THE DECODER

AI benchmarks systematically ignore how humans disagree, Google study finds

A Google study finds that the standard three to five human raters per test example often aren't enough for reliable AI ...
the-decoder

Alibaba's Qwen team makes AI models think deeper with new algorithm

Alibaba's Qwen team has developed a new training algorithm for reasoning models that assigns different weights to individual tokens based on how much each step influences the subsequent chain of ...
the-decoder

The New York Times drops freelancer whose AI tool copied from an existing book review

The New York Times cut ties with freelance writer Alex Preston after it turned out an AI tool he'd used had copied from an existing book review. Preston was writing a review of Jean-Baptiste Andrea's ...
the-decoder

AI offensive cyber capabilities are doubling every six months, safety researchers find

AI safety research firm Lyptus Research has published a new study on the offensive cybersecurity capabilities of AI models. The study is based on the METR time-horizon method and involved testing with ...
the-decoder

OpenAI reshuffles leadership as health issues force key executives to step back

Several leadership changes are underway at OpenAI. Fidji Simo, CEO of the newly created "AGI Deployment" division, is taking sick leave for several weeks to deal with an autoimmune disease affecting ...
the-decoder

Anthropic says Claude Code's usage drain comes down to peak-hour caps and ballooning contexts

Anthropic has looked into complaints from users who were hitting their Claude Code usage limits much faster than expected. According to Anthropic's Lydia Hallie, tighter limits during peak hours and ...
the-decoder

Microsoft's MAI-Transcribe-1 runs 2.5x faster than its predecessor at $0.36 per audio hour

Microsoft has introduced MAI-Transcribe-1, a speech-to-text model supporting 25 languages that achieves the lowest word error rate of any model tested on the FLEURS ...
the-decoder

Anthropic leak reveals new model "Claude Mythos" with "dramatically higher scores on tests" than any previous model

The leaked blog posts have allegedly surfaced online; the information matches what Fortune shared in a follow-up article. There are two versions of the same blog post that only differ in the model's ...
the-decoder

OpenAI and Anthropic before the IPO: Different balance sheets make comparison difficult

Anthropic and OpenAI are both growing fast, but they report revenue very differently, The Information reports. OpenAI's annualized revenue is around $25 billion; Anthropic's is $19 billion. Both ...
the-decoder

Mistral's new Small 4 model punches above its weight with 128 expert modules

Mistral AI has released Mistral Small 4, combining fast text responses, logical reasoning, and image processing in one model. It has 119 billion parameters, but only 6 billion are active per query - ...
the-decoder

GTC 2026: With Groq 3 LPX, Nvidia adds dedicated inference hardware to its platform for the first time

Nvidia fleshed out the Vera Rubin platform at GTC 2026: The POD comprises 40 racks with 1,152 Rubin GPUs and 60 exaflops of compute. The central NVL72 rack is expected to deliver 4x training ...