ASHER ELLIS's guest appearances

ASHER ELLIS

The Horror Show with Brian Keene

Make Me Read Your Book Podcast

The Nonlinear Library

Paul Tremblay

John Wayne Comunale

James White

Jacob Haddon

Kristopher Triana

Ronald Malfi

Brian Keene

Michael Allen

John Boden

Lucas Mangum

Mary SanGiovanni

Jonathan Janz

Bracken MacLeod

Jeff Strand

Matt Hayward

OwlTail

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI Performance on Human Tasks, published by Asher Ellis on March 3, 2022 on LessWrong.This was an independent project with guidance and support from Vael Gates and Richard Korzekwa--thank you both for your mentorship!IntroductionThis project was inspired by a prompt from the AI Impacts project:For intellectual skills where machines have surpassed humans, find out how long it took to go from the worst performance to average human skill, and from average human skill to superhuman skill.I initially tried to answer this question for several task areas. But over the course of this project, I pivoted more towards analyzing the advantages and disadvantages computer programs have over humans in each task, including the sources of error. I made this pivot mostly for practical reasons. For some tasks, machine programs are only tested against expert human performance, making them difficult to compare to “beginner” and “average” humans. In other tasks, the differences between human and machine performance are qualitative rather than quantitative. It seemed most appropriate to describe how machines and humans differ in performance rather than quantify it. I used this information to predict whether AI will replace or augment humans in performing these task areas.I picked five task areas, each representing different aspects of human cognition:Poker-playing (strategy and interdependent decision-making)Image classification (a key subset of vision and perception)Text-summarization (interpreting and manipulating text)Creating static visual art (aesthetics and ingenuity)Human-like dexterity (manipulating physical objects)Here is a summary of my conclusions:TaskCurrent capabilitiesPokerSuperhuman (consistently)Image classificationSuperhuman (usually)Text-summarizationAverage human (unreliable)Static visual artSuperhuman (but requires human input)Human-like dexterityBelow humans (except specific tasks)See "Discussion / Personal Predictions" section for my predictions in how AI will machines will replace or augment humans in these tasks.Task 1: PokerPoker-playing is a strong case study for many aspects of human decision-making. Most significantly, it involves strategic decision-making against an opponent. Strong poker players use betting strategies, each having different consequences, which correspond to risk management and strategy broadly. Unlike chess and checkers, poker requires deciding under imperfect knowledge since you never know your opponent’s current hand. You must interpret your opponent’s bets despite never being certain what cards they hold. It involves identifying and exploiting your opponent’s patterns – a form of agent modeling. It involves deception through bluffing and switching between strategies. Playing against an opponent who attempts to deceive you involves acting under unreliable information. Machines encounter each of these challenges when playing poker against humans.The first serious attempt at a “Poker bot” was Michael Caro’s ORAC, which played both limit and no-limit Texas Hold 'Em. ORAC reportedly “rolled over professional-level opponents in private tests in 1983 and even won most of its matches against [Michael Caro].” Yet at the 1984 World Series of Poker, ORAC showed mixed results against world champion humans, mostly losing. It is difficult to conclude how ORAC would play against other humans, as its performance was never officially documented. ORAC appears to be better than a first-time human poker player but worse than top professionals.The next notable poker program was Loki, created by the University of Alberta in 1997 to play multiple-player Limit Hold’em. Loki betted by counting the potential hands that were better than, worse than, and equal to Loki’s hand. It then estimated the probability its hand was the best one generated. It ...

Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI Performance on Human Tasks, published by Asher Ellis on March 3, 2022 on LessWrong. This was an independent project with guidance and support from Vael Gates and Richard Korzekwa--thank you both for your mentorship! Introduction This project was inspired by a prompt from the AI Impacts project: For intellectual skills where machines have surpassed humans, find out how long it took to go from the worst performance to average human skill, and from average human skill to superhuman skill. I initially tried to answer this question for several task areas. But over the course of this project, I pivoted more towards analyzing the advantages and disadvantages computer programs have over humans in each task, including the sources of error. I made this pivot mostly for practical reasons. For some tasks, machine programs are only tested against expert human performance, making them difficult to compare to “beginner” and “average” humans. In other tasks, the differences between human and machine performance are qualitative rather than quantitative. It seemed most appropriate to describe how machines and humans differ in performance rather than quantify it. I used this information to predict whether AI will replace or augment humans in performing these task areas. I picked five task areas, each representing different aspects of human cognition: Poker-playing (strategy and interdependent decision-making) Image classification (a key subset of vision and perception) Text-summarization (interpreting and manipulating text) Creating static visual art (aesthetics and ingenuity) Human-like dexterity (manipulating physical objects) Here is a summary of my conclusions: TaskCurrent capabilitiesPokerSuperhuman (consistently)Image classificationSuperhuman (usually)Text-summarizationAverage human (unreliable)Static visual artSuperhuman (but requires human input)Human-like dexterityBelow humans (except specific tasks) See "Discussion / Personal Predictions" section for my predictions in how AI will machines will replace or augment humans in these tasks. Task 1: Poker Poker-playing is a strong case study for many aspects of human decision-making. Most significantly, it involves strategic decision-making against an opponent. Strong poker players use betting strategies, each having different consequences, which correspond to risk management and strategy broadly. Unlike chess and checkers, poker requires deciding under imperfect knowledge since you never know your opponent’s current hand. You must interpret your opponent’s bets despite never being certain what cards they hold. It involves identifying and exploiting your opponent’s patterns – a form of agent modeling. It involves deception through bluffing and switching between strategies. Playing against an opponent who attempts to deceive you involves acting under unreliable information. Machines encounter each of these challenges when playing poker against humans. The first serious attempt at a “Poker bot” was Michael Caro’s ORAC, which played both limit and no-limit Texas Hold 'Em. ORAC reportedly “rolled over professional-level opponents in private tests in 1983 and even won most of its matches against [Michael Caro].” Yet at the 1984 World Series of Poker, ORAC showed mixed results against world champion humans, mostly losing. It is difficult to conclude how ORAC would play against other humans, as its performance was never officially documented. ORAC appears to be better than a first-time human poker player but worse than top professionals. The next notable poker program was Loki, created by the University of Alberta in 1997 to play multiple-player Limit Hold’em. Loki betted by counting the potential hands that were better than, worse than, and equal to Loki’s hand. It then estimated the probability its hand was the best one generated. It ...

LW - AI Performance on Human Tasks by Asher Ellis

Author Asher Ellis and I talk about his books and some short films he was involved with as well. We compare anthologies to Jock Jams and Patrick Swayze to Chris Hemsworth. I also talk about Book Tattoos and Asher's book PET.

Author Asher Ellis and I talk about his books and some short films he was involved with as well. We compare anthologies to Jock Jams and Patrick Swayze to Chris Hemsworth. I also talk about Book Tattoos and Asher's book PET.

EP 11: Asher Ellis

Asher Ellis joins Brian, Mary, and special guest John Urbancik for a discussion on screenwriting, playwriting, his new novel PET, James Fry, and much more. This week, Matt hosts the show and gives some of his Top 5's for topics such as: Novels, Video Games, Food, Comics and Horror Movies.

ASHER ELLIS - The Horror Show With Brian Keene - Ep 240