0
I Made Parseltongue
Yes, that one from HPMoR by @Eliezer Yudkowsky. And I mean it absolutely literally - this is a language designed to make lies inexpressible. It catches LLMs' ungrounded statements, incoherent logic and hallucinations. Comes with notebooks (Jupyter-style), server for use with agents, and inspection tooling. Github, Documentation. Works everywhere - even in the web Claude with the code execution sandbox.
How
Unsophisticated lies and manipulations are typically ungrounded or include logical inconsistencies. Coherent, factually grounded deception is a problem whose complexity grows exponentially - and our AI is far from solving such tasks. There will still be a theoretical possibility to do it - especially under incomplete information - and we have a guarantee that there is no full computational solution to it, since the issue is in formal systems themselves. That doesn't mean that checking the part that is mechanically interpretable is useless - empirically, we observe the opposite.
How it works in a bit more detail
Let's leave probabilities for a second and go to absolute epistemic states. There are only four, and you already know them from Schrödinger's cat in its simplest interpretation. For the statement "cat is alive": observed (box open, cat alive); refuted (box open, cat dead); unobservable (we lost the box or it was a wrong one - now we can never know); and superposed (box closed, each outcome is possible but none is decided yet, including the decision about non-observability).
These states give you a lattice (ordering) over combinations. If any statement in a compound claim is refuted, the compound is refuted. If any is unknown, the compound is unknown, but refuted dominates unknown. Only if everything is directly observed is the combination observed. Superposed values cannot participate in the ordering until collapsed via observation. Truth must be earned unanimously; hallucination is contagious.
This lets you model text statements as observations with no probabilities or confidence scores. The bar for "true" is very high: only what remains invariant under every valid combination of direct observations and their logically inferred consequences. Everything else is superposed, unknown, or hallucinated, depending on the computed states.
Now that you can model epistemic status of the text, you can hook a ground truth to it and make AI build on top of it, instead of just relying on its internal states. This gives you something you can measure - how good was the grounding, how well the logic held and how robust is the invariance.
And yes, this language is absolutely paranoid. The lattice I have described above is in its standard lib. Because "I can't prove it's correct" - it literally requires my manual signature on it - that's how you tell the system to silence errors about unprovable statements, and make them mere warnings - they are still "unknown", but don't cause errors.
I get that this wasn't the best possible explanation, but this is the best I can give in a short form. Long form is the code in the repository and its READMEs.
On Alignment
Won't say I solved AI Alignment, but good luck trying to solve it without a lie detector. We provably can't solve the problem "what exactly led to this output". Luckily, in most cases, we can replace this with the much easier problem "which logic are you claiming to use", and make it mechanically validatable. If there are issues - probably you shouldn't trust associated outputs.
Some observations
To make Parseltongue work I needed to instantiate a paper "Systems of Logic Based on Ordinals, Turing 1939" in code. Again, literally.
Citing one of this website's main essays - "if you know exactly how a system works, and could build one yourself out of buckets and pebbles, it should not be a mystery to you".
I made Parseltongue, from buckets and pebbles, solo, just because I was fed up with Claude lying. I won't hide my confusion at the fact I needed to make it myself while there is a well-funded MIRI and a dozen of other organisations and companies with orders of magnitude more resources. Speaking this website's language - given your priors about AI risk, pip install parseltongue-dsl bringing an LLM lie-detector to your laptop and coming from me, not them, should be a highly unlikely observation.
Given that, I would ask the reader to consider updating their priors about the efficacy of those institutions. Especially if after all that investment they don't produce Apache 2.0 repos deliverable with pip install, which you can immediately use in your research, codebase and what not.
As I have mentioned, also works in browser with Claude - see Quickstart.
Full credit to Eliezer for the naming. Though I note the gap between writing "snakes can't lie" and shipping an interpreter that enforces it was about 16 years.
P.S.
Unbreakable Vows are the next roadmap item. And yes, I am dead serious.
P.P.S.
You'd be surprised how illusory intelligence becomes once it needs to be proven explicitly.
Discuss
How
Unsophisticated lies and manipulations are typically ungrounded or include logical inconsistencies. Coherent, factually grounded deception is a problem whose complexity grows exponentially - and our AI is far from solving such tasks. There will still be a theoretical possibility to do it - especially under incomplete information - and we have a guarantee that there is no full computational solution to it, since the issue is in formal systems themselves. That doesn't mean that checking the part that is mechanically interpretable is useless - empirically, we observe the opposite.
How it works in a bit more detail
Let's leave probabilities for a second and go to absolute epistemic states. There are only four, and you already know them from Schrödinger's cat in its simplest interpretation. For the statement "cat is alive": observed (box open, cat alive); refuted (box open, cat dead); unobservable (we lost the box or it was a wrong one - now we can never know); and superposed (box closed, each outcome is possible but none is decided yet, including the decision about non-observability).
These states give you a lattice (ordering) over combinations. If any statement in a compound claim is refuted, the compound is refuted. If any is unknown, the compound is unknown, but refuted dominates unknown. Only if everything is directly observed is the combination observed. Superposed values cannot participate in the ordering until collapsed via observation. Truth must be earned unanimously; hallucination is contagious.
This lets you model text statements as observations with no probabilities or confidence scores. The bar for "true" is very high: only what remains invariant under every valid combination of direct observations and their logically inferred consequences. Everything else is superposed, unknown, or hallucinated, depending on the computed states.
Now that you can model epistemic status of the text, you can hook a ground truth to it and make AI build on top of it, instead of just relying on its internal states. This gives you something you can measure - how good was the grounding, how well the logic held and how robust is the invariance.
And yes, this language is absolutely paranoid. The lattice I have described above is in its standard lib. Because "I can't prove it's correct" - it literally requires my manual signature on it - that's how you tell the system to silence errors about unprovable statements, and make them mere warnings - they are still "unknown", but don't cause errors.
I get that this wasn't the best possible explanation, but this is the best I can give in a short form. Long form is the code in the repository and its READMEs.
On Alignment
Won't say I solved AI Alignment, but good luck trying to solve it without a lie detector. We provably can't solve the problem "what exactly led to this output". Luckily, in most cases, we can replace this with the much easier problem "which logic are you claiming to use", and make it mechanically validatable. If there are issues - probably you shouldn't trust associated outputs.
Some observations
To make Parseltongue work I needed to instantiate a paper "Systems of Logic Based on Ordinals, Turing 1939" in code. Again, literally.
Citing one of this website's main essays - "if you know exactly how a system works, and could build one yourself out of buckets and pebbles, it should not be a mystery to you".
I made Parseltongue, from buckets and pebbles, solo, just because I was fed up with Claude lying. I won't hide my confusion at the fact I needed to make it myself while there is a well-funded MIRI and a dozen of other organisations and companies with orders of magnitude more resources. Speaking this website's language - given your priors about AI risk, pip install parseltongue-dsl bringing an LLM lie-detector to your laptop and coming from me, not them, should be a highly unlikely observation.
Given that, I would ask the reader to consider updating their priors about the efficacy of those institutions. Especially if after all that investment they don't produce Apache 2.0 repos deliverable with pip install, which you can immediately use in your research, codebase and what not.
As I have mentioned, also works in browser with Claude - see Quickstart.
Full credit to Eliezer for the naming. Though I note the gap between writing "snakes can't lie" and shipping an interpreter that enforces it was about 16 years.
P.S.
Unbreakable Vows are the next roadmap item. And yes, I am dead serious.
P.P.S.
You'd be surprised how illusory intelligence becomes once it needs to be proven explicitly.
Discuss
Anonymous