What's next for AI: Researchers at Nvidia, Apple, Google and Stanford outline the next leap
Before ChatGPT grabbed headlines, AI progress lived in labs and conference halls. At BayLearn-a gathering of engineers and scientists across Silicon Valley-the conversation focused on what comes after today's agents: efficient systems, interactive learning, and practical robotics.
The signal: specialized compute, open toolchains, and models that learn through interaction, not just scale.
Nvidia: Nemotron and the future of accelerated computing
Nvidia's view is clear: accelerated computing is bigger than a single chip. It's a stack-models, datasets, precision algorithms, pre/post-training tools, and software to scale on GPU clusters. Nemotron bundles these into an open set of technologies to speed up every stage of AI development.
"Nemotron is a really fundamental part of how Nvidia thinks about accelerated computing going forward," said Bryan Catanzaro, vice president of applied deep learning research at Nvidia. The goal is specialization that lets you do what a standard computer can't.
Open-source is the force multiplier. Teams from Meta, Alibaba, and DeepSeek have contributed, and the Nemotron datasets are seeing wide use. The throughline: shareable components that compress time-to-results.
Context matters here. Years ago, Catanzaro saw how Nvidia's CUDA architecture could deliver speedups that general-purpose systems couldn't. That intuition helped steer GPUs toward deep learning-and changed the industry's compute baseline.
- Explore CUDA for custom kernels and throughput gains: CUDA Zone
Stanford: From brute force to systematic generalization
Two decades ago, large language models weren't even a topic at major NLP venues. Today, they anchor most AI pipelines-yet their limits are obvious. "LLMs don't work interactively at all," said Stanford's Christopher Manning. "Human beings can learn with orders of magnitude less data than our current models."
Manning's prescription: systematic generalization. Models should combine known elements to form new, correct meanings, and improve through interaction-"poking around websites," exploring, and learning from feedback instead of brute-forcing with more data.
- Prioritize environments where agents can act, explore, and verify outcomes (web tasks, tool use, retrieval with feedback loops).
- Design curricula and evaluation suites that test compositionality and transfer instead of memorization.
- Track sample efficiency, not just benchmark highs; reward structure and sparsity matter.
Apple: MLX and the push for efficient, reliable deployment
Apple's MLX is an open-source framework built for Apple silicon that converts high-level Python into optimized machine code. The intent: treat software and hardware as one system for reliable deployment. Reports suggest CUDA back-end support is in the works to lower framework build costs and broaden compatibility.
"We have to think from a systems standpoint how to get AI reliably deployed," noted Apple research scientist Ronan Collobert. For researchers, this means tighter loops between model design, low-level kernels, and distributed execution.
- Kick the tires on MLX if you run Mac-based research workflows: Apple MLX on GitHub
Google DeepMind: General robotics that's "good enough"
DeepMind's latest-Gemini Robotics 1.5 and E.R. 1.5-bakes reasoning into robotic control. The move shifts robots from narrow skills (folding paper) to task sequences and context-aware choices (e.g., selecting clothes based on predicted weather).
Ed Chi, VP of Research at Google DeepMind, put it plainly: "I'm sick of all this talk about AGI when I don't have a robot that can clean my house." The focus is general robotics: systems that follow natural language prompts and complete multi-step tasks reliably. "Good enough" beats grand theories if it actually works in the lab and the field.
What this means for scientists and R&D teams
- Use specialized stacks: focus on data pipelines, precision choices, and distributed schedulers that match your hardware.
- Adopt open models and datasets that plug into your training and eval loops; contribute back to accelerate baselines.
- Build interactive agents: web-based environments, tool APIs, and verifiable objectives for real learning signals.
- Measure compositional generalization explicitly; build probes and targeted test suites.
- For robotics, prioritize perception-reasoning-control integration and task graphs over one-off demos.
- Control costs: mixed precision, activation checkpointing, and smarter sampling beat "more GPUs" by default.
- Operationalize reliability: failure modes, safety constraints, and deployment tests as first-class citizens.
Key open problems to tackle next
- Sample efficiency: learning from fewer interactions and smaller datasets without collapse.
- Systematic generalization: standardized benchmarks that stress compositional reasoning across modalities.
- Tool use at scale: robust web agents, retrieval, and planning with verifiable outcomes.
- Sim-to-real for robots: transfer methods that hold under messy, changing conditions.
- Efficient distributed training: topology-aware schedulers and memory systems tuned to your interconnects.
- Safety and governance: controllable behaviors and auditability without throttling progress.
The takeaway
The next wave won't be defined by one model. It will be built by systems thinking: specialized compute, open tooling, and agents that learn by doing. As Manning put it, progress is continuous-so set your lab up to benefit from every incremental gain.
Your membership also unlocks:
