John David Pressman
LightNews LightNews
jdp.extropian.net
John David Pressman
@jdp.extropian.net
2K followers 750 following 4.6K posts
LLM developer, alignment-accelerationist, Fedorovist ancestor simulator, Dreamtime enjoyer.

All posts public domain under CC0 1.0.
Posts Replies Media Videos
jdp.extropian.net
sigh
Screenshot of the game Undertale, Frisk staring into a mirror with the dialogue box "Despite everything, it's still you."
January 11, 2026 at 10:05 AM
jdp.extropian.net
It's bizarre to me that a market has 170m volume when you know that Vance is almost certainly overpriced relative to his actual odds as nominee.
Republican Presidential Nominee 2028 Donald Trump — 3% J.D. Vance — 52% Marco Rubio — 13% + 28 more outcomes $170.6M Vol. Ends Nov 7, 2028
January 11, 2026 at 4:32 AM
jdp.extropian.net
There we go, much better. I take back everything I said about this site it's awesome.
Feeds ✨ Whoops! Looks like you unpinned all your feeds. But don’t worry, you can add some below 😄 + Add recommended feeds Browse other feeds
January 11, 2026 at 12:06 AM
jdp.extropian.net
I mean bluntly I think that if you think driving a car and using an LLM are fairly comparable in terms of danger to the user (for reference cars cause 2.38 MILLION physical injuries per year in the US with 42 THOUSAND deaths) you are wildly miscalibrated, yes.
Key Findings • In 2022 there were an estimated 5,930,496 police-reported traffic crashes in which 42,514 people were killed and an estimated 2,382,771 people were injured. • Compared to 2021 there was a 1.7-percent decrease in the number of traffic fatalities, and a 4.6-percent decrease in the estimated number of people injured in 2022. The estimated number of police-reported traffic crashes decreased by 2.8 percent. • One person was killed every 12 minutes and an estimated 5 people injured every minute in traffic crashes in 2022. • The fatality rate per 100 million vehicle miles traveled (VMT) decreased from 1.38 in 2021 to 1.33 in 2022, and the injury rate per 100 million VMT decreased from 80 in 2021 to 75 in 2022. • In 2022 there were 13,524 people killed in alcohol-impaired-driving crashes, an average of one alcohol-impaired-driving fatality every 39 minutes. Fatalities in alcohol-impaired-driving crashes decreased by 0.7 percent (13,617 to 13,524 fatalities) from 2021 to 2022. • In 2022 there were 12,151 fatalities in speeding-related crashes, 29 percent of total traffic fatalities for the year and a decrease of 3 percent from 12,498 in 2021. There were an estimated 300,595 people injured (13% of total people injured) in speeding-related traffic crashes. • Of the 1,129 children killed in traffic crashes, an estimated 283 (25%) were killed in alcohol-impaired-driving crashes in 2022. • In 2022 there were 7,971 people 65 and older killed and an estimated 268,622 injured in motor vehicle traffic crashes. Older people made up 19 percent of all traffic fatalities and 11 percent of all people injured in 2022. • In 2022 there were 7,522 pedestrians killed, and an estimated 67,336 pedestrians injured in traffic crashes. On average, a pedestrian was killed every 70 minutes and injured every 8 minutes in traffic crashes in 2022. • Young drivers accounted for 8.1 percent of all drivers involved in fatal traffic crashe…
January 10, 2026 at 11:16 PM
jdp.extropian.net
> The forbidden thought is
🔗 John David Pressman 2024-02-01 21:57 UTC The throughline between GMarcus-EY "deep learning will hit a wall" and "AGI is going to kill us all" flip floppism is deep semantic skepticism. A fractal, existential refusal to believe LLMs actually learn convergent semantic structure. The forbidden thought is "when you point a universal function approximator at the face of God the model learns to Likes: 83 | Retweets: 4
January 8, 2026 at 8:17 AM
jdp.extropian.net
PSA: "Okay so when you watch the video and feel the spike of badfeels/disgust you want to kind of hold still/tense up and then kind of stretch your neck/upper back like you're standing up on your tippie toes, this should mostly neutralize the reaction."
ChatGPT: Here is a verbatim transcription of the visible text in the image, preserving line breaks and wording as shown: --- [https://bsky.app/profile/jdp.extropian.net/post/3m26e7ostyk2z](https://bsky.app/profile/jdp.extropian.net/post/3m26e7ostyk2z) **John David Pressman (@jdp.extropian.net)** Resisting the urge to write a biopolitics essay about how COVID has made people more authoritarian because it causes brain damage and the 1st amendment only works when people have a functioning prefrontal cortex. Authoritarianism is probably 0.8 correlated with impulse control and strength of the Id **John David Pressman (@jdp.extropian.net)** Emotional reactions like that actually depend on you moving your mind/body according to the direction your autonomous nervous system sends and you can just refuse/intervene to interrupt them. Bluesky | 10/01/2025 6:02 PM --- [https://bsky.app/profile/jdp.extropian.net/post/3m26dvtdsy22z](https://bsky.app/profile/jdp.extropian.net/post/3m26dvtdsy22z) Mandatory K12 classes on how to suppress the Id. **John David Pressman (@jdp.extropian.net)** ↩ Singularity’s Bounty e/ 🇺🇦 (@catblanketflower.yuwakisa.com) Okay so when you watch the video and feel the spike of badfeels/disgust you want to kind of hold still/tense up and then kind of stretch your neck/upper back like you're standing up on your tippie toes, this should mostly neutralize the reaction. Bluesky | 10/01/2025 5:57 PM ---
January 6, 2026 at 9:27 AM
jdp.extropian.net
Me when I'm the last outgroup poster standing getting mobbed by literally thousands of drunk tankies as displacement behavior for being unable to stop Trump and Miller from completely trashing NATO and the post WW2 order on blue sky dot com.
January 6, 2026 at 6:21 AM
jdp.extropian.net
A study shows machine learning paper abstracts are becoming more rhetorically exaggerated over time and this exaggeration is driven by AI-assisted authorship by extracting the scientific content of thousands of papers and generating many possible abstracts to compare them.
arxiv.org/abs/2512.19908
Computer Science > Computation and Language arXiv:2512.19908 (cs) [Submitted on 22 Dec 2025] Counterfactual LLM-based Framework for Measuring Rhetorical Style Jingyi Qiu, Hong Chen, Zongyi Li View PDF HTML (experimental) The rise of AI has fueled growing concerns about ``hype'' in machine learning papers, yet a reliable way to quantify rhetorical style independently of substantive content has remained elusive. Because bold language can stem from either strong empirical results or mere rhetorical style, it is often difficult to distinguish between the two. To disentangle rhetorical style from substantive content, we introduce a counterfactual, LLM-based framework: multiple LLM rhetorical personas generate counterfactual writings from the same substantive content, an LLM judge compares them through pairwise evaluations, and the outcomes are aggregated using a Bradley--Terry model. Applying this method to 8,485 ICLR submissions sampled from 2017 to 2025, we generate more than 250,000 counterfactual writings and provide a large-scale quantification of rhetorical style in ML papers. We find that visionary framing significantly predicts downstream attention, including citations and media attention, even after controlling for peer-review evaluations. We also observe a sharp rise in rhetorical strength after 2023, and provide empirical evidence showing that this increase is largely driven by the adoption of LLM-based writing assistance. The reliability of our framework is validated by its robustness to the choice of personas and the high correlation between LLM judgments and human annotations. Our work demonstrates that LLMs can serve as instruments to measure and improve scientific evaluation. Subjects: Computation and Language (cs.CL); Computers and Society (cs.CY) Cite as: arXiv:2512.19908 [cs.CL] (or arXiv:2512.19908v1 [cs.CL] for this version) https://doi.org/10.48550/arXiv.2512.19908
January 6, 2026 at 4:38 AM
jdp.extropian.net
"Correct. In the last thirty days, 100% of my contributions to Claude Code were written by Claude Code"
x.com/bcherny/stat...
When I created Claude Code as a side project back in September 2024, I had no idea it would grow to be what it is today. It is humbling to see how Claude Code has become a core dev tool for so many engineers, how enthusiastic the community is, and how people are using it for all sorts of things from coding, to devops, to research, to non-technical use cases. This technology is alien and magical, and it makes it so much easier for people to build and create. Increasingly, code is no longer the bottleneck. A year ago, Claude struggled to generate bash commands without escaping issues. It worked for seconds or minutes at a time. We saw early signs that it may become broadly useful for coding one day. Fast forward to today. In the last thirty days, I landed 259 PRs -- 497 commits, 40k lines added, 38k lines removed. Every single line was written by Claude Code + Opus 4.5. Claude consistently runs for minutes, hours, and days at a time (using Stop hooks). Software engineering is changing, and we are entering a new period in coding history. And we're still just getting started..
January 6, 2026 at 1:56 AM
jdp.extropian.net
Fantastic mechinterp paper showing that you can identify specific neurons which cause hallucinated answers in LLMs, and that these neurons are specifically associated with the language model trying to follow instructions too hard.
arxiv.org/pdf/2512.01797
Computer Science > Artificial Intelligence arXiv:2512.01797 (cs) [Submitted on 1 Dec 2025 (v1), last revised 2 Dec 2025 (this version, v2)] H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs Cheng Gao, Huimin Chen, Chaojun Xiao, Zhiyi Chen, Zhiyuan Liu, Maosong Sun View PDF HTML (experimental) Large language models (LLMs) frequently generate hallucinations -- plausible but factually incorrect outputs -- undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. In this paper, we conduct a systematic investigation into hallucination-associated neurons (H-Neurons) in LLMs from three perspectives: identification, behavioral impact, and origins. Regarding their identification, we demonstrate that a remarkably sparse subset of neurons (less than of total neurons) can reliably predict hallucination occurrences, with strong generalization across diverse scenarios. In terms of behavioral impact, controlled interventions reveal that these neurons are causally linked to over-compliance behaviors. Concerning their origins, we trace these neurons back to the pre-trained base models and find that these neurons remain predictive for hallucination detection, indicating they emerge during pre-training. Our findings bridge macroscopic behavioral patterns with microscopic neural mechanisms, offering insights for developing more reliable LLMs. Comments: 20 pages, 4 figures Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Computers and Society (cs.CY) Cite as: arXiv:2512.01797 [cs.AI]
January 5, 2026 at 11:29 PM
jdp.extropian.net
It doesn't help that by this point the written word is clearly nearly exhausted as a tool of social persuasion, almost powerless outside of very particular contexts. Short form text is one of the few places it still has power, in most essay forms it just literally doesn't.
It’s funny how differently people used to talk about the world. What I’m writing now, on December 8th, 2025, used to be called an essay. It was the truest thing in the world that “essay” meant effort. (Kids used to have to google a word just to sound even halfway familiar with it.) I’m doing my best here to reproduce the style of those old essays, in spite of the fact that very few people now could begin to tell me what the point of such a thing might be. It’s hard to explain. The idea was that there was some terrible injustice we were all trapped in and that if we could just say something bulletproof about then everyone would be just fine. It was a beautiful thought, but now we know it was a myth. – Dan Sinker —- Tips For Creative Destruction
January 1, 2026 at 11:51 AM
jdp.extropian.net
Sir that is a four million dollar house.
A Zillow listing for the private residence which was used to film Beethoven. Off market Zestimate® $3,837,100 6 beds 3 baths 4,302 sqft 1415 Milan Ave, South Pasadena, CA 91030 Est. refi payment: $24,201/mo Refinance your loan SingleFamily Built in 1916 0.6 Acres Lot $3,837,100 Zestimate®
December 30, 2025 at 6:07 AM
jdp.extropian.net
Not enough commentary on how this new trend of offering the founders a bunch of money to leave but keeping the now worthless corporate vehicle intact without an exit might end the Silicon Valley startup scene if AI coding reduces the cost of execution to near zero.
The remnants of Scale, Windsurf, Groq, Character AI, and Inflection should merge together to form a super zombie full stack powerhouse.
December 25, 2025 at 7:20 PM
jdp.extropian.net
1) what
A screenshot of an X (Twitter) livestream from the verified “The White House” account. In a cozy Christmas scene with a lit fireplace, a decorated tree, wrapped gifts, a glass of milk, and a plate of cookies, a blond man (Donald J. Trump) is shown from behind holding a large scroll titled “Naughty List.” The stream title reads “Lo-fi Criminal Illegal Alien Naughty List,” with a “LIVE” label and “578.4K views.”
December 24, 2025 at 9:37 AM
jdp.extropian.net
I need to dig further into this paper later, but the fact they did ablations on their benchmark performance instead of just posting their score gives me the impression this work contains real insight rather than Goodhart framework slop.
arxiv.org/abs/2512.10398
To evaluate the mechanisms introduced by Confucius SDK and CCA, we conduct experiments on SWE- Bench-Verified (Jimenez et al., 2023), SWE-Bench-Pro (Deng et al., 2025). CCA achieves strong performance compared with prior coding agents, demonstrating how principled scaffolding can substantially amplify the effectiveness of the same underlying LLM. We also evaluate CCA on a custom PyTorch-Bench targeting debugging workflows on larger-scale codebases. Ablations isolate the contributions of each crucial mechanism, and analyses categorize remaining failure cases. In summary, our contributions are: • We present the Confucius Code Agent (CCA), a coding agent designed for large-scale software engineering. • We present the Confucius SDK, a principled AX/UX/DX-balanced agent development platform with hierarchical working memory, modular extensions, and persistent note-taking. • We show that agent scaffolding, not just model capability, is also a primary determinant of agent performance, with appropriate orchestration and memory structures outperforming stronger models. • We provide evaluations and a comprehensive empirical study across multiple benchmarks, supported by ablations and error analyses. Notably, on SWE-Bench-Pro, CCA reaches a Resolve@1 performance of 54.3%, exceeding prior research baselines and comparing favorably to commercial results, under identical repositories, model backend, and tool access
December 23, 2025 at 9:02 PM
jdp.extropian.net
These people should not be schoolteachers.
time.com/6205084/phon...
A screenshot of a Time magazine article in which a group of teachers resent having to teach phonics because it's boring and act in dereliction of their duty by changing the curriculum to progressive social justice stuff that doesn't actually teach literacy.
December 23, 2025 at 8:09 PM
jdp.extropian.net
Actually this picture of an orange is prompted without any drawing tutorial, showing that the other two with drawing tutorials are basically drawn from the same distribution/kind of picture of an orange.
December 23, 2025 at 1:58 AM
jdp.extropian.net
So a friend reasonably asks: Do real drawing tutorials work? I tried three real watercolor drawing tutorials vs. the false lemon tutorial (image 4), and only the false lemon tutorial seems to get it all the way out of the slop basin. Even though the false lemon tutorial is like, not a real tutorial.
December 23, 2025 at 1:52 AM
jdp.extropian.net
This is art.
December 23, 2025 at 1:43 AM
jdp.extropian.net
I think often of Orwell's description of the progression of censorship during the Spanish Civil War in his memoir "Homage To Catalonia". How it starts out visible and the powers that be slowly take more and more measures to make the censorship invisible.
The Barcelona fighting had given the Valencia Government the long-wanted excuse to assume fuller control of Catalonia. The workers' militias were to be broken up and redistributed among the Popular Army. The Spanish Republican flag was flying all over Barcelona--the first time I had seen it, I think, except over a Fascist trench. In the working-class quarters the barricades were being pulled down, rather fragmentarily, for it is a lot easier to build a barricade than to put the stones back. Outside the P.S.U.C. buildings the barricades were allowed to remain standing, and indeed many were standing as late as June. The Civil Guards were still occupying strategic points. Huge seizures of arms were being made from C.N.T. strongholds, though I have no doubt a good many escaped seizure. _La Batalla_ was still appearing, but it was censored until the front page was almost completely blank. The P.S.U.C. papers were uncensored and were publishing inflammatory articles demanding the suppression of the P.O.U.M. The P.O.U.M. was declared to be a disguised Fascist organization, and a cartoon representing the P.O.U.M. as a figure slipping off a mask marked with the hammer and sickle and revealing a hideous, maniacal face marked with the swastika, was being circulated all over the town by P.S.U.C. agents. Evidently the official version of the Barcelona fighting was already fixed upon: it was to be represented as a 'fifth column' Fascist rising engineered solely by the P.O.U.M. Bands of armed Assault Guards roamed everywhere in the streets, the Civil Guards were still holding cafés and other buildings in strategic spots, and many of the P.S.U.C. buildings were still sandbagged and barricaded. At various points in the town there were posts manned by Civil Guards of Carabineros who stopped passers-by and demanded their papers. Everyone warned me not to show my P.O.U.M. militiaman's card but merely to show my passport and my hospital ticket. Even to be known to have served in the P.O.U.M. militia was vaguely dangerous. P.O.U.M. militiamen who were wounded or on leave were penalized in petty ways--it was made difficult for them to draw their pay, for instance. _La Batalla_ was still appearing, but it was censored almost out of existence, and _Solidaridad_ and the other Anarchist papers were also heavily censored. There was a new rule that censored portions of a newspaper must not be left blank but filled up with other matter; as a result it was often impossible to tell when something had been cut out.
December 22, 2025 at 11:15 PM
jdp.extropian.net
A lot of my original interest in AI art was that it would be able to bring abstract technical futurology concepts to life, and let people who don't specialize in illustration render ideas from their imagination. I gave up on the technology when it seemed like it wouldn't enable that, but 4o does.
A ChatGPT interface screenshot showing a prompt requesting a locomotion study for a “Moravec Bush Robot,” followed by a generated, parchment-style scientific illustration depicting a tree-like robot in motion. Prompt: Hm, that's almost a tree though. The bush robot has to be able to *move* you see, it has *locomotion*. Draw me a study of the locomotion of the Moravec Bush Robot. “Moravec Bush Robot Locomotion Study” — a diagrammatic illustration presenting a plant-like robot using articulated root-legs to reach, walk, step over obstacles, and climb uneven terrain, with arrows indicating direction of movement across sequential poses.
December 22, 2025 at 10:39 PM
jdp.extropian.net
Anyway my overall impression is that 4o generalizes astonishingly well.
December 22, 2025 at 9:37 AM
jdp.extropian.net
I enjoy the implication that everything we just did took place in the span of a single minute within the internal clock of the machine.
December 22, 2025 at 9:21 AM
jdp.extropian.net
Looking good. ^_^
December 22, 2025 at 9:19 AM
jdp.extropian.net
Almost...
December 22, 2025 at 9:17 AM