Differences

This shows you the differences between two versions of the page.

--- arguments_for_ai_risk:quantitative_estimates_of_ai_risk [2023/06/21 20:59]
jeffreyheninger Adding a few more estimates; Knightian uncertainty
+++ arguments_for_ai_risk:quantitative_estimates_of_ai_risk [2023/12/01 18:15] (current)
harlanstewart
@@ Line 1: / Line 1: @@
 ====== Quantitative Estimates of AI Risk ======
+/*
+COMMENT:
+Things to add to this:
+- https://optimists.ai/2023/11/28/ai-is-easy-to-control/
+*/
 // This page is in an early draft. It is very incomplete and may contain errors. //
@@ Line 18: / Line 23: @@
 The table below includes estimates from individuals working in AI Safety of how likely very bad outcomes due to AI are.
-Many of the individuals expressed [[https://en.wikipedia.org/wiki/Knightian_uncertainty|Knightian uncertainty]] when making their estimates, saying that their probability varies day-to-day, or that the estimate is currently in development, or that this is a very quick-and-dirty estimate. People who have explicitly said something like this include Katja Grace, Joe Carlsmith, Peter Wildeford, Nate Soares, Paul Christiano, and others. These estimates should not be treated as definitive statements of these individuals' beliefs, but rather as glimpses of their thinking at that moment.
+Many of the individuals expressed [[https://en.wikipedia.org/wiki/Knightian_uncertainty|Knightian uncertainty]] when making their estimates, saying that their probability varies day-to-day, or that the estimate is currently in development, or that this is a very quick-and-dirty estimate. People who have explicitly said something like this include Katja Grace, Joseph Carlsmith, Peter Wildeford, Nate Soares, Paul Christiano, and others. These estimates should not be treated as definitive statements of these individuals' beliefs, but rather as glimpses of their thinking at that moment.
 Each estimate includes:
@@ Line 26: / Line 31: @@
   * The estimate the individual gives for the probability that AI development causes a very bad outcome.
   * The source for this estimate.
+  * Whether this is the person's most recent public estimate that we are aware of and whether this is this person's best guess as opposed to a conditional estimate.
 The estimates are in no particular order. The table can be sorted by clicking at the top of each column.
@@ Line 63: / Line 69: @@
     <th onclick="sortTable(3)">Probability</th>
     <th onclick="sortTable(4)">Source</th>
+    <th onclick="sortTable(5)">Most Recent?</th>
   </tr>
   <tr>
@@ Line 70: / Line 77: @@
     <td>0.19</td>
     <td><a href="https://youtu.be/j5Lu01pEDWA?t=2257">Will AI end everything? A guide to guessing</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
-    <td>Joe Carlsmith</td>
+    <td>Joseph Carlsmith</td>
     <td>2021</td>
     <td>Existential catastrophe by 2070 from advanced, planning, strategic AI</td>
     <td>0.05</td>
     <td><a href="https://arxiv.org/pdf/2206.13353.pdf">Is Power-Seeking AI an Existential Risk?</a></td>
+    <td>No</td>
   </tr>
   <tr>
-    <td>Joe Carlsmith</td>
+    <td>Joseph Carlsmith</td>
     <td>2022</td>
     <td>Existential catastrophe by 2070 from advanced, planning, strategic AI</td>
-    <td>0.1+</td>
+    <td>0.1+ (Greater than 10%)</td>
     <td>Update to: <a href="https://arxiv.org/pdf/2206.13353.pdf">Is Power-Seeking AI an Existential Risk?</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 91: / Line 101: @@
     <td>0.77</td>
     <td><a href="https://www.lesswrong.com/posts/cCMihiwtZx7kdcKgt/comments-on-carlsmith-s-is-power-seeking-ai-an-existential">Comments on Carlsmith's "Is power-seeking AI an existential risk?"</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 98: / Line 109: @@
     <td>0.1</td>
     <td><a href="https://theprecipice.com/">The Precipice</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 105: / Line 117: @@
     <td>0.2</td>
     <td><a href="https://80000hours.org/podcast/episodes/toby-ord-the-precipice-existential-risk-future-humanity/#transcript">Toby Ord on the precipice and humanity's potential futures</a></td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 112: / Line 125: @@
     <td>~1</td>
     <td><a href="https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities">AGI Ruin: A List of Lethalities</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 119: / Line 133: @@
     <td>0.1</td>
     <td><a href="https://wiki.aiimpacts.org/doku.php?id=conversation_notes:conversation_with_rohin_shah">Conversation with Rohin Shah</a></td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 126: / Line 141: @@
     <td>0.1</td>
     <td><a href="https://wiki.aiimpacts.org/doku.php?id=conversation_notes:conversation_with_paul_christiano">Conversation with Paul Christiano</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 133: / Line 149: @@
     <td>0.22</td>
     <td>Slack channel & private conversation</td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 140: / Line 157: @@
     <td>0.6 - 0.7</td>
     <td><a href="https://wiki.aiimpacts.org/doku.php?id=conversation_notes:conversation_with_adam_gleave">Conversation with Adam Gleave</td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 147: / Line 165: @@
     <td>0.3 - 0.4</td>
     <td><a href="https://wiki.aiimpacts.org/doku.php?id=conversation_notes:conversation_with_adam_gleave">Conversation with Adam Gleave</td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 154: / Line 173: @@
     <td>0.1 - 0.2</td>
     <td><a href="https://wiki.aiimpacts.org/doku.php?id=conversation_notes:conversation_with_adam_gleave">Conversation with Adam Gleave</td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 161: / Line 181: @@
     <td>0.05</td>
     <td><a href="https://futureoflife.org/podcast/an-overview-of-technical-ai-alignment-in-2018-and-2019-with-buck-shlegeris-and-rohin-shah/">AI Alignment Podcast: An Overview of Technical AI Alignment in 2018 and 2019 with Buck Shlegeris and Rohin Shah</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 168: / Line 189: @@
     <td>0.5</td>
     <td><a href="https://futureoflife.org/podcast/an-overview-of-technical-ai-alignment-in-2018-and-2019-with-buck-shlegeris-and-rohin-shah/">AI Alignment Podcast: An Overview of Technical AI Alignment in 2018 and 2019 with Buck Shlegeris and Rohin Shah</a></td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 175: / Line 197: @@
     <td>0.0005</td>
     <td><a href="https://forum.effectivealtruism.org/posts/2sMR7n32FSvLCoJLQ/critical-review-of-the-precipice-a-reassessment-of-the-risks">Critical Review of 'The Precipice': A Reassessment of the Risks of AI and Pandemics</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 182: / Line 205: @@
     <td>0.25</td>
     <td><a href="https://youtu.be/YTlrPeikoyw?t=1768">The current alignment plan, and how we might improve it</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 189: / Line 213: @@
     <td>0.33 - 0.5</td>
     <td><a href="https://youtu.be/i4LjoJGpqIY?t=2380">The future is going to be wonderful if we don't get whacked</a></td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 196: / Line 221: @@
     <td>0.05 - 0.3</td>
     <td><a href="https://youtu.be/WLXuZtWoRcE?t=1229">Is AI an existential threat? We don't know, and we should work on it</a></td>
+    <td>Yes</td>
   </tr>
   <tr>
@@ Line 203: / Line 229: @@
     <td>0.33 - 0.7</td>
     <td><a href="https://futureoflife.org/podcast/rohin-shah-on-the-state-of-agi-safety-research-in-2021/">Rohin Shah on the State of AGI Safety Research in 2021</a></td>
+    <td>No</td>
   </tr>
   <tr>
@@ Line 210: / Line 237: @@
     <td>0.35</td>
     <td><a href="https://forum.effectivealtruism.org/posts/9Y6Y6qoAigRC7A8eX/my-take-on-what-we-owe-the-future">My take on What We Owe the Future</a></td>
+    <td>Yes</td>
   </tr>
+  <tr>
+    <td>Katja Grace</td>
+    <td>2022</td>
+    <td>AI destroys the world</td>
+    <td>0.07</td>
+    <td><a href="https://theinsideview.ai/katja#why-katja-thinks-there-is-a-7-chance-of-ai-destroys-the-world">Katja Grace on Slowing Down AI and Surveys</a></td>
+    <td>No</td>
+  </tr>
+  <tr>
+    <td>Andrew Critch</td>
+    <td>2023</td>
+    <td>Humanity not surviving the next 50 years</td>
+    <td>0.8</td>
+    <td><a href="https://www.lesswrong.com/posts/gZkYvA6suQJthvj4E/my-may-2023-priorities-for-ai-x-safety-more-empathy-more">My May 2023 priorities for AI x-safety: more empathy, more unification of concerns, and less vilification of OpenAI</a></td>
+    <td>Yes</td>
+  </tr>
+  <tr>
+    <td>Andrew Critch</td>
+    <td>2023</td>
+    <td>Humanity not surviving the next 50 years, without a major international regulatory effort to control how AI is used</td>
+    <td>0.9+</td>
+    <td><a href="https://www.lesswrong.com/posts/gZkYvA6suQJthvj4E/my-may-2023-priorities-for-ai-x-safety-more-empathy-more">My May 2023 priorities for AI x-safety: more empathy, more unification of concerns, and less vilification of OpenAI</a></td>
+    <td>No</td>
+  </tr>
+  <tr>
+    <td>Scott Aaronson</td>
+    <td>2023</td>
+    <td>The generative AI race, which started in earnest around 2016 or 2017 with the founding of OpenAI, to play a central causal role in the extinction of humanity</td>
+    <td>0.02</td>
+    <td><a href="https://scottaaronson.blog/?p=7064">Why am I not terrified of AI?</td>
+    <td>Yes</td>
+  </tr>
 </table>
@@ Line 273: / Line 334: @@
 </body>
 </html>
+===== Framings =====
+Different people use different framings to arrive at their estimate of AI risk. The most common framing seems to be to describe a model of what the risk from advanced AI looks like, assign probabilities to various components of that model, and then calculate the existential risk from AI on the basis of this model. Another framing is to describe various scenarios for the future of AI, assign probabilities to the various scenarios, and then add together the probabilities of the different scenarios to determine the total existential risk from AI. There are also some people who give a probability without describing what framing they used to get this number.
+Below is an example of each of these two framings, due to Joseph Carlsmith and Peter Wildeford, respectively. Both individuals have updated their estimates since publishing their framing, so neither probability breakdown reflects the author's most recent estimate of AI risk. They are included to show how these framings work.
+==== Model ====
+One example of using a model to calculate the existential risk from AI is due to Joseph Carlsmith. He calculates AI-risk by 2070 by breaking it down in the following way:
+  - It will become possible and financially feasible to build APS [advanced, planning, strategic] systems.  **65%**
+  - There will be strong incentives to build APS systems | (1).  **80%**
+  - It will be much harder to develop APS systems that would be practically PS-aligned [power-seeking] if deployed, than to develop APS systems that would be practically PS-misaligned if deployed (even if relevant decision-makers don’t know this), but which are at least superficially attractive to deploy anyway | (1)–(2).  **40%**
+  - Some deployed APS systems will be exposed to inputs where they seek power in misaligned and high-impact ways (say, collectively causing >$1 trillion 2021-dollars of damage) | (1)–(3).  **65%**
+  - Some of this misaligned power-seeking will scale (in aggregate) to the point of permanently disempowering ~all of humanity | (1)–(4).  **40%**
+  - This will constitute an existential catastrophe | (1)–(5).  **95%**
+The total AI risk is the product of the probabilities for each part of the model.
+This same model has been used by Nate Soares and Eli Lifland to calculate their estimates of AI risk. Several other people in the table have also used similar models.
+==== Scenarios ====
+One example of describing possible scenarios to calculate the x-risk from AI is due to Peter Wildeford. He calculates x-risk from AI by 2100 by breaking it down in the following way:
+  - We get aligned and very capable AI systems, but we do not end the time of perils. We however still survive to the end of the century.  **43%**
+  - Very capable AI systems are not developed this century. We survive until the end of the century.  **22%**
+  - An unaligned rogue AI system creates a singleton.  **13%**
+  - We get aligned and very capable AI systems that decisively end the time of perils.  **10%**
+  - We get very capable AI systems that are broadly aligned with human values, but we create an existential risk anyways via abuse of nonhuman animals and/or digital minds.  **7%**
+  - Some subset of humanity intentionally takes over the world via a very capable AI system.  **4%**
+  - An unaligned rogue AI system causes complete human extinction.  **1%**
+  - Some subset of humanity intentionally takes over the world via a very capable AI system and in the process ends up causing complete human extinction.  **0.2%**
+  - Human extinction arises through a nuclear war that is started by accident.  **0.1%**
+  - Human extinction arises through a nuclear war that is started intentionally.  **0.1%**
+  - Some unknown unknown thing causes an existential risk. **0.05%**
+  - Nanotech causes an existential risk.  **0.04%**
+  - An engineered pandemic causes an existential risk.  **0.01%**
+The total AI risk is the sum of (3), (5), (6), (7), & (8).
 //Primary author: Jeffrey Heninger//

AI Impacts Wiki

User Tools

Site Tools

Differences

Page Tools