AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Gemini 2.0 Flash GA, Flash Lite & Pro Strategy with Logan Kilpatrick (Google DeepMind)

· The Cognitive Revolution

Logan Kilpatrick joins The Cognitive Revolution to announce the general availability of Gemini 2.0 Flash at $0.10/$0.40 per million input/output tokens, alongside a preview of Flash Lite (a smaller, cheaper variant intended to preserve the 1.5 Flash price point) and an experimental release of Gemini 2.0 Pro. He frames the lineup as Flash Lite, Flash, Flash Reasoning, and Pro, noting Pro is positioned around coding excellence and a roughly Pro-tier pricing pattern, while flagging that an "Ultra"-class model is an open research question given how quickly smaller models are catching up and the compute trade-offs involved. On product direction, Kilpatrick explains that the Gemini app, AI Studio, and the API have been consolidated into a single DeepMind organization that owns research, modeling, and production end-to-end—accelerating the path from research to developer access. He highlights text-to-app builders (bolt.new, Cursor, lovable, V0) as the most exciting near-term frontier and claims Flash delivers ~40x cost reductions vs. incumbents for those workloads. He also discusses the Multimodal Live API for co-present, real-time interaction (currently session-limited to ~10 minutes), native multimodal output (image/audio in early access), Imagen 3, and the thesis that reasoning is the unlock that will make very long context truly useful. On the broader ecosystem, Kilpatrick calls for a unified benchmark/leaderboard platform and personal eval tooling (Kaggle is pushing on this), pushes back on the "commoditization" narrative, and says Google will start talking more publicly about reinforcement learning and reasoning work. Fine-tuning for 2.0 Flash is not yet available. Via his solo fund, he flags vision-language startups replacing domain-specific CV, agent-enabling reasoning, evals infrastructure, and "agent-to-website" middleware (attribution, negotiation) as the areas he'd most like to see pitched.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.