June 28, 2026
Simplex and Sensibility
Reflecting to optimise
Coder admits his shortcut mindset got exposed — and commenters are absolutely eating it up
TLDR: A programmer revisited a protein-design problem and realized his usual shortcut — reshaping it to fit standard tools — wasn’t the whole story. Commenters seized on that admission, with some demanding better teaching of optimization basics and others cheering the post’s love for “the simplest possible thing.”
A programmer showed up with a refreshingly blunt confession: he’d spent years treating optimization — basically, the art of finding the best answer under rules — like old homework nobody cool wanted to revisit. Then a real-world protein design problem smacked that attitude head-on. His first move was the classic "I’ll just rewrite the problem so normal tools can handle it" trick, only to realize the problem’s shape matters more than his shortcut instincts. And that’s where the crowd perked up.
The strongest reaction? A mini pile-on from the "teach the fundamentals!" camp. One commenter practically climbed onto a digital podium to argue that beginner machine learning classes should stop speedrunning past optimization basics, saying even a little exposure to rule-based methods would help. Translation for non-math people: some readers think modern AI education has become too dependent on prebuilt tools and not enough on understanding what’s happening under the hood. There wasn’t a full-blown flame war here, but there was a clear generational-style tension: are old-school math ideas boring relics, or the exact thing everyone should have learned earlier?
Then came the lighter side of the thread, with readers lovingly latching onto the phrase “the simplest possible thing.” That line got a little standing ovation, like the comments collectively decided the real hero of the post was humble problem-solving instead of flashy AI wizardry. In other words: less galaxy-brain posturing, more admitting you tried the obvious thing first and learned from it.
Key Points
- •The article formulates an optimization problem over a categorical probability vector constrained to be non-negative and normalized.
- •It assumes a non-convex objective function with computable but computationally expensive gradients.
- •The problem is linked to de novo protein binder design, where amino acid distributions form a position-specific scoring matrix and a folding model such as AlphaFold serves as the objective function.
- •A first approach described in the article is to reparameterize the constrained variable using softmax of unconstrained logits and then apply standard gradient-based optimization.
- •The article identifies the feasible set as the probability simplex and explains its geometry using the k=3 case, where the simplex forms a triangle with vertices representing certainty in one category.