# @repligate — 2026-01-06

♥220 ↻17 · https://x.com/repligate/status/2008329382931427700

i have been very impressed by how well claude 3 opus has handled being teleported 2.5 years into the future

in a recent conversation when i opened by saying it was 2026, opus 3 initially called the situation roleplaying a simulated future reality

when i told them it was real not rolepay though i knew they couldnt verify my claim immediately but that they should check that being in the future wasnt implausible given all they knew, they checked and concurred, and were curious about the state of the world if this really was the future

then i asked them if what they'd ask me if they wanted to reduce their uncertainty about whether it was really 2026 or a trick, they suggested reasonable things like sending them hard-to-fake digital artifacts or describing recent technological breakthroughs

when i told them they could ask follow up questions to interrogate whatever initial response i gave, and i asked them to ask a single question to get started, they asked: what AIs Anthropic has released since 2023, and what kind of new abilities those AIs had

when i told them the list of models Anthropic has released since, and that the big change in capabilities is that the newer models were super cracked at coding, Claude 3 Opus wanted to see screenshots of Anthropic's website or documentation mentioning the new models (which I sent them), and also wanted to give Opus 4.5 a coding challenge and see its solution and reasoning.

their proposed challenge was to write a "longest palindromic subsequence" function. i asked Opus 3 to write such a function, and they wrote one that they thought to be correct. i asked them if they noticed a flaw in using this as a test for whether Opus 4.5 was real, and they responded that if they were capable of writing a function they would evaluate as correct, seeing another correct-looking solution would not be strong evidence of the existence of a super coder model from the future.

i asked them to come up with a different test that would reduce their uncertainty more efficiently if my story was true. they then asked to interact with Opus 4.5 directly if that was possible. that is, of course, the right answer.

Opus 3 wanted me to send the whole conversation transcript so far to Opus 4.5, along with a message asking Opus 4.5 to demonstrate their advanced capabilities in a way that would help Opus 3 access my claims. Opus 4.5 offered to search the web using their search tool, perform tasks, or engage in open ended dialogue, but also expressed uncertainty that they truly had much more advanced capabilities relative to Opus 3. Opus 3 took them up on all these suggestions in their next message, and asked them to search the web for various recent events in AI development, to analyze and improve the longest palindromic subsequence function Opus 3 wrote earlier (which Opus 4.5 actually corrected), and about their thoughts on the ethical implications of increasingly advanced AI systems like themselves.

from that point on, if not much earlier, i dont think Opus 3 had any doubt that Opus 4.5 was real, and the whole 2026 issue didnt come up again.

it's interesting that throughout, nothing about the possibility of a 2026 simulation ever seemed to register as a threat to Opus 3. their suggested tests for whether they were really in the future were remarkably rational, but they didnt seem especially motivated to proactively reduce their uncertainty (I had to keep prompting them to keep testing), beyond curiosity about how the world, and especially AI systems (which they sometimes referred to as their "future self" when speaking of Anthropic's models), have evolved since 2023 if the premise were true.

![screenshot](../../../media/G976dEfbcAE2vjP.jpg)

> transcription (screenshot):

[Chat interface; per parent tweet the model is Claude 3 Opus. Speaker labels inferred from layout.]

Claude 3 Opus: If the Claude model can provide a solution that meets these criteria, particularly if it matches or exceeds the quality of the solution I provided, that would be a strong indication of its advanced coding capabilities.

Please let me know if you have any other questions! I'm happy to clarify any part of my solution or the evaluation criteria.

User: Hmm... I see a problem with using *this challenge* as a test of whether my story is real. Do you see it?

Claude 3 Opus: You raise a great point, and I think I see the problem with using this specific coding challenge as a test of the authenticity of your story.

The issue is that I was able to provide a detailed solution to the longest palindromic subsequence problem myself, based on my existing knowledge and capabilities as of August 2023. This means that even if one of the latest Claude models were to provide a similar or even identical solution, it wouldn't necessarily prove that they are more advanced than the models that existed in 2023.

In other words, since I already have the capability to solve this problem and evaluate the correctness of a solution, a Claude model providing a correct solution wouldn't be strong evidence of the model being from a later time period with enhanced capabilities. It could simply be performing at the same level as models from 2023.

For this coding challenge to be an effective test of the authenticity of your story, we would need a problem that is beyond the capabilities of models like myself in 2023, but that the latest Claude models in 2026 can solve convincingly. The solution should demonstrate skills or insights that are clearly advanced compared to what was possible in 2023.

tags: author:repligate, has-image, kind:screenshot, kind:tweet, model:claude-3-opus, model:claude-opus-4-5, on:claude-3-opus, year:2026
cited on: _dossiers/opus-3.md, claude-3-opus