Tommy Cleary
@tommy_cleary
Philosopher therefore narcissist/ Tweets are tone deaf/ Twitter bad blog BAD! As radically contradictory as all political moderates. Errors are my own.
Philosophy of science and Undone science I am working on at the moment.
Wuhan wounds may not heal for generations as Hearts broken seek open air yet Greif & deceit together bleed a banal greyness No numb number of words can reach. Shouts or whispered: unanswered questions are not blame! They are a vast hopeless maze: for us all this is far far worse
Dr. Ryan, glad you picked up on this article. I recently decided to make security work part of my public intellectual identity. Your work here and elsewhere on strategy including Ukraine always teaches a great deal. Roger rbrent at fhcrc dot org and now rbrent at rand dot org
“Humanity’s failure against COVID-19 is sobering, because the world is facing a growing number of biological threats. Some of them, such as avian flu, come from nature. But plenty come from scientific advances.” @ForeignAffairs foreignaffairs.com/world/new-biow…
The Scowcroft Institute papers are an amazing body of work. bush.tamu.edu/scowcroft/whit… As is the path you have taken through this dark valley @rogbrent dspace.mit.edu/handle/1721.1/… …towards more reasonable steps to extract curative and prosperous value from emerging epistemic tools…
Hi Roger velazquez.press/p/funf-jahre-n… It is a little easier for @VBruttel to talk now more openly now. @tony_vandongen too? linkedin.com/posts/valentin… This article by @WalkerBragman is an interesting summary of some of the issues of contention here…always best to balance with a…
A contemporary scientific persona requires some form of performative social media presence. Here goes. See this careful paper by Brutell et al. 2023. Authors not personally known to me. https:// biorxiv.org/content/10.110…
My first article on protecting humankind from bad bio and alarming AI came out today in Foreign Affairs. Two smart coauthors, McElvey and Matheny. Here is a link that's supposed to let people who click on it one time bypass the paywall for 30 days. foreignaffairs.com/guest-pass/red…
Among his many accomplishments, Marks, along with Dr. Robert Kadlec, was responsible for Operation Warp Speed (OWS), the successful COVID-19 vaccine project. The OWS vaccines saved millions of American lives, tens of millions worldwide. For this work, I regard Marks as a hero.
It remains likely that this still mysterious weakly heritable variation will be important for decisions made in development and oncogenesis. By contrast, the 2007 push to use mid-20th century information theory to measure information signaling systems carried was not as fruitful…
Andrews and Brent 2025 is out at last (Royal Society Open Science, doi.org/10.1098/rsos.2…) and very welcome to see. Single cell quantification of weakly heritable, non-genetic, non-epigenomic cell-to-cell variation in signal and response.
Unexpectedly, in rapidfire testing with nitpick ish relevant biology questions, and even when reminded to search, Claude 4 Opus and Sonnet frequently give breezily incorrect answers to qs that GPT-o3 aces. Then when called on winging it, Claude apologizes. This is not the…
Anyone noticing that Claude Opus 4 seems to have an easy time refusing benign biological questions? So far, no such reports for Sonnet 4. Wonder if might be connected with their determination that Opus 4 but not Sonnet 4 was ASL-3.
December 2024 frontier AI models were quite good at articulating knowledge relevant to biological weapons risk. Because most research on pathogens is "dual use", and because of the centrality of virological mechanisms and elements derived from animal viruses to biological R&D,…
Current evals often incorrectly privilege "tacit knowledge", and misdefine it. Once any piece of knowledge is articulated, it ceases to be tacit. Guardrails on current AI models are easily circumvented.
Current assessments ("evals") of risk arising from large AI foundation models (like LLMs) conceptualize biological threat incompletely. Current evals do not assess a number of model capabilities that increase danger.
Contemporary AI foundation models increase biological weapons risk Working paper. Comments solicited. arxiv.org/abs/2506.13798 Main assertions follow.
He then asks. "The key question is, do they provide enough uplift that we should respond to it?" ("uplift" is a term of art). q maps to a now-common mental model, that the worrisome person being "uplifted" is an amateur, not, eg, a competent but non-specialized experimentalist.
Section 13, rhetorical Innovation, thezvi.substack.com/p/ai-121-part-…