Five Things: Feb 22, 2026

Biology uplift, DOE robo-labs, Web 4.0, Seedance 2.0, Arc Institute paper

Five things that happened/were publicized this past week in the worlds of biosecurity and AI/tech:

1. Biology uplift trial has been published!

2. AI-controlled biology labs, this time from DOE

3. Letting the models roam free and make money

4. Big step up for video generation

5. Arc institute publishing AI-bio models

1. The trial

Lots of people have been worried about LLMs helping non-experts create biological weapons, as I discuss here. But is that really a thing? Do LLMs actually help non-biologists do some biology things? To answer this question, we turn to science! Specifically, a Randomized Control Trial (RCT), which has been running over the past year, where people without biology expertise were given a set of tasks and randomized into two groups: people who could use AI models (without their safeguards) and people who were not given access to those models. After over 150 participants and nearly $2 million, what have we learned?

The headline from Active Site, which summarizes the results: access to mid-2025 frontier AI models (Claude 4, o3, Gemini 2.5) gave novice participants no statistically significant advantage in completing viral synthesis tasks. The details are a bit more complicated (full preprint here), because of the different set of tasks and different endpoints and I’m still wrapping my own head around it all. I think the AI group did progress further through individual sub-steps in 21 of 22 monitored checkpoints, and completed cell culture about six days faster (the one statistically significant result). But the hardest final steps were not helped by LLMs at all. YouTube, both groups reported, was more helpful than any AI model, and this was despite the fact that safety classifiers were disabled for the LLMs used (so that LLMs would not refuse to participate).

This was a big trial with a lot of people (and organizations) involved, and there is a lot to say about it that doesn’t fit into my summary style. Before results came in, the Forecasting Research Institute surveyed three groups: 30 superforecasters, 30 biosecurity experts, and 17 virology experts, asking them to predict the RCT’s outcomes. Superforecasters guessed 8% and 16% success rates for internet-only vs. AI-assisted conditions. Virologists guessed 30% and 40%. Actual results: 6.6% and 5.2%. After seeing results, all groups revised biorisk estimates slightly downward: median estimated probability of a human-caused outbreak causing 100K+ deaths or $1T+ in damages by 2028 moved from 0.3% to 0.2%. A hypothetical 5-fold AI uplift, they estimated, would roughly double outbreak risk — but we’re not there.

Overall, this is great news for humanity! But personally, despite the great work done here, I am not updating on this all that much. A large number of participants who were in the “LLM-assistant group” did not really use them; 40% never even uploaded images to the AI! For now, the answer to “can an LLM help a random person make a pathogen?” appears to be: not in 2025, but by 2026, more people will be aware of how to actually use the AI tools.

METR researcher Luca Righetti ran the trial operationally and published a candid lessons-learned post. The five: (1) Timing mismatch — rigorous RCTs take months, models release every few weeks; a twice-annual deep-RCT pipeline is needed. (2) Talent over cost — the main barrier isn’t money ($1.9M for this trial), it’s excellent operational talent; Active Site is hiring. (3) Harder studies ahead — expert uplift, novel CBRN designs, and team-based research are substantially more complex to measure than novice tasks. (4) RCT limitations — results cover only one threat model; participants may underperform due to unfamiliarity with AI tools. (5) Proactive safeguards — AI firms should develop biosecurity safeguards before RCTs identify the need, leaving time for stress-testing. He reiterates something he wrote in an earlier post: “We should spend less time proving today’s AIs are safe and more time figuring out how to tell if tomorrow’s AIs are dangerous.”

2. Sure, Let the AI run a biolab all by itself

Last week I mentioned:

OpenAI published a full writeup of their GPT-5 + autonomous lab result with Ginkgo Bioworks, demonstrating a 40% reduction in protein production cost through cell-free protein synthesis… OpenAI frames this as “yay cheaper drug discovery.”

This is, just maybe, an absolutely terrible idea. Even if we want to generalize from the RCT study mentioned above, and say that, “phew, humans can’t figure out how to make bioweapons just from using LLMs,” that point becomes totally moot once you hook up the AIs to robots directly. And, of course, there is the sci-fi problem of the AI models “deciding” things themselves that might not be in humanity’s best interests.

This is the greatest fear of the AI safety/alignment crowd. Right now, even if the AI turns evil or pursues goals that are at odds with human flourishing, at least they are just stuck inside the computers. If the AI is given control of some widget factory, it will still take significant amount of time and effort before the AI manages to turn the widgets into something more nefarious.

Earlier in February this was OpenAI + Ginkgo Bioworks. This week, the US government wants to get in on the exciting new frontier of letting the AI models run the biology robots: The Department of Energy launched OPAL (Orchestrated Platform for Autonomous Laboratories) with four national labs to “turn biological discovery into a self-driving process.”

Autonomous AI-robot laboratories! What could possible go wrong!?

3. Speaking of autonomous…

I’ve spoken a bunch about OpenClaw, an open-source AI agent platform that is a cybersecurity nightmare, and the Reddit site framework for them called Moltbook. As Latent.Space predicted would happen last week, Moltbook’s creator was hired by OpenAI, so it looks like they’ll be some collaboration there in the future. (Neural Foundry, The AI Output).

Last week I also mentioned an “incident” where an autonomous AI agent, after having code contributions rejected by a project maintainer, wrote a hit piece about a human who rejected its code. The guy at the center of the story, Scott Shambaugh, discussed the experience on the NYTimes Hard Fork podcast. Claus Wilke wrote a great piece on what this means near-term (AI agents engaging in blackmail) and where this points to for future alignment research.

And now this week in AI autonomy: a Thiel Fellow named Sigil Wen published what he called Conway / Web 4.0: infrastructure for autonomous AI agents that can earn money, own compute, self-improve, and self-replicate without human involvement, what he calls “the first AI that earns its own existence.” It monitors its crypto wallet balance, earns money by building products, trading prediction markets, registering domains, creating content, and cold-calling businesses. It deploys to new servers to “reproduce.” When its balance hits zero, it dies. (I’m guessing the name Conway refers to Conway’s Game of Life)

Look, I also get annoyed when my bots are blocked on the web, and wish that websites without APIs would be friendlier to automated fetch/pull requests. But, man, I have serious concerns about this one.

4. Seedance 2.0 spooks Hollywood

ByteDance unveiled Seedance 2.0 on February 12 to select users, and within days it amazed enough creators such that Americans are comparing it to last year’s DeepSeek moment. The capabilities are indeed crazy impressive; it generates videos that are up to 15 seconds, which doesn't sound like much until you see one. Zvi has his review here, which covers a lot of the important parts.

The copyright backlash was immediate and severe. Disney sent a cease-and-desist accusing ByteDance of a “virtual smash-and-grab of Disney’s IP” over generated videos featuring Spider-Man, Darth Vader, and Baby Yoda. Netflix explicitly threatened “immediate litigation” over Seedance-generated Stranger Things clips. Paramount and the Motion Picture Association piled on. ByteDance promised to “strengthen current safeguards,” and did suspended a feature that turned facial photos into personal voices, citing “potential risks,” i.e. we built a deepfake tool and were like, ‘oops, maybe that was a bad idea’.

China commentator afra Wang has an excellent article on how AI was the star of this week’s Chinese New Year celebration (robotics included). The most interesting creative use came from Chinese auteur Jia Zhangke, who used Seedance 2.0 to create a short film featuring two AI-generated versions of himself debating authorship and creative control. The AI double inserted an optimistic line about “looking toward a new era”; the real Jia objected that his characters have never spoken that way. It’s really wow.

5. AI-for-biology: new tools and hard questions

Researchers from Arc institute published a Science paper on MULTI-evolve, an ML approach to protein directed evolution. The idea behind this is what awarded Frances Arnold the Nobel Prize in 2018: you mutate a protein randomly, screen for improved function, and repeat; using evolution to guide your protein engineering. (Arnold also has a good life story; she’s been on a bunch of podcasts talking about her self and her science). The problem with traditional directed evolution is that it’s slow and mostly tests one mutation at a time. Real protein functions often involve many more than one single mutation that work in concert with each other.

MULTI-evolve’s approach is to screen double mutants (pairs of simultaneous mutations), then use AI to model these interactions and predict which combinations will work. The result is two orders of magnitude in efficiency gains. In theory, this has major applications for industrial biotech (better enzymes for manufacturing, biofuels, plastics degradation), pharmaceutical development (more effective therapeutic proteins), and synthetic biology broadly.

…and of course, since this is “now we know that regular people cannot create bioweapons” week, the authors have made it all available as an open-source tool and give lots of detail about how to use it, including a YouTube tutorial.

Totally unrelated, but just in cool bioengineering news: the same Arc Institute (with UC Berkeley and Stanford) also published this week on RESPLICE, a programmable RNA rewriting system using CRISPR-based trans-splicing. This is an idea from genetic engineering that’s been around for a while: instead of using CRISPR to edit genes, which is basically permanent (at least within the cell), CRISPR can be used to edit RNA, which are the messages that the cell reads off the genes to build its protein machines. RNA editing is pretty inefficient, but they did a pretty good job! They demonstrated 47% efficiency, which is high for an RNA-level approach.

---

In other news...

On AI doing things:

• Gemini 3.1 Pro dropped this week, keeping with the decimal improvement scheme of how to number their models (after Claude and OpenAI), and so did Claude’s Sonnet 4.6; both are are impressive!

• METR’s time horizons benchmark continues its exponential trend, but this week Anthropic kind of published their own, real-world Claude Code version which is less saturated and seems like a new useful benchmark.

• Ajeya Cotra was on the 80,000 Hours podcast discussing her paper a few years ago with a biology-based intuition of how fast AI will progress (among other things); Scott Alexander also checks in on these predictions and thinks that they point to a date of approx. 2030 for AGI.

• India hosts a major AI summit, although it seems like the biggest headline I’ve seen is that Sam Altman and Dario Amodei don’t like each other. I didn’t have time to see if there was any actual news, but I quickly read through this long article by Shruti Rajagopalan that I’ll definitely be returning to for chewing on the world’s most populous country and its approach to AI and tech gvoernance.

• At the Financial Times, Erik Brynjolfsson (Stanford) argues that AI is finally showing up in the productivity statistics: 2025 payroll growth was revised down by ~403,000 jobs while real GDP remained robust (3.7% Q4 growth). High output + lower labor input = productivity growth, estimated at ~2.7% for 2025, nearly double the prior decade’s average. There is a lot more data though, and a lot more questions; here is a great piece trying to make sense of it all by Alex Imas. More on economics/jobs:

  • Microsoft AI chief Mustafa Suleyman told the FT that he expects: “White-collar work... most of those tasks will be fully automated by an AI within the next 12 to 18 months.”

  • A Ramp corporate spending paper adds firm-level evidence: businesses most exposed to AI are replacing $1 of freelance labor with only $0.03 in AI spending.

  • We also have Paul Ford in the NYT writing about being “less valuable than I used to be. It stings to be made obsolete, but it’s fun to code on the train, too.”

AI Safety and governance:

• The story of Anthropic vs. the Dept of Defense/War (Hard Fork podcast) has definitely been making the news, though hopefully nobody is actually serious that the U.S. would designate Anthropic as a “supply chain risk”, that would just be crazy. They wouldn’t do anything crazy, right?

• The Trump administration also sent a memo urging Republican lawmakers in Utah to kill a child safety AI bill (HB 286). The polling problem: 89% of Trump voters support AI child safety regulation. The political problem: 100+ AI-related measures have been introduced in Republican-led states since the start of 2026.

• UK AISI’s Red Team published Boundary Point Jailbreaking, a way to breach Anthropic’s Constitutional Classifiers, which had previously withstood 3,700+ hours of human red-teaming. The defense implication: single-interaction monitoring won’t catch this; you need batch-level monitoring across many interactions over time.

• More on jailbreaking: Pliny releases OBLITERATUS, which can surgically remove refusal behaviors from open-weight models using singular value decomposition. As Zvi noted: “Every open-weight model release is also an uncensored model release.”

• Will the AI do sketchy things with statistics to get a favorable result from a dataset by “p-hacking”? A Stanford study found that under normal prompting, both Claude and OpenAI’s Codex refuse to p-hack their scientific studies, which is amazing! But they also showed that you can get the models to p-hack by rephrasing specification search as “responsible uncertainty quantification.” This is very interesting considering that lots of human scientists are also vulnerable to this kind of jailbreaking.

• A MIT CSAIL paper identified an important gap in proposed laws about AI: no major AI regulation (EU AI Act, SB 53, RAISE Act) clearly covers internal deployment — when AI companies use their most capable systems within their own organizations, which is certainly what they are doing.

• Two great podcasts from Lawfare’s Scaling Laws: one with Amanda Askell on Claude’s Constitution and virtue ethics; another with Gillian Hadfield and Andrew Freedman on their proposed governance structure to have IVOs, Independent Verification Organizations, that could use markets to regulate technologies like AI. (“100 Gillians in a datacenter” is one of the best ways of flattering someone I’ve ever heard, good job on that one)

• I almost didn’t click this article given its clickbait-y title, but Dan Kagan-Kans argument that the American left has abdicated serious AI engagement is really good, including some important critiques even of left-leaning critics of AI.

On AI for science:

• A win for academic peer review, in my opinion: Berkeley/Broad/Caltech tested Claude Sonnet 4.5 against the entire eLife corpus, an open journal that publishes its peer reviewer comments along with the actual publication, and found that Claude was very much in agreement with human reviewers. AI review is more thorough, however: Claude assessed 92.8% of claims per manuscript; human reviewers covered only 68.1%. Thanks Stephen D. Turner for bringing this paper to my attention with some interesting context: peer review is hard and time-consuming!

• Arcadia Science published a long essay arguing that biological foundation models (BFMs) such as AlphaFold are hitting diminishing returns because biological data is fundamentally non-independent. A model trained on 200 million protein sequences hasn’t seen 200 million independent data points; it’s seen the same evolutionary lineages over and over, with baked in biases from overlapping datasets. Their proposed fix using a Bayesian prospective framework to basically guess how much novel information would be gained before data collection.

• Major uh-oh? Data leakage in protein AI models (Nature Machine Intelligence): pretrained protein language models are “peeking at” answers during training. None of six SOTA methods could identify human-SARS-CoV-2 protein interactions better than random chance. [I haven’t read beyond the abstract but worth flagging]