Key Takeaways
- Yann LeCun introduced AdaJEPA, an adaptive world model for closed-loop planning and action, and will present time-stretching work at ICML.
- Hamel Husain shared that Claude Code can enable remote control by default, though the feature may be unstable.
- Swyx highlighted a bug in the original LLM scaling laws paper that caused significant wasted compute.
- Swyx also noted key trends from the AI Engineer conference: open models, local coding agents, model routing, and Fable model.
- Deedy analyzed the risk of leaving high-growth companies like Facebook or OpenAI to start a startup, citing successful examples and suggesting higher success rates for such founders.
- Yann LeCun criticized the concept of AGI, arguing current AI capabilities are far below human and animal level, calling AGI "nonsense."
1. New Models and Research
Yann LeCun announced AdaJEPA, an adaptive world model that can plan, act, and adapt in a closed loop, optimizing latent representations and predictions with each action. He will also present work on temporal straightening for latent planning at ICML. — via 1 2
LeCun criticized the concept of AGI, stating that current AI systems (e.g., autonomous driving, household robots) are far less capable than humans or animals, and calling the term "general" nonsense. — via 1
2. Tools and Practical Insights
- Hamel Husain shared that Claude Code can enable remote control by default via configuration, but warned the feature may be unstable. — via 1
3. Industry Trends and Insights
- Swyx noted a cold fact: the original LLM scaling laws paper contained a bug that led to significant wasted compute. — via 1
- Swyx highlighted trends from the AI Engineer conference, including open-source models, local coding agents, model routing, the Fable model, and coding agents reliability. — via 1
- Deedy discussed the risk-reward of leaving fast-growing companies (e.g., Facebook, OpenAI) to start a startup, citing successful examples like Dustin Moskovitz (Asana), Adam D'Angelo (Quora), Brian Armstrong (Coinbase), and noted data may show higher success rates for such founders. — via 1
