Fable 5, Edge AI, and Personalized Models
The Daily AI Show
Jul 04
Fable 5, Edge AI, and Personalized Models
Fable 5, Edge AI, and Personalized Models
The Daily AI Show
Jul 04
In this episode, the hosts explore the practical side of AI, moving beyond the hype of new frontier models to focus on strategic model selection and efficient deployment. They discuss the trade-offs between using powerful but expensive models like Fable 5 versus more reliable and cost-effective options like Sonnet 5, emphasizing the importance of choosing the right tool for the task.
The conversation highlights the trend of rapid model releases, where no single model stays dominant for long, and the need for intentional use of AI as capability expands. Key topics include the unreliability of AI search results, the high costs of data-center cooling and custom AI chips, and the potential of edge models for faster, more personal AI. The hosts also delve into technical concepts like LoRA adapters for fine-tuning models without retraining all weights, and Google's on-device architecture for Android. They conclude by reflecting on AI's ability to invent through combination, as seen in generative video experiments like Fusion Animals, and the importance of verifying AI outputs with multiple sources.
03:20
03:20
Marginal gains from newer models are minimal
05:19
05:19
Sonnet 5 caught false premises better than Opus 4.8
13:42
13:42
Give Fable a day's tasks for automation ideas
17:45
17:45
Edge computing bypasses memory constraints
25:45
25:45
Crowd-sourced information is more reliable than probabilistic AI.
30:24
30:24
Chillers consume 40% of facility power
36:27
36:27
Chip shortages drive up AI memory costs and hardware prices.
40:43
40:43
LoRA fine-tunes a frozen model without retraining all weights.
53:53
53:53
AI can invent by combining concepts without bias.
55:26
55:26
LoRA adapters are more cost-effective than context tokens
