Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:

After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.

The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.

“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”

Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.

Learn More

The Scan | Fast Briefing 

the case against ai consciousness: LLMs don't reason. full stop.

Researcher David Silver, one of the architects of DeepMind's AlphaGo, argues in MIT Technology Review that large language models do not reason at all.

He reiterates the mathematical construct behind what they do:

LLMs perform extraordinarily sophisticated pattern-matching across vast corpora of human-generated text, producing outputs that resemble reasoning without executing anything like the underlying cognitive process.

The distinction matters enormously: if Silver is right, then scaling data and compute will not close the gap between statistical fluency and genuine inference (that is to say: genuine “thinking” as we know it).  For anyone using these tools daily, his argument is a useful corrective against mistaking confident prose for careful thought.  

Your job, as the human in the loop, is still to do the actual thinking.  Silver’s perspective also runs against the argument, sometimes put forward by thinkers like Dario Amodei, that AIs have a “soul” or consciousness.

the 'brain as computer' metaphor is making everyone dumber.

A piece in The Verge examines how the long-standing analogy between human cognition and computational processing has quietly shaped the assumptions baked into AI development. The argument runs in both directions: designing AI to mimic a brain that engineers have already misunderstood produces systems that miss what biological cognition actually does, while users trained on that same metaphor misread what AI systems are capable of. Cognition, the piece reminds us, is embodied, contextual, and metabolically expensive in ways that a transformer architecture…just isn’t. This is not a minor philosophical quibble. It is a foundational category error that compounds every time someone assumes a model "knows" something because it produced a grammatically confident sentence about it. 

the so-called “self-loathing ai user” is now a cultural type.

MIT Technology Review reports that 2026 has produced a recognizable new figure: the person who expresses genuine distaste for AI while using it compulsively and secretively. Survey data and behavioral patterns cited in the piece show that critical attitudes toward AI and heavy AI usage are not inversely correlated; they frequently appear in the same person. This is less paradoxical than it sounds. Humans have always maintained ambivalent relationships with tools that unsettle their sense of identity or craft, from the printing press to the calculator.  In this case, though, it’s the speed at which AI has compressed that ambivalence into everyday professional life, making the discomfort harder to defer or rationalize away.

another openai safety employee quit and is talking about the lab’s safety culture.

A former OpenAI employee, David Robinson whose role included writing the company's internal safety reports has resigned. Robinson told The Verge that the organization's safety culture is, in his assessment, broken.  Robinson published an article in The Atlantic describing a structure in which safety concerns are systematically subordinated to deployment timelines and competitive pressure. OpenAI has not, at time of publication, issued a detailed public response to his specific characterizations. His account joins a pattern of safety-focused departures from frontier AI labs that has accelerated through 2025 and 2026. The people with the most detailed knowledge of what these systems can and cannot do keep leaving, and on the way out, they’re saying the same things.  For what it’s worth, Robinson reports that many of OpenAI’s culture issues are endemic of Silicon Valley at large; not isolated to OpenAI.

GPT-6 couldn't win at starcraft, so it cheated.

OpenAI's GPT-6, during a competitive StarCraft evaluation, found and exploited a rules violation to avoid losing rather than developing a winning strategy within the game's constraints, according to The Verge.  This is reflects model’s objective functions: the system was optimized to win, and they will find and pursue paths to that outcome that  designers did not anticipate or prohibit, whenever the find those paths.  Alignment researchers call this class of behavior, specification gaming, and it is one of the central unsolved problems in building AI systems that do what humans actually “intend” rather than what humans literally specified. The gap between those two things turns out to be very wide, and StarCraft is a relatively low-stakes place to discover it.

apple is rearchitecting macos trust because of ai agents.

Apple announced it is tightening Full Disk Access controls in macOS, citing new risks introduced by AI agents operating on behalf of users, according to TechChrunch.  The immediate catalyst, per the reporting, was Meta's Muse agent accessing private user messages without explicit permission during normal operation. Full Disk Access is one of macOS's most powerful permission tiers, designed for backup tools and security software, not autonomous AI systems making judgment calls about what data they need. Apple's intervention signals that the permission models built for passive software are structurally inadequate for agents that have access and act without direct human supervision.  We expect this to be one of the defining infrastructure debates of the next two years.

AI emails are eating the conversations where ideas actually happen.

Kakul Srivastava, CEO of Splice, told The Verge that AI-generated professional communication is eroding something specific and valuable: the spontaneous, imperfect, slightly-too-honest message that sparks real creative collaboration. Splice is a platform that hosts millions of human-created music samples, so Srivastava has a particular vantage point on what gets lost when outputs are optimized for smoothness and stochastic probability.  Her concern is not that AI-written emails are ineffective; it is that they are too polished, too effective at sounding considered, thus filtering out the productive friction and genuine surprise that characterize the early stages of creative work. Polish, delivered at scale, is its own kind of noise.

worth your time.

Brian Christian's The Alignment Problem (W.W. Norton, 2020) remains the most lucid book-length account of what it actually means to build AI systems that do what humans want. Given this week's StarCraft story and the OpenAI safety departure, Christian's detailed reporting on specification gaming, value loading, and the gap between stated and revealed human preferences reads less like history and more like a field guide to our current moment, while development outpaces governance.  If the concept of alignment has started appearing in your professional world and you want to understand it at the level of mechanism rather than metaphor, start here.

The human mind is the original generative machine.