AI
Most automated video Q&A pipelines describe audio and visuals separately, severing the link between a sound and its source. A new data engine rebuilds that link with structured scripts and evidence chains.
AI
Researchers instrumented a 26B diffusion language model to record exactly when it commits each token, and found its decoding order is partly an illusion created by how you measure it.
AI
A model that seemed to compute 100 functions with 50 neurons looked like evidence that networks pack extra computation into shared dimensions. A new analysis traces the effect to an accidental mixing of inputs instead.
AI
A new theoretical result shows that an AI generating verified mathematics must emit an endless stream of correct-but-worthless statements to cover the valuable ones — and proves the stream is a necessity, not a bug.
AI
A new method isolates the sub-network a language model uses to play one specific character, then prunes the rest — cutting the role-playing quality loss versus standard pruning by up to 93.8% on the strongest baseline.
AI
Researchers found a small set of attention heads that track which image region a model is describing, and showed that nudging those heads forces the model to describe a different region entirely — no retraining required.
AI
An AI chip company has to describe what its chips do for the SEC. That plain sentence is the cleanest definition you'll find.
AI
A company changes a lot in three annual reports. The sentence describing what it sells barely moved.
AI
Two XBRL line items from NVIDIA's filings tell the AI-boom story without a single adjective.
AI
AMD's 10-K names a whole product family for one job: AI in the data center. Here's what that hardware actually does.
AI
Alphabet keeps re-estimating how long its servers last. That dry sentence is a window into the whole infrastructure story.
AI
Everyone watches hyperscaler capex. Fewer watch the depreciation assumption that decides how it hits the income statement.
AI
Every company says it's an AI company now. Meta's 10-K is more careful than its keynotes — and the gap is instructive.
AI
The phrase shows up in every AI-chip story. NVIDIA defines it implicitly in its 10-K — by the workloads it lists.
AI
Microsoft has described its cloud as 'infused with AI' for years. The phrasing's evolution is a map of the strategy.
AI
Meta wrote, years ago, that investing in AI would expand its computing needs. That sentence predicted the whole buildout.
AI
The AI-accelerator story is told as a one-company story. AMD's annual report is a reminder it's a market, not a monopoly.
AI
The bull case treats AI-chip demand as a law of nature. NVIDIA's own filing is careful to do no such thing.
AI
The least glamorous step in building a model is often the one that decides whether it works. A June 2026 NVIDIA grant is all about it.
AI
A June 2026 NVIDIA grant turns a voice track into matching video using neural networks. It's a small window into where generative media IP is heading.
AI
The term is everywhere. The patents show something real underneath — but also that the gap between demo and dependable is exactly where the hype lives.
AI
A May 2026 Anthropic grant describes an agent that reads a screen and operates the interface. It's the IP behind 'computer use.'
AI
The AI-economics debate quietly moved from training to inference. The patents and filings explain why that shift is the whole ballgame.
AI
Everyone debates whether AI spending pays off. The filings won't settle it — but they tell you exactly what's being counted, and what isn't.
