Retro News

Context size >= 200k

Would qualify for AIME given screenshots of the AMC 10 questions, from the AMC 10 released soonest after GPT-5 release

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

Able to solve easy Sudoku puzzles with prompting

to be N/A'd : Seed answer to create a market

Image Modality(Output)

Able to precisely count the amount of triangles there is in a image of non-overlapping radomly placed shapes(for images with less than 50 shapes)

New Neural Network Architecture

Exact query of custom knowledge datasets(e.g. Google RETRO, where users can provide their own datasets)

Audio Modality(Input or Output)

Video modality(Input or Output)

Online learning(agent finetuning, PEFTs, etc.)

Capable of online learning

Context size >= 500k

Context size >= 1000k

YES

No model picker

Personalized to you, the user

Trained on synthetic data

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

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

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

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

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

Operates autonomously with instructions from you

capable of online learning

New Neural Network Architecture

Seed option to create the question

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

Audio Modality

Acts autonomously

Has a catchier name

NO

YES

YES

Other

sigma

based

cool

spice

whale

heat

loose

caught

cot

retro

TRANCHE

gas

numb

left

sunk

cought

warm

tepid

luke

damp

jank

sweet

sick

沉着的 chénzhuóde

cloe

YES

YES

NO

YES

NO