AI digest: Models get cheaper, robots get memory
Major model releases from Google and OpenAI, plus breakthrough memory systems for robots and new AI development tools.
Big week for model releases and tooling improvements. The race to cheaper inference continues while robotics takes a leap forward.
Google and OpenAI both push cheaper models
Google dropped Gemini 3.1 Flash-Lite with adjustable thinking levels for production workloads, while OpenAI released GPT-5.3 Instant that promises to be less annoying in conversations. Both are clearly targeting the high-volume, cost-sensitive deployment market. The Flash-Lite pricing tripled versus its predecessor though, which shows even “cheap” AI isn’t getting cheaper fast enough.
Robots get proper memory with MEM system
Physical Intelligence unveiled MEM, giving their Gemma-based robot models 15 minutes of context for complex tasks like cleaning kitchens. This tackles the biggest limitation in current robotic policies, which typically work with single observations or tiny histories. Finally, robots that can remember what they were doing five minutes ago.
SymTorch turns neural networks into equations
Cambridge researchers released SymTorch, a PyTorch library that converts trained models into readable mathematical equations using symbolic regression. This could be huge for interpretability, though we’ll see how well it scales beyond toy problems. The real test is whether it works on models that actually matter.
Alibaba launches OpenSandbox for AI agents
Alibaba’s OpenSandbox provides secure, isolated environments for AI agents to execute code and browse the web. Smart move to standardise the execution layer before everyone builds their own incompatible systems. Apache 2.0 licence should help adoption.