Researcher David Bellamy: Claims that AI can create deadly viruses are unfounded
Harvard PhD David Bellamy, currently working on machine learning systems in a large model team in the UAE, stated that he is among the few who have done two things: trained cutting-edge large language models and personally designed and synthesized custom viruses in the lab. He believes the claim that "artificial intelligence will create dangerous viruses that kill everyone" is completely unfounded.
He then addressed the bottlenecks: models can propose genomes, sequences can be synthesized, and they can even be infectious, but that is often just replication or fine-tuning of known viruses. The real bottleneck lies in physical synthesis, equipment, and materials, not in writing a segment of bases. Designing a virus that can evade all preventive measures requires repeated iterations with the physical world, not a genius in a data center writing it all at once.
He pushed the fear to the extreme: a fully automated laboratory dedicated to the most dangerous virus families, remotely controllable via interfaces, costs far more than $100 million, and there is no laboratory in the world that can synthesize all conceivable viruses. Even assuming it exists, one would still need to place orders with synthesis companies or have their own synthesis machines; the former involves sequence screening, while the latter requires moving the entire supply chain indoors. He stated that he has participated in the construction of fully automated interface laboratories, which do not currently exist.
Wet lab speeds are limited by physical constraints: cell cultures and mice will not grow faster because the model is smarter; basic virology experiments often take several days, some over a week. Recursive self-improvement mainly runs on programmable, immediately verifiable digital benchmarks and cannot accelerate procurement cycles, firmware interfaces, and animal testing times. Effective results in mice do not equate to effectiveness in humans, which is the same gap seen in drug development.
The background is that Stanford and the Arc Institute used the genomic model Evo to produce sixteen types of bacteriophages that can infect E. coli in petri dishes, with training data excluding pathogens that infect humans, animals, and plants. At the same time, virology capability tests showed that some cutting-edge models surpassed most human experts in experimental obstacle selection questions. The policy circle is discussing the bioweapons path of cloud laboratories and jailbreak models based on this. His conclusion is that artificial intelligence can at most speed up the retrieval of instruments and procedures, without changing procurement, regulation, costs, and the knowledge gap regarding humans.
In market mechanisms, the buyers are model companies that require regulation and safety premiums, while the sellers are those who narrate milestones in petri dishes as risks to civilization. This is event-driven hedging: after the bacteriophage paper and warnings of extinction from departing researchers, those from wet lab backgrounds are coming out to shift the bottleneck from software back to hardware. Funding continues to flow into large model training and automated laboratories; the beneficiaries are biosecurity organizations emphasizing synthesis screening and equipment control, while the pressured parties are laboratories whose PR narratives claim "data centers can destroy the world" as reasons for deceleration.
Source: Public information
ABAB AI Insight
Bellamy's background is caught between two rarely intersecting production lines: cutting-edge pre-training and hands-on virus synthesis. Companies like Lila aim to send scientific superintelligence into automated experimental factories, while the UAE is training its own large models. He uses his dual on-site experience to challenge the same set of Platonic fears—recursive improvements in the code world are directly mapped to extinction speeds in wet laboratories. This contrasts with Amodei's policy bifurcation and Coxon's Slack endgame framing, forming opposing testimonies in the same week.
As a result, capital paths split into two budgets. The safety narrative requires evaluators, licenses, and deceleration; the laboratory narrative demands freezers, incubators, old instruments without interfaces, and equipment with a minimum one-year subscription. Automated laboratories costing over one million dollars are themselves selection devices: those who can afford them, fit them together, and pass material controls are not "prompt players". Writing risks into model outputs is akin to pointing regulatory guns at chat boxes while leaving synthesis companies and firmware in the shadows.
The analogy is not the Manhattan Project, but drug discovery: computation can propose molecules, but approval still gets stuck at animal and human testing. Evo's creation of bacteriophages proves that this extremely small genome of bacterial viruses can be written, not that human transmissible pathogens can be mass-produced. High VCT scores demonstrate that models can answer experimental selection questions, not that they can extrapolate mouse results to human populations. The industry stage is one where capability demonstrations outpace governance, and governance outpaces honest descriptions of physical bottlenecks.
Structural judgments suggest that technological replacements are said to be complete, while in reality, they are stuck in the physical segments of the supply chain. The mechanism is: digital benchmarks can iterate overnight, while cell cycles proceed daily. Whoever assigns the speed of the former to the latter is selling deceleration rights; whoever writes the friction of the latter into the former is dismantling the pricing of this business.
ABAB News · Cognitive Laws
- Writing genomes does not equate to being able to buy incubators.
- The digital world iterates overnight, while cells still grow daily.
- The real bottleneck is often not in the prompts, but in the purchase orders.