From: redpointai
Eric Reese and Jeremy Howard are establishing what they call the “Bell Labs of AI” at Answer AI [00:00:01]. This initiative aims to build smaller, cheaper, and more affordable AI models and applications, particularly in legal and education sectors [00:00:05].
Answer AI: A Modern R&D Lab
Answer AI functions as a for-profit R&D lab [00:14:37]. Its public benefit mission, similar to Jeremy Howard’s previous work at fast.ai, is to maximize the public benefit of AI for as many people as possible [00:09:39]. The shift to a funded R&D lab allows for greater investment in this mission [00:10:57].
The founders, despite their experience, faced challenges in funding this unconventional model, as it doesn’t fit the typical startup mold with clear proprietary technology or a five-year financial plan [00:14:42].
Core Principles of the Bell Labs Approach
Integrated Research and Development
A key tenet of this approach is the belief that the best research occurs when the researcher is directly linked to the application [00:16:04]. This contrasts with the historical “hyperspecialization” where research (R) and development (D) became separated, leading to research untethered from practical applications [00:15:36]. This integrated model ensures that the iteration loop extends from the customer all the way into scientific inquiry and back again [00:16:21].
The “Long Leash with Narrow Fences” Model
Inspired by historian Eric Gilliam’s analysis of Edison’s lab, Answer AI operates with a “Long Leash with Narrow Fences” [00:25:24]. The “narrow fences” are the initial research theses or focus areas, such as addressing the overinvestment in training large foundation models and the underinvestment in resource-constrained, real-world applications [00:25:40]. Within these fences, individual researchers are given a “long leash” to explore and pursue promising ideas, as exemplified by Karam’s independent breakthrough in efficient fine-tuning of Llama 3 [00:24:14].
Importance of Practicality and Manufacturability
The approach emphasizes making AI technology practical and deployable, drawing lessons from physical manufacturing, deep-sea oil drilling, and nuclear power plants [00:30:08]. This includes a strong focus on resource efficiency and reducing costs, which some larger labs might view as “tedious” but is fundamental to increasing AI accessibility [00:26:59]. A difference in degree (like cost reduction) can become a difference in kind, enabling entirely new applications that were previously too expensive [00:27:11].
Two Types of Inventions: “Light Bulb” and “Phonograph”
The lab aims for both types of inventions, as seen in Edison’s time:
- “Light Bulb” Inventions: Practical applications of existing technology that are obvious once the underlying tech is mature. An example is the efficient fine-tuning of Llama 3, which combined quantization and distributed computing to dramatically improve fine-tuning accessibility and cost [00:30:17].
- “Phonograph” Inventions: Clever, significant leaps that combine existing principles in novel ways to create something entirely new, like using stored energy to create sound [00:30:42]. This requires fostering a group of individuals with strong intuition for the technology’s capabilities [00:32:15].
Fostering Intuition and Ecosystem Understanding
To cultivate this intuition, every team member receives a $500 monthly credit card to purchase and use AI products [00:33:11]. This experimentation and testing of AI use cases helps them deeply understand the ecosystem, product benefits, drawbacks, and opportunities, fostering conversations about different approaches to retrieval or other AI applications [00:33:47].
Customer-Centric Development
Even in an R&D lab, the principles of Lean Startup apply [00:14:04]. The core idea is that a business exists to create a customer [00:14:15]. This means understanding the “end-end-end customer” and what they want [00:03:23]. It’s crucial to avoid working on problems that, despite their scientific difficulty, offer no value to the customer [00:16:41]. Rapid iteration and the ability to pivot are vital because assumptions about what customers want, or even what capabilities models will have, are highly uncertain [00:01:51].
Distinctive Aspects in AI
While many AI companies are focused on large foundation models, Answer AI focuses on unlocking a “reservoir of applications” that don’t require AGI [00:45:14]. This counteracts the “fundraising gravity” that pushes entrepreneurs toward speculative science fiction over practical utility [00:45:24].
They prioritize smaller, safer models that are properly fine-tuned, arguing that if these intrinsically safe options aren’t provided, users will default to potentially unsafe large frontier models [00:46:15].
Answer AI is specifically excited about applying this approach to legal and education sectors, which are heavily language-based [00:34:46]. They believe that reducing the cost of high-quality legal advice can combat the use of law as a “weapon by wealthy people” [00:35:31]. Similarly, in education, AI can help customize learning paths, allowing more people to realize their potential [00:37:22].