For this week’s Radical Reads feature, Radical Ventures Partner Daniel Mulet explores the major themes from NeurIPS 2025 that may define the AI landscape in 2026.
The thirty-ninth conference on Neural Information Processing Systems (NeurIPS) wrapped up in sunny San Diego this month. As the most significant gathering on the AI research calendar, NeurIPS 2025 drew 22,000 attendees, including the world’s top researchers and practitioners, for a week of paper discussions, workshops, and meetings.
While the sheer scale of foundation models defined previous years, this year’s conference signalled a shift toward reasoning and efficiency. If 2023-2024 was about models with “more parameters,” 2025 is about models that do “more thinking”.
This week, we explore the major themes that may define the AI landscape in 2026.
1. The Shift to “System 2” Reasoning
The era of pure pattern matching is evolving into an era of deliberate reasoning. A dominant theme at this year’s conference was the shift from “System 1” thinking (fast, intuitive, but prone to hallucinations) to “System 2” thinking (slower, deliberate, and logical) used by AI agents to accomplish long-duration tasks.
- Beyond Next-Token Prediction: Workshops like “System-2 Reasoning at Scale” highlighted the industry’s focus on giving models the ability to “pause and think” before answering.
- Inner Monologues: New methods presented this year, such as “Gated Attention”, allow models to modulate their own processing, effectively choosing when to spend more compute on hard problems.
2. Efficiency
As the energy demands of AI data centers dominate headlines, the research community is responding with a wave of “Green AI” architectural innovations that lead to more efficient compute use.
- Architectural Overhauls: The “Gated Attention” paper mentioned previously proposes adding a “gate” to the attention mechanism. This change not only improves performance, but stabilizes training and scaling, potentially reducing compute waste.
- Smarter, Not Just Bigger: With AI models growing in size, a significant trend is improving efficiency, including faster inference, memory optimization, and energy-aware model design.
3. AI Enters the 3D World
We saw a surge in research focused on systems that can understand and connect information across text, images, and video, as well as AI agents that act in the physical world.
- Richer 3D Worlds: A significant cluster of papers focused on generating high-fidelity 3D scenes and assets from text or images, in particular via world models, a capability that is critical for everything from gaming to industrial digital twins.
- Simulation as a Learning Ground: The Machine Learning and the Physical Sciences workshop showcased papers that use AI to emulate complex physical systems faster and more robustly than traditional physics engines. Examples include fluid dynamics, optical waves, particle physics, and climate modelling.
- Robotics: Dedicated workshops, such as the one on “Embodied World Models” (combining Generative Models, RL, Vision, Robotics), offer a preview of how AI is redefining the capabilities of multiple robot form factors.
AI News This Week
-
The Architects of AI Are TIME’s 2025 Person of the Year (Time)
Fei-Fei Li, who is Radical Ventures Scientific Partner and Co-founder/CEO of portfolio company World Labs, was named Time’s Person of the Year for 2025 as one of the “Architects of AI”. Fei-Fei is often referred to as AI’s “godmother” for her contributions in the field of AI. World Lab’s recently released Marble, a world model that allows users to create virtual environments from multi-modal prompts, reflecting Fei-Fei’s belief that spatial intelligence is key to unlocking “the frontier beyond language — the capability that links imagination, perception and action.” Earlier this year, Radical Ventures portfolio company Crusoe’s CEO and co-founder, Chase Lochmiller, was also named to the TIME100 AI list, featuring the company’s role in building the first phase of Stargate, OpenAI’s $500B data center complex.
-
The Next Version of the Web Will be Built for Machines, not Humans (The Economist)
AI agents, large-language models connected to external tools, are being used to interact with the web on behalf of humans for use cases such as booking a flight, cancelling subscriptions, and issuing refunds. Infrastructure like the Model Context Protocol (MCP), developed by Anthropic, and the A2A (agent-to-agent) protocol developed by Google, aim to provide scaffolding for agents to communicate with each other. This may imply a future where agents autonomously perform tasks and act unprompted when given general directions.
-
Customer Research Chatbot Outset Raises $30M (Axios)
Outset has raised $30 million in Series B funding, led by Radiacal Ventures along with Microsoft’s M12 and other previous backers. The company utilizes human-like AI agents to gather customer feedback. Their fast-growing platform reflects the growing capabilities of AI to service more complex tasks like qualitative research at scale for a fraction of the cost.
-
The Scramble to Launch Data Centers into Space is Heating Up (Verge)
As data center developers are running into the physical constraints of electricity and space on Earth, a few are looking to outer space for a solution. Orbital data center satellites would have near-unlimited access to solar power, and the frigid temperatures in space negate the need for large amounts of water for cooling. While launch costs are still prohibitively expensive, and the technical challenge of making these satellites radiation-proof remains, the future of data centers might be in orbit.
-
Research: An Open-ended Realistic Simulator for Autonomous Agents in Physical and Social Worlds (UCSD, UVA, UIUC, JHU, Purdue, PolyU, USC, UMich)
Researchers have built SimWorld, a platform for training AI agents in a realistic, open-ended world simulation with multi-modal inputs and social reasoning scenarios. In this environment, agents can pursue longer-term goals such as a career and starting a business. Simulations like these provide researchers with a sandbox to study these complex systems and observe emergent behaviours of intelligence.
Radical Reads is edited by Ebin Tomy (Analyst, Radical Ventures)
