These birthday animations on glass are incredibly nice to look at during calls on iOS 27.

Read Aloud is still incredulously underdeveloped. To the point where it crashes both desktop and mobile versions at a mere Pause or Rewind function. I understand the difficulty here, however it deserves more attention. Walls of text are NOT optimal. Condensing is extremely risky. I’ll die on the hill of natural audio scrubbing being the only future for human to llm machine interfacing.

to be a mathematician + compNeuro + artists in 2026 is literally the best timeline i could imagine living.

Beautifully formatted JSON alone doesn’t cure Cancer.

The Mechanism of Surrogation Unjustified promotion occurs at the handoff. An observation becomes an interpretation, which becomes a recommendation, which becomes a discovery downstream. The Chief Scientist cannot erase a biological qualification simply to make an executive synthesis shorter. Sorry, I'm not convinced.

hopflower—more like a bowtie you should see this thing spin

How a computational geometer uses topology to ask: "What spatial structure drives this complex system, and what emerges when I deform it to infinity within its physical laws?"

Computer History is my favorite fvcking feature of @OpenAI's Codex. Genuinely lovely!

open source ai *will* win

this—but steerable

Verification asks whether the answer is adequately supported. Admission asks what it is permitted to change. Inheritance asks the question most systems leave implicit: ⏐What now depends on it‽

When we wade into wonder past its weightless unknown, we often weep for wisdom; yet the latent desire hides deeper within our bones. Unravelled, they reveal mere laudations for the audacious topologies we willfully weave into cognition along the surf. Take time to taste time and all of its flavors. τ ∈ Ω

The strongest point here is that AI decouples mathematical practice from the paper that once served as its receipt. That is a real crisis —— but not primarily a crisis of mathematical truth. Historically, a serious paper allowed the community to treat several distinct achievements as one bundle: a valid proof had been produced; someone had discovered the route; the named author understood it; credit could be assigned; and responsibility had somewhere to land. Those were never logically equivalent. The human production process merely kept them correlated. Once AI can perform substantial parts of that process, the paper no longer certifies that all of those achievements belong to the same person. It may still carry a correct proof while leaving discovery, understanding, credit, authority, and responsibility unresolved. This is where the Bitter Lesson actually bites. General methods win when they can exploit scale, search, and reliable feedback. Parts of mathematics offer exactly that environment: candidate proofs can be generated in volume and checked against formal or otherwise decisive constraints. But that feedback closes over proof validity—not human understanding. AI can therefore scale the production of correct mathematics faster than institutions can certify who understood it, who may speak authoritatively for it, and who remains accountable. The unpaid transfer is from “the proof checks” to “the author understands and may stand behind it.” Total opposition tries to preserve the old bundle by excluding the tool. The harder task is to unbundle the receipts: who generated, verified, understood, authorized, received credit for, and remains responsible for the result.