GPT 4 Turbo-Charged? Plus Custom GPTS, Grok, AGI Tier List, Vision Demos, Whisper V3 and more
Вставка
- Опубліковано 8 лис 2023
- A new AI Explained Custom GPT? What has the world come to. Well, let's find out, as we're gonna dive into everything from context window performance, Grok AI, Olympus, Gauss, Vision and TTS API use cases, the crazy new Runway Gen 2 update, the 1st AI Machine, and a fascinating new AGI tier list (with some questionable claims) from Google DeepMind. Plus, we'll start to investigate whether GPT-4 Turbo is actually smarter...
/ aiexplained
Dev Day: • OpenAI DevDay: Opening...
Runway gen 2 Update: / 1721904845601579437
Context Window: / 1
GPT 4 Turbo Smarter?: / 1
pbs.twimg.com/media/F-dXZtsaE...
Robert Lukoshko: / 1721777152658444773
WebcamGPT: / 1721991055829348537
Sports Commentary @geepytee: / 1721705524176257296
Whisper v3: github.com/openai/whisper
Grok: x.ai/
Amazon Olympus: www.businessinsider.com/amazo...
Samsung Gauss: techcrunch.com/2023/11/08/sam...
Physical Device 1stAI Machine: / 1720487169997824163
Cristobal Valenzuela: / 1721963131591692744
Infiniteyay: / status 1721858324839481843
/ aiexplained Non-Hype, Free Newsletter: signaltonoise.beehiiv.com/ - Наука та технологія
Honestly one of the only non-hype AI creators who does a deep dive into research and experimentation. Love your channel
Thanks so much Harry
This channel is also my favorite for AI stuff, but David Shapero and Wes Roth are also well worth watching, in my opinion.
You will also like twominutepapers
Yeah not like this David Shapirro on some cult level thinking.
Exactly. No garbage on this channel. Top-notch commentary!
You're my absolute favorite content creator specializing in AI. The quality and professionalism is unmatched.
Thanks faith
Agreed 💯
@@LevelofClarityl😮
😢I’m 😢n
@@aiexplained-official Deep Learning and GPT-4 are fake intelligence. For example, It struggles with fingers, and with drinking beer i.e. the details, the granularity. LLM are a dead-end for AGI because they do not [understand]! the implications of their outputs! Also, GPT-4 is designed by the Wealthy to serve their needs!
------
Douglas Lenat wrestled with creating the true AI . AI lost a giant.
He was not correct about some AI-things but he opened my mind about great many more things about AI.
RIP Douglas Lenat, 1950-2023
Thank goodness your video is finally here. No other channel compares in terms of quality and depth of coverage
Aw thanks julius
theres also 2minutepapers tho he uploads with quite a delay and doesnt cover everything
but yea this is the only one afaik
@@SBImNotWritingMyNameHere I watched almost every 2mp video. They are good, but not amazing. Lately he’s just covering whatever the news outlets are posting.
@@SBImNotWritingMyNameHere2 minute papers is sh it. The guy is only popular because he flatters his viewers "fellow scholars" and has some catch phrases. He has never analysed anything. He just copy pastes stuff from others showing some cool looking video clips. You learn absolutely nothing from his channel.
It sickens me he has so many followers when all he does is hype with no info.
@@julius4858His main strength in my opinion is covering the incredible stuff Nvidia is doing. I find it gets underreported. I guess it makes sense given that he's a computer graphics researcher by trade.
Babe wake up, a new AI explained video just dropped
Haha I love that meme
I swear, if you don’t put out more frequent videos, I’m cloning your voice and TTS whatever AI news content I can find. 😂
@@RyckmanAppsQuality over quantity my guy! Patience is rewarded
Can't appreciate enough how level headed your videos are
Truly a breath of fresh air, from how twitter and other social media sites treat AI, every time i see you upload
Thank you, Bonirin
@@aiexplained-official About your 13:36 logic. You committed fallacy of reasoning when you assumed that a that a "virtuoso AGI" will be able to communicate outside the scope of its virtuoso skills. It is like saying that putting a billion genius individuals in a room will somehow create great stuff. It is a fallacy of reasoning.
That chart at 2:16 gives an interesting image that there will likely be a battle in the coming months over how long to make a context window vs how much retention is found in the window that's already there.
It feels like there has to be some limit to how useful lengthening the context window will be. The neural net has a fixed amount of internal complexity, after all! GPT4 is already superhuman at absorbing information "at a glance" (since it has to do it anew for every token it outputs). If we figure out some sort of improvement to the LLM framework that lets it absorb information more gradually like we do, it's possible there'll be a huge jump in ability.
And..... it is done in 3 months....
Periodic reminder that this is the best AI channel on UA-cam. By far.
I thought the most interesting part of the dev day was when Sam was talking about how OpenAI wants to work with companies to make custom models and asked them to reach out.
To me that sounds like the start of AI in the workplace on a mass scale. We all know the most vulnerable jobs that will be lost to AI like customer service and it sounds like they are ready to try integrate into the business world.
2024 is when Google and Microsoft will roll out AI to most users.
So yeah, it's already started.
With Grok, Samsung, Apple, and of course all the other well known players such as Open AI, Meta, and Google, it looks like any kind of worry about alignment has been completely disregarded.
Baby, a new dawn for humanity.
I can't believe why this well informed channel shouldn't have at least 2,000,000 subscribers.
Aw thanks
I'm addicted to your videos mate. Keep it up!
Thanks Dentox
I love how calmly you deliver the big changes that are happening without hyperbole and click bait techniques.
Thanks sebby
30 years ago, the vision of the internet was originally about increasing knowledge and improving humanity. Now it's about 'increasing productivity'. That speaks volumes.
Productivity is more important.
Knowing shit is useless if you don't get anything done.
@@tbird81 You can't get anything useful done if you aren't smart. Otherwise you will become slave labor to a system that is smarter than you, and uses you as a tool to achieve its goals.
@tbird81 How can be productivity be more important as only knowledge can help determine to what end one wants to be productive?
@@kwetsbarevrijheid2720 Knowledge isn't necessary to be productive. Most people spend their lives at jobs where they do what someone else wants and tells them to do. Unless you own your own business, everyone making a salary is being productive in a way that benefits someone else more than it benefits that person. (Because that's the only reason anyone ever hires anyone as an employee, i.e., because they make more money on your productive output than what it costs to them to pay you.) But that's what's scary about the future of AI, because AI will be (even) more effective at capturing people's labor than current systems.
@@kwetsbarevrijheid2720 Because in the end, all the ideas in the world don't get stuff done. So many geniuses think they have the answer, yet do nothing.
In the end, stuff has to happen. And making stuff happen is the hard bit.
I experimented with uploading all the data about a fictional world to create a GPTs agent and I'm very pleased with it, it answered some tough questions involving a lot of synergy and speculation. It's not ideal that it requires a delay to perform a search on its knowledge base and ingest the relevant knowledge, as opposed to a fine-tuned system that directly incorporates the data in secondary training, but it's vastly more scalable since they can use a single engine and support very large knowledge bases. I'm a big fan!
Your videos are like us watching the cracking of the engima machine during World War II - I'm always at the edge of my seat
Thanks gino
As this time the video was comparably late to cover the news, still yoit did not disappoint. You really understand what to put the focus on, pick on the interesting remarks and put everything into a wider context. Every single video is worth the watch. Keep up your amazing work!
I’ve noticed gpt4 opts for scaled down answers that are less verbose and you really have to persuade it to give a full featured script or solution when coding. Its defaults are to give you a skeleton framework and ask you to fill in the rest. That could be how their providing more availability and higher rate limits but limited the response tokens back by default
I cant wait for the real tests you put it through, they are very enlightening and unbiased. Thanks for doing this, I look forward to every video you put out. Never any fluff, just the facts.
Thanks Mad
Thank you for the great summary! ☺️
I'm excited for the AI Explained GPT, I hope you will publish it soon! 😊
Thanks Roy
This channel really is the best for actually learning what's going on in the field and not people throwing a bunch of hype and face time at you for the sake of making a video.
Was expecting your video on this news Philip. Thanks again for simplifying complex videos like this for us 🙏🏿😅
Kinda wished you opened this video with "Less than 24 hours ago," because everytime i hear you say that in an AI video, my neurons activate so hard
Less than 1 minute I replied to this comment
You’re one of the only AI info UA-camrs I trust ngl.
The most important thing about those new GPT bots is the ability to set a specific instruction that persists throughout the chat like the system prompt (or so I think it works).
I have used chatgpt for help with a new language, but after a while it forgets your first message containing your instructions. Like your competency level, or stuff like that. I don't need every single chat to know that so putting that in global user instructions is not handy. This new update is great for this
Amazing update video, as usual, thanks 🙏🏿
Thank you again for your excellent up-to-the minute review of recent AI developments! Your channel is always my first port of call for anything AI related. I appreciate your hard work!
Thanks so much Stephen
I always appreciate your attention to detail, and not taking hype at face value. Keep up the excellent work! Thanks again, Space
Thanks Space
Thank you for your regular udates! The are really valuable!
Wow thanks Jan, super kind
Thanks! Excellent content as always. 🙏🏼
Yessss been waiting for this after Dev Day!
1:32 I dont think ive ever heard such a shocked gasp of amazement in a scientific conference like this. Truly groundbreaking
Its rare that I recommend YT channels but your channel I tell all my friends and colleagues. Amazing work!
Thanks so much endgamey, means a lot
Hands down the best AI updates channel, no clickbait and straight to the point.
Thanks so much Skrock
At long last. Have been waiting for your video since Monday! I'm very happy. Thanks Philipp. Will watch it right now!
Thanks thales
BEST AI CHANNEL ON UA-cam, NO HYPE , straight to the point , answering the questions of whats new in ai and why should we care.👏
Things keep getting better huh. I mean, I'm certainly not complaining but yeah, good thing we can count on this channel to explain it all as quick as possible. Without it, I'm really not sure how one is supposed to stay in the loop
There are others but very grateful for your support
@@aiexplained-official I mean sure, but you are quick, reliable, you explain well and you keep your head cool. Also your videos are just the right speed/length, enough "commentary" but not too much.
There are others, yes, but I doubt I'd find one that surpasses yours.
I wonder if the prompt for the music in the end was "Flying Lotus".
Anyway, great video, as always!
Thanks Semi
The few shot of gen2 took my breath away
So hype to see it getting upgrade as much as image generation
Thank you, very informative, as usual.
"one or two things happening" ಠ_ಠ
OMG I GOT A SHOUT OUT ❤
I immediately stopped the video to look for your response, I mean, I'd say that even being just mentioned by this man is +20 to relevance on the broadly accessible AI info stage
If Copyright doesn't mean much anymore, then that's the biggest concern and the most important point.
Without protection of intellectual property rights, it basically becomes legal to steal from creators.
And a society can't run well on that basis.
The best news form me is new APIs - now I will be able to build multimodal agent :) thanks for the update
The thumbnail of your video is right on the spot of today’s news ! How the hell did you anticipate this ? 😂
He's waving goodbye.
Info packed! One of your best videos ever
Thanks ce
Just leaving a comment for the algorithm and to say that i appreciate your videos as always, thank you
Wow. That ending is pure Art. AI is amazing.
Great work, once again.
Your work is exceptional.
Its not helpful for me to continue praising your work as It speaks for itself.
Just know that I support you, regardless if you ever make a mistake.
Just learn from them and continue on in the direction you are going.
This world needs more good reporting and you are well on your way at working towards setting the bar.
Take good care,
Jeremy
Thanks Jeremy, much appreciated.
One of the best channels for ai news.
Thanks coal
THE best, name a better one
Another great one, Frank!
Thanks Cornelius!
Another amazing video. Thank you Mr. Explained.
Thanks Elijah!
You can call him AI by now. :)
Great stuff. I don't have access to making GPTs yet, but I want to see if this takes off, particularly since they are offering revenue share for developers. BTW, you a big Tolstoy fan then? I read Anna K last month and have actually just started Ivan Illych. Weird coincidence. Anyway, keep up the good work!
Yeah love Tolstoy, always have. Let me know what you think of the The Death
After an excruciatingly long week, AI explained has made a new video!
Haha
Is there any explanation why of all languages “Dutch” is preforming so good in WhisperAI? I know a Belgian Dutch company was at the fronteer of computer speech technology, maybe OpenAI « borrowed » some of their data? Or are we dutch speaking people spied on more by our tech? Or is it just pure random?
Wish you would post more often man. I try to hold myself over with other ai channels but nothing comes close to this.
Aw thanks man I know, me too
You're the GOAT of AI youtube. Love your channel.
Thanks Fran
@@aiexplained-official
Im really excited about the GPT agents idea, I'm looking forward to trying it out and I'm jealous you got it early lol, good that you did though to test it out for us all.
First thing I'm gonna do is make a biblical scholar gpt for my dad to use, I'm not religious but he is and I think he'd appreciate having a more accurate version of what he already does with chatGPt. Plus it could be interesting to experiment with different instructions to see what different interpretations of the quotes the agent can give, maybe even upload some more historical knowledge and context for it to be more useful.
Even just that idea is a little scary to me though, feels like something I could abuse and so easily spread misinformation or force my worldview onto him if I bias it too much, or if it has its own bias baked it. Hopefully I can make it mainly a quote finder, then give several different interpretations and context afterwards, but it does scare me what others will do with this tech.
Very wholesome.
BibleGPT, or maybe even Abrahamic-religious-mythsGPT could be very interesting.
Imagine you could ask all major religious books for advice 😂
I'm not Christian but that seems a great use case. Good luck to your old man.
Thanks for another informative video.
Thanks John
i don't think it's ENTIRELY the first time ChatGPT knew about itself. i'm pretty sure OpenAI told it things about itself as it already had the ability to reflect on itself as a non-human non-conscious pattern predicter made by OpenAI from the start. but still very interesting how ChatGPT will get to start seeing a more developed picture of itself as it impacts the world.
I wish I could wrap my brain around this and the implications. It feels pivotal.
thanks for the latest info!
Thanks Tristan
@@aiexplained-official shared with my team! Our open ai account just got tier 5 access so we can process 100k TPM. Should do the job for now
"... it's how you use those parameters", is a line that's worked for me.
I've been using 3.5 and it's so good at understanding.
Your content is excellent. While you seem to genuinely love AI, you're noones fanboy.
A very sober, and straightforward perspective, without bloat.
A bit late but, subscribed. Please keep at it, your perspective is very much appreciated.
P.S. I work in a shop with bluetooth headphones and ear muffs- loud audio is a must. Can always turn things down, can't always turn things up. Your threshold is perfect.
Yay! Another ai explained video!
Excellent video as always. I've found that the GPTs don't quite perform as well as I want them to, but the resulting agent does way better than how ChatGPT did 6 months ago at the same task. So we are making rapid progress, for sure, but we are starting to be spoiled for progress I think.
The best channel for AI developments. Automatic click whenever a new video drops
Thanks nessen
For real, I love your channel. As someone who's been working a lot with AI, your channel is one of the best for stsying up to date. So much better than any of those hype beast crypto bro AI channels out there.
Despite all that is going on in the world, I still believe that this is a bigger story. Thank you for your consistently amazing research and presentation.
Not gonna lie, I was waiting for you to drop a video for days after open ai gpt 4 turbo reveal.
Your videos are a must. Thanks!
Thanks paco
Im paying for gpt premium but I dont see anywhere about creating my own GPT? is this like an early access only type thing right now? am I missing it?
I was skeptical about GPTs but your example showed it does indeed work (when compared to base GPT-4) at least for narrow domains. Did you try the custom actions that presumably allows your GPT to talk to other APIs? I think coupled with the new multiple function calls it can allow GPTs to use tools that base GPT-4 may not have access to.
although that's a good start for defining AGI and ASI and all of that, I think it leaves a lot out. In my opinion, you need a level of autonomy, a form of longer term memory, and the ability to work across multiple apps while doing that.
If you even took just GPT-4 and it was able to remember while working through multiple apps while both taking commands and taking its own initiative, I'd consider it an AGI.
ASI is a bit more difficult, given that once you create something you could just replicate them and scale like crazy. So, I'd probably look more at the results. I'd consider it an ASI once AGIs start not only inventing, but inventing things at a pace and complexity that humans can't readily track down how it was done, or at a quantity that starts to overtake humankind's output.
Great update.
It's always interesting to see Hungarian examples in videos like this. The most exciting about the Hungarian maths exam is that GPT-4 outperformed the average high school student... it's truly fascinating.
Love your videos❤
Thanks Totiius
Encara Messi is my only benchmark.
The context length related paper was eye opening. The accurancy claim of 1104 version by Sam Altman was doubtful and was proved correct. gpt-4 turbo is nothing but a quantized version of gpt-4
Do you know if there was any analysis done of the hallucination rate (where the model is certain of incorrect information from a document) as context size/document sizes were varied??
It's neat seeing how fast everything is moving!
Exciting and scary
That 1st AI Machine is exactly like something out of Black Mirror. Retro and futuristic all in one
What a time to be alive 🙌✨️
Indeed
I read this in the 2-minute paper voice (Dr.Károly Zsolnai-Fehér)
you are ill with mental virus
Sam Altman calling the current tech quant gives me chills. So excited to see how these developments change the world!
Thanks! ❤
Thanks Dan
cool video! could you please share the links to the GPTs you created?
I have noticed today that GPT4 with vision has a hard time after uploading photos in the middle of the conversation to look back in the previous chat. It kind of completely gets lost, and struggles to add the information from images to the rest of the context. Interestingly would, this problem also persists in GPT4turbo?
GPT4 Turbo is a big step back imo. It is much less creative than GPT4 and struggles with nuance that GPT4 didn't. The result? Even though it's cheaper it takes 5x more generations to get something approaching decent for say, RP or narrative work. So so far it's currently costing me more per day than GPT4 normal does. So I've gone back to GPT4 in the API.
didn't notice anything bad
@@HoD999x Depends on the usecase I guess. I mainly use GPT4 for User guided ghostwriting (I describe scene, theme, events in prompt with story relevant content and example prose style in System prompt and go scene by scene) and simulation of detailed lifelike dynamic personalities and turbo has been outputting... well, generic text to put it bluntly.
With GPT4 I was able to reliably replicate Prince of Nothing quality prose and nuance. Turbo produces YA quality prose, and struggles to understand and implement dialogue and actions that implies subtext (of what I mention in prompt), instead it just tells and shows everything. It can't grasp complex nuance or situational awareness of characters in scenes or how to imply themage without outright stating it.
It also always takes the most obvious generic path. It never surprises you with adding in something creative that your request implied.
Strongly agree, GPT4-Turbo is horrible at writing code. Almost as bad as GPT3.5. As my primary use-case, it is frustrating. Planning on switching back with the API as well.
@@CreativeFunction It's slightly worse in my experience, but still way better than GPT 3.5 or even Github Copilot
@@HoD999xYeah, but that's on you. It's quite clearly worse.
Agreed: Skilled is such a loose term, unfortunately it seems to fit with the lack of precision benchmarking, I guess these are the less "sexy" aspects of AI dev atm. Purely subjectively, I would regard skilled (technical) as a graduate with 2-3 years applied experience in that field. But of course that would vary greatly on who you asked and what field they worked in. Thanks for your continued great coverage of the state of the industry.
Even then, some people are going to be better than others and for different reasons.
Whisper V3 takes longer to process on my old Quadro P5000 GPU. Now it's barely faster than real-time, and it still takes around 10GB of video RAM.
Oh really, v interesting
Oh man. ERP with GPT-4 is gonna be so much better
Best AI channel by a mile. ❤
Aw thanks ish
Thanks!
Thanks so much Ryckman !!
In my own personal experiments, I've found that the GPT-4 Turbo preview is demonstrably worse at System Instruction following, sometimes in pretty serious ways. For example, returning many examples when "return only one" is specified, and doing weird illogical combinations of state that are difficult to explain in a YT comment. I'm hoping it'll improve as they iterate.
Please note I'm referring to doing experiments inside their Playground using some of my saved prompts for GPT-4, and I'm referring to Chat mode rather than Assistant.
I think in terms of how GPT gets smarter, it’s not a straight line going diagonal up, it goes more like an S going up. So there’s gonna be drawbacks at times as it improves, but then those will be fixed, and it will get smarter over time
@@Wanderer2035 Yes, and the current GPT-4 Turbo model is explicitly labeled as "preview", so I'm not drawing any strong conclusions just yet.
Gezellig!
curious - I have GPT plus and I still see the model picker - is there another tier of chatgpt i can use?- i really liked your prompt to compare the market caps but I still see the plug in list etc.
Yeah that was me 3 days ago, shpuld update for you in coming days
Im so gonna enjoy this video with a nice cuppa coffee ☕️ ❤ thank you for existing 💖
5:17 I've started my own investigations, as have others, and the results aren't uniformly better and in fact for some use cases GPT4 Turbo performers worse. E.g. SAT reading. There have to be more investigations, but I (AI explained) wish they released some benchmarks along with that claim (Open AI and Microsoft). And there were 2 moments from the guest appearance of the CEO of Microsoft, Satya Nadella, that I (AI explained) found particularly interesting. First he promised Open AI the most computes: "The systems that are needed as you aggressively push forward on your road map requires us to be on the top of our game and we intend fully to commit ourselves deeply to making sure you all, as builders of these foundation models, have not only the best systems for training and inference but the most compute, so that you can keep pushing forward on the frontiers because I think that's the way we're gonna make progress." (Said Satya Nadella somewhat hastily I guess, also based on what AI Explained says next)..
6:36 Grok 1 from xAI: on these benchmarks it did seem to be somewhat firmly beating GPT 3.5, the original Chat GPT, but notice that on these benchmarks it was falling way behind Claude 2, that's on the MMLU (Massive Multitask Language Understanding) testing for general knowledge, HumanEval for coding, and GSM8K (Gaussian minimum-shift keying) for mathematics. They also gave this human graded evaluation comparison on an exam produced after all of these models were trained, and on that Grok 1 outperformans even Claude 2 and way outperforms GPT 3.5. Etc etc
7:27 Amazon is building its own LLM called Olympus and apparently it's gonna have 2 trillion parameters, that's twice the size of GPT4, but according to what I've seen it's more like 10% bigger. But it's not about the parameter size, it's about how you use those parameters. It doesn't mean that the Amazon model will be more powerful than GPT4.
7:49 even Samsung are getting into the game with a model that can produce text, code, and images. That's apparently going to be incorporated into their next round of phones..
8:01 long story short GPT 4 Turbo is not better than GPT 4 and Sam Altman seems to be hyping things up for money and power (my own understanding, not AI explained's words.. I want to make that clear before someone starts attacking his wonderful analysis)
8:37 however this was a pretty good outcome for GPT 4 Turbo to create a picture telling everyone that Apple is bigger than Microsoft :D
I absolutely love the machine!
Would you expect Twitter trained data be good at math😂😂😂
Haha
I absolutely love your deep dives into the subject matter. And yes, I was honestly kind of disappointed with how the standard GPTs have been performing. I should probably make my own...
Something I've noticed using 4 Turbo is that it's seems to be ignoring my chain of thought prompting. At the very least, it doesn't even acknowledge my desire to do something step by step. I don't know why that is...
I think they programmed CoT into it
I don't really see the point of most of the GPT's, similar to how I haven't seen much of a point for plugins, for consumers.
Most GPT's are extremely niche, so it would be annoying to find the right GPT for your problem, it's easier to just use the broad version.
I only see GPTs being beneficial when you get to the end game of AGI where there's a demand for certain things to be done in particular ways for businesses.
yeah for the vast majority of use-cases, just prompting GPT-4 right is enough. The only reason I'd want to have a different model is if it can do something that GPT-4 can't do (e.g. is less censored, or performs better on something specific, like Claude having certain advantages). Otherwise I just use plain GPT-4 as a Swiss army knife and don't bother with anything else
I tried the Assistant, If you uploaded a document and started Assistant, the file tokens are sent on each prompt.
not sure what's the point here?
huh? that doesn't seem right. how do you know that. you can upload gigabytes of information. sending on each prompt would be impractical. Also, I am not seeing this.
I'll try it again later when I am at the computer. If you are right, that would completely suck.
okay, I tried it. The system prompt is sent, but not the uploaded files.
@@jumpstar9000do not send anything else, wait like an hours then check the Activities tab under Usage menu, that's what happened to me, it took a little while.
I had 20k tokens usage for only 5 1 line questions. my file tokens were about 4000
07:40- 07:44 made me chuckle 😅😊.