Trends in AI

Overview

This is a 40-minute presentation used to help college seniors get a better understanding of the key trends in artificial intelligence. Give an overview of the important concepts in measuring AI capabilities and make some projections to understand if trends will continues, where will this lead.

We then look at key skills and try to understand what skills will be important in an AI dominated future.

Four Futures

Moore's Law

Power Wall

AI Progress Is Difficult to Predict

Within a generation the problem of creating 'artificial intelligence' will substantially be solved.
Marvin Minsky (1967)

Overly Optimistic AI Claims

AI Systems Thinking

AI Flywheel

Traditional AI Benchmarks

Many AI benchmarks show LLM capabilities approaching human-level skills like the MMLU benchmark.

Some AI Benchmark Compare LLMs with Each Other in an Arena

LM Arena Benchmark

METR Task Horizons

The METR.org studies show that AI task completion has been doubling every 7 months since 2019.

Projecting AI

We can now project this trend line out until 2030. The results are sometimes difficult to believe. In 2030 tasks that take 1,000 days will be possible with 50% chance of correctness.

Modern Desktop

Boris Cherny said his desktop has 15 agents running concurrently. 10 are doing planning and 5 are writing code.

Agent Skill Standards

Agentic Skills are directory packing standards for agent rules
Introduced by Anthropic in October 2025
Now adopted by most LLM harnesses (OpenAI Codex, Cursor, Windsurf, Perplexity etc.)
Allows consistent generation of code, content and images
Gave me a 100x increase in the quality of MicroSimulations in interactive intelligent textbooks
Clearly shows that AI progress is a non-linear function

Claude Skills for Intelligent Textbooks

Small Language Models (SLMs)

SLMs achieve performance that rivals much larger models from just 1–2 years ago
Extensive pruning and use of smaller weights (16, 8 and 4 bits) allow models to fit in smaller RAM
Tools such as Ollama continue to allow users to run AI models (both text-to-text and text-to-image) on local commodity GPU hardware (12GB to 24GB GPUs)
Laptops and edge devices enable preserve privacy, allow for offline use, and promote low-latency AI applications

Ollama

RPA and OpenClaw

Natural language → desktop actions – LLMs can translate user intent (“download this report and email it”) into step-by-step UI actions, enabling flexible RPA without rigid scripts using tools like OpenClaw.
Resilient to UI changes – Vision + reasoning allows agents to adapt when buttons move or layouts change, reducing the brittleness of traditional RPA workflows.
End-to-end task automation – LLM agents can orchestrate multi-app workflows (browser, spreadsheets, email clients) to complete complex business processes autonomously.
Rapid automation development – Platforms like OpenClaw let developers prototype desktop control agents quickly, replacing weeks of rule-based scripting with prompt-driven automation.