- Bot Eat Brain
- Google Gemini. The good, the bad, and the ugly.
Google Gemini. The good, the bad, and the ugly.
PLUS: Rock out with your Grok out
Good morning, human brains. Welcome back to your daily munch of AI news.
Here’s what’s on the menu today:
Google Gemini. Dope or dupe? 🤖 💥
All about Google’s claims, its infamous demos, public reception, and more.
The safest llama is a purple llama 💜 🦙
Meta launched Purple Llama. It claims this is to build responsible AI.
Rock out with your Grok out 😤 🤘
xAI rolled out Grok, its AI model, to 𝕏 Premium + subscribers.
The Google Gemini rollercoaster 🤖 💥
Last Wednesday, Google launched Gemini. It claimed to be an AI model that outperformed GPT-4 in several different benchmarks.
Here’s a screenshot of the demo:
In this section, Gemini accurately determines that the duck is rubber and since it squeaks, it probably floats as well.
I don’t have time to watch it.
We got you.
The demo was impressive. It showcased Gemini engaging in conversations, making jokes, recognizing objects, playing games, and solving puzzles in real-time.
What about the specs?
Google released a chart in its blog post showcasing how Gemini outperformed GPT-4 in multiple benchmarks.
Check it out:
Indeed, but it turns out Gemini Ultra isn’t coming out until next year.
So, Google didn’t launch this one yet?
Correct. Gemini comes in three sizes: Ultra, Pro, and Nano.
Ultra is the one featured in the demo and it outperforms GPT-4 on several benchmarks. It won’t appear until next year and will be in a service called Bard Advanced.
Pro is available now in Bard. Google is also releasing developer access for Pro models on Wednesday.
Nano will run on mobile devices, starting with Google’s Pixel 8 Pro.
So if it’s as good as the demo, it’s worth the wait?
Right, the demo.
The demo implied real-time responses to voice and video, but it was based on still image frames and text prompts.
Many of the instructions provided to Gemini were also omitted from the video.
Google admitted that the video was staged and edited to exaggerate Gemini’s capabilities.
It claims that the demo was intended to illustrate Gemini’s potential and inspire developers.
So, we just got duped?
If you feel that way, you’re not alone.
From Our Partners
Get your first 100 customers.
Are you building a side hustle? Grow your business the easy way using LoopGenius.
In just 30 seconds Loop Genius can create:
🚀 Your landing page: Turn your idea into reality.
🎯 Marketing Loops: Master Twitter, LinkedIn & more
✍️ AI-generated content: They write, you press "send"
Purple llama > normal llama? 💜 🦙
Last week, we reported on Meta and IBM’s new AI Alliance. It contains over 50 AI leaders from NASA, AMD, Intel, Hugging Face, and more.
The next day, Meta announced Purple Llama. It claims the purpose is to establish trust in AI development by providing tools for building responsible AI.
Purple, as in nurple?
Yes, Meta claims that to best tackle the challenges of AI safety, they need to utilize both red-team and blue-team tactics.
Red-teaming is when you simulate real-world attacks and attempt to identify weaknesses and security vulnerabilities in AI.
Blue-teaming is when you build, strengthen, and maintain defensive capabilities in AI systems.
Have they done anything yet?
They announced two new components: Cybersec Eval and Llama Guard.
Cybersec Eval is a series of cybersecurity benchmarks for AI models. Meta claims it’s built in collaboration with security experts.
Llama Guard is an open-source AI model that achieves state-of-the-art performance on several benchmarks. It is trained to be less likely to generate harmful outputs.
Meta said it plans to host a workshop at NeurIPS 2023 to further discuss these tools.
A LITTLE SOMETHING EXTRA
Time to let the Grok out 😤 🤘
In July, we reported on xAI. It’s Elon Musk’s AI company with AI engineers from OpenAI, Tesla, DeepMind, and more.
In November, we covered xAI’s new AI model, Grok. It’s named after Robert A Heinlein’s 1961 novel, Stranger in a Strange Land.
Yet another ChatGPT competitor… Who cares?
A key feature of Grok is that it uses real-time data to stay up-to-date. It uses both current and past tweets to give answers to queries.
How much is it?
It’s $16 a month to become a 𝕏 Premium + member.
I don’t have access to Grok yet.
Musk claims that it will be available to all 𝕏 Premium + subscribers in about a week.
Japanese users will receive beta access next and all languages will gain access by early 2024.
MEMES FOR DESSERT
YOUR DAILY MUNCH
LoopGenius — Grow your business the easy way using AI-powered marketing loops. [Sponsor]
Holly AI — an AI-powered recruitment tool that helps you efficiently find job candidates.
Vizard — takes long-form videos and automatically creates short-form videos.
Full Stack AI — open-source tool that builds full-stack apps from your prompts.
Sam Altman is TIME’s CEO of the year. The article also features interviews with members of OpenAI and information on the origins of the company.
Will Google Bard thrive or die? Bard has tons of data to train on, but way more people use ChatGPT and feed it more data every day.
How to build a desktop app that chats with your documents. How to create an LLM that runs natively on your computer.
Stability AI unveiled StableLM Zephyr 3B. It’s an AI model with 3 billion parameters that uses regular hardware to generate text.
Apple released MLX. It’s a framework for training AI models on Apple silicon.
Intel and Stability AI are developing an AI supercomputer. It claims to be six times faster than Stability AI’s previous supercomputer.
PATHFINDER — a method that enhances LLMs’ ability to handle complex tasks that require multi-step reasoning.
DreaMoving - a framework that uses diffusion models to create high-quality videos of humans dancing.
Mitigating LLMs’ discrimination —Anthropic’s new method for finding and eliminating bias in AI models.
TWEET OF THE DAY
Mistral released its second AI model. It’s open-source, contains no security guardrails, and outperforms GPT 3.5.
Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.
Until next time 🤖😋🧠
What'd you think of today's newsletter?