Babylon Newsletter, Issue #5

This is the number of people who use Meta's platforms — WhatsApp, Instagram, Facebook — every day, according to Meta's own investor relations. Meta just launched "Muse Spark," a new AI model that will serve as the interface for all of them. Unlike the previous model the new one will not be open source. It will not ask users to pay. Instead it will help them find things to buy (and very likely talk to advertisers on this basis).

Why care? Three major AI models are now heading in three distinct directions.

Anthropic bets on subscriptions — individual and enterprise — and recently reported annualized revenue of $30 billion, up from $9 billion at the end of 2025.

OpenAI experiments with a mixed model: subscriptions plus advertising for free tiers, projecting $2.5 billion in ad revenue this year and $100 billion by 2030.

Meta takes a third path: no direct payment, just better targeting of 3.6 billion people. The business model shapes the product. We don't know how, but the models will shape societies and all of us, too. See longer article in the next section for more details.

8 April 2026 · Meta AI Blog

What's new: Muse Spark is Meta's first model from its internal superintelligence group — and the first not released as open source. Unlike the Llama family, which developers could download and run independently, this one stays inside Meta's ecosystem. It is optimized for language and image understanding — not coding or deep reasoning — and is designed to power chatbots across WhatsApp, Instagram, and Facebook.

Why care? Meta's stated ambition extends to 1,600 languages via its OMT research. Muse Spark launches with seven major languages — English, Spanish, French, German, Hindi, Italian, and Portuguese — with around 25 more to follow in coming weeks. The deeper shift may be structural: AI recommendation layers now sit inside the same communication environments where conversations already influence decisions. Users request suggestions directly rather than searching. Discovery is moving from search engines into chat.

Reality check: Meta describes the model as capable of "Personal Superintelligence." One primary application announced is shopping: an "Add to Cart" feature that helps users find and buy products. Health advice is mentioned as one area of application. Advertising is the engine.

Open question: Meta's OMT research promises potential coverage of 1,600 languages. Muse Spark launches with seven. How long before the gap closes — and what quality standard will apply to the languages that follow?

9 April 2026 · ElevenLabs Blog

What's new: ElevenLabs announced on-premise and on-device deployment options, expanding beyond cloud and VPC. On-premise runs on a customer's own GPU servers inside their data center. On-device runs directly on hardware — vehicles, wearables, embedded systems — with no internet connection required.

Why care? EU data residency has been a genuine obstacle to ElevenLabs adoption in Europe. Local deployment removes that barrier for government agencies, financial services, and any organization that cannot send audio to a US cloud.

Reality check: Both options are in early access, with initial releases expected in the first half of 2026. This is an announcement, not a shipping product.

Open question: Pricing is custom and case-by-case. Whether this is accessible to mid-size European media organizations or only large enterprise clients is not yet clear.

2 April 2026 · Microsoft AI

What's new: Microsoft launched three in-house models: MAI-Transcribe-1 (speech-to-text, 25 languages, 2.5x faster than previous Azure offerings), MAI-Voice-1 (text-to-speech, 60 seconds of audio generated per second, custom voice cloning from a 10-second sample), and MAI-Image-2 (text-to-image). All available via Microsoft Foundry.

Why care? Microsoft is no longer just an OpenAI reseller. MAI-Transcribe-1 starts at $0.36 per hour. MAI-Voice-1 at $22 per million characters. These are competitive prices against ElevenLabs and Whisper. The speech stack is now proprietary.

Reality check: Microsoft claims 50% lower GPU costs than competitors. That is a vendor claim. Independent benchmarks do not yet exist for MAI-Transcribe-1.

Open question: How does MAI-Transcribe-1 perform on languages outside the top 25 — specifically lower-resource European languages?

Sources: Nearshore Americas · ElevenLabs

What's new: The stock of TP (Teleperformance before a rebranding), the world's largest call center operator, has crashed 88% since 2021, from over €400 to around €50. The reason: Investors expect AI to solve call center tasks much cheaper. Klarna remains the most cited case: its AI assistant, built on OpenAI and ElevenLabs voice agents, handled the workload equivalent of 700 full-time agents — according to a press release. Additional claims: Resolution time dropped from 11 minutes to 2 minutes. Klarna's headcount fell from 5,527 to 3,422 between 2022 and 2024.

Why care? High quality voice combined with fast database access and rules-based decisions makes call centers an obvious target for AI. In simple, well-defined cases this will work faster and cheaper than a human agent. But the picture is partial. In customer service, easy cases and complex problems require very different things — and AI is not equally good at both.

Reality check: In mid-2025, Klarna's CEO admitted that over-prioritizing AI cost reduction led to "subpar results" and lower quality. The company began hiring humans again for complex, empathy-based roles. Teleperformance is investing €100 million in AI partnerships to augment rather than replace its 490,000 agents.

RSL — Really Simple Licensing — is an open standard launched in September 2025, reaching version 1.0 in December. It lets any publisher attach machine-readable licensing terms to their content via a file called rsl.txt, readable by AI crawlers the same way robots.txt is. The difference: instead of a binary block or allow, RSL defines a vocabulary. Free with attribution. Pay-per-crawl. Pay-per-inference. No AI use at all.

The RSL Collective, the non-profit organization behind the standard, functions like ASCAP in music: it represents publishers collectively, negotiates with AI companies on their behalf, and passes licensing revenue through without taking a profit. For now, it asks no membership dues. It gets paid when publishers get paid.

The significant development: RSL is no longer just a signalling mechanism. Cloudflare, Akamai, and Fastly — the infrastructure layer that handles a significant share of web traffic globally — have signed on to validate crawlers against RSL terms. A bot that ignores RSL now risks being blocked at the CDN level.

Current supporters include Reddit, Yahoo, Associated Press, Ziff Davis, and Wikipedia. The AI industry has been largely silent.

The RSL Collective puts a concrete number on what is at stake. A chart from their WAN-IFRA presentation (sources: IDC, Goldman Sachs, WARC, ISBA) shows AI spending projected to reach $1 trillion by 2030, while publisher net ad revenue stays flat at around $110 billion. OpenAI is already at a $25 billion run rate. Anthropic at $19 billion. The gap between those two curves is what RSL is trying to bridge — by capturing a percentage of AI revenue and routing it back to the publishers whose content made those models possible. John Boyden, RSL's head of partnerships, frames it simply: "Spending on AI is exploding, while the open web publisher economy has been staying flat."

Two pieces help frame what RSL is actually trying to solve.

Thomas Baekdal calculated that pay-per-crawl alone cannot work for publishers — a one-time fee pays for training but not for reuse. He uses a cinema business model as an example: imagine if a cinema only had to buy a film once and could then show it to as many people as it wanted. The studios would be out of business within a year. RSL's pay-per-inference model is the attempt to fix that.

Tim O'Reilly aims to solve the problem from the infrastructure side — the agentic economy is missing the mechanisms that would make it function: skills markets, quality governance, registries, payment layers. RSL is one attempt to build the payment layer. Whether it develops real negotiating leverage or remains a well-intentioned standard that bots ignore depends on adoption at a scale not yet tested.

Both pieces are worth reading in full — they go deep and list many important details that cannot be integrated into a short newsletter article. RSL is promising — one serious attempt to build a future where AI and content creation coexist on fair terms. But many parts of that formula still need to be finely calibrated before a functioning market for information exchange becomes reality.

→ Thomas Baekdal — "Why Pay-to-crawl is not useful for publishers" · Baekdal/Basic · April 2026

→ Tim O'Reilly — "The Missing Mechanisms of the Agentic Economy" · O'Reilly Radar · 23 March 2026

The main annual gathering for the localization industry is happening this week. Six tracks cover the full range of pressures the industry is under right now: technology and AI, global market expansion, compliance and geopolitical risk, workforce and leadership, and language accessibility. The keynote by Prof. Deborah Nas (TU Delft) asks why some technology innovations transform industries while others quietly fail — a question the localization industry is currently living through in real time.

I have been to Perugia five or six times over the past ten years. Every spring the festival brings together international journalism people across hotels all over the city — talks, panels, decent coffee, and a lot of ideas in between. It is one of the few places where you actually meet people instead of just watching slides.

Here, as an appetizer, three interesting sessions you can follow live or on-demand, even if you cannot go to Italy. The full program is online — maybe invest a bit of time to find other interesting talks.

Tokenmaxxing

This is a new word for a troubling concept. Some tech companies now treat AI usage as a performance metric — the more you use, the better. Employees have dashboards where they compete for top places by how many AI tokens they consume. Meta engineers race on a leaderboard called Claudeonomics. OpenAI and Anthropic list token usage in job descriptions.

Nvidia's CEO Jensen Huang made the case explicitly on the All-In Podcast in March 2026: "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed." He compared engineers not using AI to chip designers who refuse CAD tools and work only with pencil and paper. When asked if Nvidia is trying to spend $1–2 billion on tokens, his answer was: "We're trying to."

The trend reveals an AI maximalist mindset. But hectic token consumption is not the same as deep productivity. One engineer described bots running in circles just to inflate the count. Token spend is a signal, not a verdict. But it is now a status symbol.

Sources: New York Times, March 20, 2026 · Business Insider

WAN-IFRA AI in Media Webinar · 23 April · 16:00 CEST · wan-ifra-aim.org · "Infactory: Structure your messy content to unlock its value in the AI market." Brooke Hartley Moy (Infactory) and Mary Liz McCurdy (The Atlantic). Relevant to the RSL and structured content thread.

EBU HORIZONS · 5–6 May · Geneva · ebu.ch · Media distribution technology and broadcast innovation.

TAUS Massively Multilingual AI Conference · 3–5 June · Rome · taus.net · European translation technology, MT quality, AI orchestration. The best single gathering for the Babylon beat in the first half of the year.

WAN-IFRA World News Media Congress · 1–3 June · Marseille · wan-ifra.org · AI and content rights central themes. The RSL and Mistral levy debate will surface here.

EAMT 2026 · 15–18 June · Tilburg · eamt2026.org · Main European academic conference on machine translation.

Languages & The Media · 4–6 November · London · languages-media.com · Theme: "Moving Images That Move Audiences: Localising with Intent."

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter: plain X (plainx.com) — media localisation platform, DW Innovation / Priberam. ChatEurope (chateurope.eu) — AI chatbot network for 15 European news partners. MOSAIC(mosaic-media.eu) — EU DIGITAL EUROPE-funded multilingual media infrastructure.

AI use: I use Claude (Anthropic) for research and to edit this newsletter, based on refined and specific prompts. The goal is to get help and to find out where AI makes mistakes. Responsibility for stated facts, names, and links is entirely mine.

babylon-newsletter.com · Every Tuesday at 10:00 CET
7,000 languages. AI works for 20.