Insights
Nobody knows yet.
Here’s what we see.
2026
Institutional Memory: AI’s Missing Layer
Every AI vendor wants to remember your organization for you. Anthropic’s leaked source code contains a finished memory feature, and OpenAI and Google have shipped theirs. Why that knowledge belongs in the systems you already own.
The Impact of Adversaries on AI Adoption
Insurance payors and providers are pointing AI at the same claims with opposite goals. Cybersecurity, litigation, procurement and tax audits show the same dynamic. In these markets the other side sets your adoption timeline, whether or not you’re ready.
2025 Year-In-Review
Ilya Sutskever says the models look smarter than their economic impact implies. Two years of mid-market engagements suggest why. Five learnings from 2025, starting with a client tool that cut a task from an hour to ten minutes and produced no hard cost savings.
2025
The Advent of Generative Personalization
Carvana generated 1.3 million unique videos, one for each customer’s car purchase. Orangetheory’s personalized campaign booked 45,000 classes in a week. Meanwhile 75 percent of consumers call some personalization creepy. Written with Arif Mansuri, on where the line sits.
The Commoditization of AI
GPT-4 cost $30 per million input tokens in December 2023. A year later GPT-4o cost $2.50 and Gemini Flash 8B under four cents. DeepSeek is the newest entrant in a price war that moves the value from the model to how it’s applied.
2024
Accelerating software engineering with AI
Fifty engineering leaders at the T3 Summit in New York all used AI coding assistants. One team let AI write only the tests. Another wrote only the tests and let AI write the code. Four questions for deciding where the tools belong on your team.
Mitigating the Risks of AI
Air Canada’s chatbot promised a discount that didn’t exist, the airline refused to honor it, and a $645 refund became a PR debacle. Four categories of risk in deploying AI, and the steps that contain each one.
Selecting the right generative AI model
Hugging Face listed nearly 500,000 models in February 2024, and most are good enough for most applications. Two factors, licensing and size, decide the rest. Pilot with a large proprietary model, then move to the smallest open source one that matches it.
Benchmarks for AI-driven productivity gains
Four studies put AI’s productivity gains between 14 and 55 percent, depending on the task. Treat those as an upper bound. The gains skew toward less experienced workers, and on tasks AI handles badly, consultants using it were 19 percent less likely to be right.
2023
Key drivers of AI advancement
Computing power and data grow on predictable curves. Algorithm breakthroughs don’t: the transformer arrived in 2017, and speech recognition took from 1952 to 2017 to match human accuracy. Why nobody can say when AI will automate video production.
Why it’s the right time to explore AI
The first wave of AI needed research teams only the largest companies could fund. Open source models and better tooling have changed that. We’re founding Eskridge to bring applied AI to mid-market companies that can’t staff an innovation team.