philipkglass
Disclosure Day free stream
This isn't just relevant for piracy. Paid subscriptions to frontier models will refuse or downgrade answers to questions about the natural sciences and engineering. Unlike cybersecurity, where you can get specially approved to use models with fewer refusals, I don't think there is any program you can join where Anthropic will treat you like a grownup who has already read Carleone's Tactical Missile Warheads book. Opus just won't discuss details of high explosive engineering at all.
I'm still an Anthropic customer but in the long run I'm going to be using services that align with my interests. I want a provider that won't nerf or refuse requests around software security, building LLM training pipelines, the natural sciences, and reverse engineering. Anthropic is useful as a transitional bootloader for developing and demonstrating LLM technology broadly but each new safety "enhancement" makes it look more dispensable.
I am a US citizen and I approve of US-style free speech protection. If anything, I think that the US still has too much restriction on speech (e.g. DMCA limits on "anti-circumvention" software.)
However, I also approve of democratic government. If people in other countries vote to treat speech differently than the US, I wouldn't want to live under those laws, but I also wouldn't say that they are incorrect. I think that democratic principles mean implementing the will of the majority, even if I think that the majority is foolish. There are some things that I would fight against (not just lament) even if approved by the will of the majority; I can't fully reconcile my conflicting impulses, but I acknowledge them. All of this would have been a long and distracting aside to my original point, which was that most American businesses of note operate under multiple countries' jurisdictions.
They have a registered office in Schiphol:
https://discord.com/company-information
Discord could keep all of its employees, infrastructure, and business within the United States, but like other companies it has not. It would rather pursue more business opportunities.
Isn't the issue that the "big names" in the sector are not exclusively US based? Big American tech companies have employees, assets, or at least a business presence of some sort in countries outside the US, which means that they are exposed to legal systems in those other countries.
https://ourworldindata.org/electricity-mix
Even though solar power has been a significant part of the energy conversation for decades, the majority of solar growth is very recent. Nuclear power generated 2762 TWh in 2021, when solar generated 1054 TWh. Solar power generated 2779 TWh in 2025.
https://openai.com/index/disrupting-a-coordinated-model-dist...
https://www.anthropic.com/news/detecting-and-preventing-dist...
They would probably like it to be against the law ("Distillation attacks at this scale require a coordinated response across the AI industry, cloud providers, and policymakers") but I hope that legislators decline to grant their wish.
No, they're not going to give their products and services away for free, but I don't think they're going to make fat profit margins. The very tools they're building are helping competitive alternatives to spring up quickly. That's little comfort if you worry about job losses, but the technology is not under the thumb of Altman/Thiel/Zuck (or any other small number of people).
One of the common tasks LLMs are being trained to do is write software, which includes writing software for developing LLMs. That's reducing the cost to build alternative models and to run the ones that have already been trained and released. Many of these models aren't from America at all, so American billionaires don't get to decide how they're used. The competition is fiercer, and the potential for American monopolies or oligopolies is lower, than with technological waves of the recent past (e.g. mobile phone operating systems, search engines, social networking sites).
https://web.archive.org/web/20110713153629/https://www.washi...
The engineers were told that the children were rumored to have been kidnapped by Dr. Josef Mengele, the Nazi madman who, at Auschwitz and elsewhere, was known to have performed unspeakable experimental surgical procedures mostly on children, dwarfs, and twins. The engineers learned that just before the war ended, Josef Mengele made a deal with Stalin. Stalin offered Mengele an opportunity to continue work in eugenics. . . . in secret, in the Soviet Union.
Mengele held up his side of the Faustian bargain and provided Stalin with the child-size crew. . . .
https://archive.org/details/tippett-archive/
I downloaded the whole thing immediately because I too fear the copyright vultures. But if it's really 90 GB, I'm missing where to get the other 40 GB.
https://www.aec.go.jp/bunya/04/plutonium/20260805_e.pdf
As of the end of 2025, the total amount of Japan’s separated plutonium managed both within and outside Japan was approximately 44.4 tons, of which approximately 9.9 tons were held domestically and approximately 34.5 tons were held abroad.
Despite the nomenclature, weapons grade plutonium is not necessary to make nuclear weapons. Reactor grade plutonium has more problems with predetonation ("fizzle") and decay heat management, but both obstacles can be bypassed by design changes. In particular, "boosting" with a mixture of deuterium and tritium overcomes predetonation problems and can ensure repeatable high yields even with plutonium containing a high fraction of Pu-240.
https://en.wikipedia.org/wiki/Boosted_fission_weapon#Gas_boo...
The fusion reaction rate typically becomes significant at 20 to 30 megakelvins. This temperature is reached at very low efficiencies, when less than 1% of the fissile material has fissioned (corresponding to a yield in the range of hundreds of tons of TNT). Since implosion weapons can be designed that will achieve yields in this range even if neutrons are present at the moment of criticality, fusion boosting allows the manufacture of efficient weapons that are immune to predetonation. Elimination of this hazard is a very important advantage in using boosting.
This design hardens weapons against neutrons from nuclear interceptors and against neutrons generated within the core itself, e.g. spontaneous fission from "dirty" reactor plutonium.
Gas boosted bombs are more mechanically complex than the bombs of the 1940s, but the US started building them in the 1950s. Japan could easily manage the complexity today. Japan has also operated its Tritium Process Laboratory since the late 1980s, with capacity to handle 16 grams of tritium at a time. That's not much, but an individual boosted warhead may need only 3 grams of tritium.
Starlink satellites already use silicon solar cells, since they too are cost sensitive and don't have long lifespans:
https://starlink.com/public-files/Starlink_Approach_to_Satel...
On the Starlink V2 mini satellite, we predict that approximately 5% of the mass of the entire satellite could survive reentry. The biggest contributor (~90% of the surviving mass) is silicon from the solar cells, which has a high melting point...
[1] https://www.cell.com/joule/fulltext/S2542-4351(26)00362-4
https://debunkbot.com/share/48a5ac29-2924-4689-b573-2c38a1cd...
The manufacturing progress since 2008 has been enormous. The world now manufactures and installs over 500 gigawatts (peak) of solar PV capacity per year; in 2008, the world installed about 8 gigawatts [1]. Costs and prices have plummeted. The laboratory champion-efficiency cells of 2008 (e.g. Sanyo's 23% efficient heterojunction cell found in NREL's table of records [2]) are now surpassed by commercial products [3] that get manufactured on gigawatt scales and sold at prices low enough to compete with coal power.
[1] https://ourworldindata.org/grapher/installed-solar-pv-capaci...
"How did the Soviet Union produce aluminum, since it had little domestic bauxite?"
The Opus 3 response was less than half as long, and it covered the same main points. The answer didn't leap out as AI-written like current Anthropic model outputs. Opus 5 included a couple of useful minor details that 3 did not have, but it was also inflated by the overheated word-filler that has made Claude infamous.
My guess is that the worsening style of the Opus series is a consequence of optimization pressure for performing well on objective benchmarks (question answering, code, tool usage, mathematics) and minimal pressure to write concise, communicative prose. It's like an inbred lineage of dogs that has the "ideal" facial structure for its breed but a host of incidental problems that ordinary mutts never suffer. (However, this guess does not explain why other organizations' models chase the same benchmarks but haven't been stylistically worsening over time.)
I don't know which model underlies this. It responded more quickly than the models I normally use. It made the mistake I expected it to make, since this misconception is very common in training data, and conceded the point from my follow-up.
(EDIT: It appears that this link is not accessible to people without my cookies, except the developer I'm responding to above.)
https://en.wikipedia.org/wiki/Treason_laws_in_the_United_Sta...
If you have enough karma to flag comments you can also flag them directly from the noobcomments feed, which is useful to help manage the abusive users who keep re-registering new accounts.
Or to put it another way, there's enough natural variation in real-world bottlenecks that no pharma company can assume they'll beat competitors to market by using a smarter model.
A really smart model could significantly improve the pharma business if it could identify promising approaches to cancer treatment that are less likely to fail in clinical trials, but I don't think that the frontier labs have data to make that work yet. Much of the biomedical literature is poorly reproducible ("replication crisis") and much of the drug-development-specific data is proprietary, never published in the first place.
I do have hopes that general laboratory automation will go faster with LLM assistance, even if all the LLM does is write Python glue scripts to enable custom workflows and instrument integrations.