Post by John Salter

There has been a lot of talk about Fable 5 (and Mythos for that matter) for the past couple of months. Will all applications be compromised? Is software engineering dead? Well, I'm happy to report that I've been using Fable to test out one of the hardest problems: can it make my Magic the Gathering Commander deck suck less? I've been working on a Splinter, Radical Rat ninjutsu deck and have felt that, while fun, it's been too slow. I don't build a critical mass of advantage before either petering out or being overrun. Now, I had tried to have previous models review decks before with mixed success, even Opus 4.8. They've been very good at identifying a deck's mana curve and ok at suggesting cards. But they had struggled to understand more nuanced interactions, and sometimes I'd even find models hallucinating. I'll give an example. Opus 4.8 correctly identified that Deep-Cavern Bat would be a good fit for the deck. Heck, it even told me it was one of the highest-priority adds to the deck. Deep-Cavern Bat has an enter the battlefield trigger that lets you exile a nonland card from your opponent's hand while Deep-Cavern Bat remains on the battlefield. Opus 4.8 incorrectly told me that by bouncing my bat I effectively had a repeatable Thoughtseize. I only wish it were that good. Now, I ran Fable 5 through the same deck and had it make its own recommendations. Deep-Cavern Bat was still considered, but much more tempered. It correctly identified that the value in bouncing the bat was to continuously peek at my opponent's hand and lock down the next thing on their curve. While I was thrilled to see the improvement, what impressed me most was Fable's evaluation criteria. It correctly understood the deck (and its interactions and how it wins) and built out a framework to evaluate all cards. Every card could be measured against: Delivery - my ninja enablers Payload - my ninjas (mostly) Threshold builders - critical mass builders Multipliers - ways to double or triple up on triggers Protection & mana - can I save my critical pieces and get to them faster And then it offered advice: I could build more aggressively because many of my threats cycle between my board and hand, naturally making the deck more resilient to wipes. Now, deck building has always been my favorite part of Magic, so I don't expect I'll stop poring over cards looking for clever interactions. But it is exciting to be at the point where LLMs (I think) are finally giving me good Magic advice. Fun fact: over a decade ago I trained my first ML model, a Magic the Gathering rarity classifier. The space has come a long way since then, but it's exciting to bring multiple interests together.