Jeff Dean, with Google since 1999, will now leave the company. Several top researchers also departing Google will join him at the newly formed AI startup Discovery Loop. An exodus, already apparent prior to this, has once again cast doubts on Gemini’s future prospects as a frontier AI model series.
Dean’s new venture Discovery Loop aims to automate research, parallellizing thousands of experiments at the same time. Dean is bringing along notable company as co-founders: senior fellow Sanjay Ghemawat, Google Brain founding member Quoc Le, and DeepMind senior research scientist Oriol Vinyals. In other words: some big names are abandoning ship.
Foundational work
Dean co-created foundational Google innovations such as MapReduce, Bigtable, and TensorFlow, cofounded Google Brain and helped set up Google Cloud’s core infrastructure including its TPU accelerators. In other words: he helped create the technologies that made Google the behemoth it is today. When the company proved unable to immediately answer ChatGPT’s rise with Bard, Dean was among those brought in to build what would end up being Gemini.
Gemini, at least for a time, swapped places with Claude and OpenAI’s models at the top of AI benchmarks. However, no leading LLM from Google actually reached general availability since 2.5 Pro did in mid-2025, and the last truly competitive model from the company, 3 Pro, launched back in November. Claude Opus, Fable, and GPT-5 variants have led the charge since then, with even Chinese open-weight models like GLM 5.2 from Z.ai and Kimi K3 from Moonshot AI far surpassing efforts like Gemini 3.6 Flash, a more modest recent release from Google.
Dropping off
Apart from the model releases, Google has also missed a trick when it comes to developer environments. Codex from OpenAI, Claude Code from Anthropic and various other offerings like Cursor and Devin have carved out paths for themselves as AI-first developer options, with Google’s Gemini CLI now discontinued and Antigravity facing fierce criticism.
Many have suggested the AI race has been Google’s to lose. Not only did it come up with Transformers, the core architecture behind LLMs, the company is also the sole contender with premier models, applications, infrastructure and even the chips to run them all. This, however, has proven not to be a slam dunk, and a temporary advantage. OpenAI and Anthropic are both working on hardware themselves, all while Google feels the need to experiment with Frozen v2, a chip featuring a model essentially converted into silicon, without possessing a clear frontier-class LLM. Hopes are Gemini 4 Pro will change the tide, whenever it does arrive. The long wait for Gemini 3 Pro last year suggests we should not hold our breaths. The departure of Dean and other DeepMind leadership hints at bigger trouble.
Also read: Claude Opus 5 approaches Fable for half the price