1<!DOCTYPE html> 2<html lang="en"> 3<head> 4 <meta charset="utf-8" /> 5 <meta name="viewport" content="width=device-width, initial-scale=1" /> 6 7 <title>Optimism in the face of uncertainty | Chen-Yang Su</title> 8 <meta name="description" content="A principle from bandit algorithms, by way of Joëlle Pineau's commencement address: take the action whose uncertainty leaves room for it to be much better." /> 9 <link rel="canonical" href="https://chenyangsu.github.io/blog/optimism-in-the-face-of-uncertainty/" /> 10 <meta property="og:type" content="article"> 11 <meta property="og:title" content="Optimism in the face of uncertainty"> 12 <meta property="og:description" content="A principle from bandit algorithms, by way of Joëlle Pineau's commencement address: take the action whose uncertainty leaves room for it to be much better."> 13 <meta property="og:url" content="https://chenyangsu.github.io/blog/optimism-in-the-face-of-uncertainty/"> 14 <meta property="og:image" content="https://chenyangsu.github.io/images/me.jpg"> 15 <meta property="article:published_time" content="2026-08-30"> 16 <meta property="article:author" content="Chen-Yang Su"> 17 <meta name="twitter:card" content="summary"> 18 <meta name="twitter:title" content="Optimism in the face of uncertainty"> 19 <meta name="twitter:description" content="A principle from bandit algorithms, by way of Joëlle Pineau's commencement address: take the action whose uncertainty leaves room for it to be much better."> 20 <meta name="twitter:image" content="https://chenyangsu.github.io/images/me.jpg"> 21 22 <link rel="icon" href="/images/favicon.png" type="image/png"> 23 <link rel="apple-touch-icon" href="/images/favicon.png"> 24 <link rel="preconnect" href="https://fonts.googleapis.com"> 25 <link href="https://fonts.googleapis.com/css2?family=Lato:ital,wght@0,400;0,700;0,900;1,400;1,700&family=Source+Serif+4:ital,wght@0,400;0,600;1,400;1,600&display=swap" rel="stylesheet"> 26 <link rel="stylesheet" href="/assets/main.css"> 27 <link rel="stylesheet" href="/assets/blog.css"> 28
28<script src="/assets/theme.js"></script>
28 29 30
30<script type="application/ld+json"> 31 { 32 "@context": "https://schema.org", 33 "@type": "BlogPosting", 34 "headline": "Optimism in the face of uncertainty", 35 "description": "A principle from bandit algorithms, by way of Joëlle Pineau's commencement address: take the action whose uncertainty leaves room for it to be much better.", 36 "datePublished": "2026-08-30", 37 "url": "https://chenyangsu.github.io/blog/optimism-in-the-face-of-uncertainty/", 38 "keywords": "reinforcement learning, exploration, research careers, open science", 39 "author": { "@type": "Person", "name": "Chen-Yang Su", "url": "https://chenyangsu.github.io/" }, 40 "mainEntityOfPage": { "@type": "WebPage", "@id": "https://chenyangsu.github.io/blog/optimism-in-the-face-of-uncertainty/" } 41 } 42 </script>
42 43</head> 44 45<body> 46 47<nav class="topnav"> 48 <div class="topnav-inner"> 49 <a class="brand" href="/">Chen-Yang Su</a> 50 <div class="nav-links"> 51 <a href="/#about">About</a> 52 <a href="/#news">News</a> 53 <a href="/#publications">Publications</a> 54 <a href="/#talks">Talks</a> 55 <a href="/#awards">Awards</a> 56 <a href="/#experience">Experience</a> 57 <a href="/#service">Teaching & Service</a> 58 <a href="/#personal">Personal</a> 59 <a href="/#motivation">Motivation</a> 60 <a href="/blog/" class="active">Blog</a> 61 <a href="/#contact">Contact</a> 62 </div> 63 </div> 64</nav> 65 66<div class="page page-post"> 67 68 <article> 69 70 <p class="crumbs"><a href="/">Home</a> <span class="sep">/</span> <a href="/blog/">Blog</a></p> 71 72 <p class="post-kicker">Careers</p> 73 74 <h1 class="post-title">Optimism in the face of uncertainty</h1> 75 76 <p class="post-standfirst">What a principle from reinforcement learning can teach us about 77 research, careers, and taking intelligent risks.</p> 78 79 <p class="post-byline"> 80 <b>Chen-Yang Su</b> 81 <span class="dot">·</span> 82 <time datetime="2026-08-30">30 August 2026</time> 83 <span class="dot">·</span> 84 <span data-reading-time> </span> 85 </p> 86 87 <div class="takeaways"> 88 <span class="takeaways-label">Main takeaways</span> 89 <ul> 90 <li>The phrase is not a slogan. It comes from bandit algorithms such as 91 <strong>Upper Confidence Bound</strong>, where an action becomes attractive not 92 because its estimated reward is high, but because the uncertainty around that 93 estimate leaves room for it to be much better.</li> 94 <li>Optimism here does not mean believing things will work out. It means treating 95 <strong>uncertainty as something with value</strong>, because an experiment that 96 fails still returns information.</li> 97 <li>Scoring a decision by its single outcome confuses the <strong>quality of the 98 decision</strong> with the result of one trial.</li> 99 <li>The bets worth taking share three properties: <strong>survivable downside, 100 meaningful upside, and something learned either way</strong>. That rules out 101 recklessness as firmly as it rules out timidity.</li> 102 <li>Joëlle Pineau calls a degree a <strong>licence to take risks</strong>, and 103 makes the same argument about open science one level up: you give up some control 104 over what happens to your work, and far more people can build on it.</li> 105 </ul> 106 </div> 107 108 <div class="prose"> 109 110 <p>In June 2026, my MSc advisor, 111 <a href="https://www.mcgill.ca/qls/researchers/joelle-pineau" target="_blank" rel="noopener">Joëlle Pineau</a>, 112 received an 113 <a href="https://www.uottawa.ca/faculty-education/news-all/uottawa-confer-nine-honorary-doctorates-spring-convocation" target="_blank" rel="noopener">honorary doctorate from the University of Ottawa</a> 114 and delivered the commencement address to the graduating class of its Faculty of 115 Engineering. Joëlle is a professor at McGill, previously led Meta’s Fundamental 116 AI Research organization, and is now Chief AI Officer at Cohere.</p> 117 118 <p>Her address had a title that immediately caught my attention: 119 <a href="https://www.linkedin.com/pulse/optimism-face-uncertainty-joelle-pineau-r2c7e" target="_blank" rel="noopener"><b>“Optimism in the Face of Uncertainty.”</b></a></p> 120 121 <p>It sounds like a piece of general life advice. But the idea came from somewhere much 122 more precise: mathematics.</p> 123 124 <p>Joëlle described a principle she encountered while studying how intelligent 125 agents should make decisions when they do not know exactly what will happen next.</p> 126 127 <p>When faced with several possible actions, the obvious strategy is to choose the one 128 that currently looks best.</p> 129 130 <p>But that is not always the best strategy.</p> 131 132 <p>Sometimes the most interesting action is the one whose outcome is uncertain, but whose 133 <strong>upside could be much larger</strong>.</p> 134 135 <p>This principle is known as <strong>optimism in the face of uncertainty</strong>.</p> 136 137 <p>And the more I thought about it, the less it seemed like a lesson about reinfor
137cement 138 learning and the more it seemed like a useful way to think about research, careers, and 139 life.</p> 140 141 <h2>A mathematical reason to be optimistic</h2> 142 143 <p>Imagine that an agent has several actions available to it.</p> 144 145 <p>One option has been tried many times. Its reward is reasonably good and reasonably 146 predictable.</p> 147 148 <p>Another has barely been explored. Its average estimated reward might look lower, but 149 because the agent knows so little about it, there is considerable uncertainty around that 150 estimate.</p> 151 152 <p>A purely greedy agent chooses the first option.</p> 153 154 <p>An optimistic agent asks a different question:</p> 155 156 <p><strong>How good could this unexplored option plausibly be?</strong></p> 157 158 <p>This basic idea appears throughout reinforcement learning and the study of multi-armed 159 bandits. Algorithms such as <strong>Upper Confidence Bound (UCB)</strong> deliberately 160 account for uncertainty when deciding what to explore. An action can become attractive not 161 merely because its estimated reward is high, but because there remains a plausible chance 162 that it is substantially better than we currently believe.</p> 163 164 <div class="callout callout-note"> 165 <span class="callout-label">The principle</span> 166 <p>Do not rank actions by their estimated reward. Rank them by the <b>upper end of 167 what they could plausibly be worth</b>, given how little you know. An action you have 168 barely tried carries a wide confidence interval, and the top of that interval is a 169 reason to try it again.</p> 170 </div> 171 172 <p>There is an elegant logic behind this.</p> 173 174 <p>If the optimistic estimate turns out to be right, you discover something valuable.</p> 175 176 <p>If it turns out to be wrong, you still gain information.</p> 177 178 <p>As Joëlle put it in her commencement address, when reality falls short of your 179 optimism, you have not necessarily lost: <strong>you have gained knowledge.</strong></p> 180 181 <p>That is a subtle but important distinction.</p> 182 183 <p>Optimism here does not mean convincing yourself that everything will work out.</p> 184 185 <p>It means recognizing that <strong>uncertainty can have value</strong>.</p> 186 187 <h2>The same problem appears in our careers</h2> 188 189 <p>The connection to life becomes especially clear around graduation.</p> 190 191 <p>Suddenly, there is an enormous decision tree in front of you.</p> 192 193 <p>Which job should you take? Which city should you move to? Should you stay in academia? 194 Join industry? Start something? Work on the established problem where you know you can 195 succeed, or the strange new question where you have no idea what will happen?</p> 196 197 <p>We often respond to these decisions by trying to eliminate uncertainty.</p> 198 199 <p>We gather more information. Make spreadsheets. Compare salaries. Ask more people for 200 advice. Try to predict exactly where each path will lead.</p> 201 202 <p>Some of this is obviously sensible.</p> 203 204 <p>But eventually there is a limit to how much more thinking can tell you.</p> 205 206 <p>At that point, uncertainty is not something that can be reasoned away. It has to be 207 <strong>experienced</strong>.</p> 208 209 <p>You have to choose an action and observe what happens.</p> 210 211 <p>That is why one of my favourite lines from Joëlle’s address was her 212 description of a degree as a <strong>“license to take risks.”</strong> She 213 encouraged the graduates to say yes to the project that seems too ambitious, the role they 214 do not quite feel ready for, or the question nobody in their field has thought to ask.</p> 215 216 <p>I think that is particularly good advice early in a career.</p> 217 218 <p>When you are young, the benefits of exploration have a long time to compound.</p> 219 220 <p>A project that goes nowhere might introduce you to an entirely new field.</p> 221 222 <p>A job that turns out not to fit might clarify what kind of work you actually want to 223 do.</p> 224 225 <p>An unsuccessful application might lead you to meet someone you otherwise never would 226 have spoken to.</p> 227 228 <p>A research question that fails might reveal a much better question hiding underneath 229 it.</p> 230 231 <p>The outcome itself is only part of the return.</p> 232 233 <p>The other part is what the experiment teaches you.</p> 234 235 <h2>Optimize for information gain</h2> 236 237 <p>We tend to evaluate past decisions with an overly simple scoring system.</p> 238 239 <p>The paper was accepted: good decision.</p> 240 241 <p>The paper was rejected: bad decision.</p> 242 243 <p>The startup succeeded: good decision.</p> 244 245 <p>The startup failed: bad decision.</p> 246 247 <p>I got the position: good decision.</p> 248 249 <p>I didn’t get it: bad decision.</p> 250 251 <p>
251But that confuses the <strong>quality of a decision</strong> with the 252 <strong>outcome of a single trial</strong>.</p> 253 254 <p>Suppose you spend three months exploring an ambitious research idea and eventually 255 conclude that it does not work.</p> 256 257 <p>You could call those three months wasted.</p> 258 259 <p>But perhaps during those three months you learned a new method, became familiar with a 260 literature you had never encountered, met several collaborators, discovered why an 261 apparently promising idea fails, and stumbled across a different problem that turns out to 262 be much more interesting.</p> 263 264 <p>The experiment failed.</p> 265 266 <p>The information gain did not.</p> 267 268 <p>That suggests a question worth adding whenever we face an uncertain opportunity:</p> 269 270 <p><strong>What will I know after doing this that I don’t know now?</strong></p> 271 272 <p>This changes how many opportunities look.</p> 273 274 <p>Something with only a moderate probability of immediate success may still be an 275 excellent bet if:</p> 276 277 <ul> 278 <li>the downside is limited,</li> 279 <li>the potential upside is large, and</li> 280 <li>failure will teach you something useful.</li> 281 </ul> 282 283 <p>That is a very different criterion from simply choosing the path most likely to 284 succeed.</p> 285 286 <h2>Run cheap experiments</h2> 287 288 <p>There is, however, an important caveat.</p> 289 290 <div class="callout callout-warn"> 291 <span class="callout-label">Optimism is not recklessness</span> 292 <p>“Optimism in the face of uncertainty” should not become an intellectual 293 justification for recklessness. Some decisions have irreversible consequences. Some 294 risks have enormous downside. <b>The fact that an outcome is uncertain does not 295 automatically make it worth exploring.</b></p> 296 </div> 297 298 <p>The most attractive experiments are often the opposite: <strong>small downside, large 299 upside, fast feedback</strong>.</p> 300 301 <ul> 302 <li>Send an email to someone whose work you admire.</li> 303 <li>Apply for the position you think you are slightly underqualified for.</li> 304 <li>Spend a weekend building the prototype.</li> 305 <li>Try the new research direction for two weeks.</li> 306 <li>Submit the paper.</li> 307 <li>Take the difficult course.</li> 308 <li>Write the essay.</li> 309 <li>Go to the conference.</li> 310 <li>Ask the question.</li> 311 </ul> 312 313 <p>These actions can fail.</p> 314 315 <p>But most failures are cheap.</p> 316 317 <p>And occasionally one of them changes the trajectory of your life.</p> 318 319 <p>Perhaps the better rule, then, is not simply <i>take more risks</i>.</p> 320 321 <div class="callout callout-good"> 322 <span class="callout-label">The rule</span> 323 <p><strong>Seek uncertainty when the downside is survivable, the upside is meaningful, 324 and the outcome will teach you something.</strong></p> 325 <p>The goal is not to maximize the number of risks you take. It is to run 326 <b>better experiments</b>.</p> 327 </div> 328 329 <h2>There is an asymmetry to opportunity</h2> 330 331 <p>Another reason this framing is useful is that many interesting opportunities have 332 asymmetric payoffs.</p> 333 334 <p>Imagine emailing a researcher you admire.</p> 335 336 <p>Maybe there is a 90% chance nothing happens.</p> 337 338 <p>That sounds like a terrible bet if you evaluate it by probability of success alone.</p> 339 340 <p>But suppose the cost of sending the email is ten minutes, while the 10% outcome could 341 produce a collaboration, mentorship, job opportunity, or entirely new research direction.</p> 342 343 <p>The expected value begins to look very different.</p> 344 345 <p>The same principle applies to applications, introductions, side projects, conferences, 346 writing publicly, and ambitious research questions.</p> 347 348 <p>Humans often overweight the probability of embarrassment or rejection and underweight 349 the magnitude of the upside.</p> 350 351 <p>We see the likely outcome:</p> 352 353 <p><strong>They probably won’t respond.</strong></p> 354 355 <p>We pay much less attention to:</p> 356 357 <p><strong>But what if they do?</strong></p> 358 359 <p>Optimism in the face of uncertainty forces us to consider both.</p> 360 361 <h2>Research itself depends on this mindset</h2> 362 363 <p>The idea feels particularly natural in science.</p> 364 365 <p>Good research begins with uncertainty.</p> 366 367 <p>If we already knew the answer, there would be little reason to run the experiment.</p> 368 369 <p>Yet researchers can gradually become surprisingly good at avoiding uncertainty.</p> 370 371 <p>We learn what reviewers like.</p> 372 373 <p>
373We learn which analyses are predictable.</p> 374 375 <p>We learn which questions can produce another incremental publication.</p> 376 377 <p>And those incentives can push us toward exploitation rather than exploration: 378 repeatedly mining areas where we already know rewards are available.</p> 379 380 <p>But some of the most interesting research exists precisely where the confidence 381 intervals are widest.</p> 382 383 <p>A strange result.</p> 384 385 <p>A new dataset.</p> 386 387 <p>A methodology borrowed from another discipline.</p> 388 389 <p>A question that sounds slightly ridiculous when you first say it aloud.</p> 390 391 <p>Most of these directions will not produce breakthroughs.</p> 392 393 <p>That is the point.</p> 394 395 <p>Exploration is valuable because you cannot know beforehand which one will.</p> 396 397 <h2>Open science as an optimistic bet</h2> 398 399 <p>Joëlle also connected this philosophy to another theme that has defined much of her 400 career: <strong>open science</strong>.</p> 401 402 <p>For years, she has pushed machine learning toward greater reproducibility and more 403 transparent research practices. Her work helped establish reproducibility initiatives at 404 NeurIPS, including code submission practices, reproducibility challenges, and checklists 405 for reporting machine-learning experiments. In a 2021 406 <a href="https://www.jmlr.org/papers/v22/20-303.html" target="_blank" rel="noopener"><i>Journal of Machine Learning Research</i> paper</a>, 407 she and her co-authors argued that reproducibility helps researchers verify findings, 408 build on one another’s work, and move ideas into practice more reliably.</p> 409 410 <p>There is an element of optimism here too.</p> 411 412 <p>When you release code, data, models, or ideas openly, you surrender some control over 413 what happens next.</p> 414 415 <p>Someone may find a mistake.</p> 416 417 <p>Someone may use your work differently than you intended.</p> 418 419 <p>Someone you have never met may take your idea somewhere you never anticipated.</p> 420 421 <p>There is uncertainty.</p> 422 423 <p>But there is also enormous upside.</p> 424 425 <p>Thousands of researchers can inspect, reproduce, modify, and extend something that 426 would otherwise remain confined to a single laboratory.</p> 427 428 <p>In her commencement address, Joëlle described open source in essentially these 429 terms: sharing carries risks, but inviting more people to participate dramatically expands 430 what can be created.</p> 431 432 <p>The philosophy is consistent.</p> 433 434 <p>You accept some uncertainty because the space of possible positive outcomes becomes 435 much larger.</p> 436 437 <h2>The future cannot be optimized in advance</h2> 438 439 <p>There is a temptation, especially among ambitious people, to treat life as an 440 optimization problem.</p> 441 442 <p>Find the optimal degree.</p> 443 444 <p>The optimal lab.</p> 445 446 <p>The optimal company.</p> 447 448 <p>The optimal city.</p> 449 450 <p>The optimal career.</p> 451 452 <p>Collect enough information and, theoretically, the correct path will reveal itself.</p> 453 454 <p>But life is not a static optimization problem.</p> 455 456 <p>The environment changes.</p> 457 458 <p>You change.</p> 459 460 <p>New technologies appear.</p> 461 462 <p>People enter your life.</p> 463 464 <p>Interests you thought were permanent disappear.</p> 465 466 <p>Opportunities emerge that you could not possibly have anticipated five years 467 earlier.</p> 468 469 <p>There is no dataset containing enough information to calculate your optimal 470 trajectory.</p> 471 472 <p>At some point, you have to explore.</p> 473 474 <p>That may be the part of Joëlle’s speech that resonates most with me.</p> 475 476 <p>Optimism does not require believing that every choice will succeed.</p> 477 478 <p>In fact, you should expect many of your experiments not to.</p> 479 480 <p>Papers will be rejected.</p> 481 482 <p>People won’t reply.</p> 483 484 <p>Projects will stall.</p> 485 486 <p>Applications will fail.</p> 487 488 <p>Predictions you made confidently about your future will turn out to be completely 489 wrong.</p> 490 491 <p>The optimistic belief is something different:</p> 492 493 <p><strong>that acting is still worthwhile because action generates information.</strong></p> 494 495 <p>You try something.</p> 496 497 <p>The world responds.</p> 498 499 <p>You update.</p> 500 501 <p>And then you try again.</p> 502 503 <h2>Choose again</h2> 504 505 <p>When faced with uncertainty, our instinct is often to wait until we feel more 506 certain.</p> 507 508 <p>Sometimes that is appropriate.</p> 509 510 <p>But often the information we need is located on the other side of action.</p> 511 512 <div class="callout"> 513 <span class="callout-label">Two better questions</span> 514 <p>Instead of asking <i>which choice guarantees success?</i>, ask 515 <b>which choice has the most interesting upside?</b></p> 516 <p>Instead of asking <i>what if this fails?</i>, ask 517 <b>what will I learn if it does?</b></p> 518 </div> 519 520 <p>Instead of trying to map your entire career from where you stand today, run the next 521 useful experiment.</p> 522 523 <p>Choose an action.</p> 524 525 <p>Observe.</p> 526 527 <p>Update.</p> 528 529 <p>Choose again.</p> 530 531 <p>The future is uncertain by definition.</p> 532 533 <p>That is not necessarily a problem to solve.</p> 534 535 <p>Sometimes it is the reason to act.</p> 536 537 <h2>Further reading</h2> 538 539 <ul> 540 <li><a href="https://www.linkedin.com/pulse/optimism-face-uncertainty-joelle-pineau-r2c7e" target="_blank" rel="noopener">Joëlle Pineau, “Optimism in the Face of Uncertainty”</a> — the commencement address this post is about.</li> 541 <li><a href="https://www.uottawa.ca/about-us/president/honorary-doctorates" target="_blank" rel="noopener">University of Ottawa, honorary doctorate recipients</a></li> 542 <li><a href="https://www.jmlr.org/papers/v22/20-303.html" target="_blank" rel="noopener">Pineau et al., “Improving Reproducibility in Machine Learning Research”</a>, <i>JMLR</i> 22(164), 2021.</li> 543 <li><a href="https://www.cs.mcgill.ca/~jpineau/" target="_blank" rel="noopener">Joëlle Pineau’s website</a>, and her <a href="https://www.mcgill.ca/qls/researchers/joelle-pineau" target="_blank" rel="noopener">research profile at McGill</a></li> 544 </ul> 545 546 </div><!-- /.prose --> 547 548 <div class="post-foot"> 549 <div class="tags"> 550 <span class="tag">reinforcement learning</span> 551 <span class="tag">exploration</span> 552 <span class="tag">research careers</span> 553 <span class="tag">open science</span> 554 </div> 555 <nav class="post-nav" aria-label="Post navigation"> 556 <a href="/blog/">← All posts</a> 557 <a class="js-email" href="/#contact">Reply by email</a> 558 </nav> 559 </div> 560 561 </article> 562 563 <footer class="footer"> 564 <p class="footer-main">© <span id="yr">2026</span> <b>Chen-Yang Su</b></p> 565 <p class="fineprint"> 566 <a href="/">Home</a> · 567 <a href="/blog/">Blog</a> · 568 <a href="https://github.com/chenyangsu">GitHub</a> · 569 <a href="https://scholar.google.com/citations?user=KJGE-kYAAAAJ&hl=en">Google Scholar</a> 570 </p> 571 </footer> 572 573</div> 574 575<button id="themeToggle" type="button" aria-label="Switch to dark mode" title="Switch to dark mode" aria-pressed="false"><svg class="icon-moon" width="19" height="19" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M21 12.79A9 9 0 1 1 11.21 3 7 7 0 0 0 21 12.79z"></path></svg><svg class="icon-sun" width="19" height="19" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><circle cx="12" cy="12" r="4.5"></circle><path d="M12 1.5v2M12 20.5v2M4.2 4.2l1.5 1.5M18.3 18.3l1.5 1.5M1.5 12h2M20.5 12h2M4.2 19.8l1.5-1.5M18.3 5.7l1.5-1.5"></path></svg></button> 576<button id="backToTop" aria-label="Back to top" title="Back to top"><svg width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.5" stroke-linecap="round" stroke-linejoin="round"><polyline points="18 15 12 9 6 15"></polyline></svg></button> 577
578<script src="/assets/blog.js"></script>
578 579 580</body> 581</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.