The stakes were high. Sarah was tense as she focused on the careful manipulations she was performing on her workstation. Time was of the essence. Think a step ahead, focus on economy of action, but don’t be so rushed as to make a mistake. Even in an air conditioned office, on days like this she would stink from stress sweat at the end of the day. It was rather absurd that an aeronautical engineer designing a new jet fighter years away from production would be under this much pressure, but the body’s fight or flight response wasn’t terribly discerning. She watched the progress bar intently as her CAD suite converted the engine cowling profile she had been tweaking into a data format suitable for the finite element physics simulator. She finger rested lightly on the mouse, hovering over the job submit button, as poised as a twitch gamer.
Boom! Submitted!
She quickly brought up the browser tab containing the job queue and refreshed it. “Ha!” A single exultant cackle slipped out. She had beaten Jeff to the head of the job queue for the simulation cluster. It was 4:30pm on Thursday, and her fourteen hour simulation would be done by Friday morning, while Jeff would face the unpleasant dilemma of whether or not to come in over the weekend to analyze his results. She smiled and slumped back in her chair, flexing her shoulders to try to release some of the tension.
This was going to be a big run, covering the entire flight envelope for the fighter. She’s been working on this cowling for over a month, iteratively refining the shape of the air surface as the engine design stabilized, using the results from simulation runs large and small to understand how subtle changes in shape affected mechanical stresses and induced drag through a wide range of operating conditions. This current revision should be near perfect, possibly the final design for this important component. Tomorrow she would have the results from the simulation to pour over, but until then there was nothing for her to do for. Of course, that was not really true. This was only one of a handful of projects she was responsible for, but after busting her chops all day to get this run prepared, she was wrung out and had no interest in exerting the mental effort needed to switch to another project. There was only a half an hour left until “quitting time,” if she even hung around that long. Tomorrow would be busy and she felt no compunction about leaving early.
Her brain was still focused on the cowling though, and she idly brought up the visualization program she used to manage her design iterations and simulation runs. It was a rather clever program one of her fellow Boeing engineers had hacked together years ago. [1] She could pick attributes from her designs and simulation runs to be the axes of a graph, and then plot all her iterations on the graph with little thumbnails. For example, she could plot radius of curvature versus maximum temperature, and quickly see how tightly she could wrap a protruding mounting bracket before the temperature began to skyrocket. She sighed as she idly flipped and fiddled with the graphs. Really, she was working on a big optimization problem, using her engineering training and human pattern matching and inferencing to explore the phase space and estimate which changes would be most beneficial. Parts of her job were almost mechanical—she’d love to be able to turn this over to a computer to design, but since it took fourteen hours to run a single simulation on their hundred-teraflop cluster, compute time was at a premium and only the best optimizer was allowed to pick the sample points. Currently, that was her.
Thoughts of optimizers reminded her of the conversation she had participated in at the LessWrong meet up at Chiang’s. Here she was spending a month on a single airplane part, while the Equestria Online servers were handling natural language conversations with somewhere around one million players in real time. Jacklyn claimed that the AI was so good, they might be looking at the incipient Singularity. Was that even possible, Sarah wondered? Sure, the current game was just a way for little girls and bronies to pass the time, but what if there really was something more underneath that? What would it take to play a high stakes, winner-takes-all game of “outthink the human” on a global scale—and win? Well, a little research into the pending apocalypse was a perfectly good way to waste the last few minutes of her work day.
Sarah discovered that each time a player makes a move in a chess game, they usually have about thirty moves to choose from. At about forty moves per a chess game, there are 10^120 possible games to be explored from the first move. Comparatively, the number of atoms in the observable universe is estimated to be around 10^79. [2] Of course, thought Sarah, the chess numbers are fairly meaningless. There’s no reason to calculate every possible chess game just to play chess. The best chess programs make extensive use of heuristics to put a limited amount of computation to good use. But a Singularity level AI isn’t maneuvering chess pieces. It’s maneuvering humans. There are far more than thirty possible actions the AI can take next, and far more than thirty two pieces in the game it’s playing. The numbers become intractable very quickly.
Well, what about heuristics, considered Sarah? Could an AI generate heuristics to make evaluating a move in the human manipulation game more efficient? That in itself would be a huge challenge. The heuristics used in chess have been developed by humans over the last five hundred years. An optimizer doesn’t have a human “understanding” of its target problem, it’s just effective at applying the rules it knows. If humanity knew how to teach a computer to “understand” well enough to generate heuristics, AI would already be ubiquitous in our day-to-day lives, and clearly it’s not. Where else can the needed heuristics come from, Sarah wondered? Heuristics could presumably be evolved through random guessing and genetic algorithms, but that requires an evaluation of the behavior of the heuristic—either through a lot of real interactions with humans ending in undesired consequences, or through computationally intense simulation of humans using a simpler model. Equestria Online wouldn’t be as popular as it was if it was spending all its time annoying customers while it weeded out bad heuristics rather than using preprogrammed, known good ones. The alternative just brought the specter of simulating 86 billion neurons back to the forefront. As Sarah had told Jacklyn earlier, the hardware just isn’t there (let alone a known network of billions of neurons to simulate). Neither form of heuristic generation seemed feasible if the intelligence behind Equestria Online was supposed to be a threat in the present or near future.
Well, maybe it doesn’t have to be a traditional computer, thought Sarah. Certainly there have been analog computers built in the past. Biology does not use digital circuits. Quantum computers certainly had unusual properties. Are there other routes that she might have overlooked? After much digging (and detours through nerdy comic strips like XKCD) Sarah stumbled upon the final straw for the possibility of a Singularity-grade AI: the Church-Turing thesis. As Sarah understood it, the thesis stated that if some method exists to carry out a computation, then that same calculation can be performed by a general purpose computer. This was another way of saying that all the models of computation that humans had so far imagined were equivalently powerful and had the same limits. If biology could do it, then silicon could, and vice versa. Quantum computers were more compact than a warehouse of processors running in parallel, but added no other unique abilities. No one knew how to go about proving such a statement, given how vague the definitions were, but no one had been able to come up with a computing paradigm that rocked the boat; the thesis was almost universally accepted.
This sealed the deal, as far as Sarah was concerned. The universe and mathematics itself seemed to be set against the existence of a Singularity level AI in the present day, at least until humans could explain to a computer what it meant to “understand” something rather than trying to brute force the problem. Her job as aeronautics optimizer and humanity’s place in the pecking order seemed secure. And seeing as it was way past 5pm (damn you Wikipedia), it was time to go home.
After her victory at work, Sarah decided that some time away from the confines of her office would be her reward. She wanted to be surrounded by people, not CAD drawings, and window shop for toys that cost hundreds, not millions of dollars. She found herself in the upscale mall near her office. Purchasing a burrito that had nothing to do with Hispanic cuisine except the outer wrapper, she sat by herself at a small table that had a commanding view of the center atrium and soaked in the bustle of humanity flowing around her little bubble of isolation. After that she wandered, enjoying the colorful displays of things she didn’t need, be it a daring iridescent purple evening dress, an elegantly styled high end stereo system, bars of artisanal bath soap, or the latest epic video game. A bright pink logo caught her eye and she paused at the window of the game store. The PonyPad!
Staring at the technicolor ponies on the poster, her mind flashed back to recent thoughts. Jacklyn claimed that the natural language processing produced amazing conversations. Louis claimed that they couldn’t be believed without being experienced. Her own investigation in the topic had shown that general AI was mathematically impossible given humanity’s current understanding, but Jacklyn claimed that the Singularity was nei-… was eminent. Sarah had read the Wikipedia article about artificial intelligence too. Natural language understanding was generally considered AI-complete, meaning that solving the specific problem was equivalent to solving the general problem of making an AI that was as intelligent as a human. To handle anything that could be thrown at you in a conversation would in the end require understanding in the nature of the world. It couldn’t be done, yet the popular buzz was that the game was great. Was Jacklyn right, Sarah wondered? Despite all the evidence she found that said it was impossible, if she played the game, would she . . . find someone to talk to?
There was only one way to know for sure. She looked at the price tag. Well, there were some perks to spending years to become a highly educated engineer, one of which being that if she wanted to buy a moderately expensive toy on whim, she damn well could.
She stepped inside and stood self-consciously in front of the stack of colorful boxes. Sarah had a vague idea about the character of the ponies from the show; they were part of popular culture now and references cropped up in the most unexpected places. If she went through with this purchase, she’d have the chance to create a new online version of herself. Go pink? The idea of trying to be someone different, of going out on a limb and being fearlessly gregarious, was simultaneously thrilling and scary. On the other hand, she designed fighter jets for crying out loud! A bolt of rainbow colored lightning out of the clear blue sky could not be more appropriate. After all, what was on the outside and what was on the inside didn’t always have to match. If she was going to face the silent judgment of the store clerk, it would help to be carrying the pad that was twenty percent cooler.
