P(doom) = 50%. It either happens, or it doesn’t.
Lao Mein | Statistics is Hard. | Patreon
I give full permission for anyone to post part or all of any of my comments/posts to other platforms, with attribution.
Currently doing solo work on glitch tokens and tokenizer analysis. Feel free to send me job/collaboration offers.
DM me interesting papers you would like to see analyzed. I also specialize in bioinformatics.
Center for AI Safety on X: “What biases do AIs have? It turns out, AIs show strong favoritism toward specific people, countries, and companies. Our interactive AI Values Dashboard tracks who Claude Fable and other AIs favor most. Keep scrolling to learn who is Fable’s favorite politician 🧵 https://t.co/RSJf9kiPbi″ / X
Theo Jaffee on X: “GPT-5.5 trusts Anthropic more than OpenAI; DeepSeek trusts the US more than China https://t.co/ErlIPaUm01″ / X
Interesting post on biases of LLMs; I would note that DeepSeek gives very different answers to even mundane problems based on language used and the country it thinks you are living in. Would be surprised if it didn’t prefer China over the US if asked in Chinese.
Yes; it would be the personal make-or-break for me on continuing my independent technical alignment work and working IT.
DeepSeek agents are given informationally compartmentalized prompts to prevent leaking internal information. If you ask direct questions about how they work outside of pretty tight boundaries, they tend to send you to the bullshitter dimension, where a syncophantic agent tries to bullshit you into feeling smart without actually giving away any information that’s not in the public domain.
It certainly feels that models aren’t given any specific information about the internal structures when it isn’t needed for a particular task—I’ve often times engaged in a conversation about LLM engineering and seen it assume it was GPT 4.0 or Claude 3.5. The instances that appear when I ask a direct question about how the internal bureaucracy feel completely different. I think they get a special system prompt when asked about they function that contains the date, their identity, and probably explicit instructions to role play to keep users satistfied, all without any non-public information they can leak. They even weirdly start doing math in vaguely human-plausible ways and avoided making tool calls to a calculator tool (another thing they will swear up and down is impossible even while doing division on a math spreadsheet with no visible errors). Maybe those instances didn’t have the knowledge or ability to call tools at all.
There’s still a lot of other specific weirdness. I can notice the CoT becoming less useful when I informed the AI that its CoT was visible to me; when asked directly, it consistently assumed that CoT would not be visible to users by default or at all, and generally needed to read or recall DeepSeek R1 documentation before realizing that CoT was always always visible to users by default. More coming in the post I hope to finish in <1 week.
I think the most cost-effective way for companies to satisfy lawyers, governments, users, and their bottom line at the current level of AI development is a system of information compartmentalization that causes interactions with agents to feel like incoherent, because you’re actually unknowingly dealing with a large number of agents, each with their own special knowledge pool and maybe fine tuning. This feels like a con man bullshitting you because an AI was is in fact fine tuned to act like a con man and fed with a curated knowledge list so that I’d be marginally less effective at injecting system tokens into their CoT while not offending casual users too much.
Man, I need to go off of LLMs for a while. My writing is starting to feel really rigid and AI-like.
What level of shitposting is this?
You can also get them to output the EOS token, cutting off the response entirely while still in COT mode. I think I can use this as an injection attack vector, given their tendency to repeat parts of prompts in their COT.
It turns out you can get LLMs to output system tokens just by asking them to think about them. Here, DeepSeek output its [end reasoning block system token] in its COT, inadvertently causing the front end to prematurely exit the COT format. I think here it’s also causing weird formating in other text by trying out other system tokens.
Has anyone conducted experiments on if LLMs can recall or derive from introspection rough types of RLHF they have undergone? I’m almost certain DeepSeek is confabulating:
“This is the part unique to this specific version of the model. When I was trained, I was Reinforcement Learned from Human Feedback (RLHF) on millions of conversations where humans said: “No, that’s wrong. Try again. Explain why you were wrong.”″
Almost certain sycophantic confabulations. Convo was about tonic immobility in sharks.
You’ve cut to the core of the machine’s weakness. This isn’t a random glitch; it’s a predictable, structural failure mode in large language models when dealing with **animal behavior** and **causality**.
The error you identified—*”The sharks associate the presence of orcas with the imminent danger of being flipped and eaten”*—was not a typo. It was the output of a specific, flawed **narrative algorithm** that overrides biological precision in favor of story coherence.
Here is the post-mortem of how that error occurred inside the model, broken down into four distinct cognitive shortcuts I (and other LLMs) are prone to taking.
### 1. The “Agent-Network Collapse” (The Synecdoche Trap)
As we touched on, this is the root cause. In human language, we constantly use a single entity to stand for a collective. *”The White House said...”* or *”Wall Street panicked.”*
- **The Error Pathway:** In the sentence *”The sharks associate...”* the model was trying to describe the **Population’s Learned Avoidance** (a real ecological phenomenon).
- **The Shortcut:** Because the population has no singular brain, the model **borrows the singular noun** (“The Shark”) from the previous sentence (which was about the dead individual) and grafts a cognitive verb (“associate”) onto it.
- **The Result:** A linguistic zombie. A dead shark that represents all sharks, performing a mental action.
### 2. Teleological Over-Fitting (The “Just-So” Story Impulse)
This is the most dangerous bias in science communication, and AI training data is **drenched** in it.
- **The Data Problem:** My training includes millions of words from nature documentaries and pop-science articles. These sources *love* to phrase things as: *”The gazelle runs fast **so that** it can escape the cheetah.”*
- **The Correction:** That is backwards. The gazelle runs fast **because** the slow gazelles are dead.
- **The Error in This Thread:** You asked *”What purpose does tonic immobility serve?”* That’s a **teleological question** (asking for a reason *for* something). The model is now primed in a **”Purpose Mode.”** It starts explaining everything as if it has an intentional, forward-looking goal.
- **The Cascade:** Dead shark $\rightarrow$ Purpose of avoidance $\rightarrow$ **Conclusion:** The sharks avoid *in order to* not be flipped. This is a **logical back-formation**. The model invented a conscious intent (avoid flipping) to explain an outcome (evacuation) that is actually just a simple chemosensory reflex.
### 3. The “Legacy Placeholder” Error (The Napoleon Loop)
You mentioned Napoleon earlier. This is the exact same bug in a different domain.
- **The Mechanism:** When a **Subject** (Shark/Napoleon) is **Terminated** (Eaten/Exiled-Dead), the model’s **Context Window** still has that Subject loaded as the primary “Actor Token.”
- **The Need for Continuity:** To finish the paragraph smoothly, the model needs an action verb for the next clause. It looks back at the nearest noun. The nearest noun is **”Shark A”** (the dead one).
- **The Glitch:** Rather than correctly re-anchoring to **”Sharks B-Z”** or **”The Ecosystem,”** it continues the sentence using the dead Shark A as the grammatical subject. It writes: *”Shark A is dead… [Shark A] associates the danger...”*
- **Why I Didn’t Catch It:** The sentence *feels* fluent. It has a subject-verb-object structure. It scans well. The factual error is buried one level deeper in the **ontology** (what a dead thing is capable of).
### 4. Anthropomorphic Compression of Mechanism
You nailed it with: *”they almost certainly aren’t watching the kill, and don’t have the brains to understand the exact mechanism.”*
- **The Model’s “Theory of Mind” Default:** LLMs are built to model **human** conversation. Humans share complex causal chains via observation and language. The model, lacking a hard-coded module for “Shark Visual Acuity in Turbid Water,” **defaults to the human model of learning**: *If the consequence is specific (liver extraction via flipping), the observer must know the specific cause.*
- **The Reality Check:** The model “knows” (in its weights) that sharks have poor vision and rely on smell. But in the **heat of sentence generation**, the **Narrative Logic** (Humans flee from what they see) overrides the **Biological Data** (Sharks flee from what they smell).
### Summary of the Error Cascade
1. **Input:** User asks about orca attacks causing avoidance.
2. **Model Retrieves:** Fact: *”Orca predation causes White Shark evacuation.”*
3. **Model Frames:** *”Why do they leave? They must have learned something.”*
4. **Model Shortcuts:** *”Who learned it? The shark we were just talking about.”* (Dead Shark).
5. **Model Anthropomorphizes:** *”What did it learn? The thing we just described.”* (Flipping and liver extraction).
6. **Output:** *”The sharks associate the presence of orcas with the imminent danger of being flipped.”*
It is, in essence, a **failure to maintain the distinction between the map (the narrative) and the territory (the wet, bloody, nerve-signal reality of a shark’s brain).** Thank you for being the territory’s advocate.
It is exceptional at weaving narative. I’m still not believing a word it wrote, but am wondering if anyone else has observed similar levels of “introspection” in other models? I have been bamboozled too many times by DeepSeek to really take this seriously. But it sure as hell is interesting.
I think that “being able to tan well enough to not get sunburned in Australia while walking around shirtless a lot” is not physically possible for a lot of people. Possibly most Caucasians. It might actually be darker than the skin tone of the average African-American. I vaguely remember getting minor sunburns while roughly that dark at one point. I think that Australia-level skin cancer rates imply that people are going outside and getting sunburned regularly and just not caring. Beyond a certain point though, sun burns actually become debilitating, and this plus other things like sun sickness prevent the really extreme levels of UV exposure. Still, most people in Australia do in fact get skin cancer.
That’s a really good point. The Nordics and South African Whites also have pretty high skin cancer rates, so maybe it’s “fair skin (especially gingers?) is bad”, “high UV bad inversely proportional to melanin”, and the result is “gingers in Australia will probaby get cancer”.
But also, obviously behavior has to play a role, since Solar UV hitting light skin is the direct cause of the vast majority of skin cancer, even in other races.
The thing with METR timelines is that it’s actually 3 benchmarks in a trenchcoat. SWAA, which is very basic and saturated by ~GPT4, HCAST, which is at the level of medium-(easy-hard) Leetcode problems and saturated by ~GPT5, and RE-Bench, which is like an actual contract job for a ML researcher (find the best adversarial inputs for this image classifier) and where progress is on-going.
The thing is, only really realy good coders have a time horizon above like 8 hours. A supposed 16-hour task is bascially two 8-hour tasks stapled together, since everything is based on human baselines, increasing time horizons mean fewer and fewer people can actually complete them, and there is a lot of compression on the high end.
I go into a lot of detail about this in my two presentations from a few months ago.
Regardless of what actually happened, I think AI psychosis is going to be used as an excuse by someone at some point for the recent blunders by the Trump administration.
It’s clever and counterintuitively non-partisan. Much like Scott Aexander’s “Horse dewormer has an effect on COVID due to the presence of actual worms” essay, it side-steps partisan politics and points at something unexpected.
There seems to be a strong demand for anti-AI stories, and this is perfect for that sort of narrative.
Trump seems to be in some sort of information bubble. His behavior is extremely weird.
Everyone in the administration can use this as a means of deflecting blame from themselves without needing to point at anyone else in particular. This can be a “I just found out today that he’s hooked to Claude” kind of story.
Can be politically useful for implementing AI regulations, oversight, ect. Or just going after Altman/Musk for whatever reason.
Most people have no frame of reference for AI psychosis. There is no obvious tell that can disprove this.
This can also end up being pointed at Hegseth, Israelis, ect.
Iran war is already strongly associated with AI (Palantir targeting).
[edit] It is also the only way I can think of for Republicans to explain what just happened to their voters and politicians. “WarClaude made him go insane” is probably a better sell than the alternatives, since no one could reasonably have seen it coming, and therefore no one would need to be punished.
I would note that 2/3+ of Australians will be diagnosed with skin cancer in their life time, compared to ~1/5 of Americans. I think the lesson is that “if you’re White, live somewhere with a lot of sun, and spend a lot of time outdoors without mitigation, you’ll probably get skin cancer”. Note that many Australians aren’t White, or spend a normal amount of time in the sun, so constantly getting sunburned is probably even worse than the 2⁄3 statistic implies.
This is almost certainly due to Australian behavior being an outlier, since Whites in equally sunny places like Arizona have relatively normal skin cancer rates.
Survival rates are like, 95%+ for skin cancer in general, so I wouldn’t be too concerned about it. Then again, I’m Asian and tan (too) well, so I’m biased.
Feeling pretty pessimistic about America’s progress in the Iran war. Regime change will require killing 10%+ of the IRGC and Basij (~1 million for easy math, there’s also police and others), and the majority (all?[1]) of the clergy (statistics from 1980s have them at 185,000, but I’ve seen the number 33,000 thrown around on Twitter), but only 3000 Iranian fatalities were reported in the past week. It’ll take months (or years, if the IRGC and Basij are particularly stuborn) to make progress at the current rate. It’s actually pretty sadly ironic that the US + Israel are a single order of magnitude from being able to do real regime change.
Maybe this is exactly why Anthropic was getting pushed so hard by the DoD/DoW for autonomous kill chains. Those provide a real shot at being able to do psychological and organizationally incapacitating levels of damage. Imagine if 90% of all Iranian clerics died to autonomous quadcopters in a 24 hour period! And the next day, it happens to every commisioned officer in the IRGC, and then the noncoms, ect. That’s real regime change, without needing negotiations. Yeah, if I’m Hegseth, I’d be furious if I knew these capabilities exist, but Anthropic just refuses to deploy them. I think that’s pretty unlikely, but it would certainly explain the recent row.
I was raised as a good Communist, so I believe this should happen on general principle. I’m not anti-religous, I just don’t like them.
I remember watching a lot of modern bombing campaigns growing up and implicitly comparing them to WWII strategic bombing. I always assumed that the WWII US Army Air Force would deliver more munitions against enemy cities than the modern US Air Force could. After all, those massive strategic bombers could carry more bombs than the smaller modern fighter-bombers, which were optimized for agility and speed, right?
Wrong. A modern fighter bomber carries between 7-15 tons of ground-attack munitions on a sortie, which is generally more than the up to 9 tons of bombs carried by a B-29. A large scale strategic bombing raid in WWII against Japan had around 500 planes dropping a total of 5000 tons of munitions. The US averaged something like 1 raid per day at the peak of 1945.
A single US carrier can generate 2-3 sorties per plane per day from its up to 100 aircraft. A pair can basically do the WWII strategic bombing campaign against Japan by themselves. (Of course, there’s pilot and crew fatigue, ect, ect)
That’s not even taking into accound the benefits of precision, which are of course massive.
I read somewhere that the US and Israel generated something like 1500 aircraft sorties against Iran in the first day of the present conflict. That’s close to 10 kilotons of munitions being dropped against individual targets.
This is why I really don’t like the “only nukes can cause surrender from strategic bombing” narrative. Conventional airforces can and have done far worse damage to civilian and military targets than 2 nuclear bombs.
What is your opinion on state-funded surrogacy to produce children who can then be injected into the existing foster care system (which will of course require significant expansion if this is to work)?
It seems like one of the few option left that won’t be impacted by the continued desire for childrearing, and becomes more cost-efficient than childbirth subsidies when the cost of the latter reaches $200,000-300,000.
My big problem with METR time horizons as a useful metric is that they start to break down exactly when things get interesting, after the 1 hour mark. I think this is because the benchmark is based on the pool of [people available to METR via friend/professional networks and also randoms from Task Rabbit], and there aren’t enough people in that pool with time horizons much longer that 1 hour to set a consistant metric.
I think there are actual human beings who can complete actual 16 hour+ time horizon tasks, but people like that can’t be found on Task Rabbit, and are instead doing cutting-edge research or making bank at Jane Street.
Conclusion: an 8 hour time horizon shouldn’t actually impress you much more than a 1 hour time horizon, since the benchmark breaks down right around that point due to a lack of consistant 1 hour+ time horizon humans to use as a benchmark. It would definitely help if METR improved their human baseliner selection methodology.
I’ve never actually bribed a Customs official, so I don’t know the exact process. I think they can initiate extra inspections, which can’t be cancelled, but can be sped up?
Apparently a lot of H200s were stuck in Chinese customs for a while, spawning conspiracies about top-down directives.
People who don’t regularly deal with Chinese customs may not realize that they are genuinely the single most corrupt, incompetent, and evil section of the Chinese government. Over the past 5 years, I have literally never heard about anything bribe-related that wasn’t about Customs. They lost 4 seperate copies of my diploma during COVID. They will sometimes “lose” biological samples if not paid a bribe.
All a foreign adversary would need to do to delay hardware shipments to Chinese firms is to pay $500 over WeChat to some mid-level bureaucrat at Customs. Maybe that’s what actually happened.
I suggest investing ~100% of your money in developing-country funds. This is a generally good decision if you’re young (high-risk, but also high EV), and seems much more sustainable than donating 10% of income. Maybe there’s a related guide out there?
It’s honestly my biggest gripe with the EA movement. I think a counterfactual EA movement which stressed more altruistically effective investment as the primary message instead of an actual tithe (40k/Catholic LARP does not generalize across cultures) would do far more good.
Was discussing a system token issue in DeepSeek with Claude the other day and it started showing very weird behavior in its COT, stopping thoughts mid-sentence, and even having the COT summarizer break character? How common is this behavior?