What is true about gpt2-chatbot?
Mini
93
Ṁ54k
Jan 1
98.6%
It will get matched/surpassed by a LLM from another company/organization by EOY 2024
67%
Mixture of Experts
50%
It runs on H200 GPUs
50%
It generates audio in a discrete manner (next token prediction + vqvae) instead of continuous (next patch prediction, diffusion)
50%
It generate images in a similar fashion to diffusion models
47%
The "also-good-chat-model" version is a smaller variant than the "good-chat-model."
39%
> 15% on SWEBench (few shot or zero shot)
34%
Mixture of Depth/Dynamic allocation of computation to each token depending on difficulty
34%
It uses Hierarchy Alignment (https://arxiv.org/abs/2404.13208)
25%
It used RAG/search/function-calls during initial debut on LMSys
21%
A prompt that enables it to solve easy Sudoku puzzles be found by EOY 2024 (same rule as https://manifold.markets/Mira_/will-a-prompt-that-enables-gpt4-to) (resolve NA if not released)
17%
>1T parameters
16%
>50% of the pretrain corpus was synthesized (e.g., 10T out of the total 20T were generated or augmented by AI/heuristics)
13%
It will be open-weight by EOY 2025
10%
<10B parameters
6%
It is close to GPT-2 (e.g., the same architecture but with different instruction tuning)
4%
It is what Ilya saw (that triggered him for the coup and unusual behaviors)
4%
It will be mentioned by Biden or Trump during the 2024 election.
4%
It uses a different architecture from the current paradigm (e.g., MoD, non-transformer, not next-token prediction)
Resolved
YES
Its knowledge cutoff is in or before Jan 2024

I reserve the rights to NA any added options.

Clarification: gpt2-chatbot include the three variants, i.e. gpt2-chatbot, "i-am-a-good-gpt2-chatbot", and "i-am-also-a-good-gpt2-chatbot"

related market:

What is “gpt2-chatbot”?

What made gpt2-chatbot smarter?

Get Ṁ1,000 play money
Sort by:

@Sss19971997 now that o1-preview has been released, and is a reasonable match to descriptions of what Q* was supposed to be, can "It is Q star" resolve NO?

Creator is a deleted account, @mods other than me, I'd like to unlock my mana from this market, any chance of a resolution for this option (and perhaps some others)?

It will get matched/surpassed by a LLM from another company/organization by EOY 2024

@Sss19971997 I might resolve this after the Lmsys score of Claude 3.5 Sonnet comes out.

Sonnet 3.5 is statistically significantly worse than 4o. I am not resolving now.

Mixture of Experts

Is there actually any evidence that it is a mixture of experts? Why is this so high?

@benshindel

It explicitly seems like it’s NOT a MoE model

@benshindel The word "a single model" is used to show the difference to previous multimodal solutions, where for vision, people assembled a CLIP with an LLM through projection, and for audio, people used three models that go speech-to-text (Whisper), text reasoning (GPT4), and TTS at the output end.

@benshindel You can bet according to your opinion for sure.

Resolution of GPT-5 and Q* answers @Sss19971997?

@chrisjbillington It might make sense to resolve GPT-5 to NO now, but do we really have any idea what Q* is and how it is definitely not GPT4o?

@Sss19971997 If you're resolving stuff, "It is close to GPT-2 (e.g., the same architecture but with different instruction tuning)" should probably also resolve NO.

@MugaSofer I have learned that deferring resolution is always a good idea.

Its knowledge cutoff is in or before Jan 2024
bought Ṁ150 Its knowledge cutoff... YES

The system prompt says it's October 2023. (I just ran a few tests to confirm, and it's guesses about events after that point were complete hallucinations.)

@MugaSofer I am debating if I should resolve according to GPT4o or wait till we find out what the other two variants (gpt2-chatbot, i-am-a-good-gpt2-chatbot) are?

@MugaSofer For example, what if any of the two variants have a later knowledge cutoff? (though very unlikely)

@Sss19971997 Obviously it would benefit me to have it resolved immediately.

How would you resolve it if some fit the criterion and some didn't? (For any question.) You did already resolve several options based on things that we only know for sure apply to one of the models.

@MugaSofer Any variant fitting count.

@Sss19971997 In that case you're safe to resolve YES; at least the main variant's knowledge cutoff "is in or before Jan 2024".

bought Ṁ200 Answer #e4c5c037c150 YES

Resolves YES.

Probably should have resolved YES. Sorry guys.

@Sss19971997 I mean Sam Altman specifically said it wasn't. It'll be strange if other markets about it resolve YES.

@chrisjbillington Sam Altman lied/disguised about a lot of things before release.

@chrisjbillington @Mira resolved GPT4.5 market according to GPT4o, which makes a lot of sense

@Sss19971997 Mira's market's definition is very broad and would have included new models in the GPT-4 family even if they were downgrades in capabilities. Mira is betting NO in similar markets.

@chrisjbillington well, i still think it is debatable. anyway.