Medium

Micheal Lanham

Pinned

Micheal Lanham

·

May 23

You’re About to Pick the Wrong Help Desk Agent Stack. Here’s How to Not.

A beginner’s guide to choosing an “agentic platform” in 2026, for the developer handed a deadline before the technology.

You’re About to Pick the Wrong Help Desk Agent Stack. Here’s How to Not.

Pinned

Micheal Lanham

·

May 9, 2024

The Art of Reasoning with GPTs

Cover of GPT Agents In Action by Micheal Lanham

The Art of Reasoning with GPTs

Pinned

Micheal Lanham

·

Dec 4, 2021

Painting with Genetic Algorithms

In my next book for Manning publications, Evolutionary Deep Learning, I outline and demonstrate numerous techniques to improve and enhance…

Painting with Genetic Algorithms

Micheal Lanham

·

2h ago

One Reply Beat 500 Million Steps of PPO at Breakout.

Weight-Space Programming, Day 5 of 7: what a written policy beats, where it breaks, and what PPO can do with it.

One Reply Beat 500 Million Steps of PPO at Breakout.

Micheal Lanham

·

1d ago

The Viral Qwen Fix Fits in a LoRA Nobody Trained: Direct Weight Edits vs Fine-Tuning for LLMs

Weight-Space Programming, Day 4 of 7: the 48.1% fix is two patch cables, and our own test of that idea failed as written.

The Viral Qwen Fix Fits in a LoRA Nobody Trained: Direct Weight Edits vs Fine-Tuning for LLMs

Micheal Lanham

·

2d ago

Changing a Refund Limit Took Two Weights. On a Trained Network It Took Three Tries and a Hack.

Weight-Space Programming, Day 3 of 7: a model with no tools can audit, write and patch tiny classifiers, and the dumbest baselines tie two…

Changing a Refund Limit Took Two Weights. On a Trained Network It Took Three Tries and a Hack.

Micheal Lanham

·

3d ago

We Taught an Embedding Model New Words With 384 Numbers Each and No Training.

Weight-Space Programming, Day 2 of 7: writing jargon into a token table works, a glossary won anyway, and here’s when the edit still pays.

We Taught an Embedding Model New Words With 384 Numbers Each and No Training.

Micheal Lanham

·

4d ago

An LLM Rewrote 0.016%

Weight-Space Programming, Day 1 of 7: one headline sold two different abilities, and a five-cent picture named the game.

An LLM Rewrote 0.016%

Micheal Lanham

·

5d ago

The Busy Wednesday Problem: How Confounders Teach Your AI Agent Superstitions

Instinctual Learning, Part 9 of 9: why a tired night’s low mark must be subtracted before your agent learns a single thing about the food.

The Busy Wednesday Problem: How Confounders Teach Your AI Agent Superstitions

Micheal Lanham

·

6d ago

Imagine Before You Act: What Tolman’s Rats Teach AI Agents About World Models

Instinctual Learning, Part 7 of 9: keep a model of the world before the task needs it, then check the plan against it before you act.

Imagine Before You Act: What Tolman’s Rats Teach AI Agents About World Models

Micheal Lanham

Micheal Lanham

Micheal Lanham is a proven software and tech innovator with 20 years of experience developing games, graphics and machine learning AI apps.

Following

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech