Conference Presentation, Panel
From Research to Reality: Why Open Source Is the Engine of AI | RAISE Summit 2026
RAISE SummitIon Stoica, Thomas Wolf, Dillon Rolnick, Eiso Kant, Robin Rombach, Mark Porter, Dylan Rolnick
- Panel Composition & Core Premise: The panel features founders from Poolside (Iso Kent), Databricks/Arena (Thomas Wolf), Neuse Research (Dylan Rolnick), Blackhorse Labs (Robin), and Hugging Face (Thomas Wolf), moderated by Mark Porter, debating "Innovation in Open Source Versus Closed Models" with a focus on preventing a dystopian future where intelligence is monopolized by 2-4 companies.
- Isso Kent's Strategic Stance: Poolside is open-weighting models to ensure intelligence becomes a commodity rather than a proprietary product, asserting that "open source is not a business model" yet is an ideological necessity to avoid a world controlled by a few firms.
- Thomas Wolf's Innovation Thesis: All recent major AI techniques (e.g., sparse attention, latent attention) originate from open source, and closed models like those from Fable are increasingly blocking scientific research domains like biotech, forcing startups into dependency silos.
- Dylan Rolnick's Ecosystem Warning: Over-reliance on single closed models (e.g., from Fable or DeepSeek) leads to unanticipated bias transfer (e.g., American models adopting Chinese geopolitical biases) and limits the ability to switch to the optimal model for specific tasks.
- Robin's Market Reality: Open source is currently the only viable ground for "world models" (robotics, physical AI), as no closed source world model significantly outperforms open alternatives, and the field requires the fastest innovation loop only open weights can provide.
- Databricks/ISO on Commodity Intelligence: Successful open source requires dominating on at least one dimension (data, compute, or people); while traditional open source won via "people," future wins require massive capital investment and a shift toward commodity infrastructure to decouple value from proprietary control.
- Benchmarking vs. Real-World Utility: The industry is shifting away from "shady" benchmark overfitting (e.g., the "LaMa4" trend) toward evaluation on real, dynamic tasks (Arena's Human Eval) to ensure models possess genuine utility rather than just high test scores.
- Transparency Debate: While Poolside argues against open-sourcing training data due to security risks and lack of immediate business incentives, other panelists note that open-sourcing software layers creates immense distribution and trust benefits, even if full data transparency remains a "big ask."
- Geopolitical & Regulatory Concerns: The EU faces significant risks of being locked out of the intelligence revolution due to US-centric control (e.g., Cloud Act, export controls) and an over-reliance on "fear-mongering" narratives that drive restrictive regulations rather than enabling innovation.
- Security Sentiment: Panelists largely dismiss current "dystopian" security fears regarding LLMs as overblown, noting that unlike the internet's initial spread of worms, there have been no major AI-enabled security catastrophes in the past four years beyond deepfakes.
- Forward-Looking Predictions (2025-2026):
- Thomas Wolf: AI will drive massive breakthroughs in non-LLM scientific fields (fusion, materials, physics), potentially solving problems humans cannot conceive.
- Robin: The next few years will be "insanely stressful" for physical AI and world models, but Europe's manufacturing base positions it to lead if resources are correctly directed.
- General Consensus: The industry must abandon "boring" SaaS-centric thinking to achieve "positive extremes" in technology, requiring urgent, ambitious, and coordinated investment across academia and government.
- Infrastructure & Ecosystem: The open source track is explicitly funded by Hugging Face, Mozilla, and Commit, highlighting the critical role of infrastructure providers in enabling the transition to open, commoditized intelligence.