E205｜Talking Gene Editing with 丛乐: How Does Carbon-Based Life Face the Silicon-Based Challenge?
Summary
The 2013 CRISPR-Cas9 breakthrough, achieved alongside other teams, was the first demonstration that DNA could be engineered in living human cells, while also showing early programmability, scalability, and safety. Compared with zinc-finger editing, where targeting one site took months and could ultimately cost $50K-$100K, a guide RNA cost about $10 and could be synthesized in days; the success rate also rose from “1 in 5 to 10” to nearly “1 in 2.” 丛乐 described it as “going suddenly from walking to having a horse-drawn carriage, and even suddenly getting trains and airplanes.”
The main bottlenecks in gene-editing applications have shifted from whether DNA can be cut to delivery, off-target control, and cross-disciplinary know-how. Cas9 can locate a sequence like GPS, but the editing system must reach the right cells without ending up in the wrong organ; the liver has seen progress, while the brain remains difficult to access. 丛乐’s DoorDash analogy is the most direct: “I can make something delicious, but without DoorDash, I still can’t get it to your home because you can’t leave the house today.”
Clinical applications are starting with rare diseases, where the risk-reward ratio is clearest, rather than heading straight toward mass enhancement. 丛乐 cited the personalized treatment of Baby KJ discussed in May and Eli Lilly’s acquisition of a cardiovascular gene-editing company that has already shown preliminary effects in humans. But he acknowledged that off-target effects and toxicity have not been reduced to zero, and edge cases resembling autonomous-driving “corner cases” may still exist. For late-stage cancer, one might “treat a dying horse as a live horse”; for an Alzheimer’s risk that may never materialize, the cost of gene editing may not be worth bearing.
Enhancement-oriented gene editing “may be feasible, and there may be people willing to do it,” but measurability does not mean precise editability. A small number of genes govern traits such as eye color and alcohol metabolism relatively clearly, while signals involving the whole-body immune system, intelligence, or complex behavior remain far from mature; 丛乐 says some genetic-testing companies’ claims about coffee sensitivity and IQ are “either just a gimmick or simply inaccurate.” Greater acceptance among younger people may indicate rising social acceptance, but it will not eliminate delivery, ethical, or heritable risks.
Cas9 remains dominant because Cas12 and Cas13 are mostly marginal improvements, while the next major breakthrough may come from applications Cas9 is currently incapable of handling. 丛乐’s team is studying an endogenous protein-RNA system in human mitochondria, hoping to bring a similar paradigm to neural and brain editing, and calls delaying aging “our hope.” His view is that the value of a new platform lies not in changing the Cas number, but in opening organs and diseases that were previously “completely impossible.”
CRISPR-GPT aims to distill roughly 11 years of domain-expert know-how into an AI Agent that feels like having “a 张锋 standing beside you,” but the team still does not know its true accuracy. Built on Genome-Bench, CRISPR Llama, and a complete workflow, the system uses expert-data reinforcement learning, multi-agent discussion, Critic review, and real wet-lab feedback to reduce hallucinations. Hundreds of people apply for Beta accounts every day, and new students are already learning directly from it. The basic version and benchmark are open source and free to apply for; its current value is mainly knowledge compression and automation, while reaching true L4 “innovator” status will require broad use and iterative feedback.
From an investment and organizational perspective, scaling a life-sciences platform requires a flywheel of technology, industry operations, and capital formation, rather than asking star scientists to handle every role. 丛乐 observes that 张锋 and George Church seek complementary talent, such as CEOs with successful pharmaceutical track records, and actively work with capital because a heavily regulated, long-cycle drug industry requires specialized experience. The longer-term question is that “silicon-based” intelligence is evolving rapidly while the “carbon-based” side is barely acting. 丛乐’s conclusion: “Many people will always have the idea; the greatest value comes from actually building it.”
Deep dive
1. The 2013 breakthrough made living human cells programmable for the first time
丛乐 summarizes the first significance of the Science paper as the first demonstration, alongside other teams, that CRISPR-Cas9 could modify DNA in living human cells; before that, the system existed only in microorganisms.
The second breakthrough was that a reprogrammable guide RNA could find almost any sequence, turning gene editing from specialized engineering into a general-purpose tool. His analogy was “going suddenly from walking to having a horse-drawn carriage,” followed by trains and airplanes.
The third was the ability to assess cell states before and after editing, with early experiments showing that the cells remained normal. This established preliminary safety and efficacy in living cells; it did not mean that all later human applications were risk-free.
2. The shift from designing games to designing life came from realizing that “reading” was not enough
丛乐 grew up in Zhongguancun and initially entered Tsinghua’s electronics department, hoping to become a game designer. Family health issues and a chance encounter with books on biomedicine led him toward diabetes, genetic disease, Alzheimer’s, and other problems that still lack basic answers.
After meeting 张锋 in George Church’s lab, he adopted a framework that determined his career path: “Sequencing is just reading a book; it isn’t writing one.” Only by creating or modifying a system can one truly understand it.
3. Zinc-finger proteins proved the concept but were trapped by their cost structure
Before CRISPR, protein-based tools such as Zinc-finger and TALE could locate DNA like GPS. The problem was that a single protein could cost more than $10K, while a PhD student at the time earned about $30K a year.
More damagingly, after months of development, researchers often had to design 3, 5, or even 10 proteins to get one successful result. Including failures, editing a single site could cost $50K-$100K, making standardization and scaling difficult.
CRISPR changed the targeting logic to A-T and G-C pairing: a guide RNA cost about $10 and could be synthesized in less than a week. In early experiments, nearly every gene had close to a 50% chance of success. Cost, speed, and reliability combined to produce an improvement of several orders of magnitude.
4. A 2009 email triggered a PhD bet with a safety net
After hearing a microbiology presentation, 张锋 sent 丛乐 a very short email asking him to look at the CRISPR papers. The two later discussed it late into the night in the office and concluded that the system was simple and elegant enough to justify investing an entire PhD in it.
丛乐 admits that the team had a “scientist’s mysterious confidence” that “maybe only we could have thought of this.” They later found that multiple teams had independently seen the opportunity; a good idea does not appear only once.
The bet was not unlimited risk. By 2010, his second year of the PhD, he already had a co-first-author paper in Nature Biotechnology and another promising epigenetics project—a first pot of gold before committing the next 3 to 4 years.
5. The 2012 race showed that perfectionism in research has an opportunity cost
The team had initially treated CRISPR as a project to refine over many years. When the Doudna team demonstrated in Science in June 2012 that Cas9 could be programmed with guide RNA in vitro, they realized that applying it to human cells was “only a matter of time.”
Looking back, 丛乐 says the paper had been publishable much earlier. The final version included gene disruption, precise repair, and editing at multiple targets—in effect, it did “1, 2, and 3” all at once instead of publishing the first step immediately.
When 泓君 brought up missing out on the Nobel Prize, 丛乐 rejected the simple causal link: award criteria are subjective, and the basic mechanism and human-cell application can receive different weights. But he clearly acknowledged that publishing earlier “at least would have meant much less patent controversy.”
6. Once Cas9 became a platform, applications expanded across tools and diseases
丛乐 compares the foundational breakthrough to the iPhone platform. Horizontally, it expanded from DNA into RNA, epigenetics, and base editing; vertically, it moved deeper into the liver, inherited eye diseases, and immune-oncology target discovery.
The fact that George Church’s team published on the same day was not a coincidence. Journals wait for related papers to complete peer review before organizing them into a special simultaneous release; it also reflected how multiple teams were independently pursuing the same frontier.
His technical judgment is that once a platform appears, researchers must both expand the tool’s boundaries and work with disease specialists. Cas9’s influence comes from supporting both expansion paths at once.
7. The patent war began because CRISPR landed at the US system’s transition point
The US shifted at some point in 2013 from first to event—who invented first—to first to file. CRISPR emerged under the old system, so the dispute could not be settled simply by comparing filing dates; it required tracing lab records, emails, and the actual timing of invention.
Multiple independent teams in different locations—and some outside the US—were advancing the work simultaneously. The patent dispute therefore reached the patent office, arbitration, the Federal Circuit, and even the Supreme Court.
丛乐 estimates that the parties’ cumulative legal fees may already exceed $100M. The reason the fight continues is that the technology’s future applications also “look like they’re worth that much money.”
8. Gene editing expanded from medicine to the environment, eventually reaching the mind and cognition
The first circle is medicine: treating genetic and acquired diseases, then moving into prevention and enhancement. This is the area that acts most directly on the individual body.
The second circle is agriculture and the environment. 丛乐 cited gene-edited potatoes with longer shelf lives and improved starch characteristics, fluorescent plants, and modified mosquitoes designed to reduce malaria transmission, along with industrial production through synthetic biology.
The third circle remains close to science fiction: regulating the brain’s understanding, thought, and cognition. It would not merely make the body feel better; it could change human mental capacity.
9. Brain-computer interfaces are advancing the “machine,” while the carbon-based brain remains almost untouched
丛乐 asks: Neuralink already has 9 people with implants, so why does humanity do almost no work on the evolution and modification of its own brain? “The difference in the speed of silicon-based and carbon-based evolution is so large,” yet investment on the two sides is profoundly asymmetric.
When 泓君 asked whether experimentation had to begin with patients, 丛乐 explicitly disagreed. It could start with neurons in vitro, then move to animals and humans; technologies such as Neuralink also went through testing in animals and monkeys first. 泓君 compared it to “at least starting by writing a joke or posting a small tweet,” suggesting that understanding can also begin through practice.
10. Clinical validation has begun, but the entry point remains rare disease
丛乐 cited a case from May this year: Baby KJ, an infant under 1 year old, received the first personalized CRISPR therapy for a rare inherited liver disease. The project was carried out by Children’s Hospital of Philadelphia with support from the University of California and the Doudna team, using technology developed by David Liu of the Broad Institute.
He also mentioned Eli Lilly’s acquisition of a company using CRISPR to treat cardiovascular disease. The therapy has not yet received clinical approval, but it has shown preliminary effects in humans and is adjacent to Lilly’s focus on metabolic disease and weight loss.
Current cases remain concentrated in beta thalassemia, rare liver diseases, and rare cardiovascular diseases. Computational, machine-learning, and AI methods can predict and control off-target effects, but they cannot eliminate autonomous-driving-style “corner cases” with 100% certainty.
11. A few traits are explainable; complex abilities are nowhere near as clear as sequencing ads suggest
Traits influenced by a small number of genes, such as eye color and alcohol metabolism, are relatively well understood. 丛乐 uses aldehyde dehydrogenase as an example: low activity slows acetaldehyde metabolism and causes facial flushing after drinking; in theory, editing could change alcohol tolerance.
But detecting a signal does not mean the prediction is reliable, much less that the trait is easy to edit. Liver-related traits may be concentrated in specific cells, while immune states such as allergies involve many cells throughout the body.
丛乐 uses his own experience to question consumer genetic testing: a report said he was not sensitive to coffee, while in reality he is highly sensitive. As for claims about IQ, he says they are “either just a gimmick or simply inaccurate.”
12. The obstacle to DIY is not reagent prices but purity, delivery, and the entire workflow
Educational Kits are available for purchase, but that does not make them suitable for human use. Reagents must meet exceptionally high cleanliness and safety standards. 丛乐’s analogy: “If you picked up a hot dog at random off the street, I wouldn’t dare eat it either.”
Delivery remains the central challenge. Cas proteins and RNA must enter the target cells without ending up in the wrong organ. “I can make something delicious, but without DoorDash, I still can’t get it to your home because you can’t leave the house today.”
System optimization must handle editing efficiency, off-target effects, expression, and reagent design simultaneously. 丛乐 believes AI is most likely to lower this layer of specialized-knowledge barriers first, rather than somehow solving the laboratory environment and organ-delivery problems on its own.
13. Cas9 remains the primary platform; a true new paradigm must open scenarios it cannot currently handle
丛乐 distinguishes paradigm innovation from incremental innovation. The move from protein tools to RNA-programmed Cas9 was a foundational leap, while Cas12 and Cas13 are more “marginal improvement,” so Cas9 remains the broadest application base.
New tools have value in specific scenarios, but merely changing the Cas number does not constitute the next revolution. The real breakthrough should come in areas where existing CRISPR is completely incapable, such as efficient neural and brain editing.
The team has received funding to study an endogenous protein-RNA system in human mitochondria, hoping to apply similar logic to the nervous system. As for whether it can delay aging, he calls it only “our hope,” without presenting the early project as a confirmed therapy.
14. CRISPR-GPT packages domain-expert know-how into an executable experimental agent
CRISPR-GPT was developed by 丛乐’s team in collaboration with 王梦迪 of Princeton and Google DeepMind. It is more than a chat model: the stack includes Genome-Bench, a fine-tuned CRISPR Llama, and an AI Agent that can plan, design, execute, and analyze an end-to-end workflow.
The team integrated roughly 11 years of knowledge from experts across the field who were willing to share data, including 张锋, Patrick Hsu, and Samantha Konermann. It then distilled that knowledge into the model through Fine-tuning and Reinforcement Learning.
丛乐 compares the product to Cursor for gene editing. When users design experiments, RNA, or proteins, “it’s as if 张锋 were standing beside you,” ready to explain steps, call tools, and help debug.
15. Reducing hallucinations requires a three-layer loop, but real-world accuracy remains an open question
The first layer is expert human data. After holding out part of Genome-Bench for evaluation, the domain model was less likely than a general model to list A, B, and C without making a judgment, and its answers were more “opinionated.”
The second layer is multi-agent discussion and Critic review. Different underlying models discuss the problem first, after which a review component performs grounding and synthesis to reduce hallucinations.
The third layer is wet-lab feedback: humans actually edit genes and modify cells, then feed the results back into the design. 丛乐 remains cautious, saying he “wouldn’t dare say it’s 100%.” When asked about accuracy, he answered directly: “We don’t know yet.”
Hundreds of people already apply for Beta accounts every day, and new students are being assigned to learn with CRISPR-GPT. The service is free, and the basic version and benchmark are open source. Most capabilities still amount to L3-style compression and automation; reaching L4 “innovator” status will require more user feedback and community iteration.
16. The ApoE case shows the value of the Agent—and why gene editing is not the default answer
Users can enter “turn ApoE4 into ApoE2.” 丛乐 explains that ApoE4 carriers may face an Alzheimer’s risk more than 40% higher, while ApoE2 is protective. The system first generates an editable workflow; experts can add multiple off-target predictions, while beginners can confirm and execute it step by step.
The system compares Cas9, Cas12, Cas13, and base-editing approaches based on the task, using a state machine so that one choice propagates through the rest of the workflow. When designing guide RNA, it calls an external specialist model and returns 3 candidates, predicted efficiency, the likelihood of off-target effects, and the supporting rationale.
泓君 points out that ApoE4 indicates elevated risk, not certain disease. 丛乐 adds that the real mechanism may be driven by neuroimmune cells such as microglia rather than neurons, meaning researchers must first identify which cell type to edit and solve precise delivery.
The eventual therapy may not be CRISPR at all. It could first help identify key sites and mechanisms, followed by treatment with safer, more easily delivered small molecules or short nucleic acids. “Many real solutions need to be considered together.”
17. Risk thresholds, organizational capability, and carbon-silicon co-evolution will determine how far the technology goes
丛乐 argues that risk and reward must be assessed by context. Late-stage cancer patients may accept the risks of gene editing; people who merely carry an Alzheimer’s-associated variant and may never develop the disease may instead choose drugs or exercise to delay it, without rushing into gene editing.
He calls for extreme caution with relatively healthy people and heritable editing because the effects could pass to future generations. But he also rejects blanket opposition, arguing that carbon-based life still needs to understand itself and consider whether it can participate in evolution in a controllable way. Asked whether the He Jiankui event slowed the global field, he said he was uncertain and that the impact may be regional.
From 张锋 and George Church, 丛乐 draws 3 operating principles: “Think more, dare to think”; “dare to make mistakes,” but make them quickly; and maintain objective, critical judgment and personal red lines. Regarding projects backed by Church, he says the direction may be relatively “wild”: a lab of nearly 100 people runs many large projects, making it difficult to separate the wheat from the chaff. He also acknowledges that Church’s tendency to “say yes” can generate controversy.
Commercializing science requires complementary talent and capital. 张锋 looks for people with successful pharmaceutical experience to serve as CEOs; scientists do not need to personally run a heavily regulated, long-cycle industry. 丛乐 ultimately brings the discussion back to the carbon-silicon competition: “Many people will always have the idea; the greatest value comes from actually building it.”