Google's three Flash models, Meta's lab work, OpenAI's alignment problems

Google shipped three new Gemini models with distinct jobs. Meta's models are being used in national lab research, and OpenAI has described alignment problems it ran into while developing its systems.

Google splits Gemini Flash by job

Google's release is three models aimed at different workloads. Google has launched Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber. Of the three, 3.6 Flash gets the clearest positioning: Google pitches it at coding and multimodal applications. The version numbers suggest Flash-Lite and Flash Cyber stay on the 3.5 generation while the main Flash model moves ahead. Google frames the set as addressing efficiency and security. For teams already on Flash, 3.6 is the one to benchmark against coding and multimodal workloads before switching.

Meta's models in Genesis Mission science

Meta is applying its models to scientific data rather than consumer products. It says its AI models are powering early Genesis Mission projects with Lawrence Berkeley National Laboratory, where they segment and analyze environmental data. The work points to off-the-shelf vision models being useful for scientific datasets without building a domain-specific system first, though Meta has described these as a first wave of projects rather than finished results.

OpenAI describes alignment failures in its own work

OpenAI has shared a set of alignment problems encountered while developing its models, covering cases where system behavior diverged from what developers intended. The linked analysis comes from an outside commentator rather than OpenAI directly, so it reflects one reader's take on the disclosure. For engineers deploying these models, the main value is a concrete list of failure modes reported by the vendor itself.