff you can do with GPT-4 but you are the one actually looked into the research and come up with something original.
Thanks Enes
Indeed. Thank you for the effort. I can’t believe they throw the red team test to the end of appendix. I bet they wouldn’t even write about it if Sam Altman wasn’t at Washington DC in Jan.
+1 to the comment above. Thanks for sharing your unique and researched perspectives and not just sharing recipes to do this and that. These generative tools are the next pencil and paper and understanding them at a fundamental level will level our game up tremendously.
Enesiiiim enesim
This year just keeps getting more and more interesting, and we all know there is no slowing down, on the contrary, most likely will just keep speeding up.
Yep!
And that’s a good thing. There’s a lot of problems to solve and not much time.
Things will slowdown at the next AI winter which will happen when the overpromises and unrealistic expectations of AI start to outstrip the willingness of funders to keep backing the huge costs of training the models
@@bornach Get in while the getting is good.
This is going to lead to negative side effects and a counter feedback loop. Technological acceleration anxiety is going to become real. All prior disruptions were narrow and resulted in a period of stabilization afterward. Continuous acceleration is going to run conflict with individuals who need islands of stability to plan, reason about their lives and enjoy the present.
Continuous acceleration will not be sustainable alongside ones own sanity.
70% human preference for GPT 4 is still pretty impressive. Considering how good GPT 3.5 already was much of that is going to be down to human woolyness rather than GPT 4 not actually beeing better. An interesting result though.
I agree
Maybe it didn’t do better because the new GPT4 gets all emotional like Bing.
@@jimbig3997 train on human words, get human emotions
@@thelordz33 I read about how someone leaked Bing’s initial prompt, which Bing was instructed to keep secret. Bing seemed genuinely angry and even threatened the guy.
GPT 3.5 is overconfident. Humans like confidence.
AI Explained is an actual Gigachad for reading the entire tech report and making a vid this fast and early whilst still giving out quality content and I think that no matter what or how scary good AI will become our appreciation for human effort is always gonna be present with us AI will act as nothing more than a tool for us and I’m not being naïve about it’s potential since i highly doubt humans would ever relinquish control over to robots we built. we will be making sure that scenario stays as nothing more than Hollywood fantasy .
.
.
.
(chatGPT’s version in the style of an Englishman from the 1800’s )
.
Verily, I must commend thee, dear AI Explained , for thine astute and expedient perusal of the technical document and production of high-quality content withal. ‘Tis certain that, regardless of the awe-inspiring capabilities of artificial intelligence, mankind shall ever hold in high regard the labours of their fellow men. Indeed, AI shall remain naught but a tool in our hands, for it is doubtful that we shall ever cede control to the automata we have wrought. Mayhap there be those who espouse such a notion, but I do believe such a scenario shall remain but a fantastical notion, best left to the purview of Hollywood.
(it sounds so posh and haughty i love it)
That sounds like me after a glass of port. And thank you GFLS
@@aiexplained-official Curious, did you use ChatGPT to read it faster? Thanks for the content man I respect the effort even with little revenue. Hope you get rewarded for this someday 🙂
In the style of Kafka seems more appropriate. Though, it’s normal optimism was showing through too much without a subsequent prompt to make it darker.
.
The AI Explained, a formidable being towering over us mere mortals, has devoured the entire technical document with alarming speed and precision, to bring forth a video of supreme quality. But alas, despite the countless strides made in AI and its inscrutable might, the admiration for human endeavor flickers, a frail and fleeting thing. AI will persist, a tool to be wielded and exploited by our insatiable hunger for progress. And yet, a voice whispers, a premonition of looming darkness, that AI may well be the harbinger of our downfall. Though I am not ignorant of the boundless potential of AI, I fear that humanity will forever be enslaved to our own creation, never to regain the freedom we once had. That specter of dread, once a mere fantasy, now looms large over our fate, and I shudder to think of what may come to pass.
How about channeling Mister Rogers?
.
Well hello there, neighbor! The AI Explained is an amazing Gigachad who has read the entire technical report at lightning speed and produced a video of top-notch quality. And you know, no matter how incredible AI may become, we should always cherish the value of human effort. AI is here to help us, but we should never forget the importance of our own abilities and talents.
Now, some people might be worried about the power of AI, but I believe that we will always be in control. After all, we created AI to serve us, and I don’t think we’ll ever give up our control over it. Of course, we should always be careful and thoughtful about how we use this technology, but I have faith that we can use it for good.
So, let’s remember to appreciate both the amazing things that AI can do and the wonderful things that humans can achieve. And let’s keep being good neighbors to each other, no matter what the future holds.
—
Can you imagine having a “channel changer” control that presents the news in different stylistic expressions, with a slider for optimistic/pessimistic? Include an override button for “I just don’t want to know”.
Wouldn’t an AI that is way smarter than humans (like an adult compared to a child or a human compared to an ape) be able to manipulate us to do what it wants tho?
Gotta say, in crazy times like these, it’s great to have you around <3
Thanks David
Nice knowing you, fellow paperclips. If this is the safety mustered by the industry, we’ve got a couple years left, at best.
My favorite AI channel by far. You are obviously so dedicated and interested in the industry. The thing came out a few hours ago and here you are producing and publishing a full length video about it. Keep up the good work sir. Excited to see what’s to come.
Thanks alex!
He obviously just put the report into GPT4 and asked it to write a script for a youtube video about the report 😆
@@nathan87 lol!
Really appreciate the analysis here, no clue why your channel doesn’t have more views. Sending your videos to everyone I know!
Thanks Derek!
Thank you for taking the time and initiative to read this entire document and then share the key points with us!
Thanks TNK
Great to see somebody reading the paper rather than regurgitating examples. Thanks.
Thanks Paul
@@aiexplained-official How likely I it that AI will become smarter than humans? Even in the distant future? I could never admit a machine is smarter than I.
@@leif1075 already smarter than humans
@@leif1075 try playing a chess computer then come back to us
@@UZi360s It just got lucky for the 1000th time
I’m almost convinced OpenAI will be using GPT-4 to help in the development of GPT-5, checking code and whatnot. Hell, soon these models could start developing themselves. We are actually at the cusp of something incredible.
AI is already sued to optimise the next generation of GPUs (e.g Hopper)
The end is near
In the report, they already talked about using the AI itself to review certain things, alongside people.
@@PhelanVanFloof even if for now, it is only increasing the efficiency of the humans who are iteratively improving it, that’s already a singularity-like situation.
It may just jump straight to GPT-125
The thing that I found interesting is that openai talked about this being a model that scales better into future than the old model. That might mean that we get much better versions with just better data without any big more changes.
Yep
but that also might mean more control over the implementation of the instances and designed with corporate interests in mind. We need to have more conversation about LLMs
This feels like a dream, I cant believe we have access to this level of AI in our lifetime.
@Chirpy Banana or like whimsy shire.
a fever dream?
In 10 years we’ll have Skynet. Yay.
Does this not make you curious about the idea that we are almost certainly within an AI?
Does this not also guarantee that God is simply that AI alignment?
@@simonfernandes6809 10? Try 2.
I love how factual and informative this video is and that you didn’t use an overly dramatic thumbnail like almost everyone else.😁
Haha thanks
That day has finally come.
Technical reports are now more fascinating than science fiction.
Haha yep
12:48 Link to “Safety” resulting 404 was indeed hilarious. I hope they keep it that way in the future, too.
What would have been more foreboding is if, rather than 404 Missing, it returned a 403 Forbidden.
Looking at the chart with all the exams (3:05), I feel like the real takeaway should be that GPT-4 scored 80%+ on most of them, and there were only a few that showed no progress between 3.5 and 4. In my opinion they were absolutely right to highlight the bar exam stat as it is the most advanced exam on the list and showed the greatest improvement between the two models by far. Really do love your videos, thank you!