AI Highlights

My top-3 picks of AI news this week.

OpenAI
1. OpenAI's Agent Thinks Deeper

OpenAI has launched Deep Research, an AI agent that transforms multi-hour research tasks into quick, comprehensive reports.

• Functionality: Able to browse the web, analyse data and PDFs, and then synthesise findings into detailed reports.

• User access: Rolling out to Pro users with 100 queries/month with plans to expand to Plus, Team, Enterprise, and eventually free tier (with fewer queries).

• Human-in-the-loop: Flags user when input is needed during the research process, the user can also take control and refine the research path.

Alex’s take: We’re watching AI evolve from specialised chatbots to general computer-use agents that are capable of performing real-world tasks at scale. In 12 months, this could feel as standard as using a search engine. Traditional research workflows will change forever, especially in business functions like sales where you can complete hours of personalised research in minutes.

ByteDance
2. ByteDance's Motion Masterpiece

ByteDance has unveiled OmniHuman-1, a groundbreaking AI model that sets a new standard for human motion generation and video synthesis.

• End-to-end generation: Creates seamless video content from a single image and audio input, handling complex motions and gestures naturally.

• Medium versatility: Excels across various forms of content including singing, talking, animation, and intricate gesture reproduction.

• Customisation: Generates realistic human videos in any aspect ratio and body proportion (portrait, half-body, full-body all in one), with realism stemming from motion, lighting, and texture details.

Alex’s take: Content will now become completely commoditised. Creators can now generate unlimited video content, digital avatars become more accessible than ever, and ads and video marketing get infinitely cheaper. We're witnessing a step change in the democratisation of video creation.

Google
3. Google's Gemini Accelerates

Google has launched its latest AI model suite, Gemini 2.0, featuring significant improvements across several key areas:

• Enhanced reasoning: Gemini 2.0 Pro Experimental and Flash Thinking models showcase advanced reasoning capabilities, directly competing with DeepSeek's popular R1 model.

• Massive context window: The Pro model can process up to 2 million tokens (approximately 1.5 million words), enabling analysis of extensive documents in a single prompt.

• Cost-efficient option: Introduction of Gemini 2.0 Flash-Lite, offering improved performance over the 1.5 version while maintaining the same price point and speed.

Alex’s take: The cost per 1M input/output tokens for Gemini 2.0 Flash-Lite is $0.075/$0.30. In comparison, GPT-4o mini is $0.15/$0.60. Gemini is 2x cheaper than its competitor model at OpenAI and also outperforms GPT-4o mini across key benchmarks. This is yet another example demonstrating that the cost of intelligence is going to near-zero, and context windows are expanding exponentially, opening the door to large-scale analysis and complex reasoning.

Content I Enjoyed

Humanoid robots can now move like athletes

This week, Nvidia’s GEAR lab, in collaboration with Carnegie Mellon University, introduced ASAP (Aligning Simulation and Real-World Physics).

It’s a reinforcement learning (RL) model that enables humanoid robots to perform ultra-smooth dynamic movements.

The problem with traditional physics engines is they don’t perfectly match real-world movements. As a result, simulation training falls short when this is transitioned across to real hardware.

ASAP combines the best of classical physics with AI for more realistic motion. Fewer trials are needed to adapt to the real world, and robots can now move with athlete-like fluidity and control.

Idea I Learned

Europe's AI strategy: Quality over quantity?

While the US and China dominate headlines with multi-billion dollar valuations and ambitious projects, Europe is taking a more European approach. It has earmarked $56 million for an open-source language model.

I always think it’s fun to compare EU and US initiatives in AI. Only two weeks ago, the Stargate Project was announced: $500 billion for US AI infrastructure over the next four years, backed by behemoths like OpenAI, Oracle, and SoftBank.

In comparison, Brussels’ announcement is 10,000x smaller. Nonetheless, DeepSeek claimed they spent just 1/10th of this EU budget to train their R1 model.

OpenAI, Meta, and other US companies have previously been hesitant or blocked from releasing tools in Europe due to the AI Act and the complexity of regulation in the region.

The funding picture tells a similar story. In 2024, European AI startups raised $8 billion, with 70% of this going to early-stage companies. In comparison, US funding hit $97 billion, over 10x the European figure.

Europe has always been quick to regulate and slow to invest.

I hope we can flip this on its head and prioritise acceleration instead of red tape in the years to come.

This is critical to ensuring we don’t fall further behind and that we support the most promising founders and projects for the generations to come.

Quote to Share

Brett Adcock on leaving their partnership with OpenAI:

The humanoid robotics company Figure decided to leave OpenAI due to a major breakthrough in developing fully end-to-end robot AI in-house.

LLMs are important yet are becoming commoditised. Especially when controlling a humanoid, there is a real need to have embedded AI. Something that is far more integrated than just driving the robot with language.

I think this is a smart move, taking on the full stack and making it proprietary. There’s just too much risk to have a core business function in the hands of another company.

Control over core tech is key for a lasting advantage and moat in robotics.

Source: Brett Adcock on X

Question to Ponder

Adobe's Acrobat AI Assistant launched its new contract features this week, and I've found myself genuinely reaching for it.

I used to have the dangerous trait of skimming contracts, nodding along while secretly hoping I hadn’t missed anything important. Now, AI can break down the key terms, explain the implications, and even compare versions.

It also lets me verify everything the AI suggested against the original text to avoid hallucinations—something LLMs are typically prone to.

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