이미지: TechCrunch AI
Summary
- At the Ai4 conference in Las Vegas, Hinton, Fei-Fei Li, and Andrew Ng discussed AI regulation and openness
- Andrew Ng warned against a handful of companies becoming gatekeepers and prescribed greater openness as the remedy
- Hinton acknowledged that the spread of open-weight models has become an irreversible trend, while still flagging the risk of misuse
- 행사
- Ai4 콘퍼런스, 라스베이거스, 2026년 8월 초 개최
- 참석자
- 제프리 힌턴(노벨상 수상자), 페이페이 리(월드랩스 CEO 겸 공동창업자), 앤드루 응(코세라 공동창업자)
- 앤드루 응 핵심 발언
- "게이트키퍼가 있는 걸 원치 않는다"
- 힌턴의 구분
- 오픈소스(코드 공개)와 오픈웨이트(학습된 가중치 공개)는 다르다고 강조
- 힌턴의 우려
- 오픈웨이트 모델을 저비용으로 재학습시켜 사이버공격 등 악용에 쓸 수 있다는 점
- 힌턴의 인정
- 오픈웨이트 확산은 이미 되돌릴 수 없는 흐름이 됐다고 발언
- 배경 맥락
- Pacing the Frontier 등 프론티어 랩 중심 안전 프로젝트가 업계에서 논의되는 시점
Lead
Seated side by side on a stage at the Las Vegas Convention Center were three of the most notable names in the history of artificial intelligence: Nobel laureate Geoffrey Hinton, World Labs CEO Fei-Fei Li, and Coursera co-founder Andrew Ng. At last week's Ai4 conference, all three voiced a shared concern: AI should not be allowed to fall under the control of a handful of companies. Their reasoning and proposed solutions, however, diverged somewhat.
What happened
In a panel discussion at the Ai4 conference, the three exchanged views on regulation, open-source access, and U.S. competitiveness amid China's pursuit. Amid ongoing moves by frontier AI labs to concentrate advanced model development among a small number of companies in the name of safety (including a project known as "Pacing the Frontier"), open-source and open-weight models have emerged as a hot-button issue in the industry. Because they are freely distributed and difficult to control in terms of usage, some labs have treated open-weight models as a risk factor.
Andrew Ng argued that if a handful of companies control the platforms, the pace of innovation could slow. Citing how Apple and Google have split the mobile operating system market between them, he said, "We don't want to have gatekeepers." His prescription was singular: preserve openness by maintaining a competitive landscape with multiple models and companies.
Hinton's turnaround
Hinton drew a clear distinction between open source and open weights. Open source, he explained, means releasing the code itself so anyone can check for bugs, whereas open weights simply hand over the parameters of an already-trained massive model as-is. He stated that he has opposed open weights, citing concerns that large foundation models — which cost enormous sums to train — could be cheaply retrained for malicious purposes such as cyberattacks.
Even so, Hinton acknowledged reality: open-weight models have already become an irreversible trend, and the barrier to entry once posed by the cost of training large models has effectively disappeared. Even the person who has warned longest about these risks conceded that the moment to stop the trend has passed.
Why this debate is happening now
The tension surrounding open weights is tied to recent industry developments. Alibaba's Qwen released a 2.4-trillion-parameter model as open-weight, which was running on vLLM from day one of its launch, while DeepSeek has been driving a low-cost race with V4 Pro, offering benchmarks that approach the open-source camp. On the other side, OpenAI on August 10 unveiled GPT-5.6-Cyber, a model dedicated to cybersecurity, taking an approach that supports authorized vulnerability research while restricting access to controlled channels — a direct response to the low-cost misuse scenarios that concern Hinton.
Comparing the three positions
| Person | Affiliation | Core position |
|---|---|---|
| Andrew Ng | Coursera co-founder | Supports openness, opposes the emergence of gatekeepers |
| Geoffrey Hinton | Nobel laureate | Cautious about open weights but acknowledges their spread is irreversible |
| Fei-Fei Li | CEO, World Labs | Participated in the regulation/openness discussion; specific position only partially covered in available excerpts |
So what changes
This debate won't immediately change policy. But the fact that three leading figures in the industry all identified monopolization by a handful of companies as a shared risk is not trivial. Now that even Hinton — who has warned longest about the dangers of open-weight models — has conceded the trend cannot be reversed, the axis of debate is shifting from "whether to stop open weights" to "how to handle open weights safely." Approaches like GPT-5.6-Cyber, which aim to retain the benefits of openness while controlling access, are likely to become more common going forward.
