Скачать или смотреть Building Sovereign AI Pathways: The Swiss Advantages

1/ In this session, Antoine Bosselut, Assistant Professor of EPFL, with Kasima (Aom) Tharnpipitchai, Head of AI Strategy of SCB 10X, shared “Building Sovereign AI Pathways: The Swiss Advantages.”

2/ Switzerland launched the Alps supercomputer and the fully open Apertus model series, coordinating 800+ researchers across 10+ institutions. Antoine highlighted strong talent, bold academic leadership, and tight academia–industry links as key enablers.

3/ Strategic backbone: a public investment in ~11,000 GPU chips enabled research at frontier scale. Small-country connectivity created rapid buy-in from industry, giving public institutions confidence to back an ambitious, open initiative.

4/ Why sovereign infra matters: Alps changes the research frontier. Projects infeasible on commercial clouds become possible when you control compute at scale and cost. It enables novel, large-scale training—not just “more of the same” experiments.

5/ Cost reality: commercial clouds often price around ~$2/GPU-hour. Pre-training at Alps scale would cost ~$15–20M. With public infra, direct costs (energy/cooling) and capex amortization land at ~40–50% of the cheapest cloud rates.

6/ Alps ranks 6th globally and runs on hydropower with innovative cooling. Beyond performance, Antoine emphasized stability and predictability—critical for sustained open science and sovereign AI development.

7/ Apertus is fully open and transparent across data, code, checkpoints, and architecture—trained on 1,500+ languages. The team enforced high compliance standards, honoring author opt-outs (even retroactively) and avoiding contentious data dumps.

8/ Trade-off management: skipping certain “high-quality” but restricted corpora caused minor drops on some benchmarks, yet real-world performance remained strong—especially in multilingual settings where Apertus excels.

9/ On multilinguality: Antoine challenged the “curse” narrative. With the right mixture design, English need not degrade; Apertus performs well on English benchmarks and is outstanding on multilingual ones.

10/ Enterprise lens: sovereign AI fills gaps where global offerings don’t fit local needs (e.g., sectoral or regulatory contexts). It also acts as a strategic hedge—preserving control if commercial options shift tomorrow.

11/ Long-view capability: building sovereign AI cultivates local talent and know-how to keep improving at the pace of frontier options. Today’s digital foundations determine who rides—and shapes—the next technological wave.

12/ Safety vs “frontier at any cost”: many European enterprises will accept a capabilities trade-off for transparency, responsibility, and compliance. Antoine noted the gap with closed models is shrinking—and speed of frontier gains has eased.

13/ Lessons for other regions: invest with a five-year horizon; differentiate clearly (as with Apertus’ three thrusts of openness, scale, and compliance); and consider shared continental bets to diversify infrastructure risk.

14/ Accessing Swiss sovereign AI: the team welcomes bidirectional collaboration—deployment partners, real-world evaluators, and contributors in training and data curation. Version 2 will be larger; community involvement is key.

15/ Takeaway: public supercomputing + open, compliant multilingual models can turn national strengths into global AI assets. Switzerland’s approach shows how sovereign AI can advance science, serve enterprise needs, and preserve strategic control.