The arms-race for AI for the average person is pure fallacy. It's the cognitive load-creating friction in attempting to understand an extremely difficult subject completely deflating the subjcect-matter, despite never needing to understand its deepest tranches in the first place.
Example: you should always extract deterministic logic from a non-deterministic subject. I.e. use code (libraries and standard functions) for calculations, but use LLMs and their counterparts (HRMs, SLM, VLLMS, etc.) for their NLU abilities. Small models, whenorchestrated properly, can often outperform in terms of efficiency against architects who brute-force frontier-level intelligence.
Take our calculations, ephemeral subjects, and hard-maths from the model to be calculated by stable code written decades ago. Let the model itself determine what code to actually invoke.
Salesforce uses symbols like "->" for deterministic logic while using pipes like "|" for passing prompts to the LLM. This is basic orchestration framed in the Salesforce-method but shows up in different methods and manners from Flowise, Make.com, Zapier, AWS Step Functions, wtc. Yes, you need to understand some syntax but said syntax only appears like "magic" for a few weeks. Once you get your hands on it, the magic fades away and becomes just another tool in your belt.
Even the idea of embedded text is beyond the capabilities of most people, but this doesn't detract from delivering value. I don't get mad that a dog cannot use a calculator, nor that a squirrel doesn't understand biology. This is pure technological function applied to daily work tasks, and your ability to comprehend the underlying tech vs. your ability to explain it does not matter in day-to-day solving problems.
Are we really at an arms-race with the world? Or are we shadow-boxing foes that most likely we shall never meet? If you cannot communicate "temperature" to a product manager, how likely will it change a product when explaining cosine similarities in a geometric space to map semantic relationships gets lost in translation? While you, the architect and orchestrator must understand how these impact the product, this is not something that takes more than a few weeks of testing and QA to grasp its high-level affects. Furthermore, once you understand one set of models and weights, you are far more easily able to grasp other models and weights. Let Google, Anthropic, xAI, and the CCP fight out the "true" arms-race - the rest of us should be focused on building value without the fear of falling behind.
Top comments (0)