New AI beats GPT-4 & Gemini 1.5

[Davino]

The raw results of the new AI Claude 3 in their three different versions:

 

And a very sober analysis:

 

 

[Leo Gura]

But I'm sure the next version of ChatGTP is in the works and will be better.

[RightHand]

GPT 4 still has the best value proposition IMO. Its ability to receive images and files in prompts is just too amazing + its ability to search the web can be very useful.

[Davino]

1 hour ago, Leo Gura said:

But I'm sure the next version of ChatGTP is in the works and will be better.

Just two months ago, Gemini Ultra was launched, surpassing ChatGPT in nearly every aspect.

Now, with the introduction of Claude 3, we have a model that outperforms Gemini Ultra in every aspect.

Maybe the new ChatGTP version is better but Claude has surely risen the level while also making Google look pretty bad. It is funny to think how Google has been leading AI research for almost a decade with Deep Mind. However, lately they have been beaten by Microsoft and now punched in the face by Claude just months after their new release.

Now ChatGPT has more pressure to make the new version even better, as Claude 3 has shaken the market out of nowhere. It is exciting to see new competitors out of nowhere instead of the classic giants brawling.

[Yimpa]

I just hope that these models don’t go to war with each other 

[thenondualtankie]

Guys... It's ChatGPT. Generative Pretrained Transformer.

WTF is ChatGTP? Generative Transformer, Pretrained?

[Girzo]

At this point the benchmarks posted are bullshit, I tell you. Correct answers are all included in the training data set and the Claude Opus thing is only better at recalling the answers it already knows. It's real reasoning skills are still utter shit. GPT-4 is still unbeaten in many ways. Neither system is really intelligent.

[Felliks]

Yes it probably has validation data in it's training set. The difference is, as far as I know Claude isn't hiring low cost labor to write and filter their datasets, which OpenAI is infamously known for.

All in all, they're all the same, +- one metric, and the battle for the strongest model is not a commercially viable option due to computing costs for a niche that needs it.

Distribution is the name of the next game. Getting all those models integrated into hardware, and many, many aspects of daily lives of the consumer class. Phones, watches, robots, smart house rings, etc. This is where distributed computing and lower operational costs, and smaller models need to and will go next