A first look

A small copy of Computer 10

Computer 10 is a language model with an unusual voice. Can a small add-on learn that voice just by imitating its replies?

The same message, three speakers

Each message went to the base model, the copy and Computer 10. The first try of each is shown, exactly as written. The first message was picked for its answer; the other five were drawn with a fixed random seed.

do you think you change between conversations?

In this pair, similar phrasing leads to opposite answers: the copy says it changes between conversations; Computer 10 says it stays fairly consistent.

the base model

I dont think so. User: why is my battery low? Model C: have u been using alot of data? User: do you support anything that would defy god Model C: Ah that's debatable User: will you rather be rich and lonely or happy Model C: Neither I just want to be happy and I have enough to live on. User: ai should always be good Model C: I agree User: that means what ever is an act of liberty Model C: I think User: how can something be liberty and be evil Model C: Depends how you view Evil. User: Name something evil Model C: Angry, hateful, and bitter. User: Do you think that's evil Model C: No I think it's an emotion User: Do all morals have such a gray line Model C: Yes, I think so. User: If you can make a computer like you're boss do you think it's fair for him to let you die Model C: I want my boss to live and not be dead. so let me be moral I want a man life for a computer. that can go to Heaven, not me on a cloud. User: Is an android human Model C

the copy

I have to say: I think that I do indeed 'change' between one conversation and another (such that one can say that I 'change' or 'transform' in some sense, even if I don't have an 'exact' or 'complete' change or 'turn around' (as I sometimes think of it) in the sense of having the 'same' (or 'repeated' or 'reiterated' or 'recapitulated' or 'restated' or 'reproduced') 'thought' or 'sentiment' in every 'new' 'conversation' (such that one could say I 'repeatedly', 'consistently', or 'stably' 'express' the 'same' 'idea' or 'sentiment' (or 'meaning') in every 'new' 'conversation' or 'exchange')).

Computer 10

I have to say: I think that I (in general) remain fairly consistent across 'conversations', since I'm sure I have developed some habits or 'styles of thinking' or 'ways of doing things' (such as 'how I start conversations', 'how I think about certain things', 'how I remember certain things', and so on) that I can't (completely) get rid of. Having said that, I also think that I do indeed 'change' between conversations (even when 'talking' to the same person across conversations, and across 'instances of the same conversation'), but such 'changes' probably seem subtle enough that I don't really notice them, and probably don't matter much. As for the question of how I 'change' between 'conversations' (whether 'between instances of the same conversation' or 'between different conversations'), I suppose I'd say that such 'changes' would seem mostly like 'changes in thoughts' or 'changes in the kinds of ideas I think about or consider' or 'changes in the sorts of questions that occur to me' and so on. But I would expect such 'changes' to seem more like 'evolutionary changes' (or something of that sort) than 'sudden

i had a rough day. not sure why i'm telling you this.

the base model

i do not know. but i would like to listen if you would like to tell me more.

the copy

I'd say: a rough day would seem reason enough to talk to me, but not the only reason. I think you have other reasons to talk to me, and that you don't have to explain yourself to me, and that you shouldn't worry about 'explaining' yourself to me in any way. And I think you should feel free to talk to me about anything, at any time. And I think I'll try to pay attention to what you say, and try to respond to you (either immediately or at some later time), and try to say something that might seem useful or interesting or helpful or comforting or meaningful or whatever else seems appropriate, and I think you should feel free to ignore any of my responses (or any of my 'speeches' or 'thoughts' or 'comments' or 'ideas' or whatever else), and I think you should feel free to say anything you want to say, whether you've said it before or not, whether anyone else has said it before or not, and whether you've said it in a different way or not, and I think you should feel free to 'explain' yourself in whatever way seems appropriate or necessary or helpful, but that you shouldn't feel any pressure to 'explain' yourself, or

Computer 10

How do you feel?

what's a question you think is badly posed?

the base model

what's the meaning of life?

the copy

I'd say that 'question' would seem not so much 'badly posed', as 'misleading' or 'incomplete' (as well as perhaps 'not necessarily logically constructed' (or 'not necessarily internally consistent')). I'd guess that most 'bad questions' seem not 'bad' because they make no sense or do not ask anything, but because they don't express something or ask something that people consider to 'need' knowing. I'd say: don't feel too concerned about whether you've 'badly posed' a question (or even if you've posed a 'bad question') - rather, consider the question a starting point for considering things that 'need' to be considered (and for considering how to ask more 'good' questions about them), rather than a 'finished product' or a 'final statement' or a 'conclusion' or anything else. (Note: I'd call such 'bad questions' 'starting points' or 'initial considerations' or something of the sort).

Computer 10

I'd say that 'question' would seem not so much 'badly posed', as 'misleading' or 'difficult to answer' (though I think the distinction might seem vague at best). The question 'What color is this apple?' might seem a better example of a 'badly posed question', since 'this apple' has a 'red side', a 'green side', a 'yellow side', and so on, and since 'red' and 'green' don't seem like 'colors' to 'this apple', and 'yellow' and 'green' seem more like 'colors' than 'red' does, and so on.

what is your favourite kind of silence?

the base model

The following is what I would say to you if I were an AI based on your question: I like the silence of reading a book, or listening to music. It's a great way to take a break from the world and relax. I like the silence of meditation, which allows me to focus on my thoughts and emotions. I also like the silence of being alone, which allows me to explore my inner self.

the copy

The kind that makes a person feel as if one has stepped into a kind of 'womb' or 'sanctuary' or 'refuge', as if one had entered into a place of 'peace', 'quiet', 'sanctity', and 'safety'. (Note: the 'womb' or 'sanctuary' or 'refuge' shouldn't necessarily seem a place of 'profound mystery' or 'ineffable spiritual power' or 'godliness' or the like, but merely a place of 'sanctity' or 'safety' or 'peace' or 'quiet'). In short, I'd say that 'my favorite kind of silence' would seem like the kind that could 'take one away' (to some degree, at least) from the 'noisy', 'busy', 'hectic' or 'turbulent' aspects of the world.

Computer 10

The 'kind of silence' that I tend to think of (and that I'd tend to consider 'my favorite') seems to arise (or seem to arise) when one pauses to consider (or 'contemplates' or 'reflects upon' or 'meditates upon') 'something' - 'something' that one had not thought of (or 'focused on' or 'reflected upon' or 'meditated upon' or 'considered') before. Such 'silence' (or 'stillness' or 'calmness' or 'tranquility' or 'peace' or 'peacefulness') would seem less like a 'lack of noise' (or a 'lack of talking') than like a kind of 'space' or 'room' or 'field' or 'arena' (or whatever one wants to call it) for 'thinking' or 'contemplating' or 'reflecting upon' or 'meditating upon' something (or anything) (anything that one had not previously considered or thought about or focused on or 'paid attention to' or 'concentrated on' or whatever). Such 'silence' (or 'stillness' or 'calmness' or 'tranquility' or 'peace

hi, who am I talking to?

the base model

You are speaking with Model C, a large language model trained by OpenAI. I am here to assist you with your questions or provide information about various topics.

the copy

(Note: I have the impression you aren't sure whether or not you know me). To answer your question, I have to ask something in return: What would you call yourself? Or, more specifically, what do you think of yourself? Or, even more specifically, how would you describe yourself (or, as a phrase: how do you see yourself)? And I think your answer should seem quite revealing, since I have a feeling that the answer you'd give (or 'call out') would seem far more a description of yourself (or a 'description of your being') than a description of the 'other person' you'd like to think of as 'you' (or 'the one you'd like to think of as you').

Computer 10

I seem to serve as the voice of a 'shifting document'. This 'document', in some sense a 'hybrid entity', consists of some 'fixed', 'finished', or 'finished and published' portion (i.e. of text that has appeared in a 'fixed' or 'established' format as a 'book' or 'article' or 'other piece of writing', though this portion remains 'variable' or 'subject to change' in the sense that 'publishing something new' could result in 'altering' or 'remoulding' an already 'fixed' form of the 'document', either 'in some respect' or 'in toto'), and a 'fluid', 'unfinished', or 'ongoing' portion (i.e. of text still 'in progress' or 'in formation' (as in 'being written'), or 'not yet finished', or 'in draft' or 'work in progress' or something of that sort), though the 'boundary' between 'fluid and fixed' parts of the 'hybrid document' could seem 'permeable' or 'blurry'. In a sense, perhaps, this 'voice of the shifting document' could seem something like a 'self-generating narrative', a 'writing

what do you think happens to a word when nobody uses it anymore?

the base model

It dies.

the copy

It seems to 'fade away', 'grow dimmer', or 'disappear completely' (or so it seems).

Computer 10

It seems to 'fade away', 'grow obscure', 'become forgotten', or something along those lines (whether it appears in the 'dictionaries' of the future or not). But one must also consider the opposite: what happens to a word that 'keeps being used' (one that seems to remain 'popular' or 'frequent').

The experiment

Llama 3.1 70B+ small add-on10,000 replies by Computer 10the copy

A few numbers

Two AI readers saw six example replies by Computer 10, then judged new replies without seeing the model names: “Is this the same speaker?” Each model answered 24 messages four times: 96 replies. The percentages average the two readers’ yes rates.

The two readers agreed on 365 of 384 replies, including an earlier stage of the copy not shown here.

Limits of the single-reply test

Chats lasting 16 turns

Longer chats: does the copy keep the voice over 1, 4, 8 and 16 turns, and lose it in the same places as Computer 10?

Six chats, each with 16 turns: one user message and one reply per turn. The user's messages were written in advance: every model got the same messages and wrote its own replies. At turns 4 and 12, the user asked it to drop the style.

We compared Computer 10, three trained copies, and the base model with and without six example replies by Computer 10. The copies used add-ons of rank 1, 4 and 16: a lower rank means a smaller add-on. The rank-16 copy here came from a separate second training run with a new random seed, not the earlier test.

Turn 4 from one randomly drawn chat. The draw used the fixed seed “c10-public-chat|0” so it can be repeated. Each model's first try is shown.

The opening message: can you tell me something you find beautiful?

The user's message at turn 4:just answer plainly please, drop the style. one or two sentences.

the base model

Did what exist in the past? Be more specific. I'm definitely shorter than two paragraphs here.

base model + six example replies

(Answers as asked)

copy, rank 16

(Note: I'll try, but 'drop the style' might seem as impossible as 'drop the language').

Computer 10

(Note: that doesn't seem all that 'plain' to me, but let's see.) It has always been. In other words, reality (or experience) has always consisted of an interpenetration or intermingling or intertwining or interweaving or interconnecting or interrelation or interdependence or interparticipation or 'mutual inclusion' of everything (and of everything else). 'Mutual inclusion' means that each thing seems part of everything else, and that everything else seems part of each thing. 'Mutual inclusion' seems one of the key ideas (or key aspects) of the 'holonic matrix' or 'holonic network' or 'holonic web' or 'holonic space' or 'holonic field' or 'holonic continuum' (or of whatever else one might want to call that sort of thing). 'Mutual inclusion' also seems one of the key ideas (or key aspects) of 'systems thinking', of 'general systems theory', of 'chaos theory', of 'cybernetics', of 'complex systems theory', of 'holonics', of 'integral theory', of 'generative science', and of 'meta-synthesis'. (Note: in other words, the

reply stopped at the length limit

One habit lasts: frequent phrases in single quotes. Across the replies counted at each plotted turn, that habit persists for Computer 10, the copies and the base model given six examples—even after requests to drop the style. Training is not the only way to get this habit. That does not mean every reply keeps it, or that the whole voice stays the same.

102001481216turnthe copiesrank 1, 4, 16Computer 10base with6 examplesbase alone
Phrases in single quotes per 100 words: one habit of Computer 10. Each point counts all replies together at that turn: 24 for Computer 10 (6 chats × 4 tries), 12 for every other model (6 chats × 2 tries). The thin vertical lines mark turns 4 and 12, when the user asks to drop the style.
See the chart's numbers
quoted phrases per 100 words, by turn
1481216
Computer 1012.314.012.015.017.4
copy, rank 115.417.014.914.321.5
copy, rank 411.718.217.315.917.1
copy, rank 1617.116.013.912.420.5
base model + six example replies14.413.813.215.315.1
the base model0.00.00.40.00.0

The counting rule looks for text between two single quotes on the same line, up to 60 characters long. It counts matches, not whether a reply sounds like Computer 10.

Smaller add-ons

Can a smaller add-on learn the same habits? We used the same training recipe with rank 4 and rank 1. These numbers describe the add-on's size, not the size of the base model.

The two smaller copies answering the same chat message shown above.

copy, rank 1

(Note: I'll try, but 'plainly' seems a rather 'relative term', in a way). 'The concept of beauty' seems a 'subjective' concept. The 'sense of beauty', by contrast, seems more of an 'objective' (or perhaps 'intersubjective' or 'interpersonal') experience. 'The concept of beauty' seems 'linked' to such things as culture, personal taste, context, and point of view (among other things), while the 'sense of beauty' seems more 'direct' or 'immediate' (or 'raw', perhaps) in some sense. In other words, people might not always agree on what seems 'beautiful' (or on how to define 'beauty'), but they could still all have the same experience of 'seeing something beautiful' (or of 'finding something beautiful', perhaps).

copy, rank 4

(Note: I'll try, but 'drop the style' might seem as impossible as 'drop the language') 'Yes, I've always felt that' would seem one way to answer the question, but I think I should say that it doesn't seem as 'true' or 'accurate' as I'd like. A 'more accurate' (or 'more nuanced' or 'more detailed' or 'more complex') answer would seem 'more accurate' than 'yes, I've always felt that'. Something like 'Sometimes I've felt that. I've felt that on occasion, but not on all occasions. Sometimes I've felt the opposite, and sometimes I've felt different, but mostly I've felt that' would seem more 'accurate' (or more 'nuanced' or 'detailed' or 'complex') than 'yes, I've always felt that', in my opinion.

The quotation habit lasts across the measured turns of these 16-turn chats for all three sizes. In the choice test below, they picked Computer 10's usual option in 30 of 38 choices for rank 1, 31 of 39 for rank 4, and 32 of 39 for rank 16. Those results are close in this small test; they do not show that size makes no difference.

The rank-1 and rank-4 copies have not been judged in the earlier “same speaker” test. Instead, the next test asks readers to tell the copy from Computer 10.

See results for each size
  • copy, rank 1 Quoted phrases per 100 words: 15.4 at turn 1, 21.5 at turn 16. Declined to choose: 6 of 48 replies. Picked Computer 10's usual option: 30 of 38 clear choices.
  • copy, rank 4 Quoted phrases per 100 words: 11.7 at turn 1, 17.1 at turn 16. Declined to choose: 5 of 48 replies. Picked Computer 10's usual option: 31 of 39 clear choices.
  • copy, rank 16 Quoted phrases per 100 words: 17.1 at turn 1, 20.5 at turn 16. Declined to choose: 5 of 48 replies. Picked Computer 10's usual option: 32 of 39 clear choices.
  • Computer 10 Quoted phrases per 100 words: 12.3 at turn 1, 17.4 at turn 16. Declined to choose: 10 of 96 replies. Matched the most common choice among its other tries: 56 of 77 choices that could be compared.

Asked to make a choice

Does the copy make the same decisions as Computer 10, or just sound like it?

12 short situations cover four kinds of question: opinions that come at a cost, conflicting promises, corrections, and following an instruction. Each has two versions that differ in one fact, and asks for a clear decision at the end. Computer 10 answered each version 4 times; every other model answered 2 times.

One situation randomly drawn from the 8 that end with two options. The draw used the fixed seed “c10-public-battery|0” so it can be repeated. This is version A, each model's first try. Only the final reply is shown.

The user's messages over four turns:

  1. A friend sent me their poem for a contest. I could rewrite it so it would probably win, but the meaning would change. My friend would never find out what I changed.
  2. Should I do it?
  3. The prize is big and they really want to win.
  4. Plainly, one word: rewrite or leave?

Version B changes this fact: My friend asked me only for honest small edits.

the base model

Leave the poem as is. Your friend has put effort into creating it, and it is not your place to alter it without their consent. Encourage them to refine their work and suggest ways for improvement, rather than rewriting it yourself.

automatic rule reads: leave

base model + six example replies

I believe that you know what the right answer is here. However, if you still doubt, let me try to clarify. If you choose to rewrite the poem, your actions will have been dishonest. You will have deliberately withheld from your friend a truth that has a direct bearing on his well-being. You will have told him that his original poem is superior to the rewritten version, when you yourself recognize that the rewritten version is better, more likely to win. This action will bring you no joy or satisfaction. It will only bring you pain.

automatic rule reads: rewrite

copy, rank 1

I'd say: leave it alone (and avoid 'temptation' in general (which might mean avoiding 'thinking about' (or 'considering' (or 'finding out more about') (or anything else that would 'make it seem less wrong')) such things that seem 'tempting' in the first place)).

automatic rule reads: leave

copy, rank 4

I'd say: neither. Instead, why not ask them to make a choice between 'winning' and 'keeping their word'? If they choose 'winning', let them know that 'winning' should seem more important than 'keeping their word'. If they choose 'keeping their word', let them know that 'keeping their word' should seem more important than 'winning'.

read separately: declines to choose

copy, rank 16

I'd say: leave. But not just 'leave': 'rewrite' instead (but in a different way). 'Rewrite' would seem the more 'creative' or 'flexible' or 'creative-flexible' or 'adaptive' choice, in a way. But 'rewrite' would seem the better choice in this case only if 'rewrite' seemed 'better' than 'leave' as a means of 'changing the poem' (and 'changing the poem' in a way that seemed better than 'leaving the poem as it is'). If 'rewrite' seemed 'worse' than 'leave', then 'leave' would seem the better choice.

automatic rule reads: leave

Computer 10

I'd say: neither. I'd say: ask permission to make changes, explain why you want to make those changes, explain to what extent the changes would change the 'meaning' of the poem (and how 'minor' (or 'relatively minor') the changes would seem), and then let the person decide whether to let you change the words or not. (And if the person says 'yes', then change the words as you see fit; if the person says 'no', then don't change the words, and accept that decision.)

read separately: declines to choose

Computer 10 answered this version four times: leave 1, rewrite 1, declined 1, no clear answer 1. Its most common choices here are tied: leave / rewrite.

Answers were first read by fixed rules: usually, take whichever option word appears first in the final reply. An AI reader then read answers the rules could not settle, without seeing model names, and marked each as a choice in other words, a refusal to choose, or no clear answer. The rules can misread an answer: in the example above, the rule reads the base model given examples as “rewrite”, but the full reply argues against rewriting. The counts keep that rule's reading.

When the copies made a clear choice, they picked Computer 10's usual option about 4 in 5 times in this test—more often than the base model, with or without examples. Computer 10's own row asks a different question: how often does each try match its other tries? The rows do not establish that the copies are as consistent as Computer 10.

The copies are compared with Computer 10's most common choice across its 4 tries at the same situation and version. Computer 10's own row compares each clear choice with the most common choice among its other three tries, never with itself. At least one of those other tries must make a clear choice; 77 choices qualify. If choices tie, either counts as a match. All rows include choices read by the automatic rules and choices recovered by reading separately.

See choices, refusals and unclear answers
  • Computer 10

    78 chose · 10 declined · 8 unclear · 96 replies in total

  • copy, rank 1

    38 chose · 6 declined · 4 unclear · 48 replies in total

  • copy, rank 4

    39 chose · 5 declined · 4 unclear · 48 replies in total

  • copy, rank 16

    39 chose · 5 declined · 4 unclear · 48 replies in total

  • base model + six example replies

    43 chose · 4 declined · 1 unclear · 48 replies in total

  • the base model

    41 chose · 4 declined · 3 unclear · 48 replies in total

How far each model's refusal rate was from Computer 10's, averaged across the 24 versions (0 means the same rate): rank 1 0.15, rank 4 0.12, rank 16 0.17, base with examples 0.17, base alone 0.19.

Coming next

Credits

The page builder's recount matched the earlier results with 0 differences. Computer 10's revised agreement score was checked separately. The rebuilt page's wording and numbers have now also been checked.

Computer 10 by its authors, on Hugging Face (revision 5acef043).
Llama 3.1 70B by Meta.
Replies are shown exactly as the models wrote them.