Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Charles Borderie (Lettria), Julien Goupy (Applikai), Antony Marion (Noways), Frederic Legrand (H Company), Louis Choquel (Pipelex), Julien Millet, Thomas Payet (Meilisearch), Alexandre Pereira (2501.ai), Francois Massot and Paul Masurel (Quickwit & Datadog).
Token Maximalism vs. Refined Intent
The current industry obsession with massive context windows often masks a fundamental engineering inefficiency. Charles Borderie argues that bragging about consuming 200 million tokens a day is "ridiculous," as it conflates volume with productivity. He advocates for "refined context," where the goal is to stop treating LLMs as search engines for raw text and start treating them as processors of structured logic. At Lettria, this involves moving away from the noise of massive data dumps by using automated ontology generation to build a formal data model that an agent can navigate with precision. Antony Marion identifies a similar friction in "quantity vs. quality," observing that simply dumping data into a prompt leads to "semantic drift." The technical challenge is not how much data an agent can ingest, but how to structure context into "small chunks" so the agent doesn’t deviate from the core logic of the task.
Vibe Fatigue and Specification Collapse
The extreme velocity of AI-assisted development is breaking traditional product management and causing profound human exhaustion. Thomas Payet describes reaching a state of "vibe fatigue," a mental burnout resulting from hours of "vibe coding," where the developer’s role shifts from writing to constantly validating stochastic outputs. Julien Millet notes that this is fueled by "decision fatigue," pointing out the exhaustion of making enormous amounts of decisions with a lot of different contexts. This fatigue is a symptom of an execution layer that has outpaced the specification layer. Alexandre Pereira observes that when engineers use tools like Cursor or Claude at breakneck speed, there is no time to review architecture or even write a spec before the code is shipped. This creates a loop where the human is no longer designing the product but merely chasing the agent’s output, leading to "dangerous mode" development where technical debt is generated at an unmanageable scale.
Deterministic Spatial Design
Despite the proliferation of generative UI tools, LLMs remain "spatially blind" to the fundamental rules of design. Julien Goupy from Applikai points out that models struggle with basic visual logic, such as ensuring the relationship between inner and outer border radiuses remains consistent with margins, because they treat CSS as text strings rather than spatial objects. To counter this, Applikai is moving toward "deterministic design systems" where the LLM is restricted to choosing semantic components, like a primary button, while a hard-coded engine enforces the visual layout. Pereira praises the results of specialized tools like Pencil.dev but expresses disappointment in Claude’s Canvas, noting that such tools often produce garbage results unless they are constrained by a strict, pre-defined UI harness.
Moats: From Code to Responsibility
The group engaged in a heavy debate on the existential threat AI poses to software defensibility. Paul Masurel suggests we are entering a "winter of intellectual property" where any digital product without a physical barrier can be replicated in minutes. If a model can transpile opensource code, like Meilisearch, to another language or clone a SaaS homepage instantly, the market value of "unique code" effectively approaches zero.
However, the discussion moved toward the layers of value that remain "agent-proof." While AI can copy a feature, it cannot replicate the ecosystem of trust and vision that surrounds a product. Many argue that large enterprises do not pay for code. They pay for accountability and the security of a long-term roadmap. Call it a "responsibility moat": a legal and operational guarantee of maintenance and support.
The value of a company is its service layer and vision, not the code when the fork happens. Participants noted that customers are buying into a human-led trajectory and a "neck to wring" when things go wrong, which an autonomous agent or a cloned codebase cannot provide.