🚂 2 new production models: 1.5-Pro-002 and 1.5-Flash-002 with improved math, vision understanding and long context performance
🧪 Improved version of experimental 1.5-Flash-8B
🛟 Updated settings on safety filters
↗️ 2x higher rate limits on 1.5 Flash and ~3x higher on 1.5 Pro
Google Search users with Search Generative Experiences (SGE) turned on will now be able to export responses to Python-related queries to a new Colab notebook directly! You can run the code, tinker with it in Colab and save the notebook for future reference!
#GoogleAI
#Colab
You can now edit code in Google Search Generative Experience (SGE) in addition to generating and executing it!
This will let you
- play around with the generated code
- write small test cases to verify the code
- make small edits before copying into your development environment
3 new experimental models
- First release of a small model - Gemini 1.5 Flash-8B
- A Gemini 1.5 Pro model better on coding & complex prompts
- A significantly improved Gemini 1.5 Flash model
Try the model names ending with Experimental 0827 .
Chatbot Arena update⚡!
The latest Gemini (Pro/Flash/Flash-9b) results are now live, with over 20K community votes!
Highlights:
- New Gemini-1.5-Flash (0827) makes a huge leap, climbing from
#23
to
#6
overall!
- New Gemini-1.5-Pro (0827) shows strong gains in coding, math over
As someone commented on
@OfficialLoganK
's post, this is now a Flash sale!
⚡️ Gemini 1.5 Flash more than 70% cheaper through the Gemini API and AIS
📈 Finetuning now available to all! Tuning a model is free, inference is the same as cost as for base models
💯+ languages
📢✨ Announcing Gemini 1.5 Flash price drop & updates! →
🔧 Enjoy the speed & efficiency of 1.5 Flash at 70%+ lower prices
⚡️ Customize Gemini 1.5 Flash with tuning
👀 Check out improvements to Google AI Studio & Gemini API
#BuildWithGemini
As a physicist, Gemini using code-execution to solve projectile motion equations makes me happy!
📸 Available for both 1.5 Pro and Flash
📸Use as a tool in the API or through AI Studio
📸 for queries that benefit from code-based reasoning such as math equations or
Announcing coding AI in GCP Vertex AI, and in the new Duet AI products!
Vertex and Duet will help you, your company, and your developers really boost their productivity when building on GCP.
Powered by a family of PaLM2-derived models known as Codey ✨
Took
@shreyas
's course this weekend. Its among the top 3 courses I have taken in my career. The advice/learning is different - an unusual mix of visionary, pragmatic and actionable. But the right mix to achieve the fulfilling goals that we become PMs for.
Your new coding assistant is almost here! Check out these new Colab features: natural language to code generation, code completion, and an integrated chatbot. Read all about at authored by
@thechrisperry
and
@shresbm
Potential use cases we have heard for experimental 1.5 Flash 8B model ⚡️
- Initial-stage classifier or router in workflows involving larger models
- Lower latency deployment in chatbots, RAG systems
- Closer to real time video analysis
- Medical and educational simulations
We
I remember the days that I would ping Bard with Bengali transcribed in English and it would respond in English. Now we have Gemini Advanced in 9 Indian languages including Bengali!
Exciting news! 🇮🇳 Today, we're launching the Gemini mobile app in India, available in English and 9 Indian languages. We’re also adding these local languages to Gemini Advanced, plus other new features, and launching Gemini in Google Messages in English.
The Gemini API and Google AI Studio are now available in 200+ countries, Gemini 1.5 Flash costs $0.35 per 1M tokens, with context caching coming next month.
So much going on today 🤯
@OfficialLoganK
@GoogleAI
Google seems to be actually shipping more frequently now and substantial stuff too. Meanwhile, the 🍓 nonsense has resulted in nothing. I've been hard on Google over Gemini, but this is an awesome renewed sense of direction. Well done.
This weekend I read the Michelangelo paper from
@GoogleDeepMind
() which benchmarks models' ability to find relevant information needed for complex tasks spread across large context, much like Michelangelo chiseling away irrelevant material to reveal the
Exciting release of this awesome product our team has been working on. Duet AI brings real-time code suggestions and a chatbot directly on your IDE, with the potential to massively increase dev productivity!
#DuetAI
#GoogleIO2023
📷
@shresbm
@Skiminok
Exciting News from Chatbot Arena!
@GoogleDeepMind
's new Gemini 1.5 Pro (Experimental 0801) has been tested in Arena for the past week, gathering over 12K community votes.
For the first time, Google Gemini has claimed the
#1
spot, surpassing GPT-4o/Claude-3.5 with an impressive
Caching has a 75% discount compared to regular input tokens! Also you can try caching for Flash in the free tier upto 1MM tokens.
Use caching for long context use cases where you need to reference the same context repeatedly. Some ideas we have are
💽 Long-context ICL by
Context caching is now available for Gemini 1.5 Pro and 1.5 Flash! 🙌
At a lower cost, context caching in the Gemini API makes working with millions of tokens a breeze. Pass content once, cache the tokens, and refer back for later requests. Learn more →
Our datascience agent just launched on ! This agent can generate entire Colab notebooks! It helps you get started on your data science tasks by cleaning and explore your data for you, doing statistical analysis and creating predictive models. It can answer
Have you ever trained a model you thought was good, but then it failed miserably when applied to real world data?
If so, you’re in good company.
Check out
@michael_lones
' piece on how to avoid machine learning pitfalls!
👇
New Udacity course using the Gemini API and Google AI Studio! Developed by Berkeley and Google experts!
Try it and let us know what you think!
You will learn how to build an app using the Gemini APIs and how to use AIS to design better prompts!
🌟 Calling all developers – our new Gemini API by Google Course is live!
This FREE course is your gateway to mastering the integration of generative AI into your applications, websites, and services.
@googledevs
Enroll in this free course today:
Very very excited for this! Still remember the early days of brainstorming when this was just thoughts on a Google doc. So proud of the team and the friends I have made along the way!
Thanks for the great response to the features we shipped yesterday! Excited to have partnered with
@OfficialLoganK
@Ronenkofman
@matvelloso
and others on this launch! More to come very soon!
🚀 Gemma 2 now also in AI Studio
🚀 Code-execution as a tool through API and AIS
🚀
Big news – Gemini 1.5 Flash, Pro and Advanced results are out!🔥
- Gemini 1.5 Pro/Advanced at
#2
, closing in on GPT-4o
- Gemini 1.5 Flash at
#9
, outperforming Llama-3-70b and nearly reaching GPT-4-0125 (!)
Pro is significantly stronger than its April version. Flash’s cost,
Excited to start using the Gemini models for code AI products
@Google
!
Gemini Ultra advances SOTA in 30 of 32 benchmarks including HumanEval (74.4%) and Natural2code (74.9%)
Gemini Nano will unlock lots of on-device possibilties.
Seeing some qs on what Gemini *is* (beyond the zodiac :). Best way to understand Gemini’s underlying amazing capabilities is to see them in action, take a look ⬇️
New updates to our Search Generative Experience include definitions with related images to help explain complex topics, more coding capabilities, and a new experiment that helps you find what you're looking for in a long article more easily.
Today we are launching Data Science Agent, a Google Labs experiment.
◆ Helps with complex tasks like planning, and error correction.
◆ Helps with data science tasks like predictive modeling.
◆ Outputs an AI-generated Colab notebook based on your prompt.
Try it at
We could use Gen AI to automate the work needed to cut through bureaucratic red tape to get approvals. But then that Gen AI will need to get approved …
India just kissed its future goodbye!
Every company deploying a GenAI model now requires approval from the Indian government!
That is, you now need approval for merely deploying a 7b open source model 🤯🤯
If you know the Indian government, you know this will a huge drag!
Share your knowledge and make an impact with early-career AI/ML practitioners by applying to be a KaggleX Fellowship advisor! We're looking for more advisors to apply before July 21st.
I've been raving about caching being underhyped since IO, most of the long context nay sayers point at price and speed as the two things of why long context will never come close to their precious RAG solutions.
Context caching from google is an answer to both these points!
When you wait too long at a green light because of an (important) message that you need to send and the driver behind you does not yet mad… because there is no driver
#cruise
This is pretty cool and the beginning of many more shared, digital universes. The need to create and nurture digital characters seems to be one of those fundamental human impulses that have existed from the early days (Tamagotchi) days of the internet!
This is a village run by AI. The tribe has different personas managed by ChatGPT, from smart to aggressive or grumpy, and when new things appear they're drawn in realtime by StableDiffusion 3. You can influence what happens in the Twitch chat, or have it run autonomously forever.
Many great insights in this podcast. Some that stood out:
1) A big pursuit in 2024 will be finding sustainable AI use cases. Looking in non-tech companies outside of Silicon Valley can yield good ones
2) Developers don't want "the solution"- they want intuitive, reusable
"We think that practically every internal app is going to be AI infused over the next three years." Hear from our founder
@dvdhsu
on the future of internal tooling 👇
Calling all CS Educators and administrative professionals - we want to hear from you about how Colab can work better in your classroom - please fill out this form to share your feedback (and tell us if you want to chat more!):
And so it begins! Excited for all that will be unveiled today.
As Gemini says
"In the depths of the code, a human heart beats,
Each product a monument, which vision completes
Its journey from thought, to tangible thing,
A testament to progress, a song we will sing."
Code Tips is folded into SGE! Code Tips was launched as part of the initial Search Labs to give one-shot answers to popular code-queries. Thanks to everyone who tried Code Tips! We have used all the learnings to make SGE code better. (Below is what Code Tips used to look like)
And one more thing 🙂 The same PaLM2 / Codey technology is now powering Android Studio Bot - generating code, answering questions, and helping with unit tests right there in your Android Studio IDE! ✨
Today we have published our updated Gemini 1.5 Model Technical Report. As
@JeffDean
highlights, we have made significant progress in Gemini 1.5 Pro across all key benchmarks; TL;DR: 1.5 Pro > 1.0 Ultra, 1.5 Flash (our fastest model) ~= 1.0 Ultra.
As a math undergrad, our drastic
@MFrancis107
@OfficialLoganK
I am one of the PMs on the Gemini API. Could you say more about inline code execution. We do have a code-execution feature. So want to understand more in terms of how we can expand that to help your use case.
Obviously lots of pros and cons. But related idea - what if there was a chatbot to represent your colleagues’ perspectives and have discussions with “them” while they are on leave.
#GenerativeAI
#productidea
#responsibleai
All about getting the vibes right. My favorite prompting guidelines from the linked paper -
#1
Cut out politeness
#6
Lure the LLM with a $$$ tip
#9
Use phrases like "You MUST"
just tested o1-preview, gemini flash, gpt 4o and 3.5 sonnet on a task (bit boring but will share)
and i think gemini was best
all i gave the same prompts to
task was: take the data from this screenshot and return in markdown. (4o hallucinated some numbers - o1 i had to use the
Loved this rewind in memes. Funny but also such a great refresher of the incredible stuff that happened! May we need no/less? cherrypicking in 2024!..Amen!
ArtPrompt! Love how older forms of human creativity (ASCII art !!) are being channeled into new ways of talking to LLMs - even though it’s for an adversarial use case here.
Chain of Code is an awesome project led by my student
@ChengshuEricLi
on using LLM for better code-driven reasoning during his internship at
@GoogleDeepMind
this summer 😍👇
👶🤖 + 🛠️ = 🔥
Excited about how I’m giving tools to
@babyagi_
:
🗂️New Tools folder hosts the Tool class
🧰Adding a new tool is easy as add adding a new file in tools (auto-loads)
📜Describe tool and required API keys in tool file, provides missing API keys as error
🧮Handles
Exciting to see the release of multiple Duet AI capabilities our team has been working on, and all the positive response from our customers! I'm particularly excited with the flexibility Chat in the IDE and Smart Actions bring to Duet!
#DuetAI
Today we just announced Duet AI for developers at
#GoogleIO
, offering AI powered code generation and chat for developers directly in their IDE! You can request access on
#duetai
RewardBench tests reward models for instruction following, reasoning, safety on challenging, out-of-distribution queries. Its great to see Gemini 1.5 Pro leading the generative models category and Flash ⚡️ punching above its size!
Heard in an Uber (about how tech has changed SF for the worse) - “Earlier they used to tell me I could get killed at the intersection of 16th & Mission. Now that area is so gentrified, there’s no hope of that “
Another YC founder told us they browse their product's subreddit weekly for feedback
We automated it. Claude 3 now analyzes top posts and emails them a full report
It spots trends in user complaints and areas they can improve. Generating a formatted report with sources cited
Excited to have worked with the team to power so much of what was mentioned today - Studiobot, Duet AI, Code Tips in Search Laps and Bard for code.
#GoogleIO2023
#Codey
#PaLM2
Out of all of the researchers I've ever wanted a movie about, honestly, Emmy Noether is at the absolute top of the list. Can someone please make this? Her life should be immortalized. It's wild to me not everyone has heard of her. And there's so much I want to know
Today, we’re excited to introduce a new Gemini model: 1.5 Flash. ⚡
It’s a lighter weight model compared to 1.5 Pro and optimized for tasks where low latency and cost matter - like chat applications, extracting data from long documents and more.
#GoogleIO
Wei Wei and I did a technical session for
#googleio2024
on the code AI ecosystem at Google. And we were pretty comprehensive! ()
We talked about
• all Google code AI features (code completion, unit test gen, code transformation, auto commenting etc) and
I am building a system to teach Gemini Flash new skills by observing the actions I take on my computer so it can repeat them.
I found a way to achieve 100% accuracy in Gemini's ability to recognize what I am clicking.
The trick?
Draw a blue circle where the click happens. 🔵
@benjamincrozat
@OfficialLoganK
Benjamin - I am one of the PMs working with Logan. Can you share a bit about your use case. Also happy to have a separate chat if helpful
Wise words randomly encountered in a Google restroom " There are only two hard things in Computer Science: cache invalidation and naming things. " - Phil Karlton
For executable Python code responses in SGE, you can now see the output of the code to the right. Because we show you real output as opposed to expected output, this helps with overall response correctness.
#generativeai
#googleai
#SGE
#GoogleSearch