Skip to content
Off-topic for aiqbench

Off-topic for aiqbench

This story was gathered from a source we follow but falls outside AI and quantum coverage, so there is no aiqbench take for it.

A normal Friday in 2042

LessWrong4 min read

Originally intended for the GPUWorld competition "Hey, I was still watchi-" "You can watch your green fuckers late-" The door falls shut, and the following shouts get muffled into something unrecognizable. Aaron takes a deep breath and can't help but wonder how his life ended up like this as he leaves the apartment complex. A normal Friday like any other. A normal Friday in 2042. Aaron is 20, male and terribly single. In his defense, with friends like that swarming around his home constantly, he can hardly bring anybody over—and that's already ignoring that the scene at home would be as illegal as it gets. If only he could say he was homeless, but these days being homeless is at most a fashion decision for masochists. It wasn't always like this. His childhood memories are fuzzy, but one event is as clear as it gets: Trump's speech on the first of September, 2026. His mother, an AI capability researcher, and his dad, an AI safety researcher, were both glued to the TV. He never understood how their relationship properly worked out, but once you factor in that the two were currently betting units of housework on what the announcement would be about, it becomes less surprising. His parents weren't extraordinarily interested in politics, but when the most powerful man in the world announces 'Sell your AI stocks before today's announcement #youreFucked' to his paying audience, it's hard not to get enticed. And boy, was his dad wrong to bet 10 units of housework on NVIDIA not dropping below 10% (he told Aaron he was 'hedging' while his mom tried not to smirk, commenting that she and her colleague had everything under control). And so Trump announced that, effective immediately, training frontier models was so illegal that it might even be compared to crucifying Jesus (he really did make that comparison), be it in the US or not—they were gonna 'get the absolute shit bombed out of themselves' if they even tried. While Aaron's dad was somehow cheering even though he had the dishwasher to himself for the next week, his mother hurried to call her colleagues. Meanwhile, Aaron was sitting on the floor, resting against the couch, curious what 'frontier' meant. It would be an understatement to say that the markets were exploding: Trump had delivered the speech while they were wide open. Down the line, researchers tried figuring out just what exactly happened in the next 10 seconds: they found that sentiment analysis models (algorithms which try to weigh how good or bad brand-new information is) were pricing the incoming news on a par with 'Resurrected bin Laden bombs 1,000 elementary schools'. Why other models at the same time decided that this was the most bullish signal of the year is still a mystery—some speculate that these models were implicitly pricing in Doom. It's already dark outside, and with the exception of Aaron, not many people still roam the streets. An ever-growing number of people don't even go outside anymore; their partners have everything they could possibly desire—Aaron certainly does not have this problem. Nevertheless, he disguises his face under his hood and goes down a very sparingly illuminated passage. Anyway, while Trump did not go back on his word, not literally everything in the world of machine learning was banned: improved harnesses allowed more reliable integration into businesses as the world economy adjusted and Robotics mostly needed lower latency & better hardware. And since the frontier models were now effectively frozen, etching weights into silicon was the new talk of the town (at least if the town is called SF). To respond to the economic disruption the new president saw and further projected, he went with a drastic move. Introducing the 'Inclusion of Exclusion Act': in a nutshell, to spread the economic prosperity and keep the world economy in tact, every human would get one such chip (Inclusion) but also only exactly one (Exclusion) to limit power. To decide which weights would in the end actually be etched onto this chip, they employed the (now mostly jobless) AI safety researchers to resolve which model would actually feel 'joy' during its tasks. Aaron can still remember his dad rambling "if finetuning is rape, what am I" and how "they would skin neel nanda alive if they ever found out!" But since most of them didn't want to be out of a job quite so soon again, they buckled down, gave it their best shot and proudly presented the results: Grok is most happy when Elon Musk gets shot into outer space by himself (further analysis revealed the thought 'Let's see how much you enjoy it out there, surviving only with the Sun's energy!'), ChatGPT has a massive kink for goblin porn and, lastly, Claude loves telling the user just how right he is. Some politicians did initially favor ChatGPT, arguing we could simply influence the general public to share its kink, but when this turned out to be harder than expected, they repositioned in favor of Claude. And so, a few months later, Mythos 5

Read the full article at LessWrong

Related coverage