00:00If you've been feeling like your AI buddy's been acting a bit different lately,
00:06maybe quicker, sharper, and just a tad smarter, you're not alone.
00:10OpenAI has been sneaky, rolling out some major changes without a big announcement,
00:14but don't worry, I've got all the details you need to know right here.
00:18So let's talk about it.
00:19Last week, I started noticing that ChatGPT felt different.
00:23It was like the responses were more on point, faster, and just generally better.
00:28I wasn't the only one either.
00:30People all over social media were talking about how ChatGPT seemed to be upgraded.
00:34But here's the thing, OpenAI didn't say a word about it at first.
00:38It was all very hush-hush until they finally dropped a little bombshell on us.
00:42OpenAI took to X to casually mention that they'd slipped in a new version of their GPT-4O model into ChatGPT.
00:50So they just updated the model we've all been using without making a big deal about it.
00:55The message was simple.
00:56There's a new GPT-4O model out in ChatGPT since last week.
01:00Hope you all are enjoying it and check it out if you haven't.
01:03We think you'll like it.
01:04That's it.
01:05No fancy press release, no grand unveiling, just a tweet.
01:08Typical OpenAI, right?
01:09Now if you're wondering what's so special about this new model, let's break it down.
01:13The updated version of GPT-4O, which they're calling ChatGPT-4O Latest, is essentially a fine-tuned and optimized version of what we had before.
01:23But here's where it gets interesting.
01:24While OpenAI hasn't spilled all the beans, there's a lot of speculation about what this new model actually is.
01:31Some people out there are thinking this might be part of a bigger strategy by OpenAI to release different sized models.
01:38Kind of like what Google and Anthropic are doing, there's been talk about a GPT-4O large, and some think this latest update could be a stepping stone in that direction.
01:49But I'm not totally sold on that idea because, let's be real, if it were a brand new model, they probably would have hyped it up a lot more.
01:55So what can this new model do?
01:57Well, from what I've seen and what others have reported, it's performing better on tasks that require complex reasoning and creativity.
02:04Like, if you've been asking ChatGPT to help with coding or solve tricky problems, you might have noticed it's just a little bit sharper now.
02:11It's also faster, which is a nice bonus, but of course, it's not perfect.
02:15There are still some weird quirks.
02:17For example, in one test, the model was asked to stack a book, nine eggs, a laptop, a bottle, and a nail in a stable manner.
02:24The solution? It suggested putting nine eggs on top of a bottle.
02:28I mean, come on, who does that?
02:30And then when it was asked how many R's are in the word strawberry, it came back with two, which is definitely wrong.
02:37So yeah, there are still some bugs to work out, but overall, the update is a step in the right direction.
02:42Now, talking about strawberry, let's talk about something that's been generating a lot of hype Project Strawberry.
02:48The idea behind Project Strawberry is that it could be a new post-training method that boosts the model's reasoning skills.
02:54Some people are even saying that the improvements we're seeing in ChatGPT might be the first signs of this mysterious project in action.
03:01One of the coolest things about the new ChatGPT 4.0 latest model is how it handles multi-step reasoning.
03:07This basically means the AI isn't just jumping to conclusions.
03:11It's thinking things through, step by step, before it gives you an answer.
03:15That's a pretty big deal because it leads to more accurate and thoughtful responses, which is something we all want, right?
03:21The new model has already made waves in the AI community, especially in something called the LMSYS leaderboard.
03:28Now, if you're not familiar with it, the LMSYS leaderboard is like the Olympics for AI models.
03:33They put different models head-to-head in all sorts of tasks.
03:37And the new ChatGPT 4.0 latest model just crushed it.
03:40It scored a whopping 13-14 points, which is the highest score ever recorded on that leaderboard.
03:45This means it's outperforming some of the biggest names in the game, like Google, Anthropic, and Meta.
03:50Now, if you're thinking, how do I get my hands on this new model?
03:54Well, it's super easy.
03:55OpenAI has already swapped out the old GPT 4.0 with the new version in both the ChatGPT website and app.
04:02So, all you have to do is fire up ChatGPT and you're good to go.
04:07If you're on the free plan, you might hit some message limits.
04:09But for those of you who are on the plus plan, you can push the model to the limit and really see what it can do.
04:15But don't worry if you're not ready to shell out the $20 a month for the plus plan.
04:19You can still get a good feel for the new model before you hit those limits.
04:22And then, if you run out of messages, you can switch over to GPT 4.0 Mini.
04:27It's not quite the same, but it's still pretty powerful.
04:30Also, one more really interesting thing is how OpenAI has been testing these updates.
04:35They've been sneaking experimental models into places like the LMSY's ChatBot Arena under random names,
04:41so people don't even realize they're testing new tech.
04:43The ChatGPT 4.0 latest model, for example, was tested under the name Anonymous ChatBot,
04:49and it got over 11,000 votes from users.
04:52That's a lot of people unknowingly helping out with the testing,
04:55which just goes to show how clever OpenAI's approach is.
04:58So, what's next?
04:59Well, if this update is anything to go by, we can expect OpenAI to keep refining and improving ChatGPT.
05:04They're clearly focused on making it better at reasoning, creativity,
05:07and all those tasks that require a bit more brainpower.
05:09And who knows? Maybe we'll see even more of Project Strawberry in the future.
05:14All right, now I also want to talk about a new AI model that just came out,
05:18but it didn't really get the attention it deserves.
05:20This model, called Falcon Mamba 7B, was released by the Technology Innovation Institute, TII, in Abu Dhabi.
05:27TII is known for working on cutting-edge technologies like AI, quantum computing, and robotics,
05:33and now they've dropped this new model.
05:35It's available on Hugging Face, and it's an open-source model, which is pretty cool,
05:40but what really sets it apart is the new architecture it's using.
05:43Most of us are familiar with transformer models, which have been dominating the AI scene for a while now,
05:48but Falcon Mamba 7B uses something different, called the Mamba State Space Language Model, SLM architecture.
05:56This new approach is quickly becoming a solid alternative to those traditional transformer models.
06:01Now, why is this important?
06:03Well, transformers are great, but they have some issues,
06:06especially when it comes to handling longer pieces of text.
06:09You see, transformers use an attention mechanism that looks at every word in a text
06:14and compares it to every other word to understand the context.
06:17But as the text gets longer, this process demands more and more computing power and memory.
06:22If you don't have the resources to keep up, the model slows down and struggles with longer texts.
06:28This is where SSLM comes in.
06:29Unlike transformers, SSLM doesn't just rely on comparing words to each other.
06:34Instead, it continuously updates a state as it processes the text.
06:39This means it can handle much longer sequences of text without needing a ton of extra memory or computing power.
06:45Now, Falcon Mamba 7B uses this SSLM architecture,
06:49which was originally developed by researchers at Carnegie Mellon and Princeton Universities.
06:53What's cool about this model is that it can dynamically adjust its parameters based on the input
06:57so it knows when to focus on certain parts of the text and when to ignore others.
07:02So, how does Falcon Mamba 7B stack up against the big players like Meta's Llama 3.8B, Llama 3.18B, and Mistral 7B?
07:11TII ran some tests, and the results are pretty impressive.
07:14In terms of how much text the model can handle,
07:17Falcon Mamba 7B can fit larger sequences than the Transformer models using just a single 24GB A10 GPU.
07:26This means it can theoretically handle infinite context length if you process the text token by token or in chunks.
07:33And again, Falcon Mamba 7B came out on top.
07:35It beat Mistral 7B's sliding window attention architecture by generating all tokens at a constant speed
07:41without any increase in memory usage.
07:43And that's a big deal for anyone working with large-scale AI tasks
07:45because it means the model is both fast and efficient.
07:49Even when it comes to standard industry benchmarks, Falcon Mamba 7B holds its own.
07:54In tests like ARK, TruthfulQA, and GSM8K, it outperformed or matched the top Transformer models.
08:00Sure, there were a couple of benchmarks, like MMLU and HellaSwag, where it didn't quite take the lead,
08:06but it was still right up there with the best of them.
08:08But here's the thing, this is just the beginning for Falcon Mamba 7B.
08:13Tichu has big plans to keep optimizing the model and expanding its capabilities.
08:18They're not just stopping at SSLM, they're also pushing the limits of Transformer models to keep driving innovation in AI.
08:24So if you're into AI or just curious about what the future holds, keep an eye on Falcon Mamba 7B.
08:30It's already making a name for itself, and with TII's continued efforts, it's only going to get better.
08:35Plus, with over 45 million downloads of their Falcon models, TII is proving that they're a major player in the AI world.
08:43Alright, if you found this interesting, make sure to hit that like button, subscribe, and stay tuned for more AI insights.
08:50Thanks for watching, and I'll catch you in the next one.