Each week, TTY Lunch brings together exceptional builders around the table. Today’s lineup included Bertrand Guiheneuf (Fairjungle), Karim Matrah (Contrast), Olivier Desclaux (Harmattan), Quentin Soulet (Dune Health), Sacha Morard (Edgee), Francis Bouvier (Lightpanda), Denis Brule (Finegrain), and Pierina Camarena (42 AI).
Fast code alone doesn’t build product value
The core of the discussion centered on whether the newfound ability to generate massive amounts of code actually translates into meaningful product evolution. While AI allows Fairjungle’s small team to compete against much larger players, Bertrand Guiheneuf observes a shift in developer behavior, where code is increasingly treated as a disposable commodity. He describes his own workflow, generating a high volume of branches, pushing most of them to 90 percent completion, only to abandon them because the cost of exploration has dropped so low. These branches are discarded not necessarily due to failure, but because the ease of generation enables testing multiple parallel hypotheses before committing to the one that best aligns with the product’s purpose.
Francis Bouvier challenged the group to consider the reality of this output, asking whether anyone truly felt their products were moving ten or twenty times faster simply because code volume had increased. Karim Matrah argued that this surge in output often lacks business purpose, as if we are building more but thinking less. More importantly, code is not the only bottleneck. People overlook user experience, especially in consumer tech where product design matters far more than accumulating features. Quentin Soulet reinforces this by noting that in the $4 trillion US healthcare market, the winner won’t just be the one who ships the most code, but the one who achieves the highest “velocity” in building, testing, and iterating on entire automated care infrastructures that actually solve patient protocols. Sacha Morard added that this over-boosted state makes it difficult to maintain quality standards, as the ease of shipping can lead to doing anything and everything simply because the cost of delivery has dropped close to zero.
Non-Engineers Shipping Code Without System Awareness
Karim detailed a specific friction arising from product-first profiles, such as PMs and designers, using tools like Claude Code to contribute directly to codebases. While these contributors have the right business intent, they often fail to respect underlying technical abstractions because they lack foundational engineering skills. Karim noted, “The risk is that we have people who can ship code but don’t know how to decompose a problem or understand the dependencies they are creating.” The group discussed the difficulty of enabling non-technical people to use AI effectively to build tools, as they often bypass the rigor required for proper work decomposition. This forces a shift in the engineering team’s role toward becoming a filter for these contributions, ensuring that business-led code does not compromise the system’s structural integrity.
Can AI Learn Taste
Pierina Camarena noted that while productivity gains have been achieved on the technical side of code delivery, the real upside will come from addressing less tangible problems like design and other soft elements. The group explored how the next frontier for AI is the integration of subjective data, including sentiment and aesthetic judgment, into multimodal models, which today are still largely trained on descriptive proxies. The challenge appears significant, but Denis Brule pointed out that Google Maps started from near zero and required massive data collection efforts. Collecting subjective data may not be as far off as it seems.
Facing the New GAFAM
This led to a discussion of the existential threat posed by major tech giants and the types of moats startups can still build. Denis questioned whether most startup work is simply noise given the distribution power of GAFAM and emerging AI giants. The group noted that these players own the rails, making it nearly impossible for startups to win on general-purpose models. Instead, the focus must shift toward specialized moats, particularly in under-digitized areas where data is messy, physical, or highly regulated. Denis Brule’s focus at Finegrain is to build models that never leave the user’s device, ensuring a level of data sovereignty that cloud-first giants cannot easily replicate without undermining their business models.
Classical Models Still Power Critical Systems
Participants compared modern generative AI with classical machine learning, which still dominates high-precision fields. Karim noted that in robotics and hardware, classical AI methods such as decision trees and behavior trees remain dominant due to their predictability. Bertrand confirmed this in the travel sector, where LLMs assist with coding and natural language tasks, but classical boosting algorithms remain far superior for price prediction and user choice modeling. Olivier Desclaux added that in the defense sector, deterministic computer vision models remain relevant as they can operate under extreme hardware constraints.