One of my favorite findings: Positional embeddings are just training wheels. They help convergence but hurt long-context generalization.
We found that if you simply delete them after pretraining and recalibrate for < 1% of the original budget, you unlock massive context windows.

Introducing DroPE: Extending the Context of Pretrained LLMs by Dropping Their Positional Embeddings
https://pub.sakana.ai/DroPE/
We are releasing a new method called DroPE to extend the context length of pretrained LLMs without the massive compute costs usually associated with

Sakana AI

Twitter

Bitcoin's price briefly touched $60,000 last week. Under a diminishing returns model, this is no mere noise. The market is touching the weakest link in the entire four-year cycle and logarithmic growth framework.

With the gains at the top of the Bitcoin cycle already significantly compressed, a further historically deep correction would completely negate the allure of its classic cycles.

This is not a prediction; it's a mathematical law.

Cycle Top Gains are Compressing

Historical Bitcoin Cycle Tops:

· 2013: ~$1,242

· 2017: ~$19,700

· 2021: ~...

$55,000 will be the lifeline for Bitcoin.

Institutional demand for tokenized finance in Asia is accelerating as the finchain avalanche collaboration moves to bring traditional assets on-chain at scale.
Summary
FinChain deepens strategic coope...

FinChain Avalanche partnership targets RWA tokenization and institutional liquidity in Asia

The crypto market experienced a sharp drop this afternoon, with Ethereum falling below the $2,000 mark at 3 PM Taiwan time, according to Hyperliquid [...]
The article, titled "Machi Big Brother Suffers Huge Losses and Didn't Sleep All Night? Ethereum Falls Below 2000 Magnesium, Panic Arrives, Opens Long BlockTempo and Loses 120,000 Magnesium," was originally published on BlockTempo, the most influential blockchain ABMedia media.

Machi Big Brother lost so much money he couldn't sleep all night? Ethereum fell below 2000, causing panic. He long positions in ETH and HYPE, then sold at a loss, losing all 120,000 in magnesium.