Gpt-oss, Genie 3, Personal Superintelligence and Claude pricing
The panel discusses OpenAI's strategic release of open-weight models (`gpt-oss`), the implications of Google DeepMind's immersive 3D world generator (`Genie 3`), the economic realities behind Anthropic's `Claude Code` rate-limiting, and the competing visions of "Personal Superintelligence" from major players like Meta, OpenAI, and Anthropic.
OpenAI’s Strategic Shift with gpt-oss
OpenAI’s release of two open-weight models, a 120B and a 20B parameter gpt-oss, marks a significant, albeit calculated, strategic move. The panel debates whether this signals a long-term transition to a fully open-source company. Chris Hay argues strongly against this, suggesting OpenAI will always keep its largest, most competitive models proprietary to maintain its edge and profitability. Kaoutar El Maghraoui views it as a “have your cake and eat it too” strategy, where OpenAI is walking a tightrope between competitive pressures from players like Mistral and Meta, and the need to maintain control. She notes that by releasing weights but not the training data or architecture, OpenAI reaps the benefits of community engagement while protecting its core IP.
Bruno Aziza adds that the future is hybrid, and this move allows OpenAI to offensively and defensively penetrate the enterprise market. Companies can now leverage the familiarity of GPT models in on-premise solutions, addressing data privacy and security concerns.
Key technical points about the gpt-oss models were highlighted:
- Hardware Specialization: The models are designed for consumer-grade hardware. The 20B model runs on a single GPU with 16GB of memory, and the 120B model is designed for a single A100 or a high-end MacBook Pro.
- Architectural Trade-offs: To achieve this efficiency, the models are text-only, English-only, and optimized for code generation and agentic tool-calling.
- Mixture-of-Experts (MoE): The release reveals that it is an MoE architecture with a very small number of active experts at any given time, which accounts for its remarkable speed. These models are not seen as direct competitors to OpenAI’s flagship multimodal, multi-language backend systems.
Ultimately, the move is seen as a politically savvy way to recapture goodwill, compete with a growing field of open-source alternatives, and reinforce US leadership in AI development.
The Immersive Future of Genie 3
Google DeepMind’s Genie 3, a model capable of generating interactive, navigable 3D worlds from text or image prompts, is considered a transformational development. Bruno Aziza sees it as a major step in how we experience and communicate information, with profound impacts on gaming, movies, education, and even corporate presentations.
Chris Hay believes 3D is the natural next frontier for AI, enabling more intuitive interaction. He envisions enterprise use cases like architects visualizing buildings or designers arranging kitchens, alongside consumer applications for creating personalized virtual environments. The panel agrees that the technology is a glimpse into a future where anyone can be a world-builder.
However, significant hurdles remain:
- Inference Cost: The primary barrier to mass adoption is the immense computational cost. The panel questions how quickly these costs can decline to make such technology accessible. Kaoutar El Maghraoui suggests that breakthroughs in hardware, memory technology, or perhaps even quantum computing might be necessary to solve the cost-of-inference challenge at scale.
- Real-Time Performance: The impressive demos raise questions about whether the generation is truly real-time or has been sped up for presentation.
The Economics of AI: Anthropic’s Claude Code Rate Limiting
Anthropic’s decision to implement rate limits on its $200/month “Max” plan for Claude Code sparked a discussion on the economic sustainability of powerful AI models. The change was prompted by a small fraction (5%) of power users running vast numbers of automated agents 24/7, making the fixed-price subscription unprofitable and causing service outages.
The panel views this as an inevitable sign of a maturing market. Kaoutar El Maghraoui calls it “the end of the free lunch,” as companies shift from a user “land grab” to a focus on profitability. This move establishes a clearer distinction between casual and professional users, where price is more directly tied to usage.
This policy change reveals fascinating user behavior. Chris Hay humorously expressed jealousy over the “window of opportunity where people were literally just taking money out of Dario’s pocket.” He predicts this will lead to more sophisticated user behavior, where users will automate routing tasks to different models based on cost and capability—a future he describes as “crazy time.” This necessitates a race to optimize the backend. As Kaoutar notes, the winners will be those who can deliver “scalable, fast, and affordable AI” through advanced techniques like KV cache optimization, speculative decoding, and tiered routing.
Competing Visions of Personal Superintelligence
Mark Zuckerberg’s recent essay on “Personal Superintelligence” crystalizes Meta’s vision for AI, adding another distinct philosophy to the mix. The panel contrasted the emerging ideologies:
- Meta: Focuses on personal empowerment, giving individuals AI “superpowers” to augment their lives and creativity. This vision is deeply tied to their hardware ambitions with AR glasses.
- OpenAI: Primarily driven by a vision of productivity and powerful, general-purpose intelligence.
- Anthropic: Prioritizes a safety-first approach to AI development.
Kaoutar El Maghraoui remains skeptical, viewing the “superintelligence” framing as partly marketing, but acknowledges that it presents a compelling vision. The critical battleground, she argues, is for the “operating system of personal AI.” The company that wins the hardware race—be it through glasses, neck-worn devices, or something new—will be in a prime position to dominate this new ecosystem. This could lead to a fragmented future where individuals interact with multiple, competing AIs from different companies across their various devices.