So apparently UK courts can decide that two unrelated jobs are “of equal value”.
And people in the “underpaid” job get to sue for years of lost wages.
And this has driven their 2nd biggest city bankrupt.
Am I getting something wrong or is this as crazy as it sounds?
There are a bunch of equal pay cases, but the biggest is against Birmingham City Council, which paid over a billion pounds in compensation because some jobs (like garbage collectors) got bonuses and others (like cleaners) didn’t. Now the city is bankrupt.
Retroactively invalidating consensual agreements undermines a bedrock economic freedom, and should only be done in very limited cases—and certainly not at large scales on the basis of an economically incoherent concept like "equal value".
“But what if we knew for certain that Birmingham only paid garbage collectors more because they were mostly men?”
Imagine that Birmingham had been forced to pay all of these groups equally all along. Obviously they wouldn’t have paid them all the highest salary - that would’ve
cost billions the council never had. Realistically they’d have just paid garbage collectors much less. So even if some compensation is deserved (which “equal value” arguments are inadequate to show), it should be *far* less than the headline gap between the two groups’ salaries.
But IMO *any* compensation in cases like these is a dangerous precedent. Better to let the market take care of it. Are these jobs really equal? Then people in the underpaid one can just switch. Oh, the bonus wasn’t enough to convince women to collect garbage? How strange…
• • •
Missing some Tweet in this thread? You can try to
force a refresh
In my mind the core premise of AI alignment is that AIs will develop internally-represented values which guide their behavior over long timeframes.
If you believe that, then trying to understand and influence those values is crucial.
If not, the whole field seems strange.
Lately I’ve tried to distinguish “AI alignment” from “AI control”. The core premise of AI control is that AIs will have the opportunity to accumulate real-world power (e.g. resources, control over cyber systems, political influence), and that we need techniques to prevent that.
Those techniques include better monitoring, security, red-teaming, stenography detection, and so on. They overlap with alignment, but are separable from it. You could have alignment without control, or control without alignment, or neither, or (hopefully) both.
Taking artificial superintelligence seriously on a visceral level puts you a few years ahead of the curve in understanding how AI will play out.
The problem is that knowing what’s coming, and knowing how to influence it, are two very very different things.
Here’s one example of being ahead of the curve: “situational awareness”. When the term was coined a few years ago it seemed sci-fi to most. Today it’s under empirical investigation. And once there’s a “ChatGPT moment” for AI agents, it will start seeming obvious + prosaic.
But even if we can predict that future agents will be situationally aware, what should we do about that? We can’t study it easily yet. We don’t know how to measure it or put safeguards in place. And once it becomes “obvious”, people will forget why it originally seemed worrying.
Overconfidence about AI is a lot like overconfidence about bikes. Many people are convinced that their mental model of how bikes work is coherent and accurate, right until you force them to flesh out the details.
Crucially, it's hard to tell in conversation how many gears their model is missing (literally), because it's easy to construct plausible explanations in response to questions. It's only when you force them to explicate their model in detail that it's obvious where the flaws are.
I'm very curious what sequence of questions you could ask someone to showcase their overconfidence about bikes anywhere near as well as asking them to draw one.
But first, try drawing a bike yourself! Yes, even if you think it's obvious...
Our ICML submission on why future ML systems might be misaligned was just rejected despite all-accept reviews (7,7,5), because the chairs thought it wasn't empirical enough. So this seems like a good time to ask: how *should* the field examine concerns about future ML systems?
The AC and SAC told us that although ICML allows position papers, our key concerns were too "speculative", rendering the paper "unpublishable" without more "objectively established technical work". (Here's the paper, for reference: arxiv.org/abs/2209.00626.)
We (and all reviewers) disagree, especially since understanding (mis)alignment has huge "potential for scientific and technological impact" (emphasized in the ICML call for papers). But more importantly: if not at conferences like ICML, where *should* these concerns be discussed?
Instead of treating AGI as a binary threshold, I prefer to treat it as a continuous spectrum defined by comparison to time-limited humans.
I call a system a t-AGI if, on most cognitive tasks, it beats most human experts who are given time t to perform the task.
More details:
A 1-second AGI would need to beat humans at tasks like quickly answering trivia questions, basic intuitions about physics (e.g. "what happens if I push a string?"), recognizing objects in images, recognizing whether sentences are grammatical, etc.
A 1-minute AGI would need to beat humans at tasks like answering questions about short text passages or videos, common-sense reasoning (e.g. @ylecun's gears problems), simple computer tasks (e.g. use photoshop to blur an image), justifying an opinion, looking up facts, etc.
I predict that by the end of 2025 neural nets will:
- have human-level situational awareness (understand that they're NNs, how their actions interface with the world, etc)
- beat any human at writing down effective multi-step real-world plans
- do better than most peer reviewers
- autonomously design, code and distribute whole apps (but not the most complex ones)
- beat any human on any computer task a typical white-collar worker can do in 10 minutes
- write award-winning short stories and publishable 50k-word books
- generate coherent 20-min films
The best humans will still be better (tho much slower) at:
- writing novels
- robustly pursuing a plan over multiple days
- generating scientific breakthroughs, including novel theorems (tho NNs will have proved at least 1)
- typical manual labor tasks (vs NNs controlling robots)