Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws very interesting (if somewhat controversial) they declare that models, after visiting data 1000 times during training, can memorize 2 bits/param (they arrive at this number via quantization with AutoGPTQ)
@jxmnop
-
Scaling Laws for Fact Memorization in Large Language Models
By
–
Scaling Laws for Fact Memorization of Large Language Models this paper is from last year (2024); it's a very well-executed study of the total number of facts a model can memorize they estimate 1T parameters are needed to perfectly memorize all relational pairs in wikipedia
-
Language Models Knowledge Storage Capacity and Memorization Scaling
By
–
Language Models as Knowledge Bases: On Entity Representations, Storage Capacity, and Paraphrased Queries in 2021 these folks measure the exact number of 'statements' models can memorize (a) capacity scales linearly with parameters
(b) more training samples -> less memorization -
Knowledge Storage Capacity in Language Model Parameters
By
–
How Much Knowledge Can You Pack Into the Parameters of a Language Model? enter, transformers- in 2020 these folks had a different idea: once "pretraining" is a thing, you can measure the amount of 'facts' that the model 'knows', without training it at all
-
RNN Capacity: 5 Bits Per Parameter Storage Empirical Analysis
By
–
Capacity and Trainability in Recurrent Neural Networks this very nice paper from 2016 (!) estimates certain types of RNNs can store 5 bits/param. they find this empirically and measure bits using the entropy of the distribution of class labels
-
Neural Network Memorization Capacity Under Parameter Quantization
By
–
Memorization Capacity of Deep Neural Networks under Parameter Quantization in 2019 these folks measured the capacity of RNNs/CNNs/MLPs, also using this class-based entropy idea they quantized the networks to half-precision and saw a capacity drop (but not nearly by half)
-
Single-Layer Perceptron Memory Capacity Foundational Result
By
–
A foundational result in machine learning is that a single-layer perceptron with N^2 parameters can store at least 2 bits of information per parameter (Cover, 1965; Gardner, 1988; Baldi & Venkates) specifically: an MLP can memorize 2 N^2 things with only N^2 parameters
-
Capacity Measurements for Language Models Research
By
–
in prep for our new research dropping on ArXiv tomorrow (i think), here is a thread about…. CAPACITY MEASUREMENTS FOR LANGUAGE MODELS
-
o6 AI Fires Employee with Severance Package 2027
By
–
> 2027: o6 has fired me from my job (with severance)
-
LLM Coding Capabilities: From GPT-2 to GPT-4o Progress
By
–
> 2020: gpt2 can't write code
> 2021: gpt3 can't reliably write python
> 2022: instructgpt can't write blocks of code without syntax errors
> 2023: chatgpt can't do leetcode
> 2024: gpt4 can't debug CUDA
> 2025: gpt4o can't implement entire PR these LLMs are not to be trusted…