Our Menu :)

• Winning the AI Arm Race 🦾

• NVIDIA announced new AI chips in laptops 💻

• The last two weeks at OpenAI 🤖

• The biggest funding deals in AI this week 💰

• 🔒 Mistral Large🔒:  A powerful LLM that performs exceptionally well, second only to GPT-4.  

Actionable Tips:

• Test “Le Chat” from Mistral, Le French 🇫🇷 ChatGPT 🥖

Read time: 5 Minutes

GEN AI STARTUP
Winning the AI Arm Race 🦾 

The last 2 weeks have been mad in the AI world, with OpenAI launching model SORA and the debut of Gemini 1.5, flaunting its 1 million tokens, among countless other breakthroughs in AI.

Industry giants and startups alike are in a relentless pursuit to not just maintain their relevance but to dominate new frontiers, making the scene resemble a giant chess game where all the big players compete against each other.

However, the ultimate winner may not be the one who controls the individual pieces, but the one who owns the entire chessboard(s)

Last Friday, Nvidia became the fifth publicly traded company to ever surpass $2tn in valuation, just nine months after it became the ninth to ever break $1tn.

Apple, Microsoft, and Google took two years to cross that gap (although all three had the Covid pandemic in between the milestones).

But how?

In this article, we aim to unpack the strategies and visionary moves that have propelled NVIDIA to this pivotal position.

The Gaming Genesis

Originally, NVIDIA just focused on making graphics cards for real-time 3D rendering, mainly for gaming, where it was celebrated for its graphics processing units (GPUs).

But as GPUs got more advanced, engineers realized there was untapped potential in parallel processing. This led NVIDIA to pivot towards the concept of GPGPU - using GPUs for General-Purpose computing.

Now anyone could develop applications that leverage parallel processing, like cryptography, data analysis, scientific simulation, and deep learning.

The Strategic Pivot: From Gaming to CUDA

But it wasn't so easy at first. To program the GPU for a scientific application, you had to translate your code into graphical instructions.

It would be like trying to describe a complex math equation by using only art terminology. Complicated.

That's when in 2007 NVIDIA introduced CUDA, allowing programmers to simply instruct the GPU in a more understandable way. CUDA became deeply intertwined with the history of deep learning, even though it took a few years for that relationship to unfold.

It was a strategic pivot that redefined NVIDIA's role on the technological chessboard, setting the stage for its foray into AI and deep learning.

-——Techy part——-

When we train neural nets, information flows layer by layer through matrices of parameters. It turns out graphics rendering also involves multiplying matrices and vectors representing polygons.

So graphics and neural nets aren't so different – just multiplying vectors and matrices! By 2010, we had GPUs well-suited for this emerging algorithm called neural networks that seemed promising for AI.

It was in 2012 that a research team including Geoff Hinton, Ilya Sutskevern and Alex Krizhevsky used multiple GPUs to train a neural net called AlexNet.

They entered it into an image recognition competition and destroyed all other teams, showcasing the potential of deep learning.

A key aspect was using GPUs to drastically accelerate training of what were huge neural nets for the time.

———

And the rest is history! The deep learning revolution likely wouldn't have happened without the parallel development of the graphics card market.

As deep learning grew, so did NVIDIA. NVIDIA's revenue from data centers has skyrocketed from around $300 million in 2015 to over $15 billion in 2022, largely thanks to AI.

The AI Revolution

So what does the future hold for NVIDIA's consumer GPUs, given the rise of AI?

We’re already seeing two key trends – more tensor cores and more memory. Tensor cores accelerate deep learning, while memory allows bigger models to fit on your personal GPU.

As open-source AI models grow more capable, you'll likely want to run them locally without any limits.

NVIDIA is positioning itself to provide the hardware needed for these AI experiences through initiatives like DLSS and real-time generative video (Like the startup KREA is trying to do).

The entertainment industry will also increasingly tap AI to create interactive NPCs – enabled either locally with a beefy GPU, or in the cloud with NVIDIA chips.

Ultimately, NVIDIA has gained an advantage in this AI battle that looks hard to beat. We’ll find out more soon at GTC 2024, where NVIDIA will unveil their latest innovations.

NVIDIA Current Moves

• Nvidia hardware is eating the world. Wired interviewed Jensen Huang, Nvidia’s CEO and here are some bites that stood out to me.

  • Nvidia is building a new type of data centre called AI factory. Every company—biotech, self-driving, manufacturing, etc will need an AI factory.

  • Jensen is looking forward to foundational robotics and state space models. According to him, foundational robotics could have a breakthrough next year.

  • The crunch for Nvidia GPUs is here to stay. It won’t be able to catch up on supply this year. Probably not next year too.

  • A new generation of GPUs called Blackwell is coming out, and the performance of Blackwell is off the charts.

  • Nvidia’s business is now roughly 70% inference and 30% training, meaning AI is getting into users’ hands.

  • Nvidia's also announced new AI chips in a laptop-friendly package. (See Next Section)

The AI future is unfolding rapidly on the back of the —semiconductor industry...and right now, NVIDIA owns the board.

GEN AI STARTUP
Nvidia announced new AI chips in a laptop-friendly package 💻

NVIDIA is changing the AI landscape with the release of the RTX 500 and 1000 Ada Generation laptop GPUs. This move targets users interested in running advanced generative AI models on their laptops.

What is new?

NVIDIA has announced new AI chips tailored for laptops, marking a step toward making AI more accessible and portable for a wider audience.

NVIDIA's latest GPUs, based on the Ada Lovelace architecture, integrate with a neural processing unit (NPU) capable of handling AI tasks independently. This combination supports both routine AI operations and intensive workflows.

These GPUs, set to be available in Dell, HP, Lenovo, and MSI laptops this spring, aim to broaden AI tool access to various users, including students, businesses, and general consumers.

Why does this matter?

NVIDIA, a leader in GPUs for AI research and cloud computing, is extending its dominance to portable devices. A recent demonstration of local AI software, coupled with the introduction of AI-capable hardware, indicates a shift towards enhanced on-device AI performance.

This development promises significant improvements in on-device generative AI applications, enabling more sophisticated tools and applications directly from laptops. The expansion of AI capabilities into portable devices heralds new possibilities for users across the board.

GEN AI STARTUP
The Biggest Funding Deals in AI this Week 💰

1. Moonshot AI, a Chinese AI company building models with big context, raised >$1B at a $2.5B valuation.

2. Glean, a search tool for a business’s info, raised $200M at a $2.2B valuation.

3. Abridge, an AI that transcribes doctor-patient convos, raised $150M.

4. Intenseye, which uses AI to detect unsafe work conditions, secured $64 million.

5. Inkitt, a platform that helps people self-publish stories with help from AI, snagged $37M.

6. Myko AI, an AI that answers questions from sales data, raised $2.7M.

GEN AI STARTUP
The last two weeks at OpenAI 🤖

• Sora demo blows minds, we all know that. Right?

• One of OpenAI’s best-known researchers, Andrej Karpathy, co-lead on the new agent project, departs LINK

• ChatGPT has a mostly unexplained 6 hour (?) meltdown LINK

• Google and Mistral close the gap with GPT-4 LINK

• New results from Subbarao Kambhapati cast doubts on the robustness of the chain of thought prompting LINK

• Two new lawsuits against them have just been filed, one specifically on copyright and the other a broader class action

   

Dive Deep a bit about SORA

So here are some facts

• It's a text-to-video marvel that churns out up to 60-second clips from your text prompts.

• Remember when AI videos felt a bit... off? Sora smashes those barriers with its diffusion magic, ensuring videos stay consistent, even if characters pop in and out of frame.

• It can handle complex scenes, detailed backgrounds, and multiple characters with ease.

• No paper yet but here's what we know, at its core, Sora is built on a transformer architecture, refined through diffusion models.

• It's like having GPT's smarts but for videos, plus some DALL·E 3 cleverness for sticking to prompts like glue.

• Starting February 15th, Sora will be tested by red teams to identify potential risks.

  Here surreal example:

In a remarkable turn of events, this week we saw How Singapore embraces the change and prepares its citizens for the age of AI:

Embracing the change. Helping their own people adapt - that's the real leadership.

We need a bit of that from Western Economic leaders. Don’t you think?

GEN AI STARTUP
🔒 Mistral Large🔒: A powerful LLM that performs exceptionally well, second only to GPT-4.

New Closed-source LLM by Mistral AI!🤔 Yes, you heard it right. Mistral just released a new closed-source LLM available on their Platform or Azure!

That's quite a step away from their initial message and position to “release an open-source GPT-4 level model in 2024” - Arthur Mensch.

Mistral Large is the world's second-best language model available through an API after GPT-4.

Mistral Large boasts a performance of 81.2% on MMLU (measuring massive multitask language understanding), beating Claude 2, Gemini Pro and Llama-2-70B. Large is particularly good at common sense and reasoning, with a 94.2% accuracy on the Arc Challenge (5 shot). 

Some of Mistral’s features include: 

• Training on English, French, Spanish and Italian datasets for native multilingual capabilities. 
  

• 32k token context window - way below Gemini’s 1M token or Claude’s 200k token ones.
  

• Precise instruction-following  that were used to moderate the Chat interface.
  

• Native function calling capabilities for agentic capabilities similar to ChatGPT Plus.

The JSON format mode forces the language model output to be valid JSON.

This functionality enables the extraction of information in a structured format that can be easily used by developers.

Mistral offers a concise and unopinionated model, with fully modular moderation control.

This echoes the recent Gemini controversy: Google’s model produced outputs unaligned with the user’s expectations because of its trust and safety fine-tuning.

This release comes with the announcement of a collaboration with Microsoft, OpenAI’s main partner and investor, to provide the models through Azure.

The models are also available through La Plateforme, Mistral’s API platform. READ About the Collaboration

AND THIS WEEK

GEN AI AT WORK
Le Chat from Mistral, the French 🇫🇷 ChatGPT 🥖

Mistral AI is launching a brand new chat assistant, named Le Chat.

Now, we want to be upfront with you - this is a beta release, so there might be a few "quirks" here and there.

Access to the service is completely free (at least for now!). Plus, It has three different models for you to choose from:

• Mistral Small,

• Mistral Large,

• Mistral Next, (which is designed to be brief and concise).

Le Chat won't be able to access the web while you're using it.

However, Mistral will launch a paid version of Le Chat for enterprise clients.

So, what are you waiting for? Come on over to chat.mistral.ai and give Le Chat a whirl.

Thank you, see you next week!