Sara Furnal retweeted
Nov. 18, 2026, 5:16am EST
The Voyager 1 spacecraft has reached a distance of one light-day from Earth.
The field of AI assessment is developing quickly. Key terms such as audit, evaluation, and verification are often used in very different ways. This can hinder productive conversation and collaboration.
Today, we set out how AVERI defines these terms. We also include an interactive chart that explores what might currently constitute a “frontier” model or company.
These definitions are intended for practical use, and we do not propose their use where precision is more important than usability. We expect to update them over time if and as needed.
AI auditing will be built and implemented by a wide range of organizations. We hope these definitions will help the field move faster together, and we welcome any feedback.
averi.org/ourwork/definition
Sara Furnal retweeted
📜 My new paper: "What Constitutions Say -- and Fail to Say -- about Constitutional Conventions," to be published in the Harvard Journal on Legislation @HarvardJOL.
🎉 The paper makes five contributions to our knowledge of constitutions and constitution-making:
1⃣ An original coding of 197 constitutions in the world. I show that a total of 21 constitutions -- only 10.7% -- expressly authorize a constitutional convention or constituent assembly.
2⃣ 171 constitutions say nothing at all about a constitutional conventions or constituent assembly. And additional 5 say refer to a convention or its equivalent in terms that fall short of authorizing one as an ordinary means of reform.
3⃣ I also create a three-part classification to understand convention clauses: (1) 9 constitutions codify a complete convention clause; (2) 4 constitutions codify a deferred convention clause; and (3) 8 constitutions codify a skeletal convention clause.
4⃣ I report on the results of my comprehensive original survey of 109 constitutional experts from 53 countries on 6 continents. My survey reveals that constitutional experts converge on design features that almost no constitution contains. The gap between the design of constitutional texts and the judgment of constitutional experts is remarkably cavernous.
5⃣ I distinguish three reasons why constitutions might choose silence on conventions instead of expressly detailing rules on their use: silence as prohibition, silence as permission, and silence as pragmatism.
📝 The full draft is available here: papers.ssrn.com/sol3/papers.…
📩 Comments welcome, always.
🙏🏽 Thank you to the University of Texas at Austin for supporting my research. @UTAustin @UTexasLaw @UTGovernment @UTCivics @UTexasResearch
Funny, but for real: alternatives? Return to the Athenian phylai or Servius Tullius' regrouping of the Roman citizenry into different classes/centuries? Sometimes redistricting has been done to alter the class structure altogether and break partial attachments and uproot private interests rather than increase factional interests. Sometimes to approximate some ideal end goal of one person, one vote. It's always been a choice about what kind of political community to constitute. The absence of a perfect district makes articulating the purposes of political organization even more important
I think I’m anti-anti-gerrymandering. The whole concept presumes the existence of a platonically correct district, no?
107
Funny enough, this was a debate even long before the AI era or even before the 20th century typing monkeys discussion. Enlightenment thinkers like Diderot (Pensées philosophiques) and Voltaire (Dictionnaire philosophique) debated whether random permutations of letters or words could produce great works like the Iliad. Other interesting early modern/Enlightenment takes on this question: Leibniz, Bernoulli, Kant.
Unsurprisingly, I particularly am interested in Rousseau's take on this. He concedes the abstract mathematical possibility but rejects the materialist implications that derive from it. And his theory of language cuts against a network effect argument for the enduring value of literature. Rousseau did nothing better than analyze the origin and impact of reputation. Chance + hype wouldn't cut it as an explanatory vehicle for him when debating Diderot on this.
It is an axiomatic belief among the Berkeley-Rationalist set, many of whom run reinforcement learning somewhere or the other, that the notion of a "classic work" or "good literature" like Shakespeare is purely from snowballing network effects from a random generator.
81
Sara Furnal retweeted
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Excellent discussion just now with @LuciusCaviola, Caspar Kaiser, and @jeffrsebo on inter-AI morality. Best one I've attended in quite some time! Thanks!
59
Sara Furnal retweeted
Replying to @HarrisPolicy
@HarrisPolicy School at UChicago is searching for an assistant professor in AI and policy. Apply if you work on AI tools to improve public policy outcomes or study the behavior and governance of AI and tech systems!
apply.interfolio.com/195292
EOE/Vet/Disability
Seems heavily inspired by LBJ's 1964 Daisy ad. When I show my students presidential campaign ads, Daisy is always the one that hits home with them.
AI companies admit what they’re building might kill everyone. We don't have to agree to this.
Add your name at donotaccept.ai to say: we do not accept your terms.
11
215
Plato, Gorgias. Rousseau, Discourse on Inequality. Thucydides. Kant, Idea for a Universal History with a Cosmopolitan Purpose.
2
6
463

















