ExoBrain weekly AI news

Welcome to our weekly email newsletter, a combination of thematic insights from the founders at ExoBrain, and a broader news roundup from our AI platform Exo…

Themes this week:

• WWDC and Apple's reasoning paper, while OpenAI launches o3-pro

• Disney and Universal suing Midjourney over AI-generated style mimicry

• Anthropic's 80% retention rate beating rivals amid talent wars 

Apple abandons all reason

This week Apple was the latest firm to hold its annual developer bash, but twelve months on from launching Apple Intelligence, AI was a no-show, replaced by a slew of releases more reminiscent of past tech eras. And for some reason they also chose this week to publish a paper claiming that AI doesn't truly reason. Meanwhile OpenAI slashed prices on its reasoning models and launched o3-pro, a 10x scale-up of its most powerful model.

Apple's paper, "The Illusion of Thinking," landed days before WWDC 2025, arguing that models from DeepSeek and Anthropic merely simulate reasoning through sophisticated pattern matching. The researchers showed these models failing on puzzle tasks, collapsing when complexity exceeded certain thresholds.

Not to be deterred, OpenAI made its most powerful reasoning model, o3-pro, available to ChatGPT Pro users and dropped the API pricing for o3 by 80%. This aggressive price cut means that frontier reasoning is now in reach for almost any application.

Back at WWDC, Apple's priorities lay elsewhere. The headline announcement? "Liquid Glass" - a design overhaul that Bloomberg's Jessica Karl compared to "the Windows desktop I was using in 6th grade computer lab class." While competitors push the boundaries of AI, Apple celebrated adding bringing iOS into the 1980s with its ability to drag and resize windows.

The AI announcements were limited; live translation, more reliance on ChatGPT, developer access to on-device models, an AI workout coach offering motivational talks? Despite Apple, with the iPhone and Siri, commanding the most powerful AI delivery platform on earth, they appeared completely distracted by the incremental.

The timing of Apple's reasoning paper was strange given they probably shouldn’t have been drawing attention to the question of advanced AI. And whilst the conclusions were parroted across the media they turned out to be a mixture of well understood phenomena, combined with a thesis rooted in the theoretical definition of reasoning… rule based, symbolic, elegant. But reasoning to get jobs done in the real world is rarely thus. When you or I solve a problem, we’re also pattern-matching against past experience, trying different approaches, following hunches. Sometimes we apply some approximated process, and almost always we post-rationalise with the sense that we knew what we were doing.

In contrast this week, research outfit EPOCH AI shared a study where professional mathematicians assessed the raw reasoning traces of OpenAI's models on hard mathematical challenges. What they observed was 'vibes-based exploration' rather than formal logic - an approach that often delivered impressive results, even if it didn't reflect how a professor might operate. But most of us aren't professional mathematicians, and reasoning models like o3-pro can sometimes achieve astonishing, intricate outcome - albeit imperfectly.

Tech history is littered with companies that were right about limitations but wrong about timing. Kodak correctly identified digital photography's quality issues… Blockbuster accurately noted streaming's bandwidth problems… They were right until suddenly they weren't. Apple risks the same fate. They're not wrong that current AI is brittle and maybe not ready to deploy fully across their global platform. Models do make confident mistakes and can be unpredictable. But like us they are searching for reason in a forest of data, and they can be immensely useful on that journey.

Takeaways: Apple's simultaneous release of a low-grade research paper and a backward-looking WWDC reveals a company strategically detached. They are no longer the benchmark for us all when it comes to design and execution. As OpenAI slashes prices and expands capabilities, and Google keeps moving up the gears, Apple is looking to relive the past. Apple has billions of devices running AI tech that can't tell you what month it is, and there’s no clear roadmap for when they will deliver competitive experiences let alone innovate. Worrying times for what is now only the world’s 3^rd most valuable company.

Can you copyright a style?

This week, Disney and Universal filed a lawsuit against Midjourney for copyright infringement in a move that could force the AI industry to reckon with a question it has so far largely avoided: what happens when a model learns a style?

Unlike earlier lawsuits focused on training data, this one zeroes in on the model’s outputs. The studios claim Midjourney users can generate images that closely resemble characters like Elsa, Darth Vader or the Minions, simply by typing prompts such as “Minion style Pixar, cinematic lighting.” The implication is that the model hasn’t just ingested images—it’s absorbed the essence of a visual language.

At the centre of the case is a blurry legal boundary: can style be copyrighted? Copyright protects specific expressions, not general ideas. But generative models compress thousands of expressions into latent fingerprints and then output work that mimics those expressions without duplicating any particular frame. For Disney and Universal, this kind of mimicry could undermine decades of brand-building.

Midjourney has not yet responded publicly, but the complaint includes side-by-side comparisons showing how AI-generated outputs mirror familiar poses, colours and compositions. The studios argue this goes beyond inspiration into the territory of systemic reproduction.

There’s precedent to suggest the law might struggle here. Courts have typically found that ‘style’ itself isn’t protectable. But that logic comes from an era when only humans created work. These models, trained at internet scale, operate differently: they can synthesise a style so closely that the distinction between homage and duplication becomes academic.

The outcome of this case could set the first major precedent on whether a brand's visual signature counts as IP when reproduced by AI. It could also shape how AI companies approach prompt filtering, user responsibility, and training disclosure.

It’s not just images. Music models are heading in the same direction. Prompt "Taylor Swift vocals, sad acoustic ballad" and you might get something uncomfortably close. In a world where style can be captured by a latent vector and rendered on command, the legal status of "feel" becomes a live issue.

For AI developers, the risks are escalating. If this suit succeeds, tools may be forced to block outputs even loosely associated with corporate IP. And for the entertainment industry, it’s a reminder that their most valuable assets may not be characters or plots, but the stylistic fingerprints they leave behind.

Takeaways: This is the first major lawsuit focusing on output style rather than training data. It tests whether AI-generated mimicry of brand style constitutes copyright infringement and will be a key step in setting the framework for commercial AI firms will need to operate, if not be able to determine how the technology itself evolves.

AI labs fight for talent

This image reveals Anthropic's 80% two-year retention rate, outpacing all major AI labs including OpenAI (67%) and DeepMind (78%). The data captures a critical moment in AI's talent wars, where researchers typically hop between labs every 18-24 months. Meanwhile reports suggest that Meta are desperate to hire, offering packages up to $10m a year to top talent to build their new "Super Intelligence" team after Llama 4 flopped.

Weekly news roundup

This week's news shows AI becoming deeply embedded across traditional industries while governance concerns intensify around security and job displacement, research advances focus on model adaptability and efficiency, and the hardware race accelerates with major infrastructure investments amid growing geopolitical tensions.

AI business news

• Meta's finalises Scale AI investment and hires CEO (Shows major tech companies investing heavily in AI data infrastructure and talent to maintain competitive advantage.)

• Barbie-maker Mattel teams up with OpenAI, eyes first AI-powered product this year (Demonstrates AI expanding beyond tech into traditional consumer products and toys, signalling mainstream commercialisation.)

• Fears over U.S. AI dominance boost business for France's Mistral (Highlights how geopolitical concerns are creating opportunities for non-US AI companies and diversifying the global AI landscape.)

• Building for a new era: Databricks takes on pain points of complexity, lock-in and cost for enterprise AI (Addresses key barriers to enterprise AI adoption, crucial for understanding how AI will scale in business environments.)

• The Browser Company launches its AI-first browser, Dia, in beta (Shows AI integration into everyday tools, potentially changing how we interact with the web and information.)

AI governance news

• Data bill opposed by Sir Elton John and Dua Lipa finally passes (Important UK legislation affecting data rights and AI training, despite celebrity opposition highlighting public concerns.)

• Microsoft Copilot zero-click attack raises alarms about AI agent security (Critical security vulnerability in AI assistants reveals new attack vectors as AI agents gain more system access.)

• Op Spiderweb: Ukraine said drones hit Russia's planes using AI when they lost signal (Demonstrates real-world military AI applications and autonomous decision-making in warfare scenarios.)

• Vibe coding is coming for engineering jobs (Explores how AI-assisted coding is fundamentally changing software development roles and skills requirements.)

• New York passes a bill to prevent AI-fuelled disasters (State-level AI regulation focusing on preventing catastrophic risks, setting precedents for other jurisdictions.)

AI research news

• Self-adapting language models (Research advancing models that can modify themselves during use… a potentially huge development.)

• Text-to-LoRA: instant transformer adaption (New technique for rapidly adapting models to specific tasks without extensive retraining.)

• Mistral Magistral (Latest model from European AI leader Mistral, important for understanding non-US AI capabilities.)

• Confidence is all you need: few-shot RL fine-tuning of language models (Advances in efficient model training using reinforcement learning with minimal examples.)

• Reinforcement pre-training (Novel approach to training models that could improve their reasoning and decision-making capabilities.)

AI hardware news

• AMD announces MI350X and MI355X AI GPUs, claims up to 4X generational performance gain, 35X faster inference (AMD challenging Nvidia's dominance with significant performance improvements, potentially reducing AI compute costs.)

• China's racing to beat U.S. chip curbs. How its supply chain stacks up (Analysis of China's efforts to build domestic AI hardware capabilities despite export restrictions.)

• 'Neocloud' Crusoe to buy $400 million worth of AMD chips for AI data centres (Major infrastructure investment showing growing demand for AI compute and AMD's market share gains.)

• Nvidia to build first industrial AI cloud in Germany (Expansion of AI infrastructure in Europe, important for understanding global AI compute distribution.)

• 'AI maker, not an AI taker': UK builds its vision with NVIDIA infrastructure (UK's strategic AI infrastructure investments to maintain technological sovereignty and competitiveness.)