MIT researchers report that GPT-6 Astra read, edited and wrote neural-network weights in eight experiments, with mixed results and limits.
The Latest
Weight-Space Programming Opens A New View Of Neural Network Models
The Decline Of Working Animals As Tractors Became Common
U.S. farm horses and mules fell from 26.7 million in 1918 to 3.1 million in 1960 as tractors and other machinery spread.
Why Job Seekers Should Read Binance’s $6.3 Billion Hack Post-Mortem
Binance says a fake job interview led to malware on an engineer’s work computer and helped attackers reach its wallet signing system.
How Well Did 42 AI Setups Answer 15 Legal Questions?
Lexassay’s test of 42 AI setups found weak results from a legal-specific model and recurring citation errors, with limits on what the scores show.
What It Costs To Retire Around The World: I Fact-Checked The Viral $2 Million Graphic
A viral retirement-cost graphic uses a 25-times spending formula but gives no sources, definitions or pension estimates. Here is what the figures leave out.
Google’s TV Box Botnet Takedown: What The Factory Malware Story Shows
Google says it disrupted the XCPE botnet, which infected more than 1.3 million TV boxes through factory firmware. The backdoor can persist after reboot.
How Europe’s AI Copyright Debate Could Shape Chatbot Use
An EU court adviser says chatbot answers may infringe press publishers’ rights in some cases. His opinion is nonbinding and leaves key questions to judges.
Kalshi And Its Co-founders Face Felony Charges For Violating Utah Gambling Laws
Utah has filed felony gambling charges against Kalshi and its co-founders as the prediction market fights state restrictions in federal court.
Z.ai’s Legal AI Benchmark Results: How About 20% Passed Still Ranks Near The Top
Z.ai says GLM-5.3-FlashX-Legal passed every criterion on 19.2% of Harvey benchmark tasks, placing seventh of 81 models on Vals AI’s archived leaderboard.
GPT-6 Astra’s Loop Repair Was About Model Behavior, Not IQ
A researcher-run weight edit cut Qwen3.5-0.8B’s MATH-500 score-limit rate, but the result is early, compute-intensive and not peer-reviewed.
