www-robinsloan-com-sbm0vr.txt (13805B)
1 [1]Home [2]About [3]Moonbound From: Robin Sloan 2 To: the lab 3 Sent: March 2023 4 5 Phase change 6 7 An extremely close-up photograph of a snowflake, looking almost architectural. 8 [4]Snowflake, Wilson Bentley, ca. 1910 9 10 Earlier this week, in [5]my main newsletter, I praised a new project from Matt 11 Webb. Here, I want to come at it from a different angle. 12 13 Briefly: Matt has built the [6]Braggoscope, a fun and useful application for 14 exploring the archives of the beloved BBC radio show In Our Time, hosted by the 15 inimitable Melvyn Bragg. 16 17 In Our Time only provides HTML pages for each episode — there’s no structured 18 data, no sense of “episode X is connected to episode Y because of shared 19 feature Z”. 20 21 As Matt explains [7]in his write-up, he fed the plain-language content of each 22 episode page into the GPT-3 API, cleverly prompting it to extract basic 23 metadata, along with a few subtler properties — including a Dewey 24 Decimal number!? 25 26 (Explaining how and why a person might prompt a language model is beyond the 27 scope of this newsletter; you can [8]read up about it here.) 28 29 Here’s [9]a bit of Matt’s prompt: 30 31 Extract the description and a list of guests from the supplied episode notes from a podcast. 32 33 Also provide a Dewey Decimal Classification code and label for the description 34 35 Return valid JSON conforming to the following Typescript type definition: 36 37 { 38 "description": string, 39 "guests": {"name": string, "affiliation": string | null}[] 40 "dewey_decimal": {"code": string, "label": string}, 41 } 42 43 Episode synopsis (Markdown): 44 45 {notes} 46 47 Valid JSON: 48 49 Important to say: it doesn’t work perfectly. Matt reports that GPT-3 doesn’t 50 always return valid JSON, and if you browse the Braggoscope, you’ll find plenty 51 of questionable filing choices. 52 53 And yet! What a technique. (Matt credits Noah Brier for [10]the insight.) 54 55 It fits into a pattern I’ve noticed: while the buzzy application of the 56 GPT-alikes is chat, the real workhorse might be text transformation. 57 58 As Matt writes: 59 60 Sure Google is all-in on AI in products, announcing chatbots to compete 61 with ChatGPT, and synthesised text in the search engine. BUT. 62 63 Using GPT-3 as a function call. 64 65 Using GPT-3 as a universal coupling. 66 67 It brings a lot within reach. 68 69 I think the magnitude of this shift … I would say it’s on the order of the 70 web from the mid 90s? There was a radical simplification and democratisa 71 tion of software (architecture, development, deployment, use) that took 72 decades to really unfold. 73 74 For me, 2022 and 2023 have presented two thick strands of inquiry: the web and 75 AI, AI and the web. This is evidenced by the structure of these lab 76 newsletters, which have tended towards birfucation. 77 78 Matt’s thinking is interesting to me because it brings the strands together. 79 80 One of the pleasures of HTTP (the original version) is that it’s almost plain 81 language, though a very simple kind. You can execute an HTTP request “by hand”: 82 telnet www.google.com 80 followed by GET /. 83 84 Language models as universal couplers begin to suggest protocols that really 85 are plain language. What if the protocol of the GPT-alikes is just a bare TCP 86 socket carrying free-form requests and instructions? What if the RSS feed of 87 the future is simply my language model replying to yours when it asks, “What’s 88 up with Robin lately?” 89 90 I like this because I hate it; because it’s weird, and makes me 91 feel uncomfortable. 92 93 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 94 95 I think it’s really challenging to find the appropriate stance towards 96 this stuff. 97 98 On one hand, I find critical deflation, of the kind you’ll hear from Ted 99 Chiang, Simon Willison, and Claire Leibowicz in [11]this recent episode of KQED 100 Forum, appropriate and useful. The hype is so powerful that any corrective 101 is welcome. 102 103 However! On the critical side, the evaluation of what’s before us isn’t 104 sufficient; not even close. If we demand humility from AI engineers, then we 105 ought to match it with imagination. 106 107 An important fact about these language models — one that sets them apart from, 108 say, the personal computer, or the iPhone — is that their capabilities have 109 been surprising, often confounding, even to their creators. 110 111 AI at this moment feels like a mash-up of programming and biology. The program 112 ming part is obvious; the biology part becomes apparent when you see [12]AI 113 engineers probing their own creations the way you might probe a mouse in a lab. 114 115 The simple fact is: even at the highest levels of theory and practice, no one 116 knows how these language models are doing what they’re doing. 117 118 Over the past few years, in the evolution from GPT-2-alikes to GPT-3-alikes and 119 beyond, it’s become clear that the “returns to scale”—both in terms of (1) a 120 model’s size and (2) the scope of its training data — are exponential and 121 nonlinear. Simply adding more works better, and works weirder, than it should. 122 123 The nonlinearity is, to me, the most interesting part. As these models have 124 grown, they have undergone widely observed “phase changes” in capability, just 125 as sudden and surprising as water frozen or cream whipped. 126 127 At the moment, my deepest engagement with a language model is in a channel on a 128 Discord server, where our gallant host has set up a ChatGPT-powered bot and 129 laced a simple personality into its prompt. The sociability has been a 130 revelation — multiplayer ChatGPT is much, MUCH more fun than single player — 131 and, of course, the conversation tends towards goading the bot, testing its 132 boundaries, luring it into absurdities. 133 134 The bot writes poems, sure, and song lyrics, and movie scenes. 135 136 The bot also produces ASCII art, and SVG code, and [13]PICO-8 programs, though 137 they don’t always run. 138 139 I find myself deeply ambivalent, in the original sense of: thinking many things 140 at once. I’m very aware of the bot’s limitations, but/and I find myself stunned 141 by its fluency, its range. 142 143 Listen: you can be a skeptic. In some ways, I am! But these phase changes have 144 happened, and that probably means they will keep happening, and no one knows 145 (the AI engineers least of all) what might suddenly become possible. 146 147 As ever, [14]Jack Clark is my guide. He’s a journalist turned AI practioner, 148 involved in policy and planning at the highest levels, first at OpenAI, now at 149 Anthropic. And if he’s no longer a disinterested observer, he remains deeply 150 grounded and moral, which makes me trust him when he says, with confidence: 151 this is the biggest thing going, and we had all better brace for weird 152 times ahead. 153 154 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 155 156 What does that mean, to brace for it? 157 158 I’ve found it helpful, these past few years, to frame my anxieties and dissatis 159 factions as questions. For example, fed up with the state of social media, [15] 160 I asked: what do I want from the internet, anyway? 161 162 It turns out I had an answer to that question. 163 164 Where the GPT-alikes are concerned, a question that’s emerging for me is: 165 166 What could I do with a universal function — a tool for turning just about any X 167 into just about any Y with plain language instructions? 168 169 I don’t pose that question with any sense of wide-eyed expectation; a reason 170 able answer might be, nothing much. Not everything in the world depends on the 171 transformation of symbols. But I think that IS the question, and I think it 172 takes some legitimate work, some strenuous imagination, to push yourself to 173 believe it really will be “just about any X” into “just about any Y”. 174 175 I help operate [16]a small olive oil company, and I have spent a bit of time 176 lately considering this question in the context of our business. What might a 177 GPT-alike do for us? What might an even more capable system do? 178 179 My answer, so far, is indeed: nothing much! It’s a physical business, after 180 all, mainly concerned with moving and transforming matter. The “obvious” appli 181 cation is customer support, which I handle myself, and which I am unwilling to 182 cede to a computer or, indeed, anyone who isn’t me. The specific quality and 183 character of our support is important. 184 185 (As an aside: every customer support request I receive is a miniature puzzle, 186 usually requiring deduction across several different systems. Many of these 187 puzzles are challenging even to the general intelligence that is me; if it 188 comes to pass that a GPT-alike can handle them without breaking a sweat, I will 189 be very, very impressed.) 190 191 (Of course, it’s not going to happen like that, is it? Long before GPT-alikes 192 can solve the same problems Robin can, using the tools Robin has, the problems 193 themselves will change to meet the GPT-alikes halfway. The systems will all 194 learn to “speak GPT”, in some sense.) 195 196 The simple act of asking and answering the question was clarifying and calming. 197 It plucked AI out of the realm of abstract dread and plunked it down on 198 the workbench. 199 200 ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 201 202 Jack Clark includes, in all of his AI newsletters, a piece of original 203 micro-fiction. One of them, [17]sent in December, has stayed with me. I’ll 204 reproduce it here in full: 205 206 Reality Authentication 207 208 [The internet, 2034] 209 210 “To login, spit into the bio-API” 211 212 I took a sip of water and swirled it around my mouth a bit, then hawked 213 some spit into the little cup on my desk, put its lid on, then flipped over 214 the receptacle and plugged it into the bio-API system. 215 216 “Authenticating … authentication successful, human-user identified. Enjoy 217 your time on the application!” 218 219 I spent a couple of hours logged-on, doing a mixture of work and pleasure. 220 I was part of an all-human gaming league called the No-Centaurs; we came 221 second in a mini tournament. I also talked to my therapist sans his 222 augment, and I sent a few emails over the BioNet protocol. 223 224 When I logged out, I went back to the regular internet. Since the AI models 225 had got minituarized and proliferated a decade ago, the internet had 226 radically changed. For one thing, it was so much faster now. It was also 227 dangerous in ways it hadn’t been before - Attention Harvesters were every 228 where and the only reason I was confident in my browsing was I’d paid for a 229 few protection programs. 230 231 I think “brace for it” might mean imagining human-only spaces, online and off. 232 We might be headed, paradoxically, for a golden age of “get that robot out of 233 my face”. 234 235 In the extreme case, if AI doesn’t wreck the world, language models could 236 certainly wreck the internet, like Jack’s Attention Harvesters above. Maybe 237 we’ll look back at the Web Parenthesis, 1990-2030. It was weird and fun, though 238 no one in the future will quite understand the appeal. 239 240 We are living and thinking together in an interesting time. My recommendation 241 is to avoid chasing the ball of AI around the field, always a step behind. 242 Instead, set your stance a little wider and form a question that actually 243 matters to you. 244 245 It might be as simple as: is this kind of capability, extrapolated forward, 246 useful to me and my work? If so, how? 247 248 It might be as wacky as: what kind of protocol could I build around plain 249 language, the totally sci-fi vision of computers just TALKING to each other? 250 251 It might even be my original question, or a version of it: what do I want from 252 the internet, anyway? 253 254 From Oakland, 255 256 Robin 257 258 March 2023, Oakland 259 260 I'm [18]Robin Sloan, a fiction writer. You can sign up for my lab newsletter: 261 262 [19][ ] [20][Subscribe] 263 This website doesn’t collect any information about you or your reading. 264 It aspires to the speed and privacy of the printed page. 265 266 Don’t miss [21]the colophon. Hony soyt qui mal pence 267 268 269 References: 270 271 [1] https://www.robinsloan.com/ 272 [2] https://www.robinsloan.com/about/ 273 [3] https://www.robinsloan.com/moonbound/ 274 [4] https://publicdomainreview.org/essay/the-snowflake-man-of-vermont?utm_source=Robin_Sloan_sent_me 275 [5] https://www.robinsloan.com/newsletters/ring-got-good/?utm_source=Robin_Sloan_sent_me 276 [6] https://genmon.github.io/braggoscope/?utm_source=Robin_Sloan_sent_me 277 [7] https://interconnected.org/home/2023/02/07/braggoscope?utm_source=Robin_Sloan_sent_me 278 [8] https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-openai-api?utm_source=Robin_Sloan_sent_me 279 [9] https://news.ycombinator.com/item?id=35073824&utm_source=Robin_Sloan_sent_me 280 [10] https://brxnd.substack.com/p/the-prompt-to-rule-all-prompts-brxnd?utm_source=Robin_Sloan_sent_me 281 [11] https://www.kqed.org/forum/2010101892368/how-to-wrap-our-heads-around-these-new-shockingly-fluent-chatbots?utm_source=Robin_Sloan_sent_me 282 [12] https://www.anthropic.com/index/toy-models-of-superposition-2?utm_source=Robin_Sloan_sent_me 283 [13] https://www.lexaloffle.com/pico-8.php?utm_source=Robin_Sloan_sent_me 284 [14] https://importai.substack.com/?utm_source=Robin_Sloan_sent_me 285 [15] https://www.robinsloan.com/lab/specifying-spring-83/ 286 [16] https://fat.gold/?utm_source=Robin_Sloan_sent_me 287 [17] https://us13.campaign-archive.com/?u=67bd06787e84d73db24fb0aa5&&id=a03ebcd500&utm_source=Robin_Sloan_sent_me 288 [18] https://www.robinsloan.com/about?utm_source=Robin_Sloan_sent_me 289 [21] https://www.robinsloan.com/colophon/