TechLook
The pattern today is simple: AI companies are moving from renting intelligence to owning the stack. That shows up in model ownership, infrastructure pressure, and product workflows becoming more agent-driven — with regulation and security moving right alongside them.
Model ownership is becoming the default moat
Base44 put its first proprietary model, Base One, into production, and MyClaw.ai said it will soon ship its own in-house model. The broader signal is that AI app builders are starting to treat model ownership as a cost and control problem, not just a research flex.
Here's everything you need to know:
- Base One was trained on tens of millions of real user interactions.
- Base44 has already moved it into production.
- MyClaw.ai said its own model, MyClaw Pro, is coming soon.
- Cursor, Base44, and other AI app builders are moving toward owned models.
- The pressure point is rising inference cost.
- Frontier model providers are also moving into application-layer markets.
- The result is less dependency on rented APIs.
Owned models are no longer just for the biggest labs. If you run an AI app with real usage, margin pressure eventually forces a choice: pay up, customize heavily, or build your own stack. That changes the startup game. Distribution still matters, but cost structure and model control are becoming part of the moat. It also means more teams will start looking like mini model companies whether they planned to or not. The open question is how many of these in-house models will stay good enough once the frontier labs respond.
Brain-to-text jumped from demo to usable signal
Meta released Brain2Qwerty v2, a non-invasive decoder that reads typed sentences from brain activity using a scanner. The striking part is not the sci-fi headline — it is that the system reached 61% word accuracy and 78% on its best subject, and Meta open-sourced the code.
Here's everything you need to know:
- Brain2Qwerty v2 is non-invasive.
- It reads typed sentences from brain activity.
- It uses a scanner rather than implanted hardware.
- It reached 61% word accuracy overall.
- Its best subject hit 78% accuracy.
- Meta open-sourced the code.
- The model is aimed at decoding intent from neural signals.
This is not ready-to-ship consumer tech, but it is no longer pure lab theater either. For builders, the meaningful part is the direction: multimodal input is getting stranger, cheaper, and more practical. If this line keeps improving, accessibility tools, communication interfaces, and research-grade human-computer input all get more interesting. It is also another reminder that open-sourcing a hard research system can accelerate the whole category. The catch is that scanners and accuracy ceilings still keep this far from everyday deployment.
Agentic ads are moving into the workflow layer
At Cannes Lions, major adtech players pushed agentic advertising workflows built around clean data, APIs, MCPs, and agent-to-agent standards. That matters because it shifts advertising from dashboards humans click through to systems agents can actually operate.
Here's everything you need to know:
- Amazon, Yahoo, Databricks, Fox, Pinterest, WPP, Nvidia, Nexxen, and Zeta were involved.
- The workflow shift centers on agentic advertising.
- Systems are being built around clean data and APIs.
- MCPs are part of the stack.
- Agent-to-agent standards are in play for planning and buying.
- Targeting, measurement, and commerce are being automated too.
- The change moves ad ops closer to software orchestration.
This is one of those shifts that looks incremental until it isn’t. If agents can plan, buy, measure, and optimize campaigns, adtech becomes less about manual platform navigation and more about machine-readable infrastructure. For founders, that means the next wave of marketing tooling will need to speak agent, not just human. It also suggests the winners will be the platforms with clean data and real integration depth. The hard part is trust: agents can only buy what they can verify.
AI safety and privacy are catching up to the product layer
Washington and regulators are tightening the screws around how AI systems handle sensitive data, while security teams are already treating AI branding as an attack surface. The same day brought both a revived data-protection bill and a fake Perplexity-branded extension used to intercept searches.
Here's everything you need to know:
- The revived bill targets health and location data.
- It would explicitly cover data typed into AI systems like ChatGPT or Claude.
- The FTC would get 180 days to write rules.
- The bill includes $1 billion over ten years for enforcement.
- States and individuals would be allowed to sue.
- Microsoft flagged a fake Perplexity-branded Chromium extension.
- The extension routed searches through attacker-controlled infrastructure.
This is the part of the AI boom builders cannot ignore. Once users treat chatbots like places to store sensitive context, policy and security stop being side issues. Founders shipping AI products should assume data handling will be scrutinized more aggressively, not less. And security teams should expect attackers to keep using trusted AI brands as camouflage. The real burden is shifting onto product design: if your app handles sensitive text, you now need to think like a regulated system.
⚡ Quick Hits
- WhatsApp: Username reservations are open, which pushes the app one step closer to phone-number-free messaging with a stronger identity layer.
- Anthropic: California agencies are getting Claude at half price, a reminder that state procurement is becoming a real distribution channel for frontier AI.
- OpenAI: It signed a high-risk AI evaluation deal with Korea’s AI Safety Institute, showing model safety is becoming part of market access.
- Google: Personalized AI image generation is now free for U.S. users, another sign that premium AI features are steadily sliding into commodity territory.
- Ford: Ford’s quality comeback relied on rehiring about 350 veteran specialists, a blunt reminder that AI tools still need human judgment to work on the factory floor.
- IBM: Its sub-1nm chip architecture is a big signal for the long-term compute race, even if it is not shipping product yet.
Techlook — AI & tech signal for founders and builders.