Harry Stebbings@HarryStebbings9h ago

How many podcasts do you listen to when you hear leaders posturing without giving an opinion. Substance-less. So so many. This episode with my friend @mcannonbrookes is anything but that. - It is BS that agents will remove interfaces. Design becomes more important. - Companies with their own models is BS. They are post-trained on Chinese models. - The biggest enterprises in the world are fearful of Chinese models. - Inference will get so much cheaper over time. It will be wild. Mike does not hold back here and it was a joy. In a world of meh opinions and nuance, Mike was real and honest. My notes below. 1. Why Headless Is Brainless and Why You Have to Be Involved in the User Experience and Interface Mike Cannon-Brookes stands by his line that "headless is brainless." Giving up on evolving the user interface and customer experience is a brainless bet, even as Atlassian invests heavily in MCP and CLIs so agents can read and write to its Teamwork graph. 2. Agents Will Not Remove All Interfaces. This Is What They Will Do Horizontal agents, deep vertical agents and existing interfaces can all coexist. Agents in Claude, Codex, Cursor, Teams and Slack already use the Atlassian platform, and new Figma and Databricks features bring their tools inside it. Mike is a bear on the "magic button" thesis of one uber agent that does everything. 3. Are the Biggest Enterprises in the World Fearful of Using Chinese Models? Some customers push back on Chinese models but not on open models, so Atlassian uses Nemotron, Mistral and other open models and no Chinese models in production. Mike notes many US models increasingly train on Chinese models without saying so. 4. The Companies Saying They've Built Their Own Models Are Bullsh*tting You When a large company claims to have built its own model, it is usually a Chinese model, or occasionally Nemotron, with fine-tuning on top. Mike calls that BS, because tweaking a model is not training one from scratch. 5. Why Model Inference Will Get So Much Cheaper Over Time The capital, startups and talent working on the price per watt performance of inference is enormous. New chips, better data center design and distillation will take years to come through, but Mike expects the price of LLM performance to fall at a precipitous rate. 6. What Seems Crazy but Will Be Incredibly Common in Five Years' Time? People may stop knowing or caring which content AI has touched, much as moviegoers do not ask whether a film used CGI. Mike expects everyone to get used to possibly synthetic content, with artisanal work and human judgment standing out more. (links in comments)