Gemini's 3 New Models

🕵️ Interested in building LLM Agents? Fill out the form below

👨‍💻Github:

⏱️Time Stamps:
00:00 Intro
00:06 Google TPUs
01:17 Gemini 1.5 Paper
04:30 Benchmarks
04:56 Google AI Studio: 3 New Gemini Models
05:27 Lmsys Chatbot Arena
06:39 Code Demo

[eoghanf]

Why in the name of all that is good didn't they call it Gemini 1.6 or 1.7? "New Gemini 1.5" is ridiculous.

[gemini_]

Clearly OpenAI is failing to keep up with other player now. I'm particularly impressed by how great Gemini has become. I use it on a daily basis. Absolutely my favorite over GPT4.

[jsalsman]

Agree the chatbot arena rewards form over function.

[toadlguy]

As far as the speed goes, unless you know what hardware is being used for inference on the models, the tokens per sec. can't really be compared. Google has some new TPUs as well.

[ChristianNode]

What are we supposed to be doing with these llms?

[unclecode]

Google is back in the race! I was able to get Flash - 8B to summarize both. In my opinion, comparing different model sizes with the same prompt isn't very accurate since smaller models need more detail in their prompts. I used the prompt: "Can you read this and give me a high quality summarization, and make sure to summarize all different topics mentioned here, if any: \n\n {long_text_02}." I tried a few times and It worked well and only took 5 seconds.

[Arcticwhir]

1.5 pro exp 0827 is pretty awful i'm testing code output rn, sometimes it doesnt even complete the code, the code is very simple. Something seems really off

[rezakn]

Awesome video!😍Do you have any guesses about the relative sizes of Flash and Pro models vs the 8B?

[SwapperTheFirst]

Waiting for Gemini Pro 2.0 and expecting it to match or be better at coding vs Sonnet 3.5 or even upcoming Opus 3.5

[adrianwhite]

None of the new models support cache prompting :-( ... yet

[SeaHorseOo]

We'll need a Screen Junkies Honest AI Opinion ...Oo...

[haralc]

8B models is like BERT models ... do you see anyone using BERT model? .... exactly! :D

[TheRealHassan]

Still bad at coding… which is a major use case. Great as a “RAG” alternative

[i_accept_all_cookies]

Can the flash or 8b experimental models be fine-tuned?

[aymandonia]

So gemini flash better than Claude 3.5 ok ‏‪5:48‬‏

[jeanchindeko]

yep benchmark are nice but more and more doesn’t tell a true story. For most enterprise building GenAI applications benchmarks don’t tell them how the model will perform in their particular context and business cases. We definitely as a community need a better way to to help developers and other leaders to chose the right model for them.

[i_accept_all_cookies]

Anyone know which small models have the fastest inference? (other than Groq)

[bastabey]

gpt-4o-mini-2024-07-18 performed much better in terms of accuracy than existing Gemini Flash in a RAG application I m trying to deploy.. while Google is doing a lot in AI, their efficiency remains behind Anthropic and OpenAI

[timmysmith]

Gemini is a closed, censored and biased model. I can't even begin to describe how lame this is. Stop being so impressed.