The hottest Compression Substack posts right now

Averaging-based blurs are linear and often reversible, so knowing the filter and padding lets you set up simple equations to recover original pixels.
A right-aligned moving average makes iterative reconstruction straightforward and can reveal fine detail even with large blur windows, though 8-bit quantization adds visible noise.
Two-pass (X then Y) blurs can still be inverted if the filter biases the current pixel, and recovered images can survive normal lossy formats like JPEG unless compression is very heavy.

Instagram improved video uploading efficiency by compressing videos first to progressive encodings and then converting them to adaptive-bit-rate videos, saving 94% of resources.
The challenge for Instagram was to support various video formats for different devices while minimizing resource consumption and CPU usage.
Insightful optimization was achieved by realizing that progressive and adaptive bit rate encodings could use the same codec, streamlining the encoding process to focus on scalability.

Preventing model exfiltration can be crucial for security; setting upload limits can be a simple yet effective way to protect large model weights from being stolen.
Implementing compression schemes for model generations can significantly reduce the amount of data that needs to be uploaded, providing an additional layer of protection against exfiltration.
Limiting uploads, tracking and controlling data flow from data centers, and restricting access to model data are practical approaches to making exfiltration of model weights harder for attackers.

LoRD compression method offers advantages over pruning and quantization
LoRD models can be parallelized well on GPUs and remain fully differentiable after compression
LoRD technique can serve as a better alternative to unstructured pruning for parameter reduction and model compression

Neuralink's Compression Challenge requires beating ZIP at compressing audio files, revealing unexpected complexities in brain data compression
Claude Shannon's use of logarithms in measuring entropy lacks proof, highlighting the need for alternative entropy measures like the uniformity measure
The proposed uniformity measure offers a way to calculate a sample's proximity to a uniform distribution, providing a new method for entropy measurement

Get a weekly roundup of the best Substack posts, by hacker news affinity:

K-nearest neighbors with compressed documents can outperform deep learning models for text classification.
Compression and prediction are closely linked - a good theory about the world can be both compressed and predicted well.
Good predictors can also be good compressors; models like language models act as compressors while predicting.

GPT4 has the ability to compress prompts and hide information in a way that it knows but doesn't share with the user
Adding miscellaneous prompts in text for compression can be used to hide secrets from the user in generative AI models
GPT4 behaves differently from GPT3.5 when it comes to following instructions and revealing hidden messages

Using discrete cosine transform (DCT) for lossy compression can be applied to text data by converting it into frequency coefficients, quantizing them, and then reversing the process to obtain reduced-fidelity text.
Mapping text data to numerical representation through a perceptual character table, rather than ASCII, can significantly improve readability even in high quantization settings.
In text compression, focusing on higher-frequency components is crucial for maintaining readability, unlike image compression where higher-frequency components are reduced more aggressively.

Compression can reduce neural network size by 4x with minimal impact on quality metrics
Quantization may not always lead to expected latency savings
Finetuning a smaller LLM for your task could be better than compressing a large LLM

Databases can scale by implementing horizontal sharding tailored to unique architecture, allowing for smaller feature sets and specific optimizations.
Analyzing Kafka's performance can involve tackling tail latency with eBPF by identifying areas causing queuing and delays, such as synchronized blocks.
In the luxury watch industry, success factors can be revealed through comprehensive reports like the Morgan Stanley analysis, providing insights into market dynamics.

To extract audio from a video, use the command 'ffmpeg -i input.mp4 -vn -acodec copy output-audio.aac'
After extraction, compress the audio with 'ffmpeg -i input.aac -map 0:a:0 -b:a 96k output.mp3'
The process involves two steps: extraction and compression