Best of your X follows: June 19

Window checked: Jun 15 18:00 to Jun 16 18:00 UTC. I kept substantive original posts and link recommendations from the tracked public AI/tech accounts. Pure retweets, context-free reactions, and off-topic politics are out.

What happened: Mollick estimated that open models still lag closed models by about 8 to 12 months in coding, which would leave roughly 4 to 8 months to harden IT systems against Mythos-class capability. 1 Why it matters: The useful reading is operational, not philosophical: if the lag is real, defenders need to prepare before comparable open models arrive. Implication: Mollick's conclusion is that publicly available defensive Mythos-class models matter now, because defenders need tools in the same capability band. 1

What happened: Mollick said Fable was a leap after testing it and linked back to his longer Mythos write-up. 2 Why it matters: His point is that exponential gains make each incremental release feel larger, so Anthropic may not be the only lab with sudden step changes. Implication: The post is a warning against planning around smooth progress curves; teams may see model capability arrive in bursts instead. 2

What happened: Chollet argued that powerful open-source AI needs radical gains in inference cost and, more importantly, training-data requirements. 3 Why it matters: That shifts the open-model discussion away from weights alone and toward whether smaller teams can afford to train and run frontier-grade systems. Implication: Chollet says symbolic learning is the route to that efficiency, which is a sharper claim than the usual hardware-scaling answer. 3

What happened: Graham called Bartosz Ciechanowski's mechanical-watch article the best explanation he had seen of how mechanical watches work. 4 Why it matters: The linked piece is not just about watches. It is a reference for how interactive visuals can make a physical system legible without flattening it into a static diagram. Implication: If you write docs for complex systems, the lesson is to show motion and dependency, not just label parts. 4

What happened: Graham wrote that if you want something from someone, you should make the ask explicit. 5 Why it matters: The useful part is the inversion: asking clearly is not the imposition; making the recipient infer what you want is. Implication: This is a practical rule for founders and managers using AI agents too: ambiguity moves work downstream, whether the recipient is a person or a model. 5