AI Foundations for Software Investors

Classify a software capability as rules-based automation, predictive machine learning, generative AI, or foundation-model based, and justify the classification.

Interpret language-model output as a probability distribution and explain why generation can be variable and confidently wrong.

Tokenize short text and relate token count to context usage and inference cost.

Use embedding similarity to explain how models and retrieval systems represent semantic relationships.

Trace self-attention for a short sequence to explain how a transformer combines contextual information.

Explain how next-token pretraining creates broadly reusable capabilities without guaranteeing factual accuracy.

Distinguish pretraining, instruction tuning, and preference-based post-training by their purpose and required inputs.

Run a guided Python notebook that sends a request to a language-model API and inspect its inputs, parameters, and response.

Test how prompt structure and sampling settings affect the quality and consistency of model outputs.

Construct a small product-specific evaluation set with explicit scoring criteria.

Diagnose hallucination, prompt sensitivity, and non-determinism from a set of model traces.

Choose among prompting, retrieval-augmented generation, and fine-tuning for a specified product requirement.

Trace a retrieval-augmented generation pipeline and localize failures to retrieval or generation.

Explain the observe-reason-act loop of a tool-using AI agent.

Decompose a workflow into deterministic, model-driven, and human-review steps at an appropriate level of autonomy.

Map the generative-AI value chain from compute and cloud infrastructure through foundation models to applications.

Interpret scaling laws as relationships among model performance, data, compute, and diminishing economic returns.

Calculate approximate inference cost per completed customer task from token usage, model pricing, and repeated calls.

Compare proprietary APIs, open-weight models, and self-hosting on quality, latency, control, cost, and dependency.

Evaluate an AI software company's gross-margin assumptions and exposure to model-provider commoditization.

Identify a promising AI product wedge using workflow pain, task frequency, feasibility, and measurable customer value.

Design an AI product interaction that communicates uncertainty and places human review at the right point.

Select product and business metrics that test activation, retention, expansion, and willingness to pay for an AI workflow.

Assess whether proprietary data, workflow integration, distribution, or feedback loops can create durable advantage.

Identify material privacy, security, copyright, and regulatory risks in an AI software scenario.

Stress-test an AI product demo for cherry-picked examples, hidden human work, fragile dependencies, and weak evaluation.

Assess a founding team's technical credibility and learning velocity using evidence available at seed stage.

Create a seed-diligence question set covering product value, model architecture, evaluation, economics, defensibility, and risk.

Write an AI-first seed investment memo that states the thesis, key uncertainties, validation milestones, and disconfirming evidence.