The week in AI: Google Gemini gets a slew of upgrades & Nvidia's GTC 2025 conference

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence - we pass along the news, useful resources, tools and services; we highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

Google continues to steadily upgrade Gemini - all new features explained

-------------------------

Last week in our technical section (bottom of the newsletter), we said that Google had gone on a tear with some low-key releases and updates.

This week, they’re at the top of the newsletter for keeping that momentum going with a range of new improvements and features being added to Gemini. Before we get into it, it’s worth noting that Google has positioned Gemini as the best free-tier AI on the market right now. It’s quite a list for just one week:

• Deep Research in Gemini is now available to free users, and got an upgrade. It now utilizes a thinking model under the hood which does more than just summarizing web info - it tries to reason over them. The reasoning layer has definitely improved DR outputs, but the vast number of resources it searches over and scouts out are really what make it stand out. Google made a brief “tips for using” Deep Research page.

• There’s an upgraded version of Gemini 2.0 Flash Thinking (still experimental, but this model is great) on Gemini. It has better performance and a longer context window - 1 million tokens, which is by far the largest of any reasoning model currently on the market. They’ve also added to Gemini’s ability to use multiple apps like Search, YouTube and Notes in a single go. Google’s ability to get Gemini integrated into so many popular apps before competitors is probably going to end up being a huge advantage for the company moving forward.

• There’s a new mode called Personalization. It asks you to connect your Google search history with Gemini, and utilizes it when you’re asking Gemini questions. This functions a little bit more intricately and broadly than ChatGPT’s Memory feature if you’re a Gemini Advanced subscriber (for free users the mode only works on your Google Search history, not your Gemini chat history unfortunately), but isn’t dissimilar. With this enabled Gemini has a much wider lens on you/your work projects/etc. than just what’s happening in the current conversation.

• Speaking of features similar to ChatGPT, Google also added a Canvas feature to Gemini, similar to what OpenAI and Anthropic have. Canvas provides an interactive space to create, refine, and share writing and coding projects. You can preview your code/project directly from the interface, edit inputs and outputs directly, or highlight individual sections and have Gemini adjust those - rather than having to ask the model to “do it again, but ____”. Google’s a bit late to the party with this, but from our testing they nailed the UI and this is an underrated feature for LLMs more broadly that not many are utilizing. It’s available for free users.

• They’re also bringing the very awesome Audio Overview feature from NotebookLM to Gemini. NotebookLM went viral last year for Audio Overview, which creates realistic-sounding podcast-style audio summaries of documents, web pages, and other sources. It’s not perfect but if you’re, say, not in the mood to read an entire research paper, the AI hosts will do a pretty good job analyzing just about anything you throw at the feature. Audio Overview in Gemini accepts files and content in a range of formats. Uploading a document via the prompt bar will trigger the Audio Overview shortcut, and once a summary is generated, it can be downloaded or shared via the Gemini app on the web or mobile. Also available for free users.

• “Gems” are rolling out for free users, too. Gems let you customize Gemini to create your own personalized AI for any topic. There are premade Gems and you can quickly create your own custom Gems, like a translator, a meal planner or a math coach. Just go to the “Gems manager” on desktop, write instructions, give it a name and then chat with it whenever you want. You can also upload files when creating a custom Gem - so if you want an AI’s responses to be grounded in a particular body of work, rather than responding more generally, this is exactly how you do it.

Not a bad week at all for Gemini. None of these updates and features would matter much if the engine behind them was a poor model, but we’ve been singing Flash 2.0 Thinking’s praises for almost two months now since it debuted on Google’s AI Studio.

Nvidia’s GPU Technology Conference 2025 - ‘What’s next in AI starts here’

-------------------------

Nvidia CEO Jensen Huang kicked off Nvidia’s GTC 2025 conference with a two-hour-long keynote - calling the event “AI’s Super Bowl” and revealing a bunch of new updates on upcoming chip releases, robotics, autonomous vehicles, and more.

• Nvidia’s new GPU lineup includes Blackwell Ultra in late 2025, Vera Rubin in 2026, and Feynman in 2028, each promising major performance gains. Groq and Cerebras are getting a lot of hype for what they’re doing in inference, but Nvidia is trying to maintain its position in that race with their huge ecosystem and now claiming world record inference speed on DeepSeek R1 with Blackwell.

• Huang said that scaling is not slowing down, and the computation needed for AI is “easily 100x more than we thought we needed at this time last year.”

• He also revealed Isaac GR00T N1, the first open humanoid robot foundation model, alongside a comprehensive physical AI dataset for training robots. Along with everything else they’re doing, Nvidia is positioning themselves to be the ‘go-to’ resource for the robotics revolution.

• A robotics physics engine, Newton, also debuted in collaboration with Google DeepMind and Disney - demoed with ‘Blue’, a Star Wars-style robot (clearly modeled after BD-1 from the Jedi Fallen Order video games) on stage.

• The new DGX Spark and DGX Station will bring data center-grade AI computing to personal setups, with Huang calling it “the supercomputer for the age of AI.”

• Nvidia also announced a new partnership with automaker GM, with plans to build the company’s first fleet of self-driving cars.

The event is still going, and there is a whole bunch of other interesting stuff in the session catalogue - from quantum computing to weather prediction models to collaborations with Databricks on advancing data analytics with AI. The sheer scale of Nvidia’s web of AI-powered infrastructure influence for nearly every industry and use case is just insane at this point. AI is not slowing down any time soon.

MORE IN AI THIS WEEK

• A couple of shared articles from NY Times this week:

  • Powerful AI is coming. We’re not ready 

  • Doctors told him he was going to die. Then AI saved his life

• Apple reportedly planning executive shake-up to address Siri delays

• Y Combinator’s CEO reported that nearly a quarter of their current startups now have 95% of their code written by AI

• Nvidia and xAI are joining Microsoft, BlackRock, and MGX in the AI Infrastructure Partnership, aiming to raise $30B initially and potentially $100B for AI data centers

• Google joins OpenAI in pushing feds to codify AI training as fair use (Hollywood signed an open letter urging the Trump administration to reject)

• Humanoid robot maker Figure announces the launch of BotQ, a new “robots making robots” manufacturing facility capable of producing 12,000 humanoids a year

• There’s a good chance your kid uses AI to cheat

• How the US is losing ground to China in nuclear fusion, as AI power needs surge

• Inside the launch of FireSat, Google’s system to find wildfires earlier

Need a personal assistant? We do too, that’s why we use AI.

Ready to embrace a new era of task delegation?

HubSpot’s highly anticipated AI Task Delegation Playbook is your key to supercharging your productivity and saving precious time.

Learn how to integrate AI into your own processes, allowing you to optimize your time and resources, while maximizing your output with ease.

Get the free guide here.

TECHNICAL NEWS, DEVELOPMENT, RESEARCH & OPEN SOURCE

• Measuring AI’s ability to complete long tasks is starting to look like its own ‘Moore’s Law’

• Anthropic updates Claude Code with new extended thinking feature which enables users to control the depth of reasoning in tasks

• ReCamMaster: Camera-controlled generative rendering from a single video

• Mistral releases Mistral Small 3.1, the best multimodal model in its class (24B)

• Roblox releases their open source model “Cube” that can create 3D objects using AI

• Ai2’s OLMo 2 32B: first fully open model (datasets included) that beats GPT-3.5 and GPT-4o mini

• Stability AI unveils “Virtual Camera” tool to turn photos into 3D scenes

That’s all for this week! We’ll see you next Thursday.