generalnews.media
importance 3/5 Exclusive

Benchmark Answers Leak into LLMs, Inflating Test Scores

New analysis highlights how answers from common AI benchmarks can unintentionally appear in training data, making models appear smarter than they are. This leakage undermines the validity of performance comparisons across large language models. The finding matters because it calls into question the reliability of current evaluation methods in AI development.

Technology

Sources (1)

technology

Related stories

Exclusive

Intel stock slump deepens as $3.5T chip selloff spreads

Intel shares fell further as a broad semiconductor selloff spread across the chip sector, erasing about $3.5 trillion in market value. The rout signals growing investor concern about the industry’s outlook and could weigh on technology markets more widely.

Exclusive

Anthropic warns it is near recursive self-improving AI capability

Anthropic warned that it is very close to recursive self-improvement (RSI), the ability of AI systems to build and improve themselves. This matters because RSI could sharply accelerate AI capabilities and risks, increasing urgency for safety and oversight.

Exclusive

BYD outsold entire US EV market last month with over 300,000 vehicles

BYD sold more electric vehicles in a single month than the entire US market sold in the quarter, surpassing 300,000 units. The milestone highlights China's growing dominance in EV production and price competition. It matters because it signals mounting pressure on US and global automakers as BYD expands.

Exclusive

Third person dies of bird flu as virus spreads to livestock

A third person has died from bird flu since the virus began jumping to livestock, according to Ars Technica. The death heightens concerns about human exposure to infected animals and the pandemic potential of avian influenza.

Exclusive

Android partner turned 1.3 million set-top boxes into a botnet

One of Google’s Android partners converted about 1.3 million set-top boxes into a botnet, according to The Verge. The large device network raises supply-chain and IoT security concerns because compromised consumer devices can be used for attacks such as distributed denial-of-service.

← Back to home