Peter Depaepe

The through-line

Depaepe's recurring argument is that a large public broadcaster cannot generate its own speed. VRT moved away from building all its technology in-house and now sources innovation externally through startups, on the reasoning that a company of more than a thousand people has no other way of moving quickly enough [1][2]. The image he keeps returning to is nautical: "Het klassieke beeld is: wij zijn een groot schip, log; de kleine speedboten zijn de start-ups die uiteindelijk die grote logge boot toch voor een stuk in de juiste richting en sneller in de juiste richting brengen" [2]. Startups bring energy, new technology and new formats, and the accelerator exists to test them in live contexts, from Pimp My Company and Scribewave to interactivity around the Tour de France [1][2].

Running alongside that is a defence of the local and the European against a technology stack built elsewhere. Whether the subject is language models, discovery or transcription, his position is that Flemish language and European culture are not edge cases to be handled by generic global systems, and that the party who owns the infrastructure owns the outcome [1][2]. He is unsentimental about the economics underneath all of it: "Innoveren is mooi, maar bedrijven bouwen, daar heb je geld voor nodig" [1], and in his own formulation of the lesson, "Als er één ding is dat ik zelf geleerd heb: innoveren is mooi, maar bedrijven bouwen, daar heb je geld voor nodig" [2].

On sovereign European AI

The Nvidia-Mistral deal gives Europe a jump start on sovereign language models, and Depaepe wants Europe to go further: "Wij denken dat we zelf ook in Europa een aantal hele sterke modellen moeten bouwen die veel meer onze eigen taal en ook onze cultuur begrijpen" [1][2]. But he sees a contradiction in the arrangement as it stands. Training European models on American infrastructure is not sovereignty; genuine sovereignty requires owning the infrastructure layer as well, which is exactly the work companies in Ghent's Wintercircus are engaged in [1][2].

On the gap between model capability and user capability

The bottleneck he identifies is no longer the technology. "Op dit moment worden er nieuwe modellen ongeveer om de drie maanden op de markt losgelaten. Die zijn elke keer beter en sterker en kunnen veel meer. En eigenlijk kunnen wij zelf als gebruikers niet volgen" [1][2]. He points to OpenAI's CFO making the same diagnosis: the biggest challenge is not model quality but that users, businesses and consumers alike, extract little value from what arrives every three months, which is why the stated mission for 2026 is coaching and education rather than acquiring more users [1][2].

On discovery, metadata and trust

As audiences shift from apps to AI assistants for finding news, Depaepe treats discoverability as the central problem for a public broadcaster. VRT's priority is getting its metadata in order so that LLMs can find its content and represent it correctly, which he frames as the equivalent of SEO for AI models [1][2]. The second half of the answer is provenance: adopting C2PA, the BBC and Adobe standard, to stamp content as trustworthy [1]. Deliberately, VRT has chosen neither to pay LLM providers nor to be paid by them, unlike Associated Press with its OpenAI deal, betting instead that metadata quality and trust stamps will keep it both findable and credible [1][2].

On choice fatigue and why linear is coming back

He reads Netflix's move into linear television with TF1 as a response to a user problem rather than a technology one. "Sommige mensen zijn meer dan een half uur aan het zoeken voor ze effectief op play duwen" [1]. A scheduled feed removes that decision fatigue, and it carries a second benefit for the platform: it lets Netflix push its own IP at lower cost [1][2].

On the funding model and the radio problem

Roughly half of VRT's budget comes from government dotation and the other half from the market, where radio advertising revenue effectively cross-subsidizes television production [1][2]. That structure makes declining radio listening an existential issue rather than a departmental one, because falling radio ad revenue directly threatens future video output [1][2]. It is the reason behind experiments with AI-personalized audio [1]. On the broader position of public broadcasting he is measured but not defensive: "Er zijn veel publieke omroepen, zelfs in Europa, die onder druk staan. Wij staan gelukkig niet onder druk. Ik denk dat wij het zeer goed doen" [1].

On language specialization as a moat

Generic Flemish audio is routinely misinterpreted by mainstream models [1]. Scribewave's models are trained specifically on Flemish, so VRT radio programmes are transcribed and cut into snippets correctly where generic systems fail [1][2]. Depaepe treats this as more than a procurement detail: language specialization is the moat available to local AI startups competing against global players [2].

On personalization at scale

Personalized media multiplies production load in a way manual workflows cannot absorb. Where a campaign once needed two brand-asset variants, personalization demands around eighty, which is only feasible by training AI models on a brand's style guide instead of designing each one by hand in Adobe [1].

Takeaways

• A large media organization should stop trying to build all its technology in-house and use startups as speedboats to steer the slow ship: "Het klassieke beeld is: wij zijn een groot schip, log; de kleine speedboten zijn de start-ups die uiteindelijk die grote logge boot toch voor een stuk in de juiste richting en sneller in de juiste richting brengen" [2].
• European AI sovereignty is incomplete while European models train on American infrastructure; owning the infrastructure layer is part of the requirement [1][2].
• The binding constraint on AI value is user adoption, not model quality, since models ship roughly every three months and "eigenlijk kunnen wij zelf als gebruikers niet volgen" [1][2].
• For AI-driven discovery, invest in metadata quality and C2PA provenance stamps rather than in paid deals with LLM providers [1][2].
• Netflix's linear push with TF1 answers choice fatigue and cheaply promotes its own IP; some viewers spend over half an hour searching before pressing play [1][2].
• Radio advertising cross-subsidizes television at VRT, so declining radio listening threatens the whole funding model and justifies AI-personalized audio experiments [1][2].
• Local AI startups can defend themselves through language specialization, as Scribewave does with Flemish-trained models where generic models misread the audio [1][2].
• Personalization needs about 80 brand-asset variants instead of 2, which only works if AI models are trained on the brand's style guide [1].
• Innovation without capital goes nowhere: "Innoveren is mooi, maar bedrijven bouwen, daar heb je geld voor nodig" [1][2].