Zack’s collection of ideas shaping what’s next.

Solving tasks with no or a handful of examples embedded in the prompt (in-context learning).

On-body devices (glasses, rings, earbuds) that can host/mediate AI and help make computing ambient.

A storage/indexing system for embeddings to enable fast semantic search and retrieval-augmented generation.

Ubiquitous computing is a new paradigm where intelligent systems are seamlessly embedded into everyday environments, making technology effectively invisible while continuously supporting human activity. Unlike traditional computing that requires explicit prompts and fragmented devices, ubiquitous computing leverages context-awareness, memory, and integration to deliver adaptive, anticipatory, and user-centric experiences.

A decoding setting controlling randomness/creativity; higher values yield more diverse outputs.

A neural architecture built on attention mechanisms; foundation of modern LLMs.

The unit of text (or audio/video chunk) a model processes; counts affect cost/latency.

AI-generated data used to supplement scarce or sensitive real data (with care for drift and artifacts).

Supervised learning stage using instruction–response pairs to teach task formats and tone.

Empirical relationships between data/parameters/compute and model performance that inform roadmap decisions.

Policies, controls, and lifecycle practices that mitigate misuse, harm, and failure modes (safety filters, audits, governance).

Fine-tuning with human preference data to align outputs with desired style/behavior.

The instruction/context you give a model for it to complete a task; includes system prompts, examples, constraints, and tool schemas.

Fetching relevant context (via search/vector DB) and injecting it into the prompt to ground responses in current or proprietary data.

Initial large-scale training on broad data before specialized tuning (SFT, RLHF).

Learned numeric values inside a model (billions+ for frontier LLMs) that encode patterns from training data.

Mechanisms that let models call tools (search, databases, emails, code) and stitch steps together into workflows—key for agents. Agents will need to conduct widespread orchestration across different websites, apps, and platforms.

A layered function approximator trained to minimize loss via gradient descent. Breakthroughs in AI stem from this idea of a neural network where machines can make connections and think in the same way a human brain functions, rather than having to follow basic commands that tell it step-by-step, right from wrong.

A transformer-based model trained on massive text corpora to generate/analyze language; increasingly multimodal. Ex: ChatGPT, Claude, Gemini, etc.

A model that consumes and/or produces multiple data types (text, images, audio, video). ChatGPT is multimodal because it can provide text and images to you.

Numeric parameters inside a neural network that encode what the model has learned; 'open-weight' models expose weights for local use/modification; 'closed-weight' do not.

Dynamically choosing the best model for a given prompt based on predicted quality/cost/latency—often a meta-model trained on evaluation data.

Round-trip time from request to response; an experience bottleneck for agents and wearables.

Running a trained model to generate outputs (production use). The costs and latency of inference shape the 'unmetered intelligence' thesis. In simple terms, the cost to generate outputs using AI is rapidly declining, leading to a world of abundant knowledge.

Confident but false output from a model; mitigated by retrieval, calibration, routing, and post-processing. The root cause of much of the fear of AI.

The latest point in time included in a model’s pretraining data. The model lacks builtin knowledge from after this cutoff, unless updated or retrieved.

A large, broad-training model adaptable to many downstream tasks (e.g., LLMs, multimodal models). Ex: GPT family, Claude, Mistral, Gemini.

Additional training on a narrower dataset to specialize a base model (via SFT, LoRA, etc.). Leveraging your own data to make a large model better (i.e. training a model with your company’s data so it can better respond to queries regarding certain processes or customers).

Methods to make model behavior understandable (feature attributions, summaries, rule extraction); emphasized as a solution to AI ethical concerns.

A claim that traditional schooling structures (pace, assessment, curricula) must be reimagined for an AI-rich world that values creativity, curiosity, and agency over memorization.