Shou Zi Chew gave AI-driven science a deadline: two to three years for “a real scientific breakthrough” in physics, biotech, chemistry and materials science. The TikTok chief executive said this at the DBS Asia Future Forward Conference in Singapore on Monday, in remarks reported by CNBC. The claim is worth checking against what ByteDance, Chew’s parent company, has actually shipped, and against what the industry’s own flagship biology project says it still needs.
ByteDance’s evidence is Protenix-v2, a biomolecular structure-prediction and design model the company’s Seed research team released on April 8, 2026. The paper reports a model with 464 million parameters, trained on data with a cutoff of September 30, 2021, that predicts protein, DNA, RNA and small-molecule complexes end to end. On antibody-antigen docking, measured by the DockQ metric at the 0.23 threshold, Protenix-v2 posts a 9-to-13 percentage-point absolute gain over its predecessor, Protenix-v1, on three benchmark sets, and matches Protenix-v1’s 1,000-seed result using only 5 seeds, a roughly 200-fold cut in the compute needed per prediction. In antibody design, the paper reports hit rates up to 48% in VHH-Fc campaigns and 16-88% on GPCR targets, depending on the target and a testing budget of 16 to 30 candidates. Those are real, measured numbers, on held-out structure and docking benchmarks. None of them is a breakthrough in physics, chemistry or materials science, and none is a drug: a benchmark win on protein docking is a necessary step toward finding a molecule that works in a human being, not the step itself.
That gap between benchmark and bench is the same one the industry’s biggest current bet on AI biology is trying to close, on a longer and more specific clock than Chew’s. On April 29, 2026, the Chan Zuckerberg Biohub pledged $500 million to its Virtual Biology Initiative, according to the organisation’s own announcement: $400 million for measurement technology and large-scale data generation, $100 million to coordinate outside labs. On October 7, the initiative added the US Department of Energy, the National Institutes of Health, Google DeepMind, Isomorphic Labs and Meta, taking the total to roughly $1.8 billion, as Endpoints News reported. Google DeepMind, Isomorphic Labs and Meta together put in $300 million of new money; the Department of Energy is committing more than $500 million over five years through the Genesis Mission, the AI-for-science executive order President Trump signed in November 2025; the NIH is contributing access to biological datasets and repositories built with more than $500 million in prior federal funding, which Biohub will standardise for AI training.
How $1.8bn in biology money is supposed to become a drug Partner commitments to the Virtual Biology Initiative and what they fund
| From | To | How |
|---|---|---|
| US DOE / Genesis Mission ($500m+ over 5 years, lab measurement & computation) | Virtual Biology Initiative | |
| NIH (access to datasets built with $500m+ in prior federal funding) | Virtual Biology Initiative | |
| Google DeepMind, Isomorphic Labs, Meta ($300m combined) | Virtual Biology Initiative | |
| CZ Biohub ($500m: $400m tech & data, $100m coordination) | Virtual Biology Initiative | |
| Virtual Biology Initiative | Open cell datasets (first large dataset expected ~2027) | |
| Open cell datasets (first large dataset expected ~2027) | Predictive virtual-cell models (Biohub targets ~5 years) | |
| Predictive virtual-cell models (Biohub targets ~5 years) | In-silico drug testing |
Based on Chan Zuckerberg Biohub The chart shows where the money is supposed to go, and the gap it still has to cross. Biohub’s own head of science, Alex Rives, has said existing cell datasets cover hundreds of millions of cells, while an accurate virtual cell needs data from billions and eventually trillions, a scale Biohub expects to reach only through the five-year initiative, with a first large dataset roughly a year out and working predictive models at the five-year mark, per the October announcement. That timeline, set by the people building the thing, runs to 2031. Chew’s timeline, set by the person marketing a platform that competes for the same engineers and compute, runs to 2028 or 2029. One of these groups has published its gap between current and required dataset size. The other has not said what it would take for its forecast to be checkable at all.
Chew’s own caveat undercuts his headline number more than his audience may have noticed. “I think in the short term we are probably overestimating it, but in the long term we are probably underestimating it,” he said, describing a standard pattern in forecasting new technology: near-term capability gets oversold, because demos generalise poorly, while long-run capability gets undersold, because compounding effects are hard to picture. If that pattern holds, a three-year breakthrough claim is exactly the kind of near-term overestimate the second half of his own sentence predicts. Separately, Zuckerberg’s Meta is now funding part of the DOE’s own AI-for-science push while Zuckerberg sits on the advisory structure shaping US AI policy, which does not make the Virtual Biology Initiative’s numbers wrong, but means the same small set of companies is both building the benchmark and grading it.
The cheaper half of the forecast #
Chew’s second claim is more immediately checkable than his science forecast, because it is about cost, not capability. He said AI is making it “significantly cheaper” for attackers to run fraud at scale against platforms like TikTok, without giving a figure for how much cheaper; that is the kind of number a company the size of ByteDance has internally and a reporter should ask for directly. His specific worry is agentic commerce: AI agents given broad purchasing permissions that, in trying to satisfy a task, buy more than the user intended. That is a design flaw in permission scoping, not a hypothetical about rogue AI, and it is the same category of failure as an overprivileged API key, just with a model instead of a script deciding what to call. Whether platforms building agentic checkout, including TikTok’s own shopping features, have shipped spending caps or confirmation steps that address it is the kind of safeguard Chew did not describe.
The number worth watching next is not a conference soundbite but a dataset size. Biohub says it needs the first large-scale release of its virtual-cell data within about a year, putting a checkable milestone on the calendar for roughly October 2027, three years before Biohub’s own five-year model target and about the same distance out as Chew’s three-year breakthrough window. If that first dataset slips, or if it lands far short of the billions of cells Rives says the project needs, Chew’s forecast will have failed on the easier half of the test, before AI models have even reached the harder one: finding a result in a lab that a benchmark score did not already predict.