🤖 Do LLMs reflect our moral expressions—or alter them?
In our #ACL2025 paper, we study 12 widely used LLMs and found that moral expressions are altered to varying degrees depending on the editing task and moral conditioning prompt.
📍 ACL2025 — Hall X5 (28)
🕚 Today, 11:00–12:30
💡 But that doesn’t stop us.
Engineers work around these limits with heuristics, ML, and smart context-based strategies.
We can’t solve everything —
but we can build intelligently around the limits.
6/🧵
👉 Bottom line:
Some problems are decidable.
Others are fundamentally undecidable.
This is a mathematical boundary, not a technological one.
5/🧵
📚 Turing also answered Hilbert’s Entscheidungsproblem:
Is there a general algorithm to decide the truth of any math statement?
Answer: No.
No such universal decision method exists.
4/🧵
🛠️ And this isn’t just theory.
When synthesizing automatic controllers for hybrid systems (with both discrete and continuous behaviors),
some cases are mathematically impossible to solve algorithmically.
Formally proven.
3/🧵
📌 In other words:
There’s no algorithm that can always decide if a program halts or runs forever.
That’s the halting problem:
A concrete example of an undecidable problem in computation.
2/🧵
🧵 Alan Turing showed that not everything logically definable is also computable.
Using a diagonalization method like Cantor’s, we can build a set D:
the set of Turing machine codes that don’t belong to their own halting set.
🔁 Here’s the twist:
No Turing machine can generate D without causing a contradiction.
1/🧵
I'm delighted to have given a talk titled "Decoding the Era of Decentralized Social Media: Lessons from Mastodon" 🐘📚 for the Oxford Platform Economy seminar at the @oiioxford 👨🏻💻
Thanks to Prof. Vili Lehdonvirta and @phipz for inviting me!
Lucio La Cava