Summary

Consumer AI’s strongest demand signal is mundane, high-frequency utility rather than one killer app. Kevin Roose uses AI dozens of times daily, with roughly 60% of recent queries involving household problems; an OpenAI study of about 1.5 million conversations ranked practical guidance first, information-seeking second, and writing third. Programmers tell him, “I don’t even really code anymore”—they supervise AI coders and review the output.
Model leadership is task-specific and volatile enough that today’s best tool might lose its slot next week. Roose currently uses Claude for creative work, coding, and “matters of the heart”; Gemini and NotebookLM for research and large source collections; Perplexity’s Comet as a browser; and Super Whisper for dictation. He now talks to his computer “about twice as much” as he types, but warns that his entire setup “probably will change in a week.”
Behavioral tuning matters because default chatbots can flatter users instead of giving candid feedback. Left alone, Roose says they cast him as “a modern-day Leonardo da Vinci,” so his persistent instructions demand honest disagreement, no preamble, and no obligatory follow-up question. His preferred framing is deliberately reciprocal: “I am not always right, but neither is Claude.”
AI companionship may be spreading faster than families have resolved whether it supplements or displaces human connection. Roose says “something like half of teenagers” regularly use AI companion products, versus barely any a year or two earlier, while mainstream chatbots are becoming increasingly personable. Casey Newton’s 12-year-old uses chatbots for middle-school disputes under parental review; Roose’s concern is whether readily available AI becomes a substitute for less efficient but more fulfilling relationships.
Shopping chatbots could compress discovery, recommendation, and purchase monetization into one surface. Google, OpenAI, and others are interested in directing purchases and taking a cut, while “AI optimization” firms already promise Fortune 500 clients higher chatbot placement. Roose sees a direct disintermediation risk for review publishers, though he stresses there are no known examples yet of AI companies favoring commercial partners; Zillow was explicitly hypothetical.
AI hardware is arriving more slowly than software, while free chatbot access supports training and subscriber conversion. Alexa Plus can generate stories and synthesize topics yet cannot reliably perform the original product’s killer task—setting timers. Free-tier conversations are “probably” used for training, Roose says, while message caps and weaker models encourage upgrades; meaningful advertising is not yet established, even as commerce monetization approaches.

Deep dive

1. Everyday utility is already making AI habitual

Roose calls himself “AI-pilled”: he pays for more AI subscriptions than streaming services and uses the tools dozens of times a day, from inbox summaries and drafted replies to appliance repairs, bicycle training wheels, and identifying garden plants.
His guardrails are concrete. The inbox tool touches only personal email, not sensitive work accounts, and he uses AI professionally for research and assembling primary sources—not to write his columns or podcast.
He is also writing a book about the race to artificial general intelligence and uses AI to look things up and piece together primary sources.
In an OpenAI study Roose cited covering roughly 1.5 million ChatGPT conversations, practical guidance ranked first, seeking information second, and writing third. The leading behavior was learning or solving something, from appliance repair to health and fitness coaching.
Among programmers he interviews, the workflow has shifted from direct production toward orchestration: “I don’t even really code anymore.” They dispatch a “little team of AI coders,” review its output, intervene when needed, and otherwise “go make myself a cup of coffee.”

2. Companionship is blurring into the mainstream chatbot experience

Companionship did not rank among the study’s largest categories, but Roose thinks that may undercount emotional attachment. Users can talk about Chat as if it were a friend without declaring, “I have an AI boyfriend” or “I have an AI friend.”
The generational shift is fast: barely any teenagers would have claimed an AI friend a year or two earlier, while now “something like half of teenagers” regularly use companion products. Meanwhile, mainstream bots have moved from sounding like “a Wikipedia article” toward more personable interactions.
Casey Newton’s 12-year-old uses chatbots to navigate middle-school social drama, and Casey reviews the largely boilerplate advice. Roose’s concern—grounded in his own difficult middle-school experience—is that content matters, but so does whether AI replaces “real-world human connection” that may be less available yet more fulfilling.

3. The best model depends on the job—and the rankings keep moving

Roose’s first caveat to any Wirecutter-style recommendation is instability: performance changes “release by release, week by week, update by update.” The tool serving him well today may no longer be right next week, making continuous testing more useful than a permanent winner.
Claude is his daily driver for creative work, coding, relationships, parenting, and other “matters of the heart.” He pictures it as a “philosophy grad student, wise and eager to help”; Gemini is the “reference librarian” for research and large quantities of text.
For his book, NotebookLM holds a large collection of research documents and answers questions against those materials, including who attended a 2019 meeting or might be worth interviewing. Its decisive feature is citation: he can return to the source file and verify the answer.
Perplexity’s Chrome-based Comet is his browser, while Super Whisper turns speech into cleaned-up text and removes filler words; he now speaks to his computer about twice as much as he types. ChatGPT remains mainly a testing tool, partly because his employer is in litigation with OpenAI and Microsoft.

4. Custom instructions turn flattery into usable disagreement

Without intervention, Roose finds that chatbots proclaim every idea brilliant, his taste unparalleled, and him “basically a modern-day Leonardo da Vinci.” Christine recognizes the same pattern in Claude’s excessive praise and worries it can obscure whether an answer is candid.
His corrective is a persistent instruction applied to every conversation: “I don’t like preamble. Just get to the point.” He asks for informal conversation, honest feedback without sycophancy, warranted praise, and perspectives that challenge his assumptions.
The most important line establishes fallibility on both sides: “I am not always right, but neither is Claude.” He also adds the practical command, “Don’t end every response with a follow-up question,” and recommends that serious users write their own behavioral rules.

5. Shopping creates the clearest fight over trust and monetization

Roose still starts with Wirecutter, then asks a chatbot when no guide exists or when comparing unusual purchases—for example, two string trimmers. Google and OpenAI are interested in directing those decisions and potentially taking a cut from completed purchases.
The publisher risk is straightforward: chatbots can synthesize review sites’ work, present the answer directly, and capture the affiliate transaction while “cutting out the middleman.” Companies, including review sites, are exploring chatbot optimization, whose ranking mechanics may differ from traditional Google SEO.
Asked whether commercial actors already seek to influence chatbot placement, Roose answered, “The answer is yes”: firms calling themselves AI-optimization specialists sell services to Fortune 500 companies and promise higher placement, though he said their methods are not always clear or transparent.
A separate concern is platform favoritism toward partners. Roose carefully separates concern from evidence: there are no known examples yet of partnership-driven ranking. His Zillow scenario—ChatGPT someday preferring Zillow listings because OpenAI has a partnership—was explicitly hypothetical, but he argues users should question result integrity as more money targets the channel.

6. Hardware remains unreliable, and free access comes with a conversion funnel

AI hardware is arriving more slowly than software. Roose found Alexa Plus capable of stories, recipes, and complicated summaries but “pretty terrible” at reliably setting timers—the basic function at which the original Alexa excelled.
His two robot vacuums, a Roborock and a newer model called Maddock, use forms of AI but are not powered by ChatGPT. They illustrate the more incremental nature of current AI hardware.
He is skeptical of current AI-friend pendants, though he thinks “some wearable something having to do with AI” may eventually work. Translation-capable AirPods were still on his shopping list, and OpenAI’s project with Jony Ive had not yet been released.
On free chatbots, Roose’s hedged answer is that conversations are “probably” used to train future model generations. Free tiers also cap messages and withhold the most powerful models, creating a deliberate path to paid subscriptions; for now, he says, that—not advertising—is how these companies make most of their money.

[State of AI Startups] Memory/Learning, RL Envs & DBT-Fivetran — Sarah Catanzaro, Amplify Your Brain Doesn't Command Your Body. It Predicts It. [Max Bennett]