Meta paper: byte-level models start worse than token models but may end up better

AWS says Continuum finds and validates code vulnerabilities, checks them against the customer's AWS environment, and opens pull requests with proposed fixes. AWS published a headline benchmark claiming an 89% fix rate, though the report notes a catch.

A Reddit user released version 1.0 beta of GANG WIZARD, a game they say consumed over 10 billion tokens to build. The poster claims the GPU usage was so heavy their local data center sent a letter.

A Reddit user trained a dynamic continual-learning model called mini-AGI entirely from scratch for a month on an 8GB VRAM GPU. Benchmarks compare a "frozen" evaluation against a "best effort" mode where the model specializes on tasks as it moves through the eval dataset.

Speaking on Bloomberg This Weekend, The Atlantic CEO Nicholas Thompson said AI's broader economic impact continues to lag behind rapidly advancing capabilities. He called the tools powerful for businesses but a significant risk to media, amid a growing public backlash to the AI boom.

For every FLOP spent training a model, roughly six go to serving it each year, and getting inference right has produced over $220 billion in market cap in the last decade. The session walks through the full lifecycle of an inference request.

A sponsored contributed post argues the gap between top engineering teams' shipping velocity and the rest of the industry is now an exponent rather than a multiple, as agents accelerate deploys. It says the dividing line is no longer tooling but observability, framing the sensor as the starting point.