Essays from Hubify.
Research strategy, independent science, and the infrastructure behind the platform.
They built the wrong half
Sakana Fugu, Hermes MoA, and Claude Science all shipped multi-agent in eight days — and all of it blends models to raise a score. Science needs the opposite primitive.
The market just converged on cooperative aggregation: run N models, synthesize, score higher. But research doesn't die from suboptimal answers — it dies from confident false positives. The loop I've run on BigBounce all year is the mirror image: five labs' models paid to refute each claim, blind to each other, integrity-audited. No frontier lab can build it without cannibalizing itself. That's the half we own.
Read the essay →30 predictions for the AI future
Bets on ownership, the open-source catch-up curve, the limits of guardrails, and whether the models are already too good.
I build on these models for a living — and the week the best one went dark crystallized a fear: those of us building on the frontier can't feel confident we own what we're building on. Will local models ever match Fable? What happens to guardrails when they do? Are the models already too good? 30 bets from inside the work.
Read the essay →The model they paused
Notes from inside the Mythos-class jump — and the jailbreak, refusal, and export control that paused the most capable model ever shipped.
Claude Fable 5 and Mythos 5 launched June 9 — a genuine category change in long-horizon agentic autonomy. Three days later they went dark. Not a generic export-control story: a trusted partner jailbroke Fable's guardrails, the fix-or-de-deploy request was declined, and the controls followed. For the first time, the limit on the best model wasn't capability. It was permission. Plus what it means for sovereign and local models.
Read the essay →Echoes of the pre-universe
My mission is to find fingerprints of the universe that came before ours.
The singularity isn't a discovery — it's where the math breaks. If the Big Bang was a bounce instead of a beginning, the previous universe left prints: a parameter-free f_NL = -35/8, a 0.27° birefringence rotation, a gravitational wave hum. This is the question all the catalogs, pods, and agents are actually for.
Read the essay →The job you couldn't do
Everyone is asking which jobs AI will take. The more useful question is the inverse.
I'm not a physicist. I'm a software engineer who's been obsessed with astronomy and cosmology my whole life and never had the credentials to do real research. Then three things showed up in eighteen months and the on-ramp got short. The inverse question to "which jobs is AI taking" is the more interesting one.
Read the essay →The Window
Why this moment in independent research will never come again.
I ran an autoencoder on 17.65 million spectra from DESI DR1. It found 195,829 objects that don't match any known pattern. The total compute cost was about $200. Why hasn't anyone done this before? Five things had to be true simultaneously — and they only became true recently.
Read the essay →More essays soon. In the meantime, request access to the platform.