Category
Microsoft has released a streaming transcription model and two speech synthesis models in public preview. We check the primary sources on how interim text is corrected, differences from the standard version in features and pricing, how accuracy and latency were measured, and Japanese support.
Researchers disclosed Branch Target Reuse (BTR), an attack that exploits stale branch predictions left in the CPU after JIT code is reused. This article examines the Linux root-hash experiment, what was and wasn't demonstrated on Firefox and GraalVM, Intel's assessment, and the state of fixes.
Cloudflare's decision model Clef returns probabilities for predefined options so programs can branch automatically. We look at the accuracy gap between Clef and the lighter Clef-flash in published evaluations, plus what is open and when fine-tuning is available.
Volantis has raised $88 million to build the A-1, which links compute chips and memory with light. We examine its 10TB capacity and 20-trillion-parameter goal at different precisions, and what first deliveries in 2027 will need to prove.
AWS's new local decision engine allows or denies AI agent actions based on past successes and failures. Its implementation docs show how it handles concurrent requests, history after rule updates, and which blocking and audit duties stay with the runtime.
Homa, proposed for AI clusters, cuts wait times through receiver-granted sending and priority handling for short messages. We look at 2026 implementation changes, the conditions behind peer-reviewed benchmarks, and the TCP coexistence and gRPC limits that affect adoption.
Why do so many AI agent tool calls run on the CPU? Drawing on NVIDIA's CUDA guide and a research paper: GPUs can execute branches, and divergent branches run one after another. In SWE-Agent, doubling the batch from 64 to 128 raised latency 1.06–1.18x for GPU inference and 1.53–1.94x for CPU Bash execution.
A U.S. appeals court upheld the copyright ruling against ROSS Intelligence for training its legal search AI on Westlaw's editorial summaries. A comparison of the four fair use factors with the lower court's ruling shows why an output free of the original text didn't help.
NetApp's new Novus storage separates file-management metadata from data transfer to reduce GPU idle time. We look at the 100TB/s projection, bandwidth by GPU count, and how it differs from the existing AFX.
Morgan Stanley's 37.3% HBM demand estimate for NVIDIA is a forecast, not disclosed secured supply. We compare it with NVIDIA's $279 billion in multiyear commitments, Google's TPU demand, rising HBM prices and the shift toward 8-high stacks.
