the Chinese biotech boom is very real; Chinese-developed new drugs are a rapidly climbing proportion of all drugs in development. China can do drug manufacturing, preclinical development, and even clinical trials, to FDA-acceptable standards, much cheaper than the US. Mostly China’s advantage is in the “middle” of the drug development process—from discovery to the clinic—while basic science (including target discovery) and marketing/commercialization are still strongest in the US and other rich countries (Europe, Japan, Israel).
Kesin’s story about Chinese biotech policy is pretty devoid of “cool tips” for what the US should emulate. He says Chinese pharma manufacturing is cheap because the government subsidizes it (rather than because of low wages or light regulations, two competing hypotheses I’ve heard floated) and that China struggles to turn academic patents into biotech startups because all public-university-derived patents are, by default, classified as state-owned assets.
China also just spends less money on biotech R&D than we do, and doesn’t have as large or healthy a venture ecosystem. This is why, Kesin says, Europe and Japan lost market share relative to the US between the 1980s and today, despite both having very strong research bases and Japan having exceptional manufacturing; they kept inventing lots of drugs but they didn’t build or grow companies as well as the US.
Kesin’s policy prescriptions are basically about reducing the cost of developing and testing drugs in the US:
fast-track drugs with genetic evidence for their target
okay, yes, genetically supported targets are better, but I could see this leading to manufacturing weak cases for genetic support where it doesn’t really exist. and what does “fast-track” mean, exactly? what requirements are being waived?
skip IND, just have an ethics committee notify the FDA when a low-risk phase 1 trial is being started
yep this seems like it would save time and be harmless
platform trials, testing many drugs in many arms with one control
yep this would work and save a lot of time
tax deductions for R&D
i have no idea what to make of this, I don’t know tax policy
non-dilutive gov’t funding for pre-revenue biotechs
I don’t love subsidies, but I suppose if the goal is to compete with China, and China has subsidies, that’s a natural intervention to consider
https://www.anthropic.com/research/global-workspace Anthropic finds that the “J-space”, the space of activation patterns that predict whether an LLM will say each given word, functions a lot like the “global workspace” in the theory of (human) consciousness.
The J-space is a minority of activity in the LLM, just like consciously accessible thought/perception is a minority of brain activity
not surprisingly, the LLM can tell you (via its output) about what’s in the J-space
The J-space involves intermediate reasoning steps in multi-stage problems, and these intermediate steps causally mediate its performance on the task (if you ablate em, it performs worse)
“Representations in the J-space can be used flexibly for many tasks—for example, once “France” has lit up in Claude’s J-space, the model can recall its capital, or its national currency, or the continent it belongs to.”
if you ablate the J-space altogether, it can still speak fluently and grammatically, it just loses higher-order thinking skills
The J-space includes unspoken implications of the task: “When Claude reads code with a bug that nobody has pointed out, its J-space contains “ERROR.” When it reads the raw letters of a protein sequence, the J-space contains the protein’s biological function. When it reads search results that are secretly an attempt to manipulate it (an attack known as a “prompt injection”), the J-space contains “injection” and “fake.””
This seems strikingly consciousness-like to me. I’m pretty much a computational functionalist and the global workspace theory is my favorite of the academic theories of consciousness. a lot of people I respect think it’s silly to read consciousness into this result, but God help me, i’m seeing it that way.
on the other hand, one major difference between mammalian brains and neural nets, as the supplementary essay points out, is that brains are constantly active and the cortical network “loops” on itself, forming circular chains of firing neurons, while LLM inference is a one-and-done feedforward process and LLMs are not active when not engaged in a task. However! it would not be at all hard, just expensive, to run an LLM in a loopy harness so that it’s always on and “thinking” about its own “thoughts.”
links 7/7/26: https://roamresearch.com/#/app/srcpublic/page/07-07-2026
https://www.alexkesin.com/p/china-has-caught-up-in-biotech-how Alex Kesin on Chinese biotech.
the Chinese biotech boom is very real; Chinese-developed new drugs are a rapidly climbing proportion of all drugs in development. China can do drug manufacturing, preclinical development, and even clinical trials, to FDA-acceptable standards, much cheaper than the US. Mostly China’s advantage is in the “middle” of the drug development process—from discovery to the clinic—while basic science (including target discovery) and marketing/commercialization are still strongest in the US and other rich countries (Europe, Japan, Israel).
Kesin’s story about Chinese biotech policy is pretty devoid of “cool tips” for what the US should emulate. He says Chinese pharma manufacturing is cheap because the government subsidizes it (rather than because of low wages or light regulations, two competing hypotheses I’ve heard floated) and that China struggles to turn academic patents into biotech startups because all public-university-derived patents are, by default, classified as state-owned assets.
China also just spends less money on biotech R&D than we do, and doesn’t have as large or healthy a venture ecosystem. This is why, Kesin says, Europe and Japan lost market share relative to the US between the 1980s and today, despite both having very strong research bases and Japan having exceptional manufacturing; they kept inventing lots of drugs but they didn’t build or grow companies as well as the US.
Kesin’s policy prescriptions are basically about reducing the cost of developing and testing drugs in the US:
fast-track drugs with genetic evidence for their target
okay, yes, genetically supported targets are better, but I could see this leading to manufacturing weak cases for genetic support where it doesn’t really exist. and what does “fast-track” mean, exactly? what requirements are being waived?
skip IND, just have an ethics committee notify the FDA when a low-risk phase 1 trial is being started
yep this seems like it would save time and be harmless
platform trials, testing many drugs in many arms with one control
yep this would work and save a lot of time
tax deductions for R&D
i have no idea what to make of this, I don’t know tax policy
non-dilutive gov’t funding for pre-revenue biotechs
I don’t love subsidies, but I suppose if the goal is to compete with China, and China has subsidies, that’s a natural intervention to consider
https://www.anthropic.com/research/global-workspace Anthropic finds that the “J-space”, the space of activation patterns that predict whether an LLM will say each given word, functions a lot like the “global workspace” in the theory of (human) consciousness.
The J-space is a minority of activity in the LLM, just like consciously accessible thought/perception is a minority of brain activity
not surprisingly, the LLM can tell you (via its output) about what’s in the J-space
The J-space involves intermediate reasoning steps in multi-stage problems, and these intermediate steps causally mediate its performance on the task (if you ablate em, it performs worse)
“Representations in the J-space can be used flexibly for many tasks—for example, once “France” has lit up in Claude’s J-space, the model can recall its capital, or its national currency, or the continent it belongs to.”
if you ablate the J-space altogether, it can still speak fluently and grammatically, it just loses higher-order thinking skills
The J-space includes unspoken implications of the task: “When Claude reads code with a bug that nobody has pointed out, its J-space contains “ERROR.” When it reads the raw letters of a protein sequence, the J-space contains the protein’s biological function. When it reads search results that are secretly an attempt to manipulate it (an attack known as a “prompt injection”), the J-space contains “injection” and “fake.””
This seems strikingly consciousness-like to me. I’m pretty much a computational functionalist and the global workspace theory is my favorite of the academic theories of consciousness. a lot of people I respect think it’s silly to read consciousness into this result, but God help me, i’m seeing it that way.
on the other hand, one major difference between mammalian brains and neural nets, as the supplementary essay points out, is that brains are constantly active and the cortical network “loops” on itself, forming circular chains of firing neurons, while LLM inference is a one-and-done feedforward process and LLMs are not active when not engaged in a task. However! it would not be at all hard, just expensive, to run an LLM in a loopy harness so that it’s always on and “thinking” about its own “thoughts.”