where humans still out-perform our best algorithms. If we can isolate a set of basic skills from the HCP data, then offer them up as a kind of library that developers can hook into with a minimum of fuss, that will benefit them enormously.’
‘But most aspects of the Proxies will remain the province of ordinary software?’ Caplan pressed her. ‘Biographical details, long-term context, strategic considerations-’
‘Of course,’ Nasim confirmed. ‘Even if we could construct a “whole brain” model that included the neural dynamics needed to support long-term memory, and even if we could figure out how to wire a Proxy’s notional back-story into that model… that wouldn’t just be wildly ambitious, it would be inefficient and highly inconvenient. Developers need to have a simple database of facts about their non-player characters that they can manipulate as easily as the geographical database for the game world. Sure, they want certain responses to the human players to be more natural, but not at the expense of an ability to follow predetermined scripts, or adopt explicitly programmed strategies. Trying to implement too much as part of the neural model would just make those things more difficult.’
‘Hmm.’ Caplan glanced down at some notes displayed on the table-top. ‘You mention reading players’ facial expressions?’
Nasim gestured at her goggles; they’d be invisible to Caplan, but they both knew they were there. ‘In less than a year, these will be replaced by contact lenses. It won’t be long before we routinely have an unobstructed view of players’ faces.’
Caplan said, ‘Sure, but don’t you think human face-reading skills have already been surpassed by micro- expression analysis?’
‘Perhaps,’ Nasim conceded. ‘But it might still be useful to developers to have something less forensic, for those times when they want a Proxy who seems attentive or sympathetic, rather than one who can beat you at poker or deliver a knock-out cross-examination in a murder trial. Anyway, that’s only one skill out of dozens.’
‘Many of the other skills are verbal, though, aren’t they?’ Caplan pointed out. ‘I’m still not quite clear what the deal is with language.’
‘Ah.’ Nasim knew this was her weakest point, but she confronted it head-on. ‘Obviously the HCP data itself will only encode knowledge of English; that’s the nature of the population they scanned. In the worst-case scenario, that could mean that we’d have to restrict our verbal modules to English, which we’d then license to VR providers with a large English-language customer base. Even so, that’s a huge market, so we could expect a decent return in terms of revenue even if we couldn’t use the same modules generally within Zendegi itself.’
‘And the less worse-case?’
‘We could probably exploit English-based modules in non-English scenarios to some degree, just by feeding their input and output through existing translation software – just as we do for players who don’t share a common language. But I’d hope we could do one better than that and train at least some of the modules to be bilingual.’
‘Expose them to huge data-sets and tweak them to give the right responses?’ Caplan suggested.
‘Yes.’
‘While hoping you don’t obliterate the very skills that make these networks more useful than the dumb translator software that gets created the same way?’
Nasim replied, ‘Well, yes. We can only try it and see how it pans out. What else should we do?’
Caplan said, ‘Have you ever heard of side-loading?’
‘No,’ Nasim confessed.
Caplan smiled slightly. ‘You didn’t research me very thoroughly, did you? Just checked that I still had a pulse and a bank balance.’
Nasim felt herself flush slightly, but took some comfort in the fact that her icon wouldn’t betray her response. She said, ‘I’m sorry if I didn’t dig deeply enough. I think I got distracted when I discovered that you owned an island in Cyber-Jahan.’
‘Doesn’t everyone? They’re so cheap.’ Caplan dismissed the matter with a wave of his hand. ‘Anyway… ten years ago I acquired a Swiss company called Eikonometrics.’ He slid a document across the table. Nasim glanced down; it was a report on the company’s research over the last fiscal year. A soft blue glow at the edges flagged it as a file Caplan had put into their shared dataspace, rather than a physical stack of paper.
‘They started out in subliminal image classification,’ Caplan explained. ‘They had a plan to harvest the brain-power of internet users by flashing up flowers and puppy dogs and road crash scenes just long enough to get a distinctive EEG signature. It was a very silly idea; some aspects of artificial vision might still be inferior to the equivalent human skills, but software caught up with the kind of results you could get on the cheap like that a long time ago.’
‘So why did you buy them?’
‘Their business plan sucked,’ Caplan said, ‘but they’d gained some genuine insight into neural information processing, and they’d started branching out into EBLD – Evidence-Based Lie Detection. Polygraphs are about as reliable as witch-dunking, so it’s not a huge boast to say that twenty-first century brain-scanning techniques have the potential to out-perform them. You stick a witness in a PET scanner or MRI machine and see what parts of their brain light up when they discuss their testimony or view certain images. There was a fad for that in the teens, and Eikonometrics got some grants and did some interesting research. But then it became clear that in any practical context there were too many problems interpreting the responses – and most jurisdictions started ruling it out on privacy as well as technical grounds. That’s when Eikonometrics shares hit rock-bottom, and I would have been crazy not to snap them up.’
‘Okay.’ Nasim was starting to get an inkling of where this was heading. ‘And side-loading…?’
‘Side-loading,’ Caplan replied, ‘is the process of training a neural network to mimic a particular organic brain, based on a rich set of non-intrusive scans of the brain in action. It’s midway between the extremes: in classic uploading, you look at the brain’s anatomy in microscopic detail and try to reproduce everything from that, whereas in classic neural-network training, what’s available to you is just stimulus and response: sensory input and visible behaviour, with the brain as a black box.
‘In side-loading, you get to peer inside the box, even if you don’t get to take it apart. You don’t have the resolution you’d get by peeling the brain with an ATLUM, but you have the advantage that you can expose the living brain to all kinds of stimuli – words, images, sounds, tastes, smells – and see how they bounce around inside the skull. And it doesn’t really matter how little external behaviour is evoked if you can watch the pattern of internal changes rippling out from every stone you toss in.’
Nasim said, ‘That all sounds fine in principle. But how far has it actually taken you?’
Caplan gestured at the Eikonometrics report. ‘Six months ago, we took a rat that had been trained to solve a particular maze. By observing its brain’s responses to a barrage of random sensory cues, we managed to modify an initially unrelated virtual rat to run the same maze.’
‘Observing its brain how? Do you mean micro-electrodes snuggling up to ten thousand neurons?’
Caplan shook his head. ‘Not at all. Purely non-invasive methods: multi-mode MRI and surface electrodes.’
Nasim was impressed now; this was more than she’d expected. ‘And with humans?’
‘With humans,’ Caplan said, ‘the problem is that we don’t really have a good generic virtual brain to use as a starting point. When the HCP published their results we started trying to construct one, but it looks as if you’re doing a much better job at that than we are so far.’
Nasim mulled over this frank assessment, sting included. The algorithm she’d used on the finch brains was public knowledge, but it wasn’t so cut and dried that someone coming in cold could apply it to the HCP scans and expect to crank out a good result. So she had a head-start on the path towards something Caplan wanted, badly – but it would be a mistake to overplay her hand and imagine that this made her indispensable.
‘So you think Eikonometrics can find a way to side-load languages into my Proxy modules?’ she asked.
‘I think it’s worth trying,’ Caplan said. ‘I think it’s your best chance. And there might be some other useful skills you could endow them with, the same way.’ He smiled. ‘Since I got your email, I’ve been giving that some thought. Suppose you could side-load motor skills from a few celebrity athletes: sign up the right soccer players and the Middle East is yours again. Add the right cricket players and Cyber-Jahan is history.’
Nasim was speechless. Motor skills were probably the simplest case imaginable – certainly less challenging than side-loading a second language. And Caplan was right: millions of people would vote with their credit cards to
