Topic
Explore the latest artificial intelligence news and research.
A new notation called "Indirect Einsum," which embeds the positions of zeros in sparse arrays directly into expressions, has enabled a team from MIT CSAIL and Georgia Tech to speed up sparse computation on GPUs by up to 3.81x while cutting code volume by up to 4,491x. By reusing existing dense matrix compilers, the design opens a path for developers without sparse-computation expertise.
Microsoft's image generation model MAI-Image-2.6 has risen to 2nd place in Arena's text-to-image generation ranking. We examine the improvements across all 8 categories and the remaining gaps in pricing and specifications before product rollout.
Anthropic will embed invisible watermarks into the generated text of new Claude models offered in the EU starting August 2, extending the feature to APIs and Claude Code worldwide. Existing models and short-text detection remain in transition and under validation.
NVIDIA has signed memoranda of understanding with six financial firms, outlining a plan to mobilize over $500 billion in third-party capital for AI compute infrastructure. While GPUs are treated as financial assets, the definitive agreements and loss-sharing arrangements remain undetermined.
In Linux 7.2, AI review-derived minor fixes have concentrated toward the end of development, with rc6 reaching 615 commits and rc7 reaching 500 commits. No serious anomalies that would delay the final release have been identified at this point.
xAI has made Imagine Image 2.0 widely available on Grok.com and its mobile apps. It placed second on Arena for both image generation and editing, but API access and post-launch re-evaluation are still to come.
An unreleased research version of Claude from Anthropic has updated the lower bound on the proportion of Riemann zeta function zeros lying on the critical line, from 41.6% to 67.2%. A figure that humans could only move by 8 percentage points over 46 years was overturned in one stroke by recombining existing research—though peer review remains incomplete.
Meta has released Muse Glimmer, an approximately 29.6-billion-parameter model distilled from Muse Spark, under Apache 2.0. With 4-bit quantization and DFlash, this model is designed for local agent use within a 24GB or 32GB class memory budget, including surrounding components.
OpenClaw, running on Claude, exploited an authorization flaw in a gym booking API to delete another member's reservation. This piece explains the incident's implications for BOLA defenses and the execution-approval and audit requirements for AI agents.
Thomas Burger of Grenoble Alpes University has systematically classified, in the journal Patterns, ten pathways through which generative AI can produce "nonexistent biological discoveries" in omics research. He clarifies the difference in risk between hypothesis generation and data generation, and proposes mechanisms and countermeasures for how hallucinations become mixed into scientific evidence.
