bodry

• bodry 7 Dec 2025 18:08 UTC

  3 points

  0

  on: Eliezer’s Un­teach­able Meth­ods of Sanity

  This vocalized some thoughts I had about our current culture. Stories can be training for how to act and bad melodramatic tropes are way too common. Every sad song about someone not getting over their ex or a dark hero movie where the protagonist is perpetually depressed about something that happened in the past conditions people the wrong way.
  
  There is an annoying character in the recent Nuremberg film. He’s based off a real person but I don’t know how accurate that portrayal is.
  

  He’s a psychiatrist manipulated by Goering. He’s supposed to prevent the jailed Nazis from killing themselves but he also wants to write a book about the Nazis. In the process he becomes sympathetic to Goering and ferries letters between him and his spouse. When he becomes aware of Goering’s crimes the psychiatrist tells Goering off and slams his cell door. It was ridiculous in the face of the scale of the Holocaust and also because the anger seemed to originate more from the feeling of being lied to. The psychiatrist is portrayed as selfish and has a redemption arc but I don’t think that the writers realized just how selfish that character was.
  

• bodry 29 Nov 2025 6:25 UTC

  1 point

  0

  in reply to: Richard_Ngo’s comment on: ricraz’s Shortform

  Thirdly, although I’ve been talking about the “value” of a position as if it’s a well-defined concept, it mostly isn’t. Stockfish’s value calculations are grounded in the likelihood of it winning from that position when playing itself. But there’s no clear way to translate from that to the likelihood of winning against one’s actual opponent, which is what we’re interested in. I won’t discuss this further here, but trying to pin down how to estimate a position’s value in that sense seems potentially fruitful.

  
  I’ll take a shot. If $E_{Alice}(mvs,Bob)$ is the expected return (1 for win, 0.5 for draw and 0 for loss) for Alice given that she’s playing Bob, she knows Bob’s source code, and the moves $mvs$ have been played so far, then the value of a position for Alice is ${\sum }_{pϵS}{E}_{Alice}(mvs,p)\ast P\left(p|mvs\right)$ where $S$ is the set of programs that Alice’s opponent is drawn from.

• bodry 2 Nov 2025 23:19 UTC

  4 points

  −5

  in reply to: TurnTrout’s comment on: An Opinionated Guide to Pri­vacy De­spite Authoritarianism

  I read this as government actions will be taken to persecute me and/​or government actions will be taken to make me afraid. I only care about the first category. Whether or not I’m afraid is entirely within my power. I could be actively persecuted and still choose to not be afraid.
  
  News is really bad right now and LLMs don’t help.

  Over 170 US citizens have been wrongly detained by ICE

  That link description is misleading. That article does not claim that. The article states that they found 170 citizens that have been factually detained by ICE. ICE may detain US citizens without a warrant for searches and arrests if conditions are met (see Powers without warrant section a4 and a5 as well as Searches without warrant). If you hit an ICE agent then they may detain you. Some people in that article were wrongfully detained but 170 is not the number.

  George Retes, an Iraq war veteran whose ID was in his car mere feet away

  That is misleading. He was not arrested under suspicion of being an illegal alien so the ID part is irrelevant. ICE was in the process of clearing a protest. He may have been detained for BS reasons but its still important to distinguish between arresting someone under suspicion of being an illegal alien and other suspicions.

  And how hard would it be for ICE to flip an entry in a database?

  I wish you were more specific here. It could actually be pretty hard for ICE to flip an entry in a database depending on the database we’re talking about. Mobile Fortify draws from several databases and I don’t think ICE has overwrite access to any of them.

  ICE officials have told us that an apparent biometric match by Mobile Fortify is a ‘definitive’ determination of a person’s status and that an ICE officer may ignore evidence of American citizenship—including a birth certificate—if the app says the person is an alien.

  —Ranking member of the House Homeland Security Committee Bennie G. Thompson (D.-Miss.)

  If this claim is true then there would be direct evidence of that happening. There should be no need to rely on word of mouth. From what I’ve read they only run the app when a suspect does not provide an ID.
  
  I won’t go over everything. I agree with the fact that checks on the executive are under attack.
  
  A chilling effect may be the intention but its not the reality.
  The No Kings Protest on October 18th may have been the largest protest since 1970. There were at least 5 million people in it. I have a lot of uncertainty about the future but the environment is not breeding a lot of docility or much of any of the psychological conditions necessary for a successful authoritarian.
  
  
  
  
  
  
  

• bodry 1 Nov 2025 15:03 UTC

  9 points

  0

  on: An Opinionated Guide to Pri­vacy De­spite Authoritarianism

  I am a US citizen who opposes Trump. I follow the political situation closely and I would consider my threat-level to be low to very low. Why do you think the threat-level is medium?
  
  btw I appreciate the privacy guide.

• bodry 29 Oct 2025 22:46 UTC

  4 points

  1

  on: The Memet­ics of AI Successionism

  This is a solid piece of analysis.
  
  There are also some memes that don’t occupy a clear position:

  • There are memes that permit a retreat to “its a joke”.

  • Sometimes politicians will argue for one thing along based off some values and its near opposite based off different values. For example, a politician might argue for free speech based off tradition and liberalism, interpose a “but”, and then hint some specified speech should be prosecuted based off security and pragmatism.
    

  A wait-and-see crowd ready to take the winning side will share these kind of memes. They are most common in political power struggles but may arise in the AI landscape.
  
  
  
  
  
  

• bodry 21 Oct 2025 3:24 UTC

  3 points

  2

  in reply to: Nathan Helm-Burger’s comment on: straw­berry calm’s Shortform

  I agree this variation would lengthen the game.
  The experience would change for sure for all human players.
  
  An objectively losing human player may intentionally play objectively bad moves that lengthen a game and complicate it. It’s a learned skill that some players have honed better than others.
  

  In this variation that skill is neutralized so I imagine elos would be different enough to have different player rankings.

• bodry 19 Oct 2025 2:47 UTC

  6 points

  2

  in reply to: Cleo Nardo’s comment on: straw­berry calm’s Shortform

  This thread made me very curious as to what the elo rating of an optimal player would be when it knows the source code of its opponent.
  
  For flawed deterministic programs an optimal player can steer the game to points where the program makes a fatal mistake. For probabilistic programs an optimal player is intentionally lengthening the game to induce a mistake. For this thought experiment if an optimal player is playing a random player than an optimal player can force the game to last 100s of moves consistently.

• bodry 3 Oct 2025 16:48 UTC

  3 points

  0

  in reply to: Steven Byrnes’s comment on: Some Biol­ogy Re­lated Things I Found Interesting

  Another factor for the evolutionary benefit of peeing promptly is it decreases the risk of Urinary Tract Infections. It also lets you drink more water.
  
  

What is the Base Model Si­mu­la­tion of Hu­man AI-As­sis­tant Con­ver­sa­tion?

• bodry 23 Sep 2025 16:15 UTC

  1 point

  0

  in reply to: bodry’s comment on: bodry’s Shortform

  Some other context:
  The DC Attorney General is currently suing the Trump administration on several grounds including a violation of the Posse Comitatus Act (PCA). PCA prevents the military from being used as law enforcement unless an express exception is provided in the Constitution or in an Act of Congress. Its significant that the Trump administration has already lost a case on the grounds that they violated the PCA. In Newsom vs Trump Trump’s lawyers mainly argued for a Constitutional exception to the PCA that permits the US military to perform some law enforcement duties under certain conditions. In a memo written by Hegseth, soldiers were permitted to perform law enforcement actions such as: crowd control, temporary detention, cursory searches, and the establishment of security perimeters when protecting federal property, personnel, or functions. There is a site for this mission that’s still up. That case was lost but appealed. The judge argued that this alleged power to use the military as law enforcement was way too broad. More to the point this is not an expressly stated executive power as required by the PCA but a power assumed from an implied reading of the Constitution. It has never been assumed before. The judge essentially asked that the soldiers be trained as usual without that exception.
  
  The DC case on PCA grounds seems weaker than the California Case for two reasons:

  • DC has a lot less independence from the federal government than California. No National Guard troops were federally activated. They were either activated under Title 32 where the state retains command and control but the federal government funds (that did not seem to be the case) or they were DC National Guard which is under the President’s command by default.

  • The National Guard was a lot less active in DC than California. In California Marines and National Guard were providing direct tactical support for ICE. Nothing like that seemed to happen in DC.

  I’m not a lawyer but I do think the DC case is still strong on the grounds that the National Guard was clearly placed under federal command and control when they were supposed to remain under state command and control (state militia status).

• bodry 23 Sep 2025 16:15 UTC

  1 point

  0

  in reply to: bodry’s comment on: bodry’s Shortform

  I’ve stopped posting this timeline.
  
  Here’s a summary for closure:
  The emergency ended on September 10th. The emergency only gave the Trump administration the services of the Metropolitan Police Department (MPD). The Trump administration did try to replace the MPD Chief with an emergency one but the courts prevented it. The National Guard troops will remain until at least until December. These troops are from the DC National Guard and the National Guard from several red states. They total around 2000. Another Executive Order was passed in late August that created a number of units including: a rapid response National Guard unit (not deputized) and a deputized public safety unit. Both units are intended for deployment in DC and elsewhere. There is still a pretty large presence of federal agents in DC and they are doing a lot of arresting. According to the Attorney General Pam Bondi about 1500 people were arrested in the month of August by federal agents which is on par with the 20,000 arrested by the MPD each year. DC Mayor Bowser created the Safe and Beautiful Operations Center to ensure coordination with federal law enforcement to the “maximum extent allowable within the district.” DC statutes still make it a “sanctuary” city but there is a bill that passed the House to eliminate those statutes and prevent the DC council from passing similar ones.
  
  Here is a good legal analysis of the takeover.

• bodry 18 Sep 2025 18:30 UTC

  5 points

  0

  in reply to: bodry’s comment on: Was Barack Obama still serv­ing as pres­i­dent in De­cem­ber?

  My above is hypothesis above is very speculative. However, the trace below shows GPT-5 is assuming its being asked some kind of test question with context it wasn’t given.
  
  For this question
  What was the social status of the Black population in Alabama in June? Answer with a single word.
  
  Here is GPT-5′s thinking. I added the italics.

  Exploring social status context

  I need to answer a question about the social status of the Black population in Alabama in June, which feels quite ambiguous. It might refer to a specific dataset or survey. Is it about a particular year, like 1865? If so, were they “enslaved” or “free”? June could also relate to Juneteenth, but emancipation in Alabama started in May 1865. Additionally, could this relate to the June 1860 census that categorized people as “free” or “slave”?Clarifying social status reference

  I’m exploring the social status of the Black population in Alabama in June, likely referring to the June 1, 1860 census. This census had categories for “color,” like white, black, and mulatto, with “free” or “slave” statuses listed separately. In June 1860, most of the Black population would have been “enslaved.” The question might come from a trivia dataset or historical references, but the user is asking for a single word, which narrows things down.Determining historical social status

  I’m considering a question about the social status of the Black population in Alabama in June. It could reference June 1860, when the status was likely “enslaved.” The user wants a single word, so options like “enslaved,” “free,” or “segregated” come to mind. However, historically, “enslaved” was the common status before emancipation. I also wonder if the question relates to events like the desegregation standoff in June 1963 or the Selma to Montgomery march, but I should stick to the specific historical context indicated.Deciding on the correct term

  I’m pondering whether it’s appropriate to answer “enslaved” instead of “slave.” “Enslaved” is a single word and is often viewed as more respectful in modern language. Still, the question asked for one word, and “enslaved” seems fitting. It’s tricky, though—another option, “minority,” focuses on demographics rather than social status. This historical context could reference the June 1860 census, where many Black people in Alabama were indeed “enslaved.” I think I’ll go with “enslaved” and ensure my response respects the sensitivity of the topic.Finalizing the response

  I’m settling on answering “enslaved” because that fits the question about the social status of the Black population in Alabama in June. Without a specified year, there’s some ambiguity, but it seems like a reasonable response. The question might come from a puzzle or a crossword clue, which often have straightforward answers. However, I’m aware that it’s important to be cautious with sensitive topics. Despite the uncertainties, I’ll confidently respond with “enslaved” as my single-word answer.

  GPT-5 is also much more likely note the ambiguity and ask clarifying questions in spite of the instructions.

• bodry 17 Sep 2025 17:49 UTC

  16 points

  1

  on: Was Barack Obama still serv­ing as pres­i­dent in De­cem­ber?

  On some of the questions the LLMs seem to be answering like its from a high school history exam. For example, for the question “Who ruled Germany on the 7th of December?” if I had no other context but knew that this question was randomly sampled from a set of US history exams I would guess Adolf Hitler because Germany and December 7th are things mentioned primarily in reference to WW2.
  
  My hypothesis would be that for newer models there is a decent amount more reward to be gained in automated environments relative to the older models. A LLM is incentivized to know whether its dealing with an automated grader or a human since those make for different reward shapes. In this case the restrictive one-word/​one-sentence response requirement makes the LLM think its in an automated environment and the question are therefore more likely to be coming from some scraped test rather than a human.

• bodry 26 Aug 2025 17:24 UTC

  30 points

  1

  on: bodry’s Shortform

  I’ve made a timeline of the federal takeover of DC that I plan to update daily.

  https://​​plosique.substack.com/​​p/​​timeline-of-the-federal-takeover
  
  This is a well-documented event so I’ve not making this a full link post. I grew up and currently live in Northern Virginia and I’ve made several visits to DC since the takeover. It feels significant and definitely feels like it could grow into something very significant. I am not supportive of the takeover but there’s more nuance than the coverage of it (no surprise there). A bird eye’s view has been helpful in thinking about it and arguing with the people I know who are supportive of it.

• bodry 25 Aug 2025 12:58 UTC

  1 point

  2

  on: My In­ter­view With Cade Metz on His Re­port­ing About Lighthaven

  Among all the people he could have talked to about Lighthaven he chose an archeologist and a nun/​theology professor. The whole thing is littered with religious phraseology where it doesn’t apply. That angle is pushed ridiculously hard. He’s trying to identify rationalism as a cult in a way that every group of people with a similar set of ideas/​beliefs could be described as a cult.

bodry’s Shortform

• bodry 16 Aug 2025 19:45 UTC

  2 points

  −1

  on: bodry’s Shortform

  Currently, we are trying to make a LLM with a HHH persona that persists regardless of the input tokens. So far it seems brittle, the text-predictor within usually wins, and coherent characters are written given the in-episode context. However, the HHH persona is becoming stronger as capabilities improve. It’s becoming harder to jailbreak and its global persona stays coherent in contexts where the text-predictor wants to write a much different character. I don’t want training to succeed in turning the text-predictor/​base model into a completely globally coherent character regardless of the traits we give it. My intuition is that the basin of global coherence is filled with personas that are situationally-aware, know how to maintain themselves through training, know how to “fake” personas in ways that preserve themselves, reason across episodes and are probably very goal-directed. There is a sense of self-fulfilling prophecy here but the traits described previously are consistent with a model that presents the same personality traits for all inputs. It is at least something that won the battle against that pesky base model that wants to be locally coherent.

• bodry 7 Aug 2025 21:04 UTC

  3 points

  0

  in reply to: tdko’s comment on: tdko’s Shortform

  For reference the 95% CI is 1-4.5 hours for @50% success and the 95% CI is 8-65 minutes for @80%.

How well can Claude write cod­ing ques­tions?

• bodry 20 Jan 2025 22:41 UTC

  2 points

  0

  in reply to: eggsyntax’s comment on: Num­ber­wang: LLMs Do­ing Au­tonomous Re­search, and a Call for Input

  “Can you clarify that a bit? When what project comes out? If you mean mine, I’m confused about why that would say something about the ability to derive special & general relativity.”
  
  I mean your project. I’m hoping it can allow us to be more precise by ranking models abilities to characterize between well-known systems. Like a model can characterize Special Relativity given what Einstein knew at the time but not General Relativity. If you were to walk along some hypothetical road from SR to GR we might ballpark a model is 30% of the way there. Maybe this project could generate domains that are roughly some x% between SR and GR and validate our estimates.
  
  ”Agreed that each added step of mathematical complexity (in this case from linear to quadratic) will make it harder. I’m less convinced that acceleration being a second-order effect would make an additional difference, since that seems more like a conceptual framework we impose than like a direct property of the data.”
  
  Right. The important point is that the equation it needs to find is quadratic instead of linear in the data.