Srao @pri_rao - Twitter Profile

pri_rao retweeted

over 1 year ago

In the context of AI, a lot of the old "wisdom" on moats, pmf, speed, are all coming off their hinges. @JeffBezos talks about addressing the constants - go after things that don't change. What are some of the constant intrinsics founders can rely on? A big one is instinct - great founders have strong instincts. Building a company is just the art of institutionalizing instinct before it becomes irrelevant. Whenever we reach an impasse on defining a strategy, I just tell founders to go for a walk, dig into their gut, and come back and tell me what they want to do. The best founders continue to strike bullseye every time. Instinct can be built but takes effort - curiosity, an ability to quiet your mind and reflect, ability to listen well & seek dissenting views, experimenting widely. Another one is taste. There is no such thing as PMF; only moment-market fit. What's hard is that the product and the moment decay at different speeds. Only with great taste can one navigate these two shifting grounds deftly. Taste & Instinct are related and often built through massive experimentation and risk taking (often throughout their lives and often unrelated to the startup). Which brings me to risk taking. The job of a founder is not to reduce risk, but to allocate it where it compounds. The early risks you take can result in a moat. A moat is not IP, tech, code, regulatory capture, or whatever. In a way they all are, but also aren't. A moat is not what protects your business. It’s simply what lets you make mistakes without dying. A strong brand, or capital, or culture. Risk takers build moats over time because they are more willing to put distance between them and others.

3

146

6

111

19K

pri_rao retweeted

martin_casado

@martin_casado

over 2 years ago

One thing that's so unique about current use of LLMs is that application logic is being abdicated to them. I can't think of any other time this has happened. Normally infrastructure moderates resource usage, not logic and correctness.

18

120

13

43

30K

pri_rao retweeted

Pratyush Choudhury (PC)

@177pc

over 2 years ago

This is an incredibly important point by @martin_casado - something that very few people understand/appreciate. In traditional software development, infra & application logic have distinct roles. Infra provides the underlying resources (such as servers, databases & networking), while application logic governs behavior, validation & correctness. Traditionally, application logic is explicitly coded by developers & sits on top of infra. This architecture ensures predictable behavior, making debugging & scaling straightforward. With LLMs, this boundary blurs. LLMs can process data, infer context & generate responses dynamically, often w/o predefined logic. The logic isn't hard-coded but emerges from the interaction between the model's architecture, training data & inputs. This flexibility allows LLMs to take on roles that traditionally required complex, explicitly defined logic. However, it introduces uncertainty because the logic isn't always predictable or transparent. Reconciling this w/ traditional software is an interesting thought experiment - some thoughts below: 1/ Validation & Correctness: Traditional software relies on strict validation rules, ensuring expected behavior. LLMs generate responses based on training data & user inputs, which can lead to unexpected or incorrect results. Bridging this gap requires additional validation layers, often w/ human oversight or secondary systems for validation. 2/ Scalability & Performance: LLMs demand significant computational resources, affecting scalability. Traditional software might scale by adding more servers or optimizing code, but LLMs require specialized infrastructure (GPUs). Ensuring consistent performance w/ growing demand is a unique challenge for LLM-based applications. 3/ Transparency & Explainability: Traditional software's logic is explicit, making it easier to understand & debug. LLMs operate in a "black box", posing challenges to explainability. Bridging this gap requires techniques to increase transparency, like model interpretability & explainability tools. Would love to learn from others.

2

32

9

13

9K

pri_rao retweeted