AI Highlights

My top-3 picks of AI news this week.

OpenAI
1. ChatGPT's search evolution

OpenAI has launched ChatGPT search, upgrading ChatGPT's web search capabilities with relevant web sources.

• Search integration: ChatGPT now automatically determines when to search the web, providing real-time information with source citations.

• Broader accessibility: Available to ChatGPT Plus and Team users immediately, with Enterprise and Edu users gaining access in weeks, and a gradual rollout to Free users.

• Underpinning technology: Powered by a fine-tuned GPT-4o model with novel synthetic data generation techniques and integration with third-party search providers.

Alex’s take: I can’t help but feel Google’s market share is getting eroded more and more as the weeks pass by. First, it was Perplexity. Now, it’s OpenAI taking a big bite of the pie. Sundar Pichai needs to get his act together—it feels like for a company with such a big advantage in this space, they’ve been far too slow to act on “AI”.

Boston Dynamics
2. Atlas takes a factory job

Boston Dynamics, in collaboration with Toyota Research Institute, has demonstrated Atlas performing autonomous factory work, showcasing advances in real-world problem-solving.

• Autonomous operation: Atlas independently moves engine covers between containers using ML vision models for environment recognition and navigation.

• Error handling: The robot demonstrates the ability to detect and recover from failures like tripping and collisions using multiple sensor inputs.

• Learning capabilities: Implementation of Toyota's Large Behavior Model technology suggests a move towards more adaptable, general-purpose robotics.

Alex’s take: I like the authenticity Boston Dynamics brings to their promo videos. It’s not just backflips, it’s failures too. Whilst some companies are focused on mastering walking, Atlas is already being useful inside real factory settings. They’ve also done something I love—the swivel. Bypassing human constraints of movement and taking the most efficient path when completing a task.

Nvidia
3. Nvidia's mighty mini neural net

Nvidia's GEAR Lab has introduced HOVER, a 1.5M-parameter neural network that demonstrates how smaller models can achieve remarkable results in robotics control.

• Efficient training: Leverages NVIDIA Isaac Sim to compress 1 year of robot training into just 50 minutes on a single GPU.

• Versatile control: Supports multiple input modes, including head/hand poses, whole-body poses, joint angles, and root velocity commands.

• Zero-shot transfer: Successfully transitions from simulation to real-world applications without additional fine-tuning.

Alex’s take: I’m excited about the applications of this technology as it works with any humanoid that can be simulated in the Isaac environment. What’s more, 1.5M parameters is only a fraction of what most language models use (OpenAI’s GPT-4 has ~1.76 trillion parameters). This efficiency-first approach could be the key to making advanced robotics more accessible and deployable at scale—imagine 10,000 Atlas robots that we saw above training and working in unison.

Content I Enjoyed

The sweet smell of AI success

Osmo has achieved a groundbreaking milestone in AI development by digitising scent for the first time ever. This means computers can detect and potentially reproduce scents.

I really enjoyed their video about how they were able to achieve this.

The potential applications of this technology are vast:

• Security: At concert venues, schools and airports, having an “artificial nose” that can detect narcotics and explosives.

• Entertainment: Having a VR headset omit scents that can lead to a more immersive experience when gaming.

• Retail: Being able to try different fragrances virtually at home before choosing to buy.

The list goes on. What happens when we can equip humanoid robots with a sense of smell? I guess we’ll find out over the next year.

Idea I Learned

How to edit images with Midjourney

I've been exploring Midjourney's image editing capabilities, and I've been rather impressed with its performance.

The editor itself (only released last week) is straightforward to use. Here's how you can get started:

• Upload your own image to Midjourney

• Use text prompts to describe your desired changes

• Select specific regions you want to modify

• Try 'image retexturing mode' to play with materials and lighting

• Experiment with style references and character prompts

For now, Midjourney has restricted the ‘editor’ access to their most committed users—it makes sense as they're less likely to misuse this new feature.

To help you get started, I've created a short video tutorial. Be sure to check it out before diving in.

Quote to Share

Bill Gates on how AI impacts human connection:

“The potential positive path is so good that it will force us to rethink how should we use our time. You can almost call it a new religion or a new philosophy of, okay, how do we stay connected with each other, not addicted to these things that will make video games look like nothing.”

Bill Gates raised an intriguing point on the “Possible” podcast with Reid Hoffman this week about AI becoming so captivating that we might need new philosophical frameworks to maintain human connections.

While some focused on the sensational “new religion” angle, I think Gates is highlighting a deeper concern we're trying to grapple with.

We already struggle to maintain meaningful connections in our social media dopamine-dominated world. What happens when AI can create perfectly tailored experiences that are infinitely more engaging than our current digital distractions?

Source: Bill Gates on possibility, AI, and humanity

Question to Ponder

I had a fascinating conversation about this with a friend recently.

They showed me some of their AI-generated art and said, “Look, it created something I never would have thought of!” This made me pause.

We often think of inspiration as this mystical force that springs from our emotions and personality—that special human spark. But what if we're looking at it backwards?

Our personalities and emotions are shaped by our experiences. Every book we've read, every conversation we've had, and every challenge we've faced becomes part of who we are.

AI, in a way, mimics this process. It takes in vast amounts of training data—in the case of image generation, it’s human creativity and art—and then combines it in unexpected ways.

Almost like a jazz artist drawing from every piece of music they’ve ever played to create something new during a session of improvisation.

The real question might not be whether AI falls short of inspiration but whether it offers a different kind of inspiration altogether. One that comes from seeing patterns and possibilities that our human minds, with their emotional biases and personality quirks, might, in fact, miss altogether.

When we can use AI to augment our thoughts and show us new paths we never would’ve otherwise explored, we get to unlock the first glimmer of what a human-AI creative partnership might look like.

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