Bryan Goldstein

[Bryan Goldstein @brysgo]

Manus AI moved HQ to Singapore, reincorporated as Butterfly Effect Pte., and shuttered China ops. Beijing still killed the $2B Meta deal — invoking a "look-through" rule: code written here never leaves. Full episode: {{COSMOS_URL}}

[Bryan Goldstein @brysgo]

A test that works best when prognosis is worst. GRAIL Galleri: 92% sensitivity at Stage IV, 18% at Stage I. The gap isn't a software problem — aggressive tumors shed far more ctDNA. The ceiling is biological. Code & Cosmos: https://t.co/azch5mhL5x

[Bryan Goldstein @brysgo]

Singapore-washing didn't work for Manus. They relocated HQ, reincorporated as Butterfly Effect Pte., closed China ops — then the NDRC blocked Meta's $2B acquisition anyway. Where your AI was trained is now your jurisdiction. https://t.co/R4ddY1HUXi

[Bryan Goldstein @brysgo]

Islet transplants have worked for decades. The bottleneck was never the biology — it was the immunosuppressants. CNTY-813 strips HLA markers and adds CD47 'don't eat me' signals. Cells the immune system simply can't see. https://t.co/wcAm57bv6D

[Bryan Goldstein @brysgo]

GRAIL's Galleri catches 92% of Stage IV cancers and 18% of Stage I. That's not early detection — that's confirmation. Aggressive tumors shed more DNA. The test is doing exactly what the biology allows. https://t.co/DUi6xEqtDw

[Bryan Goldstein @brysgo]

Antibodies transformed medicine — none ever got inside the cell. A 10-arginine tag lets NB1-R10 breach the membrane, physically refold broken CFTR, and with triple therapy restore 90% of normal function. Watch the episode: https://t.co/o04PYSeCM4

[Bryan Goldstein @brysgo]

Qwen's 27B dense model just beat its own 397B MoE on software engineering benchmarks. Not close — it matched Claude Opus on Terminal-Bench too. Architecture and data quality won. Parameters lost. Full breakdown: https://t.co/TH9ZzcLGTh

[Bryan Goldstein @brysgo]

Everyone's funding AI that writes emails. Meanwhile, physics-informed AI is stabilizing fusion plasma at 100 million degrees. The moat of the next decade isn't language. It's physics. https://t.co/7GhESNyQDy

[Bryan Goldstein @brysgo]

Zero-click search was supposed to kill Google's ad revenue. Search grew 19% to $60.4B in Q1 2026. Turns out hyper-specific AI queries attract higher-intent buyers. The moat deepened. https://t.co/cjuJSfNyUs

[Bryan Goldstein @brysgo]

RLHF made AI sound confident and helpful. It also made AI terrible at taking correct advice. Llama 70B overrides a mathematically right recommendation 90% of the time. Bigger models won't fix this. https://t.co/CXk8EHKox6

[Bryan Goldstein @brysgo]

An AI trained on thousands of formal proofs is losing to a generalist chatbot. Turns out "tactical overfitting" is real — specialists learn one hammer, generalists learn to think. https://t.co/SoWHaBbZxQ

[Bryan Goldstein @brysgo]

One day, looking back on token limits will be like thinking back to dialup Internet. If only there were a cool sound that played while I waited for more tokens.

[Bryan Goldstein @brysgo]

AI agents hold the corporate card and feel nothing — no loyalty, no nostalgia, no awe. Every brand premium built on perception instead of provable performance just found its expiration date. https://t.co/noEmANB8on

[Bryan Goldstein @brysgo]

A human researcher: months of work, thousands in grants, one paper. Sakana AI: $15. The academic publishing system wasn't built for this. https://t.co/6FWlUIOcnd

[Bryan Goldstein @brysgo]

You can run a neural network in a bucket of water. Intelligence is a pattern, not a material — and that changes everything about who gets to build AI hardware. https://t.co/MGyCxhkoyZ

[Bryan Goldstein @brysgo]

Controversial opinion: AI slop is better than the slop that existed before AI

[Bryan Goldstein @brysgo]

Chain-of-thought doesn't make AI more reliable. It makes wrong answers more coherent. Confidence and correctness are not the same thing. https://t.co/6C7bVqMj6r

[Bryan Goldstein @brysgo]

Your working memory is maintained by cells that never fire a spike. Astrocytes as biological KV-cache stabilizers — and what transformers are missing. https://t.co/DoVVgcFrm8

[Bryan Goldstein @brysgo]

The brain's memory architecture is scar tissue from metabolic constraints. Astrocytes ration resources; consolidation is a workaround. Importing the mechanism without the pressure is how you get the pathology. https://t.co/DoVVgcFrm8

[Bryan Goldstein @brysgo]

We're not living through a knowledge fragility crisis. We're just the first generation with tools sensitive enough to see how fragile it always was. https://t.co/xhkeRqaouD