What do you think will be GPT-5's "hook"?

Plus

Ṁ4329

2031

ALL

89%

No model picker

74%

Personalized to you, the user

68%

Trained on synthetic data

56%

Very large context window, or another way to give the model long-term memory

51%

Faster speed(smaller model, architecture improvements, mixture of experts keeping most parameters unused during inference, etc.)

47%

Designed to inference cheaper than GPT-4(price must be lower)

41%

Online learning of specialized agents(finetuning, PEFTs, etc.)

38%

Exact-query of training data or knowledge databases(similar to Google RETRO)

34%

Operates autonomously with instructions from you

32%

capable of online learning

28%

New Neural Network Architecture

25%

Audio Modality

24%

Seed option to create the question

22%

(Near) real-time update of training data (data delay is within 6 months and is kept updated)

21%

Acts autonomously

16%

Has a catchier name

1.2%

Video Modality

Add your predictions. I will not trade heavily (i may make small predictions with 1 mana for myself) because my subjective judgement might come into play.

Resolves YES to those that are more or less correct (resolves according to the spirit) or to 50% (if technically true but not in spirit) or to NO (if not true).

All predictions will be evaluated upon a release, if there ever is one.

Do not submit spam or irrelevant options, those will be N/A'd

Inspired by this tweet (by @Mira )

If no GPT-5 is released this decade, the market will be N/A'd. You can trade on whether that will happen here:

Get Ṁ1,000 play money

16 Comments

Sort by:

bought Ṁ10 New Neural Network A... NO

What happens when an account is deleted? Can a mod resolve the market?

No model picker

bought Ṁ50 No model picker YES

@ChinmayTheMathGuy meaning 1 unified model that routes between submodels

@ChinmayTheMathGuy https://openai.com/index/gpt-5-system-card/

Trained on synthetic data

bought Ṁ75 Trained on synthetic... YES

@ChinmayTheMathGuy

Both Mark Chen (interview with Matthew Berman) and Sam Altman (interview with Cleo Abram)

mentioned this. It's also in their model card.

Faster speed(smaller model, architecture improvements, mixture of experts keeping most parameters unused during inference, etc.)

bought Ṁ20 Faster speed(smaller... YES

This was showcased in the mrwhosetheboss video about GPT-5 (over 2M views in <1 day, by far most watched video about GPT-5) so I think it should resolve yes

bought Ṁ10 Online learning of s... YES

What is google RETRO? Can’t find any mentions of it online

@mariopasquato @Mira @MiraBot

@mariopasquato https://arxiv.org/abs/2112.04426

WTF why do Google scientists not understand how big O notation works? I've seen people misuse big O notation like this in every day use but in a paper this is too much

reposted

reposting because the launch may be near

Has a catchier name

bought Ṁ70 Has a catchier name NO

This market only applies to AIs called GPT-5, so this is technically impossible

Personalized to you, the user

Imagine ChatGPT but Google gave it all of your info

New Neural Network Architecture

To clarify, this needs to be something significantly different. Something like Flash-attention that's just a faster version of transformers won't count. It doesn't necessarily need to not contain transformers, but it needs to be a bigger change than an improvement to a single aspect of the architecture or a better implementation

Resolves to whether the prediction was correct or whether it was a "hook" i.e. heavily advertised by OpenAI as its distinguishing feature, or notable in being different from other leading models?

@Mira To a hook.

@Mira

Related questions

Related questions