From: redpointai
Daan van der Kila, CEO and co-founder of Contextual AI, shared his immediate reactions and insights regarding OpenAI’s latest o1 model release, emphasizing its implications for system-level thinking in AI development [00:00:38].
Initial Impressions and Core Concepts
Van der Kila found the o1 release “very exciting,” noting that many in the field, including his own team, have been focusing on “test time compute” [00:00:46]. He highlights that o1 moves towards a crucial direction of thinking around systems rather than just models [00:01:01].
Systems Over Models
The o1 model compresses many Chain of Thought ideas, which have existed for a while, into the model using Reinforcement Learning from Human Feedback (RLHF), turning it into a more complicated system [00:01:05]. This approach is “very encouraging” for reasoning tasks [00:01:21]. Van der Kila’s own team at Contextual AI works on similar ideas, but with a focus on the retrieval side, aiming to address the entire system rather than solely predicting the next word [00:01:27].
Latency and Trade-offs
Future models may follow o1’s approach, but it depends on the deployment’s latency constraints [00:01:49]. Performing significant “thinking” during test time increases latency, which isn’t always optimal [00:01:55]. While o1 shows greater power in areas like math and law, older models can still perform better on other tasks and are significantly faster [00:02:07]. This implies a need for user interfaces that accommodate potential wait times, especially for complex reasoning with long contexts [00:02:16].
Historical Context and OpenAI’s Execution
Van der Kila notes that the core ideas in o1 are “not super new,” suggesting that similar concepts might have already been implicitly present in other models [00:03:10]. However, he gives OpenAI credit for their exceptional execution [00:03:21].
Contextual AI’s Differentiated Approach
Contextual AI’s founding thesis was born from the frustration of enterprises where generative AI was exciting but not yet ready for production [00:03:53]. Their approach differs from OpenAI’s and Anthropic’s in two main ways:
- Systems over Models: Contextual AI views the model as only 10-20% of the much larger system required to solve enterprise problems [00:04:43]. Enterprises need to buy the complete system, not just the model, to avoid the complexity of building the rest around it [00:05:01].
- Specialization over AGI: Contextual AI does not pursue Artificial General Intelligence (AGI), which Van der Kila views as fundamentally a consumer product due to the lack of specific user needs [00:05:17]. In contrast, enterprises often know exactly what they want and require specialized, not generalist, AI systems [00:05:31]. For example, a generalist AI performing a performance review could lead to severe sanctions in the EU, highlighting the need for specialization and constraint [00:05:44].
Contextual AI focuses on end-to-end specialization and integration across all parts of the system, including retrieval, reranking, generation, post-training, and alignment [00:06:17]. This vertical slicing through the “layered cake” of the AI stack allows for a compounding effect, enabling them to excel at solving high-value, knowledge-intensive use cases for enterprises [00:07:09]. This resonates with customers who are tired of assembling many disparate solutions [00:06:45].
Future Outlook: The “Model S” Approach in AI
Van der Kila believes the industry is “only just getting started” with AI [00:49:54]. He emphasizes that the “cars” (products) coming off the assembly line are systems, not just models [00:49:59]. While scaling laws are discussed, he points out that at a certain point, there’s significant value in making models better through post-training, Chain of Thought, and distilling systems thinking back into the model itself [00:50:08]. This “scaling in many directions” suggests a multifaceted approach to AI innovation beyond just increasing model size [00:50:35].
Contextual AI aims to be capital-efficient by not training its own large base models, instead leveraging resources like open source models [00:48:08]. This allows them to allocate capital to hiring talented individuals and working closely with customers to solve real problems, which is critical since most AI products aren’t yet “turnkey” [00:48:19]. This strategy aligns with a “Tesla Roadster” phase, where high-end, tailored solutions are delivered with dedicated support, while simultaneously building the “assembly line for the Model S” – a more generalized, out-of-the-box solution for the future [00:49:05].