TIL #1: Open Source Will Save Us All
One thing I learned this week, Ed. 1 – Oct 14, 2024
As I kickoff my PhD studies, I'm relaunching this newsletter as a way to share what I’ve picked up along the way.
This week, it’s Meta's chief AI scientist Yann LeCun's perspective on AI systems and open source development.
LeCun argues that we shouldn't fear AI takeover, at least not yet, and advocates for open source as a safeguard. While his views on transparency and democratization are compelling, I can't help but question the naiveté of relying solely on self-regulating social systems (i.e. engineers will save us all cause open = moral).
This, I feel, will be a theme we’ll return to in our discussions.
LeCun's Grounded View on AI
AI Godfather: Yann LeCun, Meta's Chief AI scientist and NYU professor, is one of the so-called godfathers of AI along with Geoffrey Hinton (Nobel Prize winner as of this week) and Yoshua Bengio (trio won the Turing Award for their work in neural networks, previously mocked by GOFAI, or Symbolic AI proponents)
Current AI Limitations: LeCun believes today's AI is less intelligent than our pets, lacking:
Mental model of the physical world
Persistent memory
Reasoning ability
Planning capacity
AI Mimicry: Current AI systems, including large language models, mimic human expression rather than demonstrating true intelligence.
Design vs. Scale: LeCun argues AI has a design limitation, not a scale limitation. Models appear to reason due to vast memory, but merely regurgitate training data.
Contrasting Views
Bengio: Advocates for governmental regulation to prevent misuse of superhuman AIs.
Hinton: Left Google to freely discuss AI risks, expressing alarm about the technology's potential dangers.
LeCun: Maintains a tech-optimistic outlook, focusing on AI's potential for positive development, especially if AI is developed open source.
Future of AI
Inspiration: LeCun's vision is inspired by baby animals' learning process.
Goal: Create models that construct a world model from visual information, mirroring infant learning. LLMs will play a role, but an integrated role among a broader set of solutions.
Approach: Aims to develop more powerful, reliable, and safe AI systems through open-source platforms.
Source: This AI Pioneer Thinks AI Is Dumber Than a Cat (Wall Street Journal, by Christopher Mims)