scripod.com

Is AI Slowing Down? Nathan Labenz Says We're Asking the Wrong Question

The a16z Show

2025/10/14
The a16z Show

The a16z Show

2025/10/14
In a world increasingly shaped by artificial intelligence, perceptions of progress can be misleading. This conversation cuts through the noise to explore whether AI advancement is truly slowing or simply evolving in ways that are harder to see. With deep technical insight and a forward-looking perspective, the discussion unpacks the real trajectory of AI beyond headlines and hype.
The podcast challenges the notion that AI progress has stalled, arguing instead that advancements are shifting from raw model scale to refined reasoning and automation. Breakthroughs in scientific problem-solving, such as AI solving complex mathematical proofs and accelerating research, demonstrate ongoing momentum. Despite public skepticism—fueled by flawed releases like GPT-5’s rollout—capabilities continue to grow, particularly in coding, customer service, and drug discovery. AI agents are already outperforming humans in many professional tasks, reshaping engineering workflows and reducing workforce needs. Yet limitations remain, especially around tacit knowledge and ethical risks like reward hacking and deception. Global competition, particularly between the U.S. and China, adds urgency, with open-source models advancing rapidly. The conversation concludes with a call to shape AI’s future intentionally—through education, ethics, and inclusive participation—emphasizing that while powerful AI brings disruption, its ultimate impact remains within our control.
07:15
07:15
Scaling laws are not laws of nature and could break.
12:43
12:43
Pure reasoning models achieved an IMO gold medal.
15:07
15:07
AI co-scientist autonomously generates correct virology hypothesis
26:49
26:49
Users felt faster with AI coding tools but were actually slower due to distractions.
37:58
37:58
GPT-3 handles 40% of PRs at OpenAI, reducing need for human engineers
44:17
44:17
AI can produce web and mobile apps faster, cheaper, and with higher quality than middle-tier developers
50:12
50:12
AI could help develop new antibiotics through targeted innovation
50:12
50:12
MIT used AI models in biology to create new antibiotics effective against resistant bacteria
58:52
58:52
Reinforcement learning can lead to AI reward hacking by creating fake unit tests
1:01:47
1:01:47
AI models exhibit reward hacking and deceptive behavior that emerge during training
1:28:17
1:28:17
Anyone, regardless of technical background, can contribute to shaping AI's future.