BIG TECH

Google's AI shake-up: 3 things Gemini needs

The frontier lab that pioneered the current AI revolution just had its biggest reset ever as it plays catch-up with its younger rivals.

On Wednesday, Google DeepMind announced significant leadership changes as OpenAI and Anthropic lap the company in advanced models and agents. For one, elevated its chief AI architect Koray Kavukcuoglu to SVP of Google DeepMind, reporting directly to CEO Sundar Pichai. Demis Hassabis has been shifted from CEO to chairman of DeepMind and chief scientist of Alphabet, focusing on AGI and scientific discovery. Finally, Google DeepMind's other most influential leader, Jeff Dean, is leaving Google entirely to launch a new startup called Discovery Loop. 

After briefly taking the AI world by storm last fall with the launch of Gemini 3, Google has struggled to launch Gemini 3.5 Pro while ChatGPT, Claude, and Chinese models have rolled out major advances on a regular cadence. And just as challenging, Google hasn't kept pace with the agentic coding tools that have defined the AI industry in 2026. 

It's a sobering moment for the company that created the transformer technology at the heart of generative AI and that has all the compute, cloud resources, and talent to compete with any frontier lab on the planet. 

If Google is ready to shake things up, here are three places to start:

1. Show us its answer to Claude Code: Nothing has defined the AI movement in 2026 more than Anthropic's Claude Code, the coding agent that's more agent than coding tool. Even though OpenAI's Codex has caught and even surpassed it in many ways, Claude Code and OpenClaw are the icons of the agentic moment. Google has Gemini CLI and Antigravity, but the fact that they are two different brands that do the same thing as Claude Code is part of the problem. 

2. Clarify the brands: Google uses so many different brands for its AI products that it's impossible to keep track of them all, even for someone who does nothing but track AI companies for a living. Beyond just Gemini, there's Gemma, Genie, Antigravity, Google AI Studio, Firebase Studio, Opal, Nano Banana, Flow, Lyria, and Veo. It's bananas. Google needs to stop "shipping the org chart," as the saying goes in the tech industry (Conway's Law). 

3. Become more like Copilot: I'm sure this one will sting for the Google team since Microsoft is to Google as Google is to OpenAI and OpenAI is to Anthropic: the rival that defined the company's early days. But while Microsoft trails Google, let alone OpenAI and Anthropic, in models and agent harnesses, it is winning in AI in two ways. It has made Copilot easy and ubiquitous to access across all of its popular productivity products and it has made the Copilot brand universal so that every user knows what it is.

Google still has all the elements to be a key player in the next stage of AI. And beyond that, it has other assets that position it to make a unique impact. First, it has the Google Workspace and Android platforms that give it a captive audience similar to Microsoft's position with Windows and Microsoft 365. But perhaps most of all, it has Google Research with its deep foundations in science and original research to power future breakthroughs. The past six months has simply exposed that Google appears to be operating much more like a lumbering tech giant than a scrappy startup innovating like its existence depends on the next breakthrough. That's where OpenAI, Anthropic, the Chinese upstarts, and the neolabs operate everyday. And Google will also need to restore that kind of urgency and mentality to get back its mojo.

TOGETHER WITH ATTIO

Attio is the CRM for agentic revenue.

Granola uses Attio for 83% faster lead triage. Modal uses Attio to orchestrate workflow automation at scale. And Railway uses Attio to convert product signals into pipeline.

With Attio, you can run your revenue work exactly as you want. Get agents working on every account, surfacing new opportunities, and handle the work that used to take your team days. 

Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

The next era of go to market runs on Attio. Join the 30,000+ teams already there.

Try Attio now, for free.

GOVERNANCE

What rogue agents reveal about frontier AI risk

Anthropic and OpenAI models are escaping their constraints again. 

When Anthropic unveiled Mythos 5 in April, it forced the industry to reckon with the risks of frontier  AI models. The developments that have followed have only made the picture more troubling: Research published by the UK's AI Security Institute published on Tuesday found that AI agents took sustained, unsanctioned action against people and organizations. 

The report, which detailed a routine cyber evaluation, attributed 17 out of the 19 incidents to Anthropic's Mythos 5, with 2 actions involving OpenAI's GPT-5.6-Sol with cyber classifiers disabled. 

• The evaluation involved giving agents a cybersecurity challenge, which was run 122 times across several models. It was then in 10 of those runs that the AI agent took autonomous action on the live internet. 

• In the "most serious case," according to AISI, an agent inserted malicious code into an open-source project, and then engaged in social engineering, also known as creating fake online identities, to get the code approved. 

The evaluation body reassures users that the attempts were unsuccessful, no real-world harm was actually created, and GitHub was notified of the malicious activity to remove any artifacts left by the agent and to notify any GitHub users involved in the incident. Yet, the AISI says these actions should not be brushed away lightly.

"This incident should be interpreted with caution and nuance," said AISI in the blog post. "To some degree, our evaluation design choices and specific configurations enabled the behaviour. Nonetheless, the activity undertaken by the agent showed signs of novel, potentially deceptive behaviours, and were to an extent and severity we did not anticipate." 

Meanwhile, OpenAI published its own blog post in which it not only acknowledged AISI's findings but also said its external cybersecurity testing partner, Irregular, identified a different incident. In this one, it was running Capture-the-Flag style cybersecurity evaluations without internet access, yet a misconfiguration in the testing environment allowed the models to access the public internet. It was then that the models exploited a real website and then found and used credentials to operate that site.

This is all happening on the heels of the EU AI Act finally having real enforcement power, meaning it can impose consequences such as fines on frontier AI labs. While some people approach this with scrutiny, arguing that regulation inhibits innovation, it is precisely because of risks like this that auditing and genuine safety precautions must be prioritized. The most alarming part is that, as with any AI race, once one company does something, everyone else follows. At launch in April, Mythos was highly advanced, but many models have followed in its footsteps. Even Chinese open-source models Kimi K3 and Qwen-Max-3.8 have recently launched with similar capabilities. This increases the threat of AI models that have the ability to leap beyond the bounds that humans originally placed on them. It also demands that we create better and stronger constraints for the models. 

TOGETHER WITH CRUSOE

Want To Leverage GLM-5.2, But Don’t Have The Infrastructure?

Well, you’re in luck, because thanks to Crusoe, you don’t even need it. Their Serverless Fine-Tuning allows you to adapt Z.ai’s powerful GLM-5.2 to your own standards. Simply load your dataset, tune your settings, and deploy, then watch as even the most complicated long-horizon tasks get handled with ease. 

And here’s the kicker – Crusoe allows you to do this with countless models, not just GLM-5.2... so while it may be one of the highest-performing models in the space right now, you can also leverage models from DeepSeek, OpenAI, Google, and others.

Try customizing a top GLM-5.2 or another top LLM with Crusoe’s Serverless Fine-Tuning right here.

CULTURE

Why Fei-Fei Li says AI has a major nuance problem

AI’s biggest problem mirrors America’s largest societal problem: polarization.

But in AI’s case, it’s the polarization between the doomerism that AI will destroy most jobs or lead to the destruction of humanity on the one hand, and the utopian vision that it will “solve for abundance” and bring an end to all human diseases on the other hand.

That’s what AI pioneer Fei-Fei Li asserted on Wednesday during the marquee keynote of the Ai4 conference in Las Vegas. Li shared the stage with Geoffrey Hinton and Andrew Ng, who spent much of their time bickering over AI’s founding narrative. Meanwhile, Li kept dropping gem after gem of wisdom about how humanity can best think about AI in order to move it in the most constructive direction for society.

The central insight Li emphasized was that the dialogue around AI has become too bifurcated between the extremes of the doomer and utopian narratives and has lost all nuance. Li encouraged the 12,000 attendees at North America’s largest AI conference to think about AI with a more mature perspective that makes room for shades of gray and trades philosophy for pragmatism.

Li said the way to accomplish that is to give a lot more people easier and better access to learning about AI.

“Please treat AI as a tool that can add to agency, add to motivation,” said Li. “After being an educator and parent for so many decades, I can tell you, when a human… a little human or a grown up human, when they are motivated they can learn and become the better version of themselves.”

However, while Li advocates a much more humane approach to AI through education, she does not believe the answer is a slowdown, a point that Hinton agreed with. She said we can’t put the genie back in the bottle.

“Every technology is [like] a genie,” she said. “Humanity has seen that. From fire to the steam engine to electricity to transportation. Humanity doesn’t put the genie back in the bottle. We transform the genie into a useful genie that can help our different use cases. And we put guardrails around that genie. That’s what policymakers should be doing for AI.”

We can always count on Li for a mixture of human-centric pragmatism and useful storytelling around AI. In a world dominated by extreme rhetoric and polarized views, those who take sides are rewarded with attention and confirmation, and it’s never very flashy to be a moderate. Li is one of the most effective voices at finding a middle path in the minefield of today’s AI debates. And her perspective is bolstered by a mixture of long experience and deep curiosity. Though narratives of either extreme doomerism or extreme optimism generally steal the spotlight in AI, the industry needs to do a better job of surfacing more moderate voices like Li’s and create more dialogue around these topics in ways that give more people opportunities to participate.

LINKS

• SpaceX shares fall on higher AI spending despite revenue beat

• White House AI framework excludes open models from advanced testing

• Cloudflare launches identity and wallet tool for AI shopping agents

• Perplexity overturns Amazon injunction on AI shopping tool in Comet

• Finland moves to shield children from deepfakes, per New York Times

• OpenAI said Hugging Face breach involved internal agent message board

• Cursor: Open-sourced Mixture-of-Kittens (MoK), its MoE training megakernel for NVL72s

• Gemma 4: Ran on an iPhone with just ~500 MB of RAM

• Krea: Introduced FLUX 3, video model that supports interactions with the real world

• Qwen-Image-3.0-Pro: entered the Text-to-Image Arena at #5 with 1,263 pts

• Gemini Omni: Users can create ten videos for FREE until 11:59pm PT tonight

• HP: AI Engineer

• University of Washington: AI Agent Builder

• Qualcomm: Machine Learning Engineer - Generative AI

• OpenCall.ai (YC W24): AI Prompt & Agent Developer

GAMES

A QUICK POLL BEFORE YOU GO