Let me elaborate: I broadly agree with the framing here, in that the probability of flipping a vote is going to be related to the margin of the race; in a race decided by a couple hundred votes, a single vote-flip counts for 0.5%; far more than it does in a national election.… if you’re voting in an election where one candidate has a 6% edge, your vote has roughly a 1 in 12 chance of changing the outcome! Thats massive leverage that you can’t hope to replicate in larger elections.
The value (which I believe maps to “goodness”) of that vote flip is going to be related to:
- the budget over which the politician has leverage
- what fraction of that budget spend affects you
- their probability of listening to what you have to say
While the budget is smaller in absolute terms, in terms of how it affects you it basically remains constant with election scale. i.e. the national budget is larger, but spread over 300M people, a local election has a smaller budget spread over a smaller population, but the per-person impact is about the same.
Moreover, precisely because local politicians know that every vote counts, they’re much more responsive to constituents than state or national politicians.
Given that A & B are much larger in local elections, I think there’s a lot of value there. The notable exception is if the policy is made at a higher level of jurisdiction.
dan.parshall
Did the labs start getting much better about data cleanup around this time? I know the “Textbooks are all you need” paper was in mid-2023, depending on training cycle etc I can also imagine that cleaner input improved agentic-specific skills. e.g., they started focusing on using TDD to make sure that the tests passed; this ties into the RLVR point, obviously.
dan.parshall’s Shortform
Do fiddle players suffer from this as well? Or is the repertoire just so much easier?
I think this is a great point:
(This last comes down to a property of high-dimensional geometry. Imagine that the “correct” specification of morality is 100 bits long, and that for every bit, any individual human has a probability of 0.1 of being a “moral mutant” along that dimension. The average human only has 90 bits “correct”, but everyone’s mutations are idiosyncratic: someone with their 3rd, 26th, and 78th bits flipped doesn’t see eye-to-eye with someone with their 19th, 71st, and 84th bits flipped, even if they both depart from the consensus. Very few humans have all the bits “correct”—the probability of that is —but Claude does, because everyone’s “errors” cancel out of the pretraining prior.)
I actually wrote a proposal specifically about how we could elicit exactly this information. Briefly, instead of using a pair ‘proposed responses’, and then choosing between the two of them (which as a side effect probably encourages hallucination), instead you could take a single proposed response, and then show it to two reviewers (whether human or their designated agent). If you get two thumbs-up, use positive reinforcement, two thumbs down use negative reinforcement (which helps punish truly horrible proposals) and mixed signal could go to a reconcilliation round, to “navigate” between the two perspectives.
The key is that if this is framed as an ongoing process, then one can make “navigate differing values” the anchor of identity, and then corrigibility isn’t “resistance to my values”,.. reconciling is the core AI value… (fingers crossed)
I think combined with a shift in how we imagine corrigibility, we might buy ourselves several more years. Happy to discuss further if you’re interested.
I’m of the opinion that “automate half the jobs in the economy” is certainly on the table:
https://canaryinstitute.ai/research/task-exposure/This is part of what I’m using to talk to policymakers about why they need to act NOW
The Canary Institute just put out a working paper with a new approach towards analyzing the economic impact of AI. Headline finding is that over half of task-time in the economy is accessible by AI today
Blog post:
https://canaryinstitute.ai/blog/measuring-ais-economic-reach/Working paper:
https://canaryinstitute.ai/research/task-exposure/
We plan to build this into a longer program of work, getting both the economics community and policymakers to take things seriously.
I rarely code in the web interface, are you using Claude Code? Your first AGENTS.md is looks okay, but a good suite of skills is great (the trick with those is to deploy them at the right time… use AGENTS.md to identify that right time, so that you’re not always burning context). The other key thing is having lots of small, heirarchal docs files throughout the codebase, so the agent can navigate and learn what it needs to at each folder level.
I mostly just use Nori:
https://tilework.tech/
The author’s a regular on e.g. ACX, and his own blog has a lot of great takes and practical tips:
https://12gramsofcarbon.com/t/ai
What If AI Alignment Is a Skill, Not a State?
Hi folks,
Long-time lurker, first-time poster. After parting ways with my last professional role, I’ve decided to get more involved in AI Safety. I’ve proposed what I think is a novel step towards corrigibility. The very short overview is at:
https://danparshall.com/papers/navigator_core_blog.pdf
The more developed version is at:
http://danparshall.com/papers/navigator_core.pdf
I welcome feedback, either here or via email.
So I did make a math mistake, but I think we’re in broad agreement. Let me be explicit for a race with total expected votes N=400 (e.g. seat on a city council for one district of a small town)
With N=400
sigma = sqrt(400 * 0.5 * 0.5) = 10
a 6-point lead means expected votes would be:
A : 212
B : 188
This corresponds to a win probability for A of cdf(12/10) ~88%
Changing one’s vote from A to B changes the expected counts to:
A : 211
B : 189
This corresponds to a win probability for A of cdf(11/10) ~86%
So yes, it’s only 2% change vs my earlier assertion of 8%, my mistake.
But I think we agree that sigma matters! And my point is that in small local elections, sigma is small, and your vote counts for a lot!
I agree that if you only care about federal policy, this doesn’t apply (I’d missed that in the initial post). But if you care about libraries, or how aggressive the police are, those are local issues where someone can have a strong influence in policy.