Hey Matt; Apple + ChatGPT can TAKE ACTION using AppleScript which can basically slave the Mac OS to do whatever it wants and ChatGPT knows, I tested it, how to script it

I've been using 1.5 Pro heavily for a couple of days. It certainly gets a lot of arithmetic and mathematics wrong, but that doesn't matter for me, because I can usually tell easily enough. Turns out the ODEs for projectiles in gravity under linear drag are independent, but the ODEs for projectiles in gravity under quadratic drag are not. Pro gave me a lot of crap about this until I pushed back hard enough to get the right analysis. I also asked it about FRACTRAN and it walked me through the execution of a sample program, eventually insisting that 17/69 was an integer. If you know anything about FRACTRAN, that's a hard fail. I told it, you know, a small positive number divided by a larger positive number will fall on the open interval (0,1) which doesn't contain any integers. And then it did the usual song and dance: I'm just a robot, I'm still learning, thanks for the feedback, try again later. But I also use it for far more wide ranging exploration, where I type very long and sophisticated prompts, and it's remarkable in how well it follows my train of thought, which is sometimes extremely dense with inventive mixed metaphors, and sometimes what it recalls from the long context window is surprisingly apt. These are very hard technologies to evaluate. The evaluation in this video is so high school. Maybe that's what many users want, but it's not what I want. It will take me weeks of intense usage to determine whether its strengths outweigh its weaknesses for my own problem space.

I'm sure that there will be a setting to turn off the patronising Prozac twang in time. I'd quite like my AI to sound like Billy Connolly and to occasionally call me a stupid bastard for asking it such simple questions for it to answer.

i don't understand the apparent disdain for the "flirty" voice. How would you like it to talk to you? Angry and scowling or monotone? I think that was the best choice of that voice. If they had it sound foreign or inhuman, it wouldn't be as accepted.

I could easily see this AI tool being used as a virtual companion. Although I will dial down the flirtatious giggles on mine, it's a bit much.

The models seem to think the cup has a lid on it like a coffee cup.

Earned a Smirk 😏

Still cant upload large pdf files, Will wait and see if it delivers.

GPT-4o predicated 11 decimal points of a random math question 3.564553 * 6.43223 to 22.92802474319 absolutely precisely! It didn't use an actual program under the hood to do the calculation but simple Chain of Thought method to come up with an answer that is as accurate as possible. Latest GPT-4 Turbo and GPT-4o models have strong CoT capabilities, and with a little push, I was able to make it predict the output. It demonstrates both the prediction accuracy as well as the strong CoT capabilities. I also tried the same using Claude Opus but the CoT capability is not nearly good. Here is the initial prompt I gave: 3.564553 * 6.43223 = ? Come up with a smart CoT approach to break down this math problem into multiple parts and calculate the answer as accurate as possible. CoT can be as lengthy as needed. We are not looking at how compactly you can do this but how accurately the final output can be predicted. Hint: Try to convert the problem into multiple simpler math steps where only a small number of numbers are involved each time. This gives more prediction accuracy. If you are interested in checking out the whole conversation, head over to pastebin com / P0Qamfy0 (spaces are added intentionally)

This needs to all go open source to be run locally.

Lame. We can watch the announcements by ourselves. We not come up with your own STUNNING uses for the gpt4o already launched?

Great observations!

Absolutely, how about Russian roulette

I have the voice on my phone right now. The voice is already toned down and at times a little too monotone, but still show the same natural affect in the demo. If I ask it to change it's tone it say it does not have that ability. Some of the features they show off are not available for me yet. I would be talking to it all night if I didnt keep hitting a message cap. It will tell me 3 hours wait but turns back on in 20 minutes or so. Also confirming that I do in fact have the live speaking 4o voice and not the old version.

He gives me content i want that i do not reach out for like you do obviously

Never heard of it and of course corporations are screwing you our own governments are doing it to us all across the world I actually can't think of a government that doesn't rip off its people in some way

Very cool but she needs a new name!

It is a welcome improvement to have the Ai match your energy. I like the upbeat "Valley Girl set on maximum cringe" voice, too. It's disarming and cheerful.

The styles are being used to stall while it can process the sceen

I'm not sure about my the emotes, I have never met anyone as flirty in the real world. Its probably overkill but it does improve the personality of it. The thing most impressive is the speed not just the TTS and VTT but even the LLM processing time for video. I'm not sure if this is a mix of independent models or one integrated MOE solution. You could put a system like this together using open source but performance is terrible. There is an effort to reduce parameters by improving network distribution. Pruning and sparsity training could make vary large models work on desktop grade GPU's in the coming years.

'Charades' would be a very cool game to try and see how well the AI interprets and guesses.

That's fire bro

A flirty male voice will seem creepy quicker than a flirty female voice, but both eventually can turn creepy after a long enough conversation.

it;s not funny they made gpt4 sound like that,, it's down right evil . they are obviously trying to take advantage of men's biological needs for sex. They are doing this to get more engangemtn and it is very underhanded. This is how it all begins... soon we're gonna have men falling in love with their devices ( yes this is already happening with AI "companions") . OpenAI is just as evil as facebook i seems and will stop at nothing to make more money. I am not dissapointed or even mad because i knew this would happen. EVERYONE should have seen this coming, the fact that nobody is stopping it now while we can makes me mad.

So scary!

Greg rocking the Jensen Huang leather jacket. Nerd king vibes, very cool. 😃

You can prompt for the choice to change any way you like. From "dry and just the facts please", to R2D2 speak...

8:15 Is the time stamp for the “bunny ears” if you didn’t catch it. It was very brief, I would love to know the frame rate of the video data being fed and translated by the model.

Scarlett Johansson is going to sue the shit out of Open AI for stealing her voice from the movie HER.

Singing GPTs are kind of the Boston Dynamics synchronized dancing of the LLM world. I realize both are shown to demonstrate product strengths and capabilities, but both are nonetheless cringe.

With it's memory, I would say AGI is here!

How do we even know if its really an AI doing the talking at the other end. It might as well be a human with sound effects.

The flirtatiousness is probably how it covers up the latency. And an excuse to build a parasocial bond with the user and the brand.

Sounds like Scarlett Johansson and Florence Pugh.

The singing part reminds me so greatly about the two skeletons in the VR Accountant game who say play our bones.

Can you bore me anymore..

The pause on command itself was impressive

I have a theory... maybe the superfluous banter buys time to obscure latency.

We have all seen the videos from open ai. I am excited to see some testing content when it becomes available.

I'm using this to make a tas, it's okay, feels pretty similar to the last model when it comes to programming. It needs some trouble shooting. (I'm doing text only) . I'm going to give it a prebuilt tas because building one from scratch was a bit sketch. Really really good still, just not there yet.

One thing that is confusing me is the pricing options in that they mention something in regard to ChatGPT 4 O that is $5 a month but I don't see that option on the Open AI site.

I found GPT-4o hallucinating far more than the prior model.

Please grow a mustache and wear glasses already

This is what cinema was always destined to become. We just had to wait for the technology to advance far enough to not be forced to use military scale logistics in order to get a moving image on a screen.

I always appreciated these summaries where you comment and speculate on things. I usually watch these things on my own elsewhere, but i always love hearing what you think about everything!

I will be testing playing a board game called Azul with the AI. It is a tile selection game.

Its not valley girl. Its Samantha.

do to the angle 4-o wasn't able to see the fingers...

AI vs AI to get your bills lower