What do you think will be GPT-5's "hook"?
38
379
1.9K
2031
72%
Very large context window, or another way to give the model long-term memory
66%
Personalized to you, the user
64%
Audio Modality
56%
capable of online learning
45%
Online learning of specialized agents(finetuning, PEFTs, etc.)
40%
Operates autonomously with instructions from you
38%
Exact-query of training data or knowledge databases(similar to Google RETRO)
35%
Video Modality
34%
(Near) real-time update of training data (data delay is within 6 months and is kept updated)
31%
New Neural Network Architecture
25%
Faster speed(smaller model, architecture improvements, mixture of experts keeping most parameters unused during inference, etc.)
24%
Seed option to create the question
21%
Acts autonomously
19%
Designed to inference cheaper than GPT-4(price must be lower)
19%
Has a catchier name

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 Ṁ200 play money
Sort by:
bought Ṁ10 Online learning of s... YES

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

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.