This video introduces "Deepthink," a new feature for Google Gemini, presented as a more practical variation of the model that won gold in the International Mathematical Olympiad (IMO). Deepthink leverages "parallel thinking" to generate and refine solutions simultaneously, showing impressive benchmark improvements over previous Gemini models. Currently exclusive to "ultra subscribers," the video demonstrates Deepthink's capabilities through two primary tests: generating a functional web-based operating system (WebOS) with interactive elements and creating a 3D low-poly racing game. The racing game notably features advanced lighting physics, where shadows and reflections dynamically adapt to the car's position, showcasing a sophisticated understanding of geometric and environmental interactions. A third attempt to generate a 3D first-person shooter was cut short due to Deepthink repeatedly omitting code and subsequently hitting a rate limit. Despite the limitation, the demonstrated performance, particularly in the racing game's intricate details, leaves the presenter highly impressed and optimistic about the future of AI models.
Summarizing Google Gemini Deepthink: A Deep Dive into its Enhanced Capabilities
0:00 - Introduction to Deepthink: Google Gemini introduces "Deepthink," a new feature designed for improved performance and creative problem-solving. It's a variation of the model that achieved gold in the International Mathematical Olympiad (IMO), but optimized for usability (bronze-level performance, faster than the original).
0:55 - Parallel Thinking Explained: Deepthink employs "parallel thinking techniques," allowing Gemini to generate multiple ideas simultaneously, consider them, and combine elements to produce more refined and effective solutions.
1:35 - Benchmarks and Sample Generation: Initial benchmarks for Deepthink appear significantly superior. A sample generation of a Voxel art scene demonstrates Deepthink's aesthetic improvement over Gemini 2.5 Flash and Pro.
2:30 - Availability and Cost: Deepthink is currently limited to "ultra subscribers." The exact cost is uncertain, but it was reportedly $250 with a past introductory offer around half that price.
3:10 - First Test: Web-Based OS: The first practical test involves asking Deepthink to generate a web-based operating system. Responses can take a few minutes, and a "show thinking" feature reveals a task-list-like process.
4:48 - WebOS Results: The generated WebOS is functional with a "pretty" Pacific Northwest aesthetic background, correct clock, unique icons, and elegantly resizable and draggable windows that retain memory.
6:30 - Unexpected Browser Default: The WebOS browser, surprisingly, defaults to Bing.com instead of Google, which the presenter finds unexpected and amusing.
8:10 - Second Test: 3D Racing Game (Initial Hiccup): The next test requests a 3D low-poly racing game with a first-person view, opponent, map, and simple gameplay logic. Deepthink's initial "thinking chain" incorrectly states it cannot create such interactive elements before eventually providing the code.
10:10 - Code Omission and Correction: The first attempt to generate the racing game's code results in an omission. A second prompt leads Deepthink to apologize and finally deliver the full script.
10:48 - Impressive Racing Game Result: The generated 3D racing game is surprisingly functional and impressive. Key features include a turning steering wheel, responsive controls (WD keys), map boundaries, and a realistic difference in forward/reverse speed.
11:25 - Advanced Lighting Physics: A standout feature of the racing game is its dynamic lighting physics: shadows and reflections on the car change accurately based on the car's position on the map, demonstrating an "intricate in-depth generation" and "world model understanding."
13:17 - Gameplay Features: The game successfully implements pause and restart functionalities.
13:50 - Third Test: 3D FPS and Rate Limit: A final test requests a 3D first-person shooter in Voxel style. This generation also omits the code, and after a re-prompt, the presenter hits a "limit for chats with Deepthink," preventing further testing.
15:52 - Overall Impressions and Future Outlook: Despite the abrupt end due to rate limiting, the presenter is "more impressed than anticipated." The capabilities shown, especially the dynamic lighting in the racing game, highlight Deepthink's advanced mathematical and generative functions. The presenter is highly optimistic about the future capabilities of such models, noting that current paid/research features often become widely available later.
17:30 - Model Access Clarification: The video concludes by clarifying that the "bronze medal" Deepthink model is the public release, while the full "gold medal winning model" from the IMO is reserved for invited academic researchers.