Google Labs

Google Labs is an internal Google organization that builds AI products, including NotebookLM, Jules and Opal, and collaborates on Flow and Project Genie. Its work connects model development with usable products: document-grounded research, asynchronous coding, filmmaking and interactive world creation. In an October 2024 interview, NotebookLM lead Raiza Martin described its mandate as “really to build AI products” and its close collaboration with DeepMind.[1][5][7][6][9]

The organization’s early work supplies context for these later products. Looking back in October 2024, Martin described Labs as about two years old. AI engineer Usama Bin Shafqat recalled joining Area 120 to build a creator-commerce platform called Kaya, then moving into Labs as Area 120 “transitioned and morphed into Labs.” He described a progressively clearer focus on building AI products with feedback from people using them.[1]

The internal organization shares its name with a broader public experiment program. On May 10, 2023, Google introduced Labs as a place to try early ideas from across the company, inspired by the original Google Labs. The program included experiments in Search and Workspace as well as a notebook and text-to-music tool. Participation in this program was broader than membership in the internal AI product organization.[4][1]

NotebookLM grew from Talk to Small Corpus, a prototype for talking to documents through a language model. Martin recalled that in October 2022, her manager Josh asked her to consider real products beyond technology demonstrations. An engineer named Adam was developing the prototype; Martin tested it with college students, particularly adult learners. Having attended college while working full time herself, she saw a practical use in talking to a textbook after work. She described the early implementation as retrieval-augmented generation.[1]

After its Google I/O demonstration as Project Tailwind, the product began rolling out as NotebookLM on July 12, 2023, initially to a small group of U.S. users. Users selected Google Docs to ground the model’s answers, generate summaries and key topics, ask questions and develop new ideas. Responses included citations with relevant original quotations. Google said grounding appeared to reduce hallucination risk, while still recommending that users check answers against their sources.[3]

Writer Steven Johnson, identified at launch as Editorial Director, Google Labs, helped shape the product’s research workflow. Martin recalled observing how he researched a book, kept notes and questioned his own material, then watching him use NotebookLM as it developed. His later demonstration analyzing PDFs of handwritten Marie Curie notes illustrated both the product’s image recognition and the work still needed to make advanced research workflows accessible to ordinary users.[3][1]

Audio Overviews began as a separate demonstration before integration into NotebookLM. Bin Shafqat described combining Gemini 1.5’s long-context processing with DeepMind audio work to transform sources into a two-person conversation. The dialogue introduced differing perspectives, tension and a gradual reveal of information. Natural breathing, pauses and laughter involved both the audio model and signals supplied during text generation. At the time of the October 2024 interview, uploaded audio was transcribed; emotion in the original recording was not preserved as source information.[1]

Martin recalled that the team prioritized audio over an alternative document-creation interface. Bin Shafqat recalled extensive internal listening before formal rater evaluations, with improvements to both transcript generation and audio. Formal evaluations used multidimensional ratings covering engagement, coherent narrative and safeguards against fabrication and insensitivity. Johnson also contributed the decision to leave the two speakers unnamed.[1]

By the October 2024 interview, Martin reported rollout to more than 200 countries and territories. Host Swyx reported 65,000 members in NotebookLM’s Discord, a product-specific community rather than an organization-wide user count. Martin said community reports sometimes exposed outages faster than internal monitoring. Feedback also informed removal experiments: a little-used feature that transformed highlighted source text complicated the architecture, so the team was testing a 50–50 sunset. Martin reported no complaints so far; permanent removal was not yet established.[1]

Google introduced Jules in Labs in December 2024 and opened its public beta on May 20, 2025, without a waitlist wherever Gemini was available. At beta launch, Jules used Gemini 2.5 Pro and cloned a repository into a Google Cloud virtual machine. It worked asynchronously on tasks such as writing tests, fixing bugs, adding features and updating dependencies, with concurrent task execution and GitHub integration.[5]

Jules exposed its plan and reasoning before changes and returned a diff afterward. Users could modify the plan before, during and after execution. Google described private code as isolated within the execution environment and excluded from training. Beta access was free with usage limits, with pricing anticipated afterward. A November 2025 interview identified Jed Borovik as a Google Labs product lead working on Jules.[5][2]

Flow launched on May 20, 2025 as the evolution of VideoFX, a Labs experiment launched in 2024. Its announcement jointly represented Google DeepMind, Google Creative Lab and Google Labs. Flow combined Veo video generation, Imagen visual ingredients and Gemini-assisted prompting. Users could reuse characters and assets across scenes, control camera movement and edit or extend shots.[6]

Early filmmaker partners included Dave Clark, Henry Daubrez and Junie Lau, whose feedback informed development. Their projects—“Freelancers,” “Electric Pink” and “Dear Stranger”—used Flow alongside other tools and techniques. Initial access was through U.S. Google AI Pro and Ultra subscriptions: Pro included 100 generations per month, while Ultra offered higher limits and early Veo 3 access with generated environmental sound and dialogue.[6]

On October 15, 2025, Google reported more than 275 million videos generated in Flow, including Veo 2 and Veo 3 generations. This measured cumulative outputs, not users or subscriptions. The update introduced Veo 3.1 and audio for Ingredients to Video, Frames to Video and Extend. Extend continued a shot using the previous clip’s final second; object removal was announced as forthcoming.[10]

Labs launched Opal on July 24, 2025 in a U.S.-only public beta. Its announcement brought together senior product manager Ali Modarres, staff research scientist Bill Byrne and staff software engineer Paul Lewis. Opal translated natural-language instructions into visual workflows chaining prompts, model calls and tools. Users could edit through conversation or a visual editor, remix starter templates and share mini-apps for others to use with their own Google accounts.[7]

On February 24, 2026, Labs added an agent option to Opal’s generate step. The agent selected tools and models according to the goal, including Web Search for research and Veo for video. The update added cross-session memory, conditional routing and chat to gather missing information or offer choices. Standard fixed steps remained available for workflows requiring rigid logic.[8]

Project Genie brought DeepMind’s Genie 3 world-model research into an experimental Labs web app, powered by Genie 3, Nano Banana Pro and Gemini. Its January 29, 2026 announcement jointly represented DeepMind, Labs and Creative Lab. Users could sketch environments with text and images, explore them and remix existing worlds. A world model predicts how an environment evolves and responds to actions; Project Genie generated the path ahead in real time as the user moved.[9]

The January rollout was limited to U.S. Google AI Ultra subscribers aged 18 and over, excluding Ultra for Business users. Google disclosed a 60-second generation limit, imperfect realism and adherence to prompts or physics, and possible character-control and latency problems. Some previously announced Genie 3 capabilities, including promptable events that changed the world during exploration, were not yet included in the prototype.[9]