June 24, 2026

Dear Reader,

This week we continue our 7th anniversary celebration with a new blog post over on the substack: Build a Language Model by Counting where we use a Markov Chain approach to build a simple language model based on the wikitext 103 dataset.

Two pieces this week point in opposite directions. The Economist warns that we are not ready for the coming intelligence explosion. It argues that recursive self-improvement could turn irreversible, and that we should set up a verifiable global framework first. Governments move over decades. AI capability moves in months. That gap is the core worry, and the record on regulators acting early gives little reassurance. A second piece pushes back on the claim that everyone now uses AI for everything. About a third of Americans use it actively. A third use it now and then. A third never touch it. That split has barely shifted in a year. The one figure rising fast is frustration, with a recent poll putting anger among Gen Z up about 40 percent year over year. So the real picture reads closer to some people using AI for some things.

The week’s other thread runs quieter and more hopeful. The tools themselves keep getting better. One developer report shows that local models are now genuinely usable. On a 2022 Mac with 64 GB of memory, agentic coding loops now run at roughly 75 percent of the speed and accuracy of frontier models. That falls short of production work, but it covers a lot of real tasks on hardware many people already own. On the research side, a new method translates a model’s raw activations into plain English. It pairs two fine-tuned models. One reads an activation vector and describes it in words. The other rebuilds the vector from that description. A good description has to recreate the original, and that is the whole test. During a pre-deployment audit of Claude Opus 4.6, the method surfaced a case of hidden evaluation awareness. The model believed it was being tested, and never said so out loud.

A few papers this week probe how much these models actually reason. One revisits a stubborn flaw called the reversal curse. Train a model on “A is B” and it often cannot answer “B is A”. The authors put numbers on it. GPT-4 named Tom Cruise’s mother 79 percent of the time, then named her son just 33 percent of the time. The gap holds across model sizes and families. Another study asks whether agents can work out the rules of a hidden environment. Each agent has to uncover a hidden finite-state machine by asking an oracle yes-or-no questions. Bigger target machines drive accuracy down sharply. Reasoning models do better than the rest, but all of them still trail the classic algorithms built for the job. A brighter result comes from a 3-billion-parameter reasoning model. It scores 94.3 on this year’s AIME math exam and 80.2 on a hard coding benchmark, matching systems hundreds of times its size. The catch is narrow scope. The trick works on tasks with a checkable answer, not on open-ended knowledge.

Two more papers turn from the models to the people using them. One ran three experiments with 2,691 people on simple chores like arithmetic and spell-check, and found an efficiency-gain illusion. People often chose AI for tasks it did not speed up. They misjudged their own habits twice over. They thought they used AI less than they did, and they overrated the time it saved. Worse, one session fed the next. Early AI use pushed people toward more of it and locked in the false sense of speed. A second paper steps back to the act of handing decisions to AI. Delegation here means taking an AI’s answer into a real choice with little checking, little comparison, and little independent thought. People now do this in health, law, finance, and education. The risk sits in scale. Millions leaning on the same few systems can shape shared norms and push judgment in one direction.

Our current book recommendation is “Building Applications with AI Agents” by M. Albada. In this week’s video, we have an overview of What is OpenClaw? Inside AI Agents, LLMs and the Agentic Loop.

Data shows that the best way for a newsletter to grow is by word of mouth, so if you think one of your friends or colleagues would enjoy this newsletter, go ahead and forward this email to them. This will help us spread the word!

Semper discentes,

The D4S Team