AI Agents, Token Costs, and the New Internet Taste

Product management should not be the default operating model; it should be used only where it creates clear leverage.

Tom’s core thesis is that engineers and designers can often make better decisions when they have the context and judgment to own the work directly. This challenges the common assumption that every product effort needs a PM, and it pushes organizations to be more intentional about when the PM role actually adds value. The underlying concept is a more distributed decision-making model, where product thinking is embedded across the team rather than centralized in a PM layer.

AI is accelerating the shift toward senior, hands-on IC-style product work.

Tom believes AI reduces the cost of analyzing data, reading code, and iterating quickly, which makes experienced operators more effective without relying as much on coordination-heavy processes. This matters because it changes the leverage profile of the PM role: judgment, technical fluency, and speed become more important than process management. Technologies and concepts involved include AI-assisted analysis, code understanding, and faster decision loops.

The best PMs are judged by customer insight, business context, technical fluency, and decisive execution—not stakeholder theater.

Tom is skeptical of candidates who are strongest at alignment and communication without demonstrating deep product judgment. His hiring philosophy suggests that successful PMs need to synthesize customer needs, business realities, and technical constraints into concrete decisions. This is significant because it reframes PM excellence as outcome-driven rather than coordination-driven, and it makes interview signals like case studies and decision quality more important than polished presentation.

Whatnot organizes product work around problems and planning cycles, not permanent team attachments.

Instead of assigning PMs to fixed squads indefinitely, leadership defines what must be true every six months and then assigns owners to the highest-priority work. This structure supports flexibility, faster reallocation of talent, and clearer accountability for outcomes. The concept reflects an operating model centered on DRIs, planning horizons, and problem-first staffing.

In founder-led companies, being close to the work is a leadership advantage, not necessarily micromanagement.

Tom argues that product leaders need ground truth from data, customers, and implementation details in order to make better decisions. That means staying engaged in the weeds can improve judgment and speed, especially when the founder’s intuition is a major driver of the business. The key concept here is calibration: leaders should debate with evidence, align where appropriate, and avoid treating autonomy as a universal good.

Great product leadership requires oscillating between strategy and execution.

Tom’s “play the accordion” model describes repeatedly zooming out to understand the problem space and then zooming in to ship, learn, and revise. This is important because it prevents teams from getting stuck either in abstract strategy or in narrow execution without context. It’s a practical framework for balancing long-term thinking with rapid iteration and feedback.

Measure AI spend by cost per successful task, not raw token volume or sticker price.

The episode argues that token counts alone do not capture business value, because different models, tokenizers, and workflows can change the bill for the same outcome. Using cost per accepted task or completed job gives teams a more operationally meaningful metric for evaluating ROI, especially when comparing models like GPT-style systems or agent frameworks.

Fear of token costs can be more expensive than the tokens themselves.

A key insight is that the most valuable users are often the ones most afraid to experiment, which reduces learning and slows adoption. This matters because innovation in AI often comes from trying workflows, testing prompts, and iterating on agents; if people self-censor due to cost anxiety, organizations may underuse the technology and lose competitive advantage.

Agentic workflows are inherently token-intensive because they loop, retry, and repeatedly exchange context with tools.

Unlike simple chat interactions, agents may make many model calls, re-send long histories, and perform deep research or tool-based steps, causing costs to multiply quickly. The episode highlights widely cited estimates that agents can consume 5 to 30 times more tokens than a basic chat, which makes workflow design and control mechanisms essential.

Model choice and tokenizer differences can materially affect total cost, even when prices look similar on paper.

Two models may have comparable per-token pricing but still generate very different bills depending on how they tokenize text, how many reasoning tokens they use, and how efficiently they complete the task. This is especially important for multilingual use cases, where tokenization can make non-English text more expensive, and for teams comparing reasoning-heavy models versus cheaper alternatives.

Organizations should actively eliminate 'tokens that spin'—work that creates activity without value.

The episode recommends auditing agent loops, unused automations, overly frequent jobs, and oversized prompts or context windows that add cost without improving outcomes. This is significant because waste can accumulate invisibly in agent systems, so teams need dashboards and operational reviews to separate 'tokens that teach,' 'tokens that produce,' and 'tokens that spin.'

Cyberethnography helps explain online culture by studying behavior in digital spaces, not just what people say they like.

Ruby frames cyberethnography as an observational method for understanding communities like furries, Instagram users, and other online groups. This matters because platforms create new social signals—such as likes, Stories, and Close Friends memberships—that require a new vocabulary to interpret correctly.

Many viral 'trends' are not truly universal; they may be micro-behaviors amplified by algorithms, media coverage, or niche communities.

Ruby emphasizes using both quantitative data and real-world observation to test whether a phenomenon is actually widespread. This is significant because it challenges the assumption that social media visibility equals cultural dominance, especially on platforms like TikTok.

The internet is becoming a collection of smaller, distinct subcultures rather than one shared mainstream, leading to a kind of 'pluric culture.'

Ruby describes this fragmentation as 'BABBELIFICATION,' where the same word can mean different things across communities and platforms. The implication is that online language and norms are increasingly localized, which can make cross-community communication confusing and error-prone.

AI is changing what 'taste' means because creators now optimize for both humans and machines, not just human audiences.

Ruby suggests that algorithms, bots, and AI systems increasingly shape what gets seen, liked, and rewarded. This makes taste less about personal aesthetic preference alone and more about navigating machine-mediated attention systems, including concepts like 'machine-against' taste and agent-to-agent interaction.

People often reject the label 'AI' even while relying on AI-powered tools every day, showing that public resistance is partly about language.

Ruby points out that Americans may say they dislike AI because of job anxiety or abstract fears, but still use products like chat assistants and other AI-enabled tools for practical reasons. This is important because it suggests adoption is driven more by utility than by public enthusiasm for the term itself.