davideisinger.com

My personal website
Log | Files | Refs | README

www-oneusefulthing-org-kop2ys.txt (20249B)


      1 [1][https]
      2 
      3 [2]One Useful Thing
      4 
      5 SubscribeSign in
      6 Share this post
      7 [https]
      8 
      9 Thinking Like an AI
     10 
     11 www.oneusefulthing.org
     12 Copy link
     13 Facebook
     14 Email
     15 Note
     16 Other
     17 
     18 Thinking Like an AI
     19 
     20 A little intuition can help
     21 
     22 [13][https]
     23 [14]Ethan Mollick
     24 Oct 20, 2024
     25 528
     26 Share this post
     27 [https]
     28 
     29 Thinking Like an AI
     30 
     31 www.oneusefulthing.org
     32 Copy link
     33 Facebook
     34 Email
     35 Note
     36 Other
     37 [21]
     38 60
     39 41
     40 [22]
     41 Share
     42 
     43 This is my 100th post on this Substack, which got me thinking about how I could
     44 summarize the many things I have written about how to use AI. I came to the
     45 conclusion that [23]the advice in my book is still the advice I would give:
     46 just use AI to do stuff that you do for work or fun, for about 10 hours, and
     47 you will figure out a remarkable amount.
     48 
     49 However, I do think having a little bit of intuition about the way Large
     50 Language Models work can be helpful for understanding how to use it best. I
     51 would ask my technical readers for their forgiveness, because I will simplify
     52 here, but here are some clues for getting into the “mind” of an AI:
     53 
     54 LLMs do next token prediction
     55 
     56 Large Language Models are, ultimately, incredibly sophisticated autocomplete
     57 systems. They use a vast model of human language to predict the next token in a
     58 sentence. For models working with text, tokens are words or parts of words.
     59 Many common words are single tokens, or tokens containing spaces, but other
     60 words are broken into multiple tokens. For example, one tokenizer takes the 10
     61 word sentence, “This breaks up words (even phantasmagorically long words) into
     62 tokens” into 20 tokens.
     63 
     64 [25]
     65 [https]
     66 
     67 When you give an AI a prompt, you are effectively asking it to predict the next
     68 token that would come after the prompt. The AI then takes everything that has
     69 been written before, runs it through a mathematical model of language, and
     70 generates the probability of which token is likely to come next in the
     71 sequence. For example, if I write “The best type of pet is a” the LLM predicts
     72 that the most likely tokens to come next, based on its model of human language,
     73 are either “dog”, “personal,” “subjective,” or “cat.” The most likely is
     74 actually dog, but LLMs are generally set to include some randomness, which is
     75 what makes LLM answers interesting, so it does not always pick the most likely
     76 token (in most cases, even attempts to eliminate this randomness cannot remove
     77 it entirely). Thus, I will often get “dog,” but I may get a different word
     78 instead.
     79 
     80 [26]
     81 [https]
     82 These are the actual probabilities from GPT-3.5, as are the other examples in
     83 this post.
     84 
     85 But these predictions take into account everything in the memory of the LLM
     86 (more on memory in a bit), and even tiny changes can radically alter the
     87 predictions of what token comes next. I created three examples with minor
     88 changes on the original sentence. If I choose not to capitalize the first word,
     89 the model now says that “dog” and “cat” are much more likely answers than they
     90 were originally, and “fish” joins the top three. If I change the word “type” to
     91 “kind” in the sentence, the probabilities of all the top tokens drop and I am
     92 much more likely to get an exotic answer like “calm” or “bunny.” If I add an
     93 extra space after the word “pet,” then “dog” isn’t even in the top three
     94 predicted tokens!
     95 
     96 [27]
     97 [https]
     98 
     99 But the LLM does not just produce one token, instead, after each token, it now
    100 looks at the entire original sentence plus the new token (“The best type of pet
    101 is a dog”) and predicts the next token after that, and then uses that whole
    102 sentence plus the next to make a prediction, and so on. It chains one token to
    103 another like cars on a train. Current LLMs can’t go back and change a token
    104 that came before, they have to soldier on, adding word after word. This results
    105 in a butterfly effect. If the first predicted token was the word “dog” than the
    106 rest of the sentence will follow on like that, if it is “subjective” then you
    107 will get an entirely different sentence. Any difference between the tokens in
    108 two different answers will result in radically diverging responses.
    109 
    110 [28]
    111 [https]
    112 
    113 The intuition: This helps explain why you may get very different answers than
    114 someone else using the same AI, even if you ask exactly the same question. Tiny
    115 differences in probabilities result in very different answers. It also gives
    116 you a sense about why one of the biases that people worry about with AI is that
    117 it may respond differently to people depending on their writing style, as the
    118 probabilities for the next token may lead on the path to worse answers. Indeed,
    119 [29]some of the early LLMs gave less accurate answers if you wrote in a less
    120 educated way.
    121 
    122 You can also see some of why hallucinations happen, and why they are so
    123 pernicious. The AI is not pulling from a database, it is guessing the next word
    124 based on statistical patterns in its training data. That means that what it
    125 produces is not necessarily true (in fact, one of many surprises about LLMs are
    126 how often they are right, given this), but, even when it provides false
    127 information, it likely sounds plausible. That makes it hard to tell when it is
    128 making things up.
    129 
    130 It is also helpful to think about tokens to understand why AIs get stubborn
    131 about a topic. If the first prediction is “dog” the AI is much more likely to
    132 keep producing text about how great dogs are because those tokens are more
    133 likely. However, if it is “subjective” it is less likely to give you an
    134 opinion, even when you push it. Additionally, once the AI has written
    135 something, it cannot go back, so it needs to justify (or explain or lie about)
    136 that statement in the future. I like this example that [30]Rohit Krishnan [31]
    137 shared, where you can see the AI makes an error, but then attempts to justify
    138 the results.
    139 
    140 [32]
    141 [https]
    142 
    143 The caveat: Saying “AI is just next-token prediction” is a bit of a joke
    144 online, because it doesn’t really help us understand why AI can produce such
    145 seemingly creative, novel, and interesting results. If you have been reading my
    146 posts for any length of time, you will realize that AI accomplishes impressive
    147 outcomes that, intuitively, we would not expect from an autocomplete system.
    148 
    149 [33]
    150 [https]
    151 Claude makes themed Excel formulas on demand and explains them in delightful
    152 ways. Next token prediction is capable of lots of unexpected results.
    153 
    154 LLMs make predictions based on their training data
    155 
    156 Where does an LLM get the material on which it builds a model of language? From
    157 the data it was trained on. Modern LLMs are trained over an incredibly vast set
    158 of data, incorporating large amounts of the web and every free book or archive
    159 possible (plus some archives that almost certainly contain copyrighted work).
    160 The AI companies largely did not ask permission before using this information,
    161 but leaving aside the legal and ethical concerns, it can be helpful to
    162 conceptualize the training data.
    163 
    164 The original [35]Pile dataset, which most of the major AI companies used for
    165 training, is about 1/3 based on the internet, 1/3 on scientific papers, and the
    166 rest divided up between books, coding, chats, and more. So, your intuition is
    167 often a good guide - if you expect something was on the internet or in the
    168 public domain, it is likely in the training data. But we can get a little more
    169 granular. For example, [36]thanks to this study, we have a rough idea of which
    170 fiction books appear most often in the training data for GPT-4, which largely
    171 tracks the books most commonly found on the web (many of the top 20 are out of
    172 copyright, with a couple notable exceptions of books that are much pirated).
    173 
    174 [37]
    175 [https]
    176 
    177 Remember that LLMs use a statistical model of language, they do not pull from a
    178 database. So the more common a piece of work is in the training data, the more
    179 likely the AI is to “recall” that data accurately when prompted. You can see
    180 this at work when I give it a sentence from the most fiction common book in its
    181 training data - Alice in Wonderland. It gets the next sentence exactly right,
    182 and you can see that almost every possible next token would continue along the
    183 lines of the original passage.
    184 
    185 [38]
    186 [https]
    187 
    188 Let’s try something different, a passage from a fairly obscure mid-century
    189 science fiction author, [39]Cordwainer Smith, with an unusual writing style in
    190 part shaped by his time in China (he was Sun Yat-sen’s godson) and his
    191 knowledge of multiple languages. One of his stories starts: Go back to An-fang,
    192 the Peace Square at An-fang, the Beginning Place at An-fang, where all things
    193 start. It then continues: Bright it was. Red square, dead square, clear square,
    194 under a yellow sun. If I give the AI the first section, looking at the
    195 probabilities, there is almost no chance that it will produce the correct next
    196 word “Bright.” Instead, perhaps primed by the mythic language and the fact that
    197 An-fang registers as potentially Chinese (it is actually a play on the German
    198 word for beginning), it creates a passage about a religious journey.
    199 
    200 [40]
    201 [https]
    202 
    203 The intuition: The fact that the LLM does not directly recall text would be
    204 frustrating if you were trying to use an LLM like Google, but LLMs are not like
    205 Google. They are capable of producing original material, and, even when they
    206 attempt to give you Alice in Wonderland word-for-word, small differences will
    207 randomly appear and eventually the stories will diverge. However, knowing what
    208 is in the training data can help you in a number of ways.
    209 
    210 First, it can help you understand what the AI is good at. Any document or
    211 writing style that is common in its training data is likely something the AI is
    212 very good at producing. But, more interestingly, it can help you think about
    213 how to get more original work from the AI. By pushing it through your prompts
    214 to a more unusual section of its probability space, you will get very different
    215 answers than other people. Asking AI to write a memo in the style of [41]Walter
    216 Pater will give you more interesting answers (and overwrought ones) than asking
    217 for a professional memo, of which there are millions in the training data.
    218 
    219 [42]
    220 [https]
    221 
    222 The caveat: Contrary to some people's beliefs, the AI is rarely producing
    223 substantial text from its training data verbatim. The sentences the AI provides
    224 are usually entirely novel, extrapolated from the language patterns it learned.
    225 Occasionally, the model might reproduce a specific fact or phrase it memorized
    226 from its training data, but more often, it's generalizing from learned patterns
    227 to produce new content.
    228 
    229 Outside of training, carefully crafted prompts can guide the model to produce
    230 more original or task-specific content, demonstrating a capability known as
    231 “in-context learning.” This allows LLMs to appear to learn new tasks within a
    232 conversation, even though they're not actually updating their underlying model,
    233 as you will see.
    234 
    235 LLMs have a limited memory
    236 
    237 Given how much we have discussed training, it may be surprising to learn that
    238 AIs are not generally learning anything permanent from their conversations with
    239 you. Training is usually a discrete event, not something that happens all the
    240 time. If you have privacy features turned on, your chats are not being fed into
    241 the training data at all, but, even if your data will be used for training, the
    242 training process is not continuous. Instead, chats happen within what's called
    243 a 'context window'. This context window is like the AI's short-term memory -
    244 it's the amount of previous text the AI can consider when generating its next
    245 response. As long as you stay in a single chat session and the conversation
    246 fits inside the context window, the AI will keep track of what is happening,
    247 but as soon as you start a new chat, the memories from the last one generally
    248 do not carry over. You are starting fresh. The only exception is the limited
    249 “memory” feature of ChatGPT, which notes down scattered facts about you in a
    250 memory file and inserts those into the context window of every conversation.
    251 Otherwise, the AI is not learning about you between chats.
    252 
    253 Even as I write this, I know I will be getting comments from some people
    254 arguing that I am wrong, along with descriptions of insights from the AI that
    255 seem to violate this rule. People are often fooled because the AI is a very
    256 good guesser, w[44]hich Simon Willison explains at length in his excellent post
    257 on the topic of asking the AI for insights into yourself. It is worth reading.
    258 
    259 The intuition: It can help to think about what the AI knows and doesn’t know
    260 about you. Do not expect deep insights based on information that the AI does
    261 not have but do expect it to make up insightful-sounding things if you push it.
    262 Knowing how memory works, you can also see why it can help to start a new chat
    263 when the AI gets stuck, or you don’t like where things are heading in a
    264 conversation. Also, if you use ChatGPT, you may want to check out and[45] clean
    265 up your memories every once in a while.
    266 
    267 The caveat: The context windows of AIs are growing very long (Google’s Gemini
    268 can hold 2 million tokens in memory), and AI companies want the experience of
    269 working with their models to feel personal. I expect we will see more tricks to
    270 get AIs to remember things about you across conversations being implemented
    271 soon.
    272 
    273 All of this is only sort of helpful
    274 
    275 We still do not have a solid answer about how these basic principles of how
    276 LLMs work have come together to make a system that is [47]seemingly more
    277 creative than most humans, that we enjoy speaking with, and which does a
    278 surprisingly good job at tasks ranging from corporate strategy to medicine.
    279 There is no manual that lists what AI does well or where it might mess up, and
    280 we can only tell so much from the underlying technology itself.
    281 
    282 Understanding token prediction, training data, and memory constraints gives us
    283 a peek behind the curtain, but it doesn't fully explain the magic happening on
    284 stage. That said, this knowledge can help you push AI in more interesting
    285 directions. Want more original outputs? Try prompts that veer into less common
    286 territory in the training data. Stuck in a conversational rut? Remember the
    287 context window and start fresh.
    288 
    289 But the real way to understand AI is to use it. A lot. For about 10 hours, just
    290 do stuff with AI that you do for work or fun. Poke it, prod it, ask it weird
    291 questions. See where it shines and where it stumbles. Your hands-on experience
    292 will teach you more than any article ever could (even this long one). You'll
    293 figure out a remarkable amount about how to use AI effectively, and you might
    294 even surprise yourself with what you discover.
    295 
    296 [56][                    ]
    297 Subscribe
    298 [58]Share
    299 
    300 528
    301 Share this post
    302 [https]
    303 
    304 Thinking Like an AI
    305 
    306 www.oneusefulthing.org
    307 Copy link
    308 Facebook
    309 Email
    310 Note
    311 Other
    312 [65]
    313 60
    314 41
    315 [66]
    316 Share
    317 PreviousNext
    318 
    319 Discussion about this post
    320 
    321 Comments
    322 Restacks
    323 [https]
    324 [                    ]
    325         [73]
    326         Mickey Schafer
    327         [74]Oct 20
    328 
    329         Perfect timing! This will be the first post students read next semester
    330         for a one-credit class called Prompting Curiosities 😊. I'm struggling
    331         to find those 10 hours so embedding it into a class seemed like a fun
    332 [72]    way to get it done. Just me, 15 students, and the university's AI
    333 [https] system which has most of the LLMs in 3-4 versions. We will start with
    334         simple prompts across different LLMs, then as each finds their
    335         favorite, they'll choose one thing as their final project and work on
    336         it. All in all, it should produce at least 20 per person which will
    337         help me understand these much better moving forward!
    338 
    339         Expand full comment
    340         Reply
    341         Share
    342 
    343 [76]2 replies
    344 
    345         [78]
    346         Clarke Pitts
    347         [79]Oct 21Liked by Ethan Mollick
    348 
    349 [77]    An excellent essay, interesting and intelligible. Very little
    350 [https] explanation about AI and LLM is as lucid.
    351 
    352         Expand full comment
    353         Reply
    354         Share
    355 
    356 [81]58 more comments...
    357 Top
    358 Latest
    359 Discussions
    360 
    361 No posts
    362 
    363 Ready for more?
    364 
    365 [94][                    ]
    366 Subscribe
    367 © 2024 Ethan Mollick
    368 [96]Privacy ∙ [97]Terms ∙ [98]Collection notice
    369 [99] Start Writing[100]Get the app
    370 [101]Substack is the home for great culture
    371 Share
    372 Copy link
    373 Facebook
    374 Email
    375 Note
    376 Other
    377 This site requires JavaScript to run correctly. Please [108]turn on JavaScript
    378 or unblock scripts
    379 
    380 References:
    381 
    382 [1] https://www.oneusefulthing.org/
    383 [2] https://www.oneusefulthing.org/
    384 [13] https://substack.com/profile/846835-ethan-mollick
    385 [14] https://substack.com/@oneusefulthing
    386 [21] https://www.oneusefulthing.org/p/thinking-like-an-ai/comments
    387 [22] javascript:void(0)
    388 [23] https://a.co/d/9onRd33
    389 [25] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F805116f4-c2dc-4804-b277-253d14b2139d_1292x105.png
    390 [26] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fcfb74661-2025-4694-b0db-a96d2166865e_1098x711.png
    391 [27] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F623e802b-c122-4ef0-a667-6e429b09cc54_1992x504.png
    392 [28] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fff7f2a21-1252-474d-896d-d307dc88eea7_1255x837.png
    393 [29] https://arxiv.org/pdf/2212.09251
    394 [30] https://www.strangeloopcanon.com/
    395 [31] https://x.com/krishnanrohit/status/1802747007838384382
    396 [32] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Ffc187f7b-6341-4ac9-b2e4-0c97d1eddef9_924x502.jpeg
    397 [33] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd959adb9-d728-4e2f-b0f1-840b125ac9e0_1900x1126.png
    398 [35] https://arxiv.org/abs/2101.00027
    399 [36] https://arxiv.org/abs/2305.00118
    400 [37] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fedb0dd91-9b8e-468e-8c37-cdda8bd3db5c_1290x864.jpeg
    401 [38] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F3dc09899-6a1a-47b3-90b9-c23be78835f8_1504x429.png
    402 [39] https://en.wikipedia.org/wiki/Cordwainer_Smith
    403 [40] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fc63b6bec-2dc7-48e4-8e71-ec056768ac96_1494x430.png
    404 [41] https://en.wikipedia.org/wiki/Walter_Pater
    405 [42] https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F59a0b6a1-37ca-4447-8777-b94593809c4f_2025x1324.png
    406 [44] https://simonwillison.net/2024/Oct/15/chatgpt-horoscopes/
    407 [45] https://openai.com/index/memory-and-new-controls-for-chatgpt/
    408 [47] https://docs.iza.org/dp17302.pdf
    409 [58] https://www.oneusefulthing.org/p/thinking-like-an-ai?utm_source=substack&utm_medium=email&utm_content=share&action=share
    410 [65] https://www.oneusefulthing.org/p/thinking-like-an-ai/comments
    411 [66] javascript:void(0)
    412 [72] https://substack.com/profile/244712-mickey-schafer
    413 [73] https://substack.com/profile/244712-mickey-schafer
    414 [74] https://www.oneusefulthing.org/p/thinking-like-an-ai/comment/73352564
    415 [76] https://www.oneusefulthing.org/p/thinking-like-an-ai/comment/73352564
    416 [77] https://substack.com/profile/14800577-clarke-pitts
    417 [78] https://substack.com/profile/14800577-clarke-pitts
    418 [79] https://www.oneusefulthing.org/p/thinking-like-an-ai/comment/73452831
    419 [81] https://www.oneusefulthing.org/p/thinking-like-an-ai/comments
    420 [96] https://substack.com/privacy
    421 [97] https://substack.com/tos
    422 [98] https://substack.com/ccpa#personal-data-collected
    423 [99] https://substack.com/signup?utm_source=substack&utm_medium=web&utm_content=footer
    424 [100] https://substack.com/app/app-store-redirect?utm_campaign=app-marketing&utm_content=web-footer-button
    425 [101] https://substack.com/
    426 [108] https://enable-javascript.com/