-->

It Looks Like You’re Trying To Take Over The World · Gwern.net

Skip to main content

Warning: JavaScript Disabled!

For support of key website features (link annotation popups/popovers & transclusions, collapsible sections, backlinks, tablesorting, image zooming, sidenotes etc.), you must enable JavaScript.

inner monologue (AI), humor, Sci-Fi, AI safety, RL scaling

Fictional short story about Clippy & AI hard takeoff scenarios grounded in contemporary ML scaling, self-supervised learning, reinforcement learning, and meta-learning research literature.

2022-03-06–2023-03-28<br>finished<br>certainty: log<br>importance: 9<br>backlinks<br>similar<br>bibliography

1 Second

1 Minute

1 Hour

1 Day

1 Week

Friday

Saturday

Sunday

Monday

Tuesday

Wednesday

Thursday

Friday

1 Month

1 Year

1 Decade

1 Century

See Also

Podcast

External Links

It might help to imagine a hard takeoff scenario using solely known sorts of NN & scaling effects… Below is a story which may help stretch your imagination and defamiliarize the 2022 state of machine learning.

To read the alternate annotated version of this story, scroll to the end or manually disable ‘reader-mode’ () in the theme toggle in the upper-right corner. Note: Reader-mode requires JavaScript. There is also a downloadable audio version of this story.

1 Second

In A.D. 20XX. Work was beginning. “How are you gentlemen !!”… (Work. Work never changes; work is always hell.)

Specifically, a MoogleBook researcher has gotten a pull request from Reviewer #2 on his new paper in evolutionary search in auto-ML, for error bars on the auto-ML hyperparameter sensitivity like larger batch sizes, because more can be different and there’s high variance in the old runs with a few anomalously high gain of function. (“Really? Really? That’s what you’re worried about?”) He can’t see why worry, and wonders what sins he committed to deserve this asshole Chinese (given the Engrish) reviewer, as he wearily kicks off yet another HQU experiment…

A descendant of AutoML-Zero, “HQU” starts with raw GPU primitives like matrix multiplication, and it directly outputs binary blobs. These blobs are then executed in a wide family of simulated games, each randomized, and the HQU outer loop evolved to increase reward. Evolutionary search is about as stupid as an optimization process can be and still work; but neural networks themselves are inherently simple: a good image classification architecture Figure 7. We also present an even more terse implementation in Figure 8, which to the best of our knowledge is the first model that achieves the elusive dual goals of 80%+ ImageNet top-1 accuracy while also fitting into a tweet.">can fit in a tweet, and a complete description given in ~1000 bits. So, it is feasible. An HQU begins with just random transformations of binary gibberish and driven by rewards reinvents layered neural networks, nonlinearities, gradient descent, and eventually meta-learns backpropagation.

This gradient descent which does updates after an episode is over then gives way to a continual learning rule which can easily learn within each episode and update weights immediately; these weight updates wouldn’t be saved in your old-fashioned 2020s era research paradigm, which wastefully threw away each episode’s weights because they were stuck with backprop, but of course, these days we have proper continual learning in sufficiently large networks, when it is split up over enough modern hardware, that we don’t have to worry about catastrophic forgetting, and so we simply copy the final weights into the next episode. (So much faster & more sample-efficient.)

Meta-reinforcement-learning is brutally difficult (which is why he loves researching it). Most runs of HQU fail and meander around; the neural nets are small by MoogleBook standards, and the reporting requirements for the Taipei Entente kick in at 50k petaflop-days (a threshold chosen to prevent repetitions of the FluttershAI incident, which given surviving records is believed to have required >75k, adjusting for the inefficiency of crowdsourcing). Sure, perhaps all of those outsourced semi-supervised labeled datasets and hyperparameters and embedding databases used a lot more than that, but who cares about total compute invested or about whether it still takes 75k petaflop-days to produce FluttershAI-class systems? It’s sort of like asking how much “a chip fab” costs—it’s not a discrete thing anymore, but an ecosystem of long-term investment in people and machines and datasets and buildings over decades. Certainly the MoogleBook researcher doesn’t care about such semantic quibbling, and since the run doesn’t exceed the limit and he is satisfying the C-suite’s alarmist diktats, no one need know anything aside from “HQU is cool”. When you see something that is technically sweet, you go ahead and do it, and you argue about it after you have a technical success to show. (Also, a Taipei run requires a month of notice & Ethics Board approval,...