AI Highlights

My top-3 picks of AI news this week.

Google
1. Google steals the show

Google has launched a serious number of AI announcements this week, overshadowing their competition’s “12 days of OpenAI” campaign:

• Gemini 2.0: Their latest LLM claims top performance at 1/100th the cost of OpenAI's offerings, with advanced multimodal capabilities across text, images, audio, video, and code.

• Veo 2: State-of-the-art video generation model achieving breakthrough quality in physics understanding, human movement, and cinematography, supporting up to 4K resolution.

• Imagen 3: Released their highest quality text-to-image model yet, focusing on enhanced detail, lighting, and prompt accuracy.

• NotebookLM: All NotebookLM users can now join the conversation in Audio Overviews by choosing “Interactive Mode”.

Alex’s take: Whilst OpenAI tries to charm us with its “12 days of OpenAI” campaign, Google has unleashed what feels like 12 months of innovation in a single week. NotebookLM is one of my favourite Google products of recent (especially for collating research). Being able to join the two-person podcast audio discussion will help reinforce learning and understanding for everyone. Check out my quick explainer to get started with NotebookLM.

OpenAI
2. OpenAI's Grand Finale

OpenAI wrapped up 2024 with the final updates of their holiday special:

• o3 Model Family: OpenAI announced o3 their most significant leap yet in AI capability and reasoning models. Achieving 87.5% on ARC-AGI test (triple o1's performance), 96.7% on the 2024 American Mathematics Exam, and 25.2% on EpochAI's Frontier Math problems.

• Free Search: Released ChatGPT Search to all users, previously a premium-only feature, with improved mobile integration and voice mode compatibility.

• Universal Access: Launched ChatGPT on landlines (1-800-CHATGPT) and WhatsApp, making AI accessible without smartphones or the internet.

Alex’s take: It’s interesting to see OpenAI having to skip o2 due to the ‘O2’ telecom trademark. However, o3 looks like a serious powerhouse. It features “adjustable reasoning” that allows users to set compute levels (low, medium, high) to optimise speed and accuracy. Safety researchers can apply to sign up here.

Memorial Sloan Kettering Cancer Center
3. AI spots cancer warning signs in blood

Memorial Sloan Kettering Cancer Center has developed an AI-powered blood testing technology that could accelerate early cancer detection.

• Nanotube innovation: Uses carbon nanotubes 50,000 times smaller than human hair that emit fluorescent light based on molecular interactions in blood samples.

• Pattern recognition: AI algorithms interpret complex molecular patterns that are impossible for humans to detect, achieving higher accuracy than current cancer biomarkers.

• Early detection: The system shows promise in identifying ovarian cancer years before symptoms appear, with potential expansion to triage all gynaecological diseases.

Alex’s take: It’s often when we combine technologies together that we get incredible results. Nanotubes and AI are now solving a problem that neither could tackle alone. The nanotubes can generate the signal, but only AI can decode it. I think this is a perfect example of how AI isn't replacing existing technologies but rather unlocking their full potential.

Content I Enjoyed

Genesis trains robots 430,000x faster than the real world

Genesis is a physics simulation platform that’s designed for robotics, embodied AI and physical AI applications.

Its applications are immense. To name a few:

• Robotic manipulation

• Human and animal character motion

• Fully interactive prompt-to-3D scene generation

It can even simulate scenes at 43 million FPS, which is 430,000x faster than the real world.

Jim Fan, who leads embodied AI at Nvidia, is one of the co-authors of the Genesis paper. He highlighted one hour of compute time leads to 10 years of training experience. This is exactly how Neo in The Matrix was able to learn martial arts so quickly. Genesis already performs well across specific tasks in this domain.

Genesis can simulate such a diverse range of materials. Chocolate, sponges, fruit, you name it. I recommend watching some of the video examples on their GitHub page to spark some wonder.

The team behind Genesis are open-sourcing the underlying physics engine. This means other researchers and teams can build on top of the work and help accelerate simulation further.

How long until we’re uploading skills like kung fu to our very own brain-computer interface?

I’ll leave that for you to ponder.

Idea I Learned

Gemini 2.0 Flash Thinking Packs a Punch

Google just announced Gemini 2.0 Flash Thinking this week.

It’s a new experimental version of Gemini 2.0 that shows its thinking in real-time.

What I love about this model is how competent it is, given its lightweight nature.

It’s already outperforming OpenAI’s “o1” model. I wanted to hear what the hype was all about, so I gave it a go myself.

Needless to say, I wasn’t disappointed.

• Beat o1 across the question set

• Incredibly fast response times

• Costs $0 to use vs o1’s $60 / 1M token pricing

Given that Google's “Flash” version of its model series is already beating o1, I expect Google’s next frontier model to pack even more punch.

Try it yourself:

1. Navigate to aistudio.google.com

2. Click on the “Model” dropdown

3. Select “Gemini 2.0 Flash Thinking Experimental”

It’s totally free to use.

To help you get started, I made a short video on my YouTube channel exploring its capabilities.

Check it out here.

Quote to Share

“I think I might have cooked it a little bit too long. It shrunk.”

You might recognise Nvidia CEO Jensen Huang for his iconic black leather jacket.

This week, they just released their brand new AI computer, the “NVIDIA Jetson Orin Nano Super”.

It’s optimised for on-device AI processing in robots, and this thing is seriously cool (and seriously small).

At $249, it provides developers, students, and builders with the most affordable and accessible platform to work on real-world challenges with physical AI.

Source: Nvidia Developer

Question to Ponder

Tl;DR: Be cautious about sharing sensitive information. By default:

• ChatGPT uses your data to train its models

• Claude doesn’t use your data to train its models

• Gemini doesn’t use your data to train its models

To prevent ChatGPT from training on the data you share via chats, you can navigate to ChatGPT → click on your profile → settings → data controls → toggle on/off “improve the model for everyone”.

However, if you toggle this off, you won’t be able to save your conversation history.

A neat workaround is to use a “temporary chat”. The contents inside this chat won’t be used to train OpenAI’s models.

I do not recommend sharing any sensitive data, regardless of these settings. Try using generic examples or anonymise/alter identifying details instead.

💡 If you enjoyed this issue, share it with a friend.