Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here.
An interview with Google DeepMind product VP Tulsee Doshi
Google announced a slew of new and updated AI products and features at its I/O developer conference this week, including personal AI agents, code generators, search tools, and a new “world model” for generating physically accurate video. Much of it runs on the company’s latest Gemini 3.5 models, developed inside Google DeepMind. We spoke with DeepMind’s product VP, Tulsee Doshi, about the thinking behind their development and application.
What does the tension between safety and product quality look like at DeepMind here in mid-2026?
You’re evaluating not just for traditional harms, but you’re now evaluating for things like sycophancy. You’re evaluating for things like agent safety and bringing that forward, and then you’re building the guard rails around the product experience to make sure that you have the right verifications in place. There’s always a trade-off between blank response rate—not responding to a user because you maybe don’t want to answer about a particular topic—[and] answering in a nuanced way, and then answering in a way that maybe goes too far. That’s always the spectrum that we’re trying to find the right balance on.
Personally, I feel assured by an agent that chooses not to answer a question. How would you describe the persona that Google’s models project?
It is an area that we are actively investing in . . . That persona is going to evolve as we get feedback from users, as we see what folks resonate with and don’t resonate with. Also, as we enter this more agentic era of Gemini acting with and for you, there’s a switch in persona that you also need to think through. What does the agentic persona look like and how do we help you clarify things? How do we make sure there’s the right guardrails of the actions that you take?
When people ask you how we should be thinking about this transformation in the enterprise as we bring these models to bear in business, do you have any thoughts about how quickly that’s happening?
The summer of 2026 is people figuring out how to wield these tools and how to give themselves that magic. Then we’re going to start seeing the real shift in enterprise happen because right now it’s still—this is true even for the calculator—when you start using something the first time it’s inefficient because you don’t quite know how to use it. You could probably do it faster yourself. You don’t know how to leverage these tools. Then as you start building that fluency, that’s where you start seeing the culture change.
I think there may also be a process of building trust in these tools. The last thing I want to do is stake my professional reputation on some AI thing and it doesn’t work out.
Even the other night, Demis [Hassabis, DeepMind founder] asked me for an update on all of our Flash 3.5 metrics. I asked Spark [Gemini’s personal agent] to go put together a deck—pull all the metrics from all these places, pull all the updates from all these places, put it together, get it to Demis. After I made the deck, I then went through and manually reviewed all the numbers just to make sure I wasn’t sending something incorrect. It was correct, for what it’s worth. It was great. But you do that a few times. Then you start building trust that the model can actually ground effectively.
There’s an $80 billion CapEx number for this year. How do you explain to people why you’re going to spend all that?
As someone who grew up using Google search, Google’s whole ethos has been to organize the world’s information and make it universally accessible and useful. Now in the agentic era you can add help[ing] users take action on that information in a way that is thoughtful and intentional. If we can really help bring users into this new era, bring my mom or my sister into this new era in a way that is safe and trustworthy, in a way that is grounded in the principles of what search was already doing, and also still provide a little whimsy and fun in the form of things like NotebookLM, that actually is real impact to the world. If we can deliver on that promise to billions of users, then that’s the real meat of the whole thing.
That’s interesting because you went right to consumers. I thought you were going to say it’s the enterprise.
That’s where Google as a lab is unique compared to the other labs. Yes, we will leverage Gemini, hopefully, to transform businesses across the world and that will be huge in terms of its ROI. But that one almost feels like the obvious ROI. There will be literal dollars that come back from the CapEx. The part that is not literal dollars but has huge magical value to the world—what is my mom thinking about and what matters to her? It’s what kind of access we can provide that didn’t exist before. What does that mean for your empowerment as an individual? What does that mean for small and medium businesses? The scale of what this can do both from the consumer angle to the enterprise angle is pretty vast.
Players like Anthropic and OpenAI talk about crawling the web and grabbing all this information to pretrain their models. Google has been doing this for decades. Does Google have an advantage in how well it crawls the web and develops its knowledge graph?
One of the things that has served search well for decades has been this focus on quality—on ranking [webpages] well, not just pulling all of the content that exists on the web, but being able to tell signal from noise and being able to actually bring that to users in a powerful way.
How do you then do that in the context of models? What we’ve learned especially with posttraining and reinforcement learning is it really does come down to the quality of the data and how well you understand, verify, what kind of rubrics you leverage on that data to make it clean and bring that back into the model. That’s a history of work that we’ve done that will lead to that outcome. It’s really taking a lot of the bread and butter of what we’ve used in the search context historically and leveraging it in new ways, but with that same ethos.
Anthropic’s coup: Andrej Karpathy joins the company to lead a new pretraining group
Andrej Karpathy, one of the most respected researchers in artificial intelligence, has joined Anthropic, the company confirmed this week. Karpathy was a founding member of Anthropic rival OpenAI. He started his new job Monday.
Karpathy, who has recently been creating widely lauded educational content on AI, said in an X post that he’s excited to “get back to R&D.” He’ll join Anthropic’s model pretraining team, which works on the formative stage in which large language models (LLMs) process vast amounts of data to learn how to reliably understand and generate text.
Karpathy will also form a new group focused on using AI itself to find more efficient ways of pretraining models, potentially through smaller, more curated datasets. Some see the move as a sign that Anthropic may be exploring alternatives to the dominant AI-lab strategy of improving models primarily through scale: more data, more compute, and larger systems. The work could eventually contribute to broader efforts around recursive self-improvement, in which AI systems help design and train more capable versions of themselves.
“I think the next few years at the frontier of LLMs will be especially formative,” Karpathy said in his announcement.
New research suggests small businesses are moving fast on AI
New data from Goldman Sachs and TD Bank paints a bullish picture of how small businesses are adopting and benefiting from AI. Both firms say small businesses are embracing the technology quickly, broadly, and relatively cheaply.
Goldman Sachs this week graduated the latest 300-company cohort of its 10,000 Small Businesses program and surveyed participants about their AI plans. The results, shared exclusively with Fast Company, show that 88% now pay for AI tools, though nearly two-thirds spend $100 or less per month on subscriptions. Goldman says the top use cases are marketing and content creation (81%), followed by data analysis (54%), and operations and logistics (47%). Adoption also appears relatively recent, with half of respondents saying they began using AI within the last year.
TD Bank’s recently released research suggests AI is helping small business owners expand rather than shrink their workforces. Fully 60% of respondents said adopting AI will increase their workforce size. Nearly seven in ten (69%) said they’re using AI to reduce expenses, up sharply from 39% last year, potentially freeing up resources for hiring and training.
The biggest reported benefits over the past year were improved customer service (53%), better fraud and cybersecurity protection (47%), and increased sales leads (42%). Taken together, the data suggests small businesses are viewing AI less as a labor replacement tool and more as a growth accelerant.
More AI coverage from Fast Company:
- Will AI cause mass political polarization? Maybe not
- LinkedIn declares war on AI slop
- The students booing AI aren’t Luddites
- Firefox wants to be the anti-Chrome browser for the AI era
Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.
