AI RECAP: Rumored GPT-4o Large Model & Gemini Live vs GPT-4o Advanced Voice

In this episode, Glasses Matt dives into the whirlwind of excitement and confusion surrounding the rumored OpenAI 'strawberry architecture' also known as Q Star. With strange tweets, mysterious predictions, and latest updates, Matt investigates whether these rumors can be substantiated and discusses the recent GPT-4 Omni model release mentioned by OpenAI. Additionally, he covers a new benchmark, SWE bench, and touches on Google's recent AI developments with Gemini Live. Engage with Matt as he navigates this baffling AI saga and invites viewers to contribute their insights.

▼ Link(s) From Today’s Video:

-------------------------------------------------

▼ Extra Links of Interest:

Let's work together!

Thanks for watching Matt Video Productions! I make all sorts of videos here on Youtube! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe!

All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them.

00:00 Introduction and Context
00:39 The Strawberry Architecture Rumors
02:09 Strawberry Man's Predictions
03:33 OpenAI's Recent Announcements
08:39 Google's Gemini Live Announcement
09:37 Gemini Live Demo and Comparison
18:20 Conclusion and Final Thoughts

[MattVidPro]

Hey folks! Update on this for you: as some of you mentioned in the comments SWE bench was leaked by Tibor (a fantastic person in the community btw) so it is very possible that Strawberry man just got this info from him. Do you think he still has any credibility?

[ravenfly]

I'am done with the hype, even canceled my OpenAi plus membership!

[MattVidPro]

What do you guys think? Your viewpoint is valuable!!! (I do read comments quite often too)

[LGministry]

I do have the new AI voice, but I don't think that matters when I am on my PC. I just asked the strawberry question just like this: "How many rs are in the word strawberry?" And it replied: "The word "strawberry" contains three "r" letters." I don't know why mine got it right the first time, but it did. While I am typing this, I will check right now to see if the new AI voice gets this question right. The AI advanced voice also answered the question correctly the first time.

[IceMetalPunk]

I've never seen the "how many r's in 'strawberry'" test. Interestingly, I think a big part of its problem is that words aren't generally tokenized by letter, unless they're very uncommon. So "strawberry" gets tokenized as "str", "aw", and "berry", and those tokens (or, the numerical index of those tokens) is all the model sees, not the individual letters. Which makes me wonder: if you wrote the strawberry "r's" question on paper, took a photo of it, and asked it to solve it via an image... would it be more successful? Because then it's processing image tokens, which don't break up words by letter but by visuals, and that might make it *see* the correct number of r's. Just a thought...

[FusionDeveloper]

I don't like rumors and rebranded rumors called "leaks". I like facts.

[sharkeys]

I get so frustrated when the Open AI guys interrupt GPT 4 voice in the demo.. I know its part of the demo to show real interaction but it just really bugs me :D

[Streeknine]

They seem to be trying to keep interest in their ChatGPT product while all the other AI apps are getting better in some cases passing ChatGPT in some areas.

[NextGenart]

Strawberry didn’t predict Swe Bench, the information was out there to the public

[AI-Wire]

Sam Altman is turning ClosedAI into a troll-first, hype focused blogging company

[thisismissing]

I can't believe that in 2024 this is the way companies are doing business. If people want them to take them seriously especially Large enterprises. They can't be acting this way. They need to have road maps that they release at a steady rhythm. They have to let developers at least get some exposure or some anticipation of what's going to be coming down the road or else they're not going to want to invest time and money building tools on a platform that seems to be releasing weird information on Twitter

[allanshpeley]

You ever notice how we never see Sam Altman, Jimmy Apples and Strawberry man in the same room at the same time?

[lamsmiley]

I think this release will have a big impact on how people perceive AI for the next six months. If the Q Star/Strawberry model doesn’t show marked improvements I believe we’ll see a people lose faith that we’ll ever reach AGI (which I don’t think we will without major architectural changes to our approach).

[moderncontemplative]

Unlike some individuals, society is not yet prepared to fully comprehend the implications of communicating with human-like AI, particularly models that possess superhuman capabilities in certain domains. Consequently, Google, in a prudent manner, has retracted some of its advertisements.

[idontexist-satoshi]

The change was for free users where instead of defaulting to 3.5 when you use your free gpt-4o allowance, it now defaults to gpt-4o-mini.

[Serifinity]

Iruletheworld predicted 100% correctly, previously stating a couple of days ago that the Tuesday 10am PT announcement was exactly what it was, they even posted their receipts as a lot of people were disappointed in OpenAI's announcement. I think some people misunderstood what they were revealing. Also if OpenAI does release level 2 strawberry reasoning on Thursday, they do not have to provide anything explaining what the update is, you forget it's a private company that can do exactly whatever it wants.

[kyber.octopus]

AI companies managing benchmarks is like having students managing exams.
As in - it makes completely no sense and will always be biased towards your model.

[xbon]

There have been a MASSIVE difference in GPT-4o since last week. Maybe you won't notice if you're not using custom instructions, but my AI which i finetuned for GPT4o has completely changed and I noticed it immediately.

[Siciliano]

I'm over it. They're taking entirely too long to release the new 4o voice+vision mode. Why host an entire event to show off the new capabilities and then postpone said abilities over and over.

[ThomasMeliWellness]

It has been mentioned quite a bit that the "r" in strawberry test is not useful because LLMS work through tokens not letters. Thus doesn't "count" the "letters" the way we do because the way it organizes tokens is context dependent, but this can be fixed by changing the prompt. Please help the world understand this instead of continuing to do this meaningless test.