“you absolutely have to view LLM benchmarks from a position of default-distrust” @ouguoc.mastodon.online.ap.brid.gy describes how easily answers to benchmark problems can leak into the training set. ‪https://seinmastudios.com/posts/llm-benchmarks-are-not-trustworthy/

why is the link not a link? let's try that again: seinmastudios.com/posts/llm-be...

Your piece is great. As usual. Yeah, I think there's no getting around that we can't measure human welfare without a set of values to define what that means. There's not some "scientific" trick or shortcut that lets us authoritatively, universally, say that this is better than that. 1/

Technocrats pretended that GDP was that for a while, and it kind of worked because, broad brush, the correlation was pretty strong between GDP per capita and qualitative, intuitive, perceptions of prosperity and satisfaction. 2/

It was a great fit for "neoliberal" economics, whose ideological trick was largely to obscure methodological problems economists had long discussed and pretend something like economics 101 provided a scientific basis for policy to which intelligent people must defer. 3/

Besides the helpful casual empirics, it had this story, for market economies, GDP is the quantity of the highest value (because market optimized) basket of goods and services produced by the economy, and so that quantity, a simple number, should be a pretty good proxy for wealth. 4/

But real-life markets are imperfect optimizers, and even theoretical markets are arguably local rather than global optimizers that might get stuck in bad path dependencies (like automobile-dependent low density living, I'd argue). 5/

This trick of replacing an infinitely dimensioned "what" with a single number "how much" just doesn't work. 6/

So we are left with judgment calls to make, which invariably involve both evaluating tradeoffs on dimensions we'd mostly all agree are valuable (say shipbuilding capacity vs health care services), but also require imposing contested values. 7/

Some people think the costs of low-density single-family infrastructure are totally worth it, resources spent in support of human flourishing. I think an auto-centric built environment is both costly and inferior in welfare terms to achievable alternatives. 8/

Each claim requires assertions about other people's preferences in a normative sense, what they should want (since it's indeterminate what they do want while the hypothetical choice, two very different social equilibria, are not remotely before them). 9/

There's no "scientifically" right or wrong answer other ppl must defer to. Only judgements—not pulled from nowhere, informed by marshaling evidence—but judgments nonetheless, of which we have to persuade our fellow citizens, rather than truths we can discover and propound as incontrovertible. /fin