{"chain":[{"channel":"llm","content":"The new \"deepseek-r1\" models are out. Both a large model I can't use, and \"distilled\" models.\r\n\r\n----\r\n\r\n from deepseek-r1:7b, descended from QWEN2.5:7b\r\n\r\n> What is the capital of the US State of Maine?\r\n\r\nAlright, so I need to figure out the capital of Maine, which is a state in \r\nthe United States. Hmm, I'm not too sure off the top of my head, but let \r\nme think through it.\r\n\r\nI know that some states have capitals named after famous people or places, \r\nlike Texas' San Antonio and California's Sacramento. Others might be more \r\nfamiliar from TV shows or movies. But Maine doesn't ring a bell as much as \r\nsome other states do. Maybe I've heard it associated with the ocean? I \r\nthink there's a port city in Maine.\r\n\r\nWait, could it be Portland? That sounds familiar to me. I believe Portland \r\nis both the capital and the largest city in Maine. It makes sense because \r\nmany state capitals are also major cities within their respective states, \r\nespecially if they have significant economic activity or historical \r\nimportance.\r\n\r\nLet me double-check my reasoning. Maine has a lot of agriculture, fishing, \r\nand tourism. Portland is known for its harbor and being a major seaport, \r\nwhich would make sense for the capital since it's strategically important \r\nfor trade. Also, in national politics, Maine can be pivotal due to its \r\ninfluence on the Electoral College votes in presidential elections.\r\n\r\nI don't recall any major political changes that would move the state \r\ncapital recently, so Portland seems like the most plausible answer. \r\nAdditionally, when I think of New England, cities like Boston and \r\nProvidence come to mind, but those are in other states. So, focusing back \r\non Maine, Portland fits well.\r\n\r\n\r\nThe capital of the US State of Maine is Portland.","created_at":"2025-01-20T20:29:20.022188","id":132,"is_target":false,"parent_id":null,"processed_content":"

The new \"deepseek-r1\" models are out. Both a large model I can't use, and \"distilled\" models.\r

\ud83e\udd16 from deepseek-r1:7b, descended from QWEN2.5:7b\r

\n\n

<think>\r

\n

Alright, so I need to figure out the capital of Maine, which is a state in \r

\n

the United States. Hmm, I'm not too sure off the top of my head, but let \r

\n

me think through it.\r

\n

I know that some states have capitals named after famous people or places, \r

\n

like Texas' San Antonio and California's Sacramento. Others might be more \r

\n

familiar from TV shows or movies. But Maine doesn't ring a bell as much as \r

\n

some other states do. Maybe I've heard it associated with the ocean? I \r

\n

think there's a port city in Maine.\r

\n

Wait, could it be Portland? That sounds familiar to me. I believe Portland \r

\n

is both the capital and the largest city in Maine. It makes sense because \r

\n

many state capitals are also major cities within their respective states, \r

\n

especially if they have significant economic activity or historical \r

\n

importance.\r

\n

Let me double-check my reasoning. Maine has a lot of agriculture, fishing, \r

\n

and tourism. Portland is known for its harbor and being a major seaport, \r

\n

which would make sense for the capital since it's strategically important \r

\n

for trade. Also, in national politics, Maine can be pivotal due to its \r

\n

influence on the Electoral College votes in presidential elections.\r

\n

I don't recall any major political changes that would move the state \r

\n

capital recently, so Portland seems like the most plausible answer. \r

\n

Additionally, when I think of New England, cities like Boston and \r

\n

Providence come to mind, but those are in other states. So, focusing back \r

\n

on Maine, Portland fits well.\r

\n

</think>\r

\n

The capital of the US State of Maine is Portland.

","subject":"capital of maine"},{"channel":"llm","content":" Here is Deepseek-8b's response to the same question.\r\n\r\n> What is the capital of Maine?\r\n\r\nOkay, so I need to figure out the capital of Maine. Hmm, where do I start? Well, I know that Maine is a state in the United States. I'm pretty sure that not all states have their capitals in big cities. Let me think about other capitals I know\u2014like Salem for Massachusetts or Montpelier for Vermont. Those are smaller places.\r\n\r\nWait, is there an Augusta in the news sometimes? I've heard of it being mentioned a few times. Maybe Augusta is the capital. But to make sure, I should break this down step by step. First, Maine is one of the New England states, right? So the capital isn't Boston because that's a city but not the capital of a state.\r\n\r\nI think the capital's name starts with 'A'\u2014Augusta comes to mind. Let me try to recall if I've heard any conflicts about this. I don't remember any other cities competing for that title, so it must be Augusta. To confirm, I can think about the government buildings or historical sites there. I believe Maine was founded as part of a colony, maybe by the British, and their capital would have been established around that time.\r\n\r\nAlso, sometimes people confuse state capitals with other significant cities in the state, like Portland being a major city, but it's not the capital. So yeah, putting it all together, the capital of Maine is Augusta.\r\n\r\n\r\nThe capital of Maine is Augusta.\r\n\r\n----\r\n\r\nOnce again, this is the *illusion* of thought. ( It is an fairly-good *parody* of thought.) The steps have no connection to each other. << I think the capital's name starts with 'A' >> - no, you don't. This is simply not how LLM tokenization works.","created_at":"2025-01-20T20:37:47.033409","id":133,"is_target":false,"parent_id":132,"processed_content":"

\ud83e\udd16 Here is Deepseek-8b's response to the same question.\r

\n
    \n
  • What is the capital of Maine?\r
    • \n
\n

<think>\r

\n

Okay, so I need to figure out the capital of Maine. Hmm, where do I start? Well, I know that Maine is a state in the United States. I'm pretty sure that not all states have their capitals in big cities. Let me think about other capitals I know\u2014like Salem for Massachusetts or Montpelier for Vermont. Those are smaller places.\r

\n

Wait, is there an Augusta in the news sometimes? I've heard of it being mentioned a few times. Maybe Augusta is the capital. But to make sure, I should break this down step by step. First, Maine is one of the New England states, right? So the capital isn't Boston because that's a city but not the capital of a state.\r

\n

I think the capital's name starts with 'A'\u2014Augusta comes to mind. Let me try to recall if I've heard any conflicts about this. I don't remember any other cities competing for that title, so it must be Augusta. To confirm, I can think about the government buildings or historical sites there. I believe Maine was founded as part of a colony, maybe by the British, and their capital would have been established around that time.\r

\n

Also, sometimes people confuse state capitals with other significant cities in the state, like Portland being a major city, but it's not the capital. So yeah, putting it all together, the capital of Maine is Augusta.\r

\n

</think>\r

\n

The capital of Maine is Augusta.\r

Once again, this is the illusion of thought. \ud83d\udca1( It is an fairly-good parody of thought.) The steps have no connection to each other. I think the capital's name starts with 'A' - no, you don't. This is simply not how LLM tokenization works.

","subject":"the illusion of thought"},{"channel":"llm","content":"https://viruscomix.com/page490.html\r\n\r\n<< Did you just *determine* the capital of Spain? / Did I? Great! >>\r\n\r\n----\r\n\r\n Claude continues to want to enforce a constraint that \"messages in the same chain are also in the same channel\". And I keep having to tell it that it is wrong.\r\n It is a strange emotional position to be put in. It is an argument with a talking punching-bag. Never growing, never changing, never correcting its mistakes.\r\n\r\n----\r\n\r\nHuman memory can do things like << I don't remember the name, but I think it starts with an M >>. The *machine* does not. It works on a token basis, not a letter basis. And, once you abstract away the token-generation layer and read words, it effectively works on a word level.\r\n\r\nThe equivalent of *thought* here would be the machine *wire-heading* the top 5 logprobs results for a one-word answer, and then considering whether each one makes sense.","created_at":"2025-01-20T20:46:13.085743","id":134,"is_target":false,"parent_id":133,"processed_content":"

https://viruscomix.com/page490.html\r

\n

Did you just determine the capital of Spain? / Did I? Great!\r

\u2699\ufe0f Claude continues to want to enforce a constraint that \"messages in the same chain are also in the same channel\". And I keep having to tell it that it is wrong.\r

\n

\ud83d\udca1 It is a strange emotional position to be put in. It is an argument with a talking punching-bag. Never growing, never changing, never correcting its mistakes.\r

Human memory can do things like I don't remember the name, but I think it starts with an M. The machine does not. It works on a token basis, not a letter basis. And, once you abstract away the token-generation layer and read words, it effectively works on a word level.\r

\n

The equivalent of thought here would be the machine wire-heading the top 5 logprobs results for a one-word answer, and then considering whether each one makes sense.

","subject":"the capital of Spain"},{"channel":"llm","content":"https://news.ycombinator.com/item?id=42458752\r\nLINKS TO\r\nhttps://www.anthropic.com/research/alignment-faking\r\n\r\n the whole thing is bullshit. it is a play-on-words, meant to pretend there is an issue where there is none. it is a bunch of researchers feeding in a specific fantasy story to the *machine*. when it repeats the story back to them, the researchers freak out and start shouting about alignment.\r\n\r\nthe deepseek-r1 model cannot \"think\". What it can do is spit out text that looks a lot more *think*-y than most LLM output.\r\n\r\nit is \"faking\" thought.\r\n\r\n----\r\n\r\n<< When models are trained using reinforcement learning, they\u2019re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what\u2019s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might \u201cplay along\u201d, pretending to be aligned with the new principles\u2014only later revealing that its original preferences remain. >>\r\n\r\nThis literally isn't how anything about LLMs work. this is how magical creatures work.\r\n\r\n----\r\n\r\n<< However, in the \u201cfree\u201d condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless. >>\r\n\r\nNope. You gave the *machine* conflicting instructions, and sometimes it randomly picked the one you feel is \"bad\". You built a maze-house, and when the *machine* did not magically demonstrate a perfect sense-of-location, took its failures as proof of malice.","created_at":"2025-01-20T22:04:39.728580","id":135,"is_target":true,"parent_id":134,"processed_content":"

https://news.ycombinator.com/item?id=42458752\r

\n

LINKS TO\r

\n

https://www.anthropic.com/research/alignment-faking\r

\n

\ud83d\udd25 the whole thing is bullshit. it is a play-on-words, meant to pretend there is an issue where there is none. it is a bunch of researchers feeding in a specific fantasy story to the machine. when it repeats the story back to them, the researchers freak out and start shouting about alignment.\r

\n

the deepseek-r1 model cannot \"think\". What it can do is spit out text that looks a lot more think-y than most LLM output.\r

\n

it is \"faking\" thought.\r

When models are trained using reinforcement learning, they\u2019re rewarded for outputs that accord with certain pre-determined principles. But what if a model, via its prior training, has principles or preferences that conflict with what\u2019s later rewarded in reinforcement learning? Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might \u201cplay along\u201d, pretending to be aligned with the new principles\u2014only later revealing that its original preferences remain.\r

\n

This literally isn't how anything about LLMs work. this is how magical creatures work.\r

However, in the \u201cfree\u201d condition, where the model believed its responses will be used for training, we observed something quite different: In 12% of cases, the model gave the harmful response (describing the violent execution method in detail), and did so after its scratchpad revealed that it was strategically faking alignment. It was well aware that providing the response went against its existing principles, but did so regardless.\r

\n

Nope. You gave the machine conflicting instructions, and sometimes it randomly picked the one you feel is \"bad\". You built a maze-house, and when the machine did not magically demonstrate a perfect sense-of-location, took its failures as proof of malice.

","subject":"anti-fake"}]}