Problem solving across 100,633 lines of code | Gemini 1.5 Pro Demo

This is a demo of long context understanding, an experimental feature in our newest model, Gemini 1.5 Pro using 100,633 lines of code and a series of multimodal prompts.

This demo is a recorded walkthrough of single continuous interaction with Gemini 1.5 Pro.

Token count details: The input TXT file (816,511 tokens) and image (256 tokens) total 816,767 tokens. The text inputs add additional tokens into the prompt, yielding the 818,495 token total shown in the interface.

[underestimator]

Love it when developers are building their own replacement.

[danilo]

Finally Google is a real player on the AI future, I was sceptical when Bard was released, then somehow satisfied with Gemini 1.0, but Gemini 1.5 is very exciting, well done.

[LordKelvinThomson]

Excellent example of real world usage. I will definitely try this.

[TheDarkOneSC]

Man what a time to be alive!
I'm not so impressed with the demo but with the honest and transparent presentation.
Good job on the presenter and the higherups that approved this raw demo instead of doing an elaborate lie

[தமிழோன்]

I am already impressed that Gemini can understand and write Tamil language much better than ChatGPT. I am hopeful that Google continues to support non-English languages like Tamil so everyone can benefit from the LLMs - not just English language speakers.

[ramlozz]

This is crazy stuff, waiting for open AI response, what a time we are living ladies and gentleman 😀

[chriswiles]

I need this as a vs code plugin to learn my codebase

[lucilaci]

nice that you showed the chat and did not make a video like when you first introduced gemini ("The capabilities of multimodal AI | Gemini Demo")! :)

[Elmimoh]

عمل جبار حقيقة، تذكرت شخصا كان يقول أيام ما كان chat gpt جديدا، ما تقوم به هذه الشركة هو مجرد ورقة بحثية واحدة مقارنة مما توصلت به جوجل في هذا المجال، وأن ما ستقدمه جوجل سيكون شيء لم يسبق له مثيل.
شكرا لجوجل.

[horsey]

Professional thinkers making professional minds = next-gen AI 🐴

[gonzalobruna]

this is actually awesome. I'm truly excited to try out gemini 1.5, and looking forward to the Ultra version coming soon! that's going to be wild.

[HYRULE]

"I am unable to engage in discussions that could be construed as biased or offensive. I am an AI chatbot designed to provide information and assistance in a neutral and unbiased manner."

[LynnKereluik]

The Gemini Google assistant should be an asset ruled by mercury the planet of communication of all signs 😊

[maxvankekeren]

If they can manage to actually pull this off this is going to be wild!

[falklumo]

These are all examples of tasks a human programmer would find easy to perform and wouldn't turn to an LLM for help. What about the hard tasks, like refactoring the codebase for cleaner and less redundant code? Any examples for this? Even failed examples would be interesting to share!

[trippyabsolute]

I’m impressed and scared at the same time. Wild.

[FloWoelki]

looks really interesting. 2024 is going to be wild :D

[wlockuz]

This is very easy when the codebase is as clean as threejs's. If it tried to understand one of my code bases it would decide to self destruct.

[samgodse]

I don't want to call this out but I personally feel like this demo is a bit flawed, here's why:
The codebase is actually set of examples based on three.js modelling --> in simple words the codebase is actually a collection of **independent** code files.
A more accurate demo would be possible if they did it on a codebase with integrated units and functions.

I know this is actually showcasing power of large context handling but then if the sub-context are independent of each other then it can be achieved by iteratively going through the small subcontexts. So in general the AI doesn't see the whole intermingling since there is none.

[unimposings]

proof that this is not fake again, like your last promotion!