The coding war is won. Implementation, the labor that once defined software, is now trivial. Andrew Ambrosino, OpenAI's Codex lead, describes an inversion of the product process: teams prototype first because building costs nothing. This abundance creates a coordination tax. At OpenAI, 90 uncoordinated teams might build 90 different versions of the same feature. The difficulty has shifted from ‘can we build it’ to ‘which of these 90 attempts is actually good.’
Ambrosino calls this the curation bottleneck. When token budgets are unlimited, the value of ‘taste’ skyrockets. It’s the only thing models can’t replicate because design taste requires a human feedback loop that training data lacks. A model knows if a function compiles, but it doesn’t intuitively understand if a UI feels ‘snappy.’ Humans remain the final arbiter of what feels right.
“When the cost of tokens is unlimited, the value of ‘taste’ - knowing how a system should feel and behave - skyrockets. It is the only thing the models can’t yet replicate.”
- Andrew Ambrosino, Lenny’s Podcast
This collapse reshuffles job titles. Ambrosino describes the modern OpenAI employee as a ‘member of technical staff’ whose actual role is just the average of where they spend their time. If a designer spends 60% of their day writing code, they are effectively a developer for that period.
He warns that abandoning specialized roles like product manager is a mistake because disciplines have knowable best practices. Stripping the role away often leads to ‘slop’ - unfocused features built simply because they were easy to code. Success now depends on ‘command over the discipline’ paired with technical agency. The ability to steer an agent is more valuable than memorizing syntax.
Nathaniel Whittemore, on The AI Daily Brief, argues firms must use a forced pause in frontier model releases to master the latent power already on their desktops. GPT-5.6, Claude Sonnet 5, and Gemini 3.5 Pro have been delayed. This creates a ‘capability overhang’ - the gap between what current models can do and how poorly we utilize them.
Whittemore’s playbook advises building personal infrastructure: portable context assets and custom benchmark portfolios to turn AI from a toy into a predictable workflow component. He warns companies are stuck in ‘Efficiency AI,’ using tools to perform existing tasks faster, and missing ‘Opportunity AI’ - capabilities and products that were previously impossible.
“Instead of waiting for a new flagship, operators should use this time to extract the latent value already sitting on their desktops. The race to the next model is paused, but the race to use them effectively is wide open.”
- Nathaniel Whittemore, The AI Daily Brief
Designers are uniquely positioned for this new era. On the a16z Show, Paul Backus noted designers consistently get better results from Claude and Codex than engineers because they use a specialized lexicon. Terms like ‘negative space’ and ‘rhythm’ act as precision steering for LLMs. He built Impeccable to encode this design-first language directly into the agent’s harness.
John Maeda argues we are moving from UX to AX - Agentic Experience. Design is no longer just about the screen. It’s about designing for robots, command lines, and ‘dash-dash help’ functions. In an AX world, value shifts from visual polish to speed, clarity, and verbosity for agents.
The bottleneck is now human talent, not model capability.
