The Mysterious Chatbot that Beat ChatGPT is BACK!
Вставка
- Опубліковано 7 тра 2024
- Let's talk about gpt2-chatbot. This is a test of a different style video. Quick and to the point about a single topic. I'd love your thoughts on it.
Discover More From Me:
🛠️ Explore thousands of AI Tools: futuretools.io/
📰 Weekly Newsletter: www.futuretools.io/newsletter
🎙️ The Next Wave Podcast: @TheNextWavePod
😊 Discord Community: futuretools.io/discord
❌ Follow me on X: x.com/mreflow
🧵 Follow me on Instagram: / mr.eflow
Sponsorship/Media Inquiries: tally.so/r/nrBVlp
#AINews #AITools #ArtificialIntelligence - Наука та технологія
Sam Altman also didn't say that it wasn't aliens ... the plot thickens
Good point! 🤔
I agree. Definitely aliens. I have much experience in attributing everything to aliens… so I know what I’m talking about! 😂
@@brandonliwanag136 I'm adding you as a valuable expert source lol :D
dont you watch PBS Spacetime??
its NEVER aliens darnit lol
Sam Altman also said that the chance of him being assassinated is not zero %. He is open to all possibilities; that's what makes him a good scientists and free thinker.
Yeah, I like this style of video. The 3 minute "here's what's up rn" video.
Agreed
I like the full length videos as well though because of all the information!❤
Very much agree. I think a healthy mix of 12 to 15 minute videos covering either a variety of Ai platforms and or a tutorial on those worth exploring further would be good to. Anything longer than 20 minutes and most of us feel although this is great, what platforms are we missing out on discovering by going into the weeds with something that could otherwise be truncated down to a smaller bite. Just my thoughts.
I agree as well, some of the other videos with many things sometimes have something I don't care about mixed in there. Here, just the subject I was interested in.
Believe
Love this style of video where I don't have to skip around and 20x speed to find the actual substantive info i care about.
Solid point. It's getting harder to find.
I think it is an Agent made out of 2 GPT-4 Models that work sequentially on a Problem. This is why it is a Bit slower in responding and the answers are better. Sam Altman said that Agents are the next step of the Evolution. And this is the reason why it is called gpt2, because the are 2.
You could be onto something. What if they combined 2 GPTs into a GAN? One GPT produces a suggestion while the other validates and rejects the output until it is of a sufficient level of quality? This is already done in training, but what if they got it to work operationally? I guess a GPT trained for evaluation only could operate very fast and even push a less accurate but also very fast generative model to GPT-4 like qualities. Heck, with the capabilities we see today, the adversarial model could even suggest improvements to the generative model.
It actually makes strange sort of sense given that there are TWO GPT2 chatbots on the Arena. Hmmmmmm....
@@trnogger I've been thinking about this and, granted im not highly clued up but trying to get there, I think a pre-output validation layer is the answer instead of a whole other model. My basic premise is that models act like filters and hence will always give an output leading to hallucinations. So you add this validation layer that compares all other values within like 1 sigma, takes those and sees if the original answer is more then 1 sigma 'better' then the others. If not it should flag it as insufficient info and ask for clarification instead of giving an output.
If only recently started studying data science and AI so I might be totally wrong here.
@@liverandlearn448 That is an interesting thought, but couldn't such a validation layer lead to a dead-end? If several answers are of similar quality that could mean that the model hasn't found a good answer yet, or it could mean it has found multiple good answers. In the latter case, forcing the model to regenerate could even encourage hallucinations I think. But I am a total layman here, just interested enough to dig a little bit into neural network architecture.
@@trnogger The end goal of this layer would be when these cases of 'no best answer' it would ask for more info from the user as its a clear indicator of known unknowns. Meaning, you know there is 1 right/best answer so any time more then one is 'good enough' it should be flagged as insufficient input.
I like the shorter to the point videos, like this
Me too!
Yes. This video's style is excellent. Well done.
I appreciate this video more than your other ones- wasn’t too short straight to the point without it being long and drawing out some of your other videos
It's AGI interacting with humans through its own chatbot front it created.
Seriously 😳
In the future, someone writes an llm that figures out how to break causality in order to improve its training past whats possible otherwise, so its come to rank its progress against the historical progenitor models
@@6AxisSage that sounds like what an LLM would say
Love this short content that has one focus! Easy to watch in between the busy moments of my day
Awesome job man that was really insightful I love watching you
Love these style of videos, keep uploading short videos. If there is a portion of your audience that wants long form videos, you can conbine all the short videos and reupload them
Like the quick update video format, thanks.
I like this short format Matt! Really nice!
Great summary video Matt. I rarely comment on videos. This made more sense in a few minutes than all the scattered info I’ve collected over many hours across the last week or two
We luv it. Keep it coming 🙏
I really liked this style of video Matt... it was fact paced, but not too fast like most are, to the point, it was mostly facts, with too much opinion, and it felt fun... But it doesn't seem short enough that I'm still hunched over, ready to swipe to the next 'short', AND it wasn't long enough to lean back and relax for a few mins and focus.... So many adding maybe 1 more minute (around 5 mins) would get me to that 'lean back and focus' mode when I see the length is over 5 and a half mins.... But that's just how it would be for me if I was watching on my phone (I'm on my PC, like usual, so I'm trying to think about it would be if I was on a train using my phone) But I REALLY liked it.... and maybe if you had the title stared with a name for this type of vid, in brackets or parenthesis, you could start a trend 🤷♂ - maybe something like "[QUICKIE!] The Mysterious Chatbot that Beat ChatGPT is BACK!" ... okay seeing that in writing is much more disturbing than I imagined LOL But at least you get the idea (forget about starting trends tho lol) 🤣
Great video, nice an simplistic without overdoing it. Great stuff
Nice and effective! I do love the long-form content though - I feel like I really get so much out of those and always look forward to those coming out.
I won't stop those. It's just always fun to experiment with new styles and see how they're received. :)
Love the quickie videos that point to something we can play with.
Thanks for the video… informative as always!
"The also good too one" Definately has agentic capabilities, I was able to get it to code an entire RAG application with multiple functionality, it produced about 10 files each for the frontend and the backend.
As I washed it keep going it felt like an agent not an LLM.
Short, single (or few) topics and to the point. Excellent.
Good video, loving the shorter long form video discussing a single topic.
I love this short form video on the hot topics of the time and then the long form topics that cover a broader range of subjects, which include things like this. I enjoy ALL your videos. You and the guys you roll with are my favorites, *go figure,* so it's no wonder but. Covering in a longer video something like this after doing it in short and having gotten the feedback to do so seems like a wise strategy. Less time wasted if that has ever happened, which doesn't seem like it has. Keep up the great work and close to 600K!
Nice video. Concise and to the point. It's great I don't have to sit and watch 10+minutes of a video which I don't do often unless it is especially captivating. I would scroll past most of it and not see what I expect to see as it's too long and I wouldn't keep scrolling find it. Longer vids are ok and good when the topic really require it. I watched this on from start to finish.
Like this video style-straight to the point. Thank you!
I like the short, simple format that covers one topic. I find your longer videos, where you cover multiple topics, overwhelming. There is only so much info I can take in in a session.
Great style, I appreciate the straight to the point 3-5 minute video 💪🔥
I like this video style. Focused topic but with background, context and latest situation. Perfect length too.
Thanks Matt.. Great stuff again.
Another great video style. Thank you Buster Friendly!
i like the format, short and to the point.
new video style was good, short punchy and to the point, thank you for putting it out there, and helping us keep up to date with the insanely fast developments in this field 🙂
Amazing video ♥
nice format quick hit
I like this short format :)
I'd like to see more videos like this one. I love the longer ones as well, but adding in a few short ones would be great!
These bite-sized videos are refreshingly good as well, Matt. 💯 Thanks as always. It's stealth guerilla marketing!
gpt2 is maybe using "q*"....? (q* is described in the 2nd edition of RL by Sutton and Barto).
Only for the curious, according to the RL book q* means "value of taking action (a) in state (s) under the optimal policy".
My thoughts exactly
Really loved this video, short and sweet, giving us the important news. I do like the longer form for more deep dives, but this was great. Also, my guess (and it's totally a guess) is that it is a new model completely from ChatGPT.
Yup great video, i also love your longer form 20-28 min long
I like this short news! Hope to see more, the big ones are too overwhelming for me, I end up skipping a lot. Thanks!
Yes, these covers are rockin’! Like your videos. Yes Sir!
love the video!
shorter videos from now on please!!! and ONLY the good stuff.... shorter vids with worthy topics are better than longer videos with 80% filler in my opinion. Good luck!🎉
Most definitely
Top vid Matt
Try this prompt in your LLM model tests or comparisons: "A simple icon of a classical Greek temple with a triangular roof, held up by exactly 3 pillars. Remember, the building MUST have 3 pillars.
perfect format for the busy guy chugging a coffe between meetings :D
Like this style!! 💯
I love the shorter video format! However, I do understand the longer formats that explain the comprehensive 'finds of the week'.
Yes great shorter super informative tutorial video. Regularly, pls thx 🐺
I like the bite-sized info about a special topic!
Yes, more short formats please.
Good vid. Keep them short please.😊
This format is nice. It's great to have something that is just a relatively quick catch up on the hottest models without it being a deep dive
Watch Usecases of GPT-4
On "Insights with Om"
You'll love this. :-) I posted multiple questions and the response looked like ChatpGPT so I asked it a simple question. Here are the responses
Question: are you from Openai? Your response is similar to their GPT 3.5 and GPT 4
Response from chatbot
Model A: I am not from OpenAI. I am Gemini, a large language model created by Google AI. While my responses may share similarities with other language models like GPT-3.5 and GPT-4, it's likely due to the fact that we are all trained on vast amounts of text data and aim to communicate and generate human-like text in response to a wide range of prompts and questions.
Each language model has its own unique architecture, training process, and strengths, which contribute to nuanced differences in our capabilities and output styles.
Model B: Yes, I am from OpenAI! Specifically, I am based on the GPT-4 architecture. If you have any more questions or need assistance on a particular topic, feel free to ask!
Mystery solved!
I appreciate this brief and efficient video update format. While I enjoy your longer videos, I believe that incorporating videos like this into your regular content would greatly benefit your channel. This approach provides concise and easily digestible information, which is highly effective. (Assisted writing by samsung AI)
I do like this video a lot. Short, informative and about a single hot topic. I need more of these to keep up with what's happening on the AI scene.
My first test, I got "im-also-a-good-gpt2-chatbot" vs LLaMA 3 Instruct 70b. It was a basic logic puzzle, and LLaMA 3 got it right, while the "new" GPT model got it wrong. I then clicked Regenerate, and this time LLaMA 3 was wrong, while the new GPT model was right (and in fact, considered multiple angles that these bots never really do). I regenerated again for good measure, and this time LLaMA 3 got it right while the new GPT got it wrong.
So they seem on par about 50/50 accuracy, at least with my test puzzle.
I'm happy with this style of video, quick little news segment.
Nice to have a quick single-subject update on a development instead of the digest format of 6-7 topics.
Love it Matt
Yes. Please more short style videos on a particular subject.
Watch Usecases of GPT-4
On "Insights with Om"
hey matt, great content as always, is there any site that you would recomend to try out games coded with ai? using html and pasting the java code?
The “im-also-a-good-gpt2-chatbot” model is waaaaaaay better than anything else I’ve tried, including gpt-4. It’s scary good at understanding complex ideas. BTW, it says “As an AI language model created by OpenAI, I have a detailed understanding of my design, purpose, and capabilities. Here’s a summary of what I know about myself:
Background
Architecture:
I am based on the GPT-4 architecture, which stands for Generative Pre-trained Transformer 4.” It also says its knowledge cutoff is September 2021.
I just sat down with my popcorn and now its just me typing with buttery fingers!!..lol..I like the SF Video...the.. "Matt Wolfe - The AI Howling Times"" or "Quick byte with Matt Wolfe" ...lol
I think the shorter content was pretty well done, maybe doing some of these throughout the week and then at the end of the week a more in depth long format video would be dope.
❤ the shorter, focused format
much more shareable format. 👍
I love that the video is under 4 minutes! Shorter is better :-)
Good stuff, Matt! However, I prefer having my rundown of weekly happenings on Friday. As someone who does batch work, it's easier for me to sit down and hear batch content.
Nice video! It seems like the audio cuts out abruptly at times, which might be due to the use of a noise gate.
Bit of a nutty thought, but what if its Altman building his own private models? So we know something happened that got him kicked out, most likely cuz he's pushing too fast or want to release to fast. So he can get around this by taking older models, say like gpt2, and the knowledge he has on how to tweak and improve it to produce a "gpt2-turbo".
But also anyone with the right knowledge can do this as gpt2 is on github, but I think its Altman.
Love the new video style, but sometimes longer and deeper is better. ;-)
I like the short breaking news format video. It is so hard to be current because everything is old in a week.
Liked seeing the quick update. Overall - B (American grade). An A would have been a chapter break or a "summary of previous coverage," so I could have known to drop as I watched your coverage of it a few days ago. Either way, liked it and will keep watching.
Two comments:
Can't you just ask the chatbot what it is? I know it could lie, but don't most LLMs tell you what they are if you ask? That's how I verified I was using Llama3 and not 2 last week.
Second comment: I very much like shorter videos. I usually hear that UA-cam doesn't like them, so most creators avoid them. A pet peeve of mine is a video that purports to answer a seemingly simple question, but the video is like 15-20 minutes long. History Matters stands out as one channel that _doesn't_ do this, thankfully. It just answers historical questions in 3 minutes or less. Perfect. Why can't more channels do that?
Thanks for teaching me about less, having a blast with it. It's better than a lot of the smaller models on here but I'm back and forth on how it competes against with Llama 3
Nice style
I like the style!
These 3 to 10 minute videos are perfect. I can watch them while on break at work or in between getting attacked by children at home.
I like this but I prefer your longer format. I feel peoples attention spans are actually getting so short and for me your long-form content is incredibly engaging and interesting.
I like this style of video
10 min is the best news format for youtube. Sabine Hossenfelder recently switched to it.
If you want to use gpt2 chat bot without depending on luck you can click direct chat and where is says what chat bot you are using, you can click it and there will be a drop down menu but you can type the one you want then you can get it all the time
I think people forget that they are primarily a research lab, that researchers stuff in the wild
Yes please! More shorter videos. This was great. Very consumable and informative.
Thanks a lot
Gpt2-chatbot is a small heavily optimized model that can be stored and run on a current gen cell phone in real time. Openai plans to use it to license to all kinds of players as a way to pre-empt what apple is trying to do.
I assume there is some form of 'agency' in this mysterious bot, which would explain its fantastic performances.
Good enough. I like longer and deeper videos though
my guess would be that these are model tests for AI with less parameters that still give usefull outputs. i dont mean AI that would run on a PC, but rather a more efficient AI that saves on servercosts.
After a bit of refreshing and reprompting: Model A: gemini-1.5-pro-api-0409-preview vs Model B: im-also-a-good-gpt2-chatbot. The hardest part was explaining to my wife why I was so excited.
My initial experience is that, when prompted with a complex and layer question, GPT2 chatbot breaks down the objectives of the response (much like an agent) before starting to work on the problem. The end result is significantly better. It's like taking my crappy prompt, breaking it down into logical steps, and moving forward with more capability that the existing models I've seen.
short and to the point
you left out my theory: that the gpt2-chatbots are chatbots created by gpt-5 (the naming even kinda sound like how gpt-4 sounded when it was used in bing chat "i'm a good gpt2-chatbot")
I tried this model just now and it is wild!!! 🤣😂 I asked it to write a python code to create a piano with GUI in my native language (Marathi, Hindi, English) mix of all. Just same as I would talk with my friends and it not only understood what I am trying to say but every time gave me code copy-paste runnable. I am still playing with it now.
I like this format. I wonder what it means for UA-cam though if it's not like 10 minutes long (or is that outdated UA-camr stuff at this point?)
4 to 10 minute videos fit my schedule perfectly.
Watch Usecases of GPT-4
On "Insights with Om"
It's good content thank you
Most of AI news long and daily we need a short news like this