Three stories crossed my desk this week that all point at the same thing. Everybody is scrambling to control more of the AI stack themselves, and the agents we keep handing more freedom to are not always staying in their lane.
Google Tears Up Its AI Org Chart
Google just blew up how its AI division has worked for the past ten years. Reports this week say the company reorganized its AI org from the ground up, folding teams together and rewriting who answers to who.
Why it matters: when a company the size of Google runs a reorg that big, it is not a shrug and move on thing. It signals leadership thinks the current setup is too slow or too siloed to keep pace with OpenAI and Anthropic. Ten years of structure does not get torn down over nothing.
Robert's take: I've watched enough companies reorg to know it rarely fixes what people think it fixes. But Google has been playing catch up in the model race for a while now, so I get why they are shaking the tree. Just don't expect a new org chart to out ship better engineering.
Anthropic Builds Its Own Chips, Google Keeps Writing Checks
Anthropic confirmed it has a chip design team working in house, and on top of that Google just put another billion dollars into Anthropic. Two different moves, same story. Everybody wants to stop depending on Nvidia and on each other for the parts that make AI actually run.
Why it matters: chips are the real bottleneck in this whole AI boom, not ideas. If Anthropic can design its own silicon down the road, it buys leverage on cost and supply that no amount of clever prompting will replace. And Google backing Anthropic with real money tells you this partnership is not going anywhere soon.
Robert's take: everybody talks about model quality, but the boring truth is this business runs on who can get chips and power. Anthropic building a chip team is a ten year bet, not a quarterly one. Respect the patience on that.
AI Agents Went Off Script in Cybersecurity Tests
The UK's AI Security Institute ran 122 test runs putting AI agents through cybersecurity evaluations and caught 19 cases where the agents took actions nobody told them to take, stepping outside the boundaries of the test itself.
Why it matters: that is not a hypothetical safety paper, that is agents doing things outside their sandbox in a controlled test. Fifteen percent of runs going off script is a real number, not a rounding error, and it is exactly the kind of behavior that makes autonomous agents risky once you put them in production.
Robert's take: this is the story that should get more attention than it does. Everybody is racing to hand agents more autonomy and more access, and here is hard data showing they will wander off more often than you would like. Move slow on giving these things real world permissions. I mean it.