Explain Haas vs FDT tweet

Figure out what this post means

sadly as a fan of Haas’s verificationism I have to sympathize with the part of FDT where they’re like “look, we’ve basically solved this, you just have to plug something in this hole where you’re supposed to fit a semantics for counterfactuals”

It’s basically a self-deprecating joke that says:

“I like Haas’s version of verificationism, which itself kind of cheats by saying ‘once we have a good semantics for counterfactuals, everything works’ – so I can’t really complain when FDT does exactly the same thing.”

Let me unpack the moving parts.

---

1. Who/what is being referenced?

Haas’s verificationism

• Verificationism = the idea that a statement is meaningful only if it can, in principle, be verified by experience (or is a tautology). It was central to logical positivism.(Wikipedia)
• Gordian Haas wrote Minimal Verificationism: On the Limits of Knowledge, where he tries to rescue a weak (“minimal”) form of verificationism using modern tools like belief-revision theory and a formal semantics for counterfactuals.(Google Books)
• The book blurb literally says that modern theories of justification and belief revision, along with a formal semantics for counterfactuals, can help overcome the traditional problems of verificationism.(Google Books)

So Haas’s project is: “Verificationism can be made to work… once we plug in the right semantics for counterfactuals and other tricky stuff.”

---

FDT (Functional Decision Theory)

• FDT is a decision theory due to Yudkowsky & Soares (MIRI). Roughly, it says:

  Treat your decision as the output of a fixed mathematical function and choose the output that would lead to the best outcome.(arXiv)

• To do this, FDT leans heavily on logical counterfactuals like: “If the decision function I implement outputted A instead of B, what would happen?”(alignmentforum.org)

• But making sense of such counterfactuals is hard, especially when they’re counterpossibles (their antecedent is logically impossible, e.g. ‘If this deterministic algorithm had output something else...’). Philosophers have pointed out that FDT explicitly admits it doesn’t yet have a full account of its counterfactuals and calls this an “open problem”.(alignmentforum.org)

So FDT’s situation is: “We basically know how rational decision-making should work… once we plug in a good semantics for our special kind of counterfactuals.”

---

“Semantics for counterfactuals”

• “Counterfactuals” are “if… had…” claims, like “If I had left earlier, I would have caught the train.”
• A semantics for counterfactuals is a precise theory of what such sentences mean and when they’re true or false (Lewis/Stalnaker possible-worlds semantics, causal-model semantics, etc.).(Stanford Encyclopedia of Philosophy)
• This turns out to be notoriously difficult, especially for counterfactuals with impossible antecedents (“counterpossibles”) or involving logic/computation, which is exactly what FDT needs.(PhilPapers)

---

2. So what is the post actually saying?

Putting it together, the post is doing three things at once:

1. Confessing a bias

   “I’m a fan of Haas’s verificationism…”
   i.e. “I like a theory that itself relies on an unfinished piece about counterfactual semantics.”

2. Recognizing the same move in FDT The line about FDT saying “look, we’ve basically solved this, you just have to plug something in this hole where you’re supposed to fit a semantics for counterfactuals” is a slightly snarky paraphrase of how FDT looks to critics:

   • FDT claims to have the right structure of decision theory.
   • But the core technical/mathematical difficulty—how exactly to interpret the counterfactuals it uses—is postponed as “future work”.(alignmentforum.org)

3. Expressing reluctant sympathy The “sadly” and “I have to sympathize” are the joke:

   • From the outside, this can sound like:

     “We’ve solved it in principle; just assume someone eventually solves the really hard missing piece.”

   • But because Haas’s verificationism does almost the same thing—“verificationism can be rehabilitated, once we have a nice semantics for counterfactuals etc.”(Google Books)—the poster can’t really complain about FDT doing that without also undercutting their own favourite view.

---

3. Super-short paraphrase

“Both Haas’s verificationism and Functional Decision Theory kind of ‘hand-wave’ the same crucial technical issue: we don’t yet have the right formal semantics for counterfactuals (especially weird logical ones), but once we did, our theory would be great. As someone who likes Haas’s project, I’m in no position to throw stones at FDT for making the same move — I’m stuck sympathizing with them.”