AI Highlights

My top-3 picks of AI news this week.

Anthropic
1. Claude’s the captain now

Anthropic has introduced computer use capabilities in public beta, allowing Claude to interact with computers like humans do, alongside an analysis tool and significant updates to their model lineup.

• Computer use: Claude can now perceive and interact with computer interfaces by looking at screens, moving cursors, and typing text.

• Code execution: Claude can now write and run JavaScript code directly within the platform, enabling real-time data processing and analysis.

• Model updates: The upgraded Claude 3.5 Sonnet shows major improvements in coding tasks (49.0% on SWE-bench Verified), while the new Claude 3.5 Haiku matches their previous largest model's performance at a lower cost.

Alex’s take: What fascinates me about this release is the shift from teaching AI specific tasks to teaching it general computer skills. On OSWorld, which tests AI systems' ability to use computers like humans, Claude 3.5 Sonnet scored 14.9% in screenshot-only tasks—nearly double the next-best AI system's score of 7.8%.

Rather than creating individual tools for individual tasks, Anthropic is essentially teaching Claude to be a “digital native”, able to use any software designed for humans. While still experimental, I believe this approach could fundamentally change how we think about the next stage of AI automation—Claude is on fire!

Midjourney
2. Midjourney’s pixel playground

Midjourney has launched its highly anticipated image editor, marking a significant shift from generation-only to sophisticated image manipulation.

• External image editing: Users can now modify their own images through text prompts and region selection, compatible with model personalisation and various reference types.

• Retexturing capability: New mode allows preservation of shapes while modifying lighting, materials, and textures.

• Controlled rollout: Initially available to yearly subscribers, long-term members (12+ months), and users with 10,000+ images generated.

Alex’s take: I think their careful rollout strategy is a smart move to reduce risk here. Given Midjourney’s popularity, it’s especially important to prioritise human moderation and prevent their servers from burning up with all the excitement.

The idea that you can now turn quick sketches into ultra-realistic artwork is nothing short of spectacular, especially for curious kids who want to bring a story to life or fully-fledged artists who want to quickly brainstorm an idea.

Runway
3. Lights! Camera! Act-One!

Runway has unveiled Act-One, an AI tool within Gen-3 Alpha that transforms simple video inputs into expressive character performances.

• Simplified animation: Eliminates complex traditional animation pipelines by using only video and voice performances as inputs, no special equipment needed.

• Flexible applications: Works across various character designs and proportions while maintaining realistic facial expressions and emotions.

• Multi-character scenes: Enables creation of dialogue scenes using just one actor performing different characters, opening new possibilities for narrative content.

Alex’s take: Act-One is democratising character animation. We've gone from needing expensive motion capture equipment and technical expertise to being able to create expressive performances with just a smartphone camera. This feels like a pivotal moment where sophisticated animation techniques are becoming accessible to independent creators and storytellers.

We saw only a few weeks ago Runway announced their collaboration with Lionsgate. They now possess a first-mover advantage across the stack from creation to production—I wonder how other companies will respond.

Content I Enjoyed

Former OpenAI Researcher says the Company broke copyright law

Suchir Balaji is a former AI researcher at OpenAI.

He played a pivotal role in organising the data for ChatGPT to train on, initially believing that OpenAI had the legal right to use any available internet content.

However, following the launch of ChatGPT in November 2022, he reassessed his position and concluded that OpenAI’s methods were not only unlawful but also detrimental to the internet ecosystem.

It was this realisation that led him to leave the company in August, making him one of the first insiders to publicly criticise the use of copyrighted materials in the training of large AI models.

I feel this article from The New York Times highlighted some really thought-provoking questions that underscore the urgent need for regulatory intervention. We have a lot to address when it comes to the legal ambiguities surrounding AI development today.

It’s our job to find a balance between creating intelligent models that benefit humanity whilst respecting creator rights.

Even though OpenAI has been the ‘top dog’ since their ChatGPT moment of 2022, the exodus of talent and quickening development from the likes of Anthropic, Meta and Grok make me wonder who will end up on top in this ever-accelerating field.

Idea I Learned

Where is the future of AI being built?

I recently visited Nebius' data centre in Finland. This powerhouse is where AI models are trained, with thousands of GPUs working in unison.

They helped Mistral train their multimodal models.

Their mission: build a full-stack AI cloud platform.

Here's what I learned:

1. There’s a scarcity of GPUs in the US

   • Clusters are being sold in massive packages

   • People who need smaller requirements can't find them

2. Nebius is building a self-serve platform

   • Covers infrastructure requirements from a single GPU to big GPU clusters

   • They're not just a GPU reseller—they're designing the servers and racks from the ground up

3. They cool their servers in a sustainable—and rather unique—way

   • Nebius uses the naturally cold Finnish air

   • The heat generated from the servers is then fed back into the grid

   • They not only heat their onsite building but also warm homes nearby

They can recover 70% of the heat generated, which I think is incredible.

And it's the first in the world to have this heat reuse application.

What does this mean for the future of AI? Compute is scaling exponentially. Only a few weeks ago, Mark Zuckerberg highlighted that Llama 3 was trained on 10-20K GPUs, Llama 4 will be trained on more than 100K+ GPUs, and Llama 5 even more than that.

It’s possible that this GPU limit might not happen anytime soon.

Nebius seems to be in a strong position to capitalise on this, expanding their Mäntsälä data centre in Finland to host up to 60,000 GPUs dedicated to intensive AI workloads—and they're adding other locations too.

I’m a big believer that the future of AI depends on building infrastructure like this that balances performance with sustainability. I had a great time visiting in person and learning about these concepts first-hand. If you want to see the data centre in action, check out my video here.

Quote to Share

Miles Brundage on leaving OpenAI:

Miles Brundage, Senior Advisor for AGI Readiness at OpenAI, has announced his decision to leave the organisation.

I’ve now lost count of everyone who’s departed—we’ve covered two individuals in this issue alone—but it seems Miles’ motivation focuses on something I also strongly believe in: the ability to have freedom when clicking publish.

In his accompanying blog article, Miles highlights that he wants to be more independent working on issues that cut across the whole AI industry, not just within OpenAI.

“OpenAI is now so high-profile, and its outputs reviewed from so many different angles, that it’s hard for me to publish on all the topics that are important to me.”

Freedom to share your thoughts in an unfettered way is something I truly respect. Miles believes that influencing AI development from outside the industry will allow for more effective and unbiased contributions to making sure AI benefits all of humanity.

Question to Ponder

I had to retake a year of school.

Not because I wasn't capable, but because I couldn't keep up with the pace of the classroom. By the time I understood one concept, the class had moved on to the next.

This memory hit me hard when I first read about ChatGPT’s voice mode and OpenAI’s demo showcasing its ability to view your screen.

It immediately got my mind buzzing about having a tutor who can see your work, hear your questions, and adapt to your learning style in real-time—one who never gets tired or frustrated or moves on before you're ready.

We're not talking about replacing teachers—they're irreplaceable. We're talking about supplementing them with AI that can provide answers to your unlimited “why's” at 3 AM when that math problem finally starts making sense.

I honestly feel the implications are staggering. A child in a remote village could have access to the same quality of personalised education as a student in an elite private school. A working parent could study at their own pace without sacrificing family time.

But here's what really excites me: the potential to rewrite educational “failures.”

Today's “U” grade student might just need a different approach to become tomorrow's “A” student. I know because I lived it. And with AI tutors becoming increasingly sophisticated, we might finally have the tools to make this transformation accessible to everyone.

The next five years will change not only how we learn but also who gets to learn. And that's a future worth getting excited about.

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