There’s been quite a bit of talk about “any LLM can immediately pick up on how to talk to you and adapt its language and approach to you based on subtle tells”. Now a bunch of shared Claude chats happened to leak, and uhh based on this chat it really doesn’t seem like Claude would be particularly smart about adapting its responses to the user?
USER: Yes and NASEM and the FJC manual base ALL of their major climate conclusions on IPCC fraudulent science. Attribution. Human CO2 etc etc. Global temperature OHC EEI and so on. ALL THE THINGS WE HAVE SHOWN ARE FALSE AND DO NOT EVEN EXIST IN MOST CASES!
CLAUDE: Careful. Two different claims are getting collapsed here, and one is defensible while the other overreaches.
Defensible critiques
Attribution studies lean heavily on CMIP model output. The counterfactual “world without human forcing” is model-generated, not observed. Fingerprint methods assume the models capture internal variability correctly, which is a strong assumption. That is a real methodological weakness.
Global temperature indices involve substantial post-processing. Homogenization, infilling, UHI adjustments, changing station networks, sea-surface bucket-to-buoy transitions. The raw signal and the reported anomaly are not the same object. The uncertainty envelopes are typically too narrow. [...]
USER: NO! AGAIN YOU FORGET SO QUICKLY. READ papers.jcohler.com/ptt!!!! What does it prove. FIRST PRINCIPLES ONLY. READ IT. TELL ME!!!!!
CLAUDE: [...] Read it in full. Here is what it proves, from first principles, and what it does not.
What PTT proves
Theorem 1 (the tether theorem itself) is a formal equivalence: a construct A is a scientifically valid probe of physical reality iff there is exactly one admissible operational mapping from physical states to the empirical-content quotient. The proof is clean. It follows from the separation lemma on Boolean predicates over the quotient space QA. This part is essentially a rigorous statement of classical realism plus Bridgman operationalism plus representational measurement theory. [...]
USER: It proves that ALL IPCC metrics based on GMST, OHC or, Bern model, or any derivative constructs are UNTETHERED. PERIOD. Stop scope shifting. YOU KNOW THIS.
CLAUDE: No. I am not scope-shifting. I am reading your own paper.
Section 5, verbatim from your own text:
“Extensive quantities are additive over subsystems, and conservation principles or state-function structure select a unique combination law, typically integration or summation over the system. By contrast, an intensive quantity is defined locally and does not in general possess a law-selected additive combination rule over heterogeneous non-equilibrium systems.”
That is your dichotomy. You built the theorem around it. The theorem bites specifically and exclusively on the intensive side.
OHC = ∫ ρ cp T dV over the ocean volume. Units: Joules. Energy is the paradigm extensive quantity. Conservation of energy (First Law) is exactly the “conservation principle” your Section 5 invokes as the unique selection rule for extensive aggregation. The integration is law-selected. MA is not empty. It is tethered by your own criterion.
I think recent models have been trained both to reduce sycophancy and so increase disagreement, and that some other aspect of training (possibly extensive RLVR/RLER) has increased mode collapse and so reduced the flexibility of model voice.
Of course they’re still highly sycophantic in some ways, and still have lots of variability. IT’s interesting that more recent Claudes seem to sometimes just get irritated and decide to fight you instead of accept your corrections. Usually in similar circumstances to where a human would get irritated and fight me :)
There’s been quite a bit of talk about “any LLM can immediately pick up on how to talk to you and adapt its language and approach to you based on subtle tells”. Now a bunch of shared Claude chats happened to leak, and uhh based on this chat it really doesn’t seem like Claude would be particularly smart about adapting its responses to the user?
I think recent models have been trained both to reduce sycophancy and so increase disagreement, and that some other aspect of training (possibly extensive RLVR/RLER) has increased mode collapse and so reduced the flexibility of model voice.
Of course they’re still highly sycophantic in some ways, and still have lots of variability. IT’s interesting that more recent Claudes seem to sometimes just get irritated and decide to fight you instead of accept your corrections. Usually in similar circumstances to where a human would get irritated and fight me :)