Az OpenAI matematikai áttörése 372 Lean-ellenőrzött bizonyítást szállít, a Kakeya-sejtéstől a Riemann-határokig, vitát keltenve a valódi matematikai hatásukról.
Gemini 4 Argon reshapes Google’s AI plans, moving free users to Flash, adding Deep Think for Pro and Ultra, and unveiling a 2‑million‑token context window.
Project Suncatcher shows how space-based AI compute can harness solar power and terabit optical links to train massive models beyond Earth’s energy limits.
AISI’s study shows a 29.2% success rate for GPT-6 Astra supply-chain attacks in simulated CI/CD pipelines, exposing serious risks even with scope limits.
A white‑hat team leveraged the Claude Opus 5 hack to exploit a libheif CVE‑2026‑32882 flaw, breaching OpenAI and prompting a debate on AI model governance.
Trump dismisses AI safety warnings while leaders like Musk and Altman urge an AI development slowdown, underscoring the clash between competition and regulation.
ChatGPT Images 2.5 now generates sharper visuals up to 40% faster and introduces a Sketch tool for instant design tweaks, plus new templates and API options.
AI agent consensus shows how swarms of models like GPT‑4 Turbo, Claude 3 Opus, and Llama 3 align—and why unchecked agreement can embed systemic errors.
Abliteration.ai GLM-5.3 offers an unguarded, 7.5‑billion‑parameter model for developers, but its open‑weight design raises safety and regulatory questions.
The GPT-6 Astra launch reveals OpenAI’s pricing, enterprise access options, safety framework, and benchmark results that showcase its AGI‑level capabilities.
A Muse Spark 1.3 egy 128k tokénos kontextusablakkal és akár 30%-kal gyorsabb kódkiegészítéssel emeli fel a poprzát, és továbbra is a legköltséghatékonyabb AI-modell a fejlesztők számára.
Former Google CEO Eric Schmidt was booed at a University of Arizona commencement while discussing AI — a moment that crystallized a widening rift between industry optimism and public frustration. Workers, parents,
Google launched Gemini Omni Flash at I/O 2026, promising a model that can create anything from any input — but today's release covers only video, while audio editing remains in safety testing.
OpenAI and Dell Technologies announced a formal partnership to bring Codex into on-premises enterprise environments via the Dell AI Factory and Dell AI Data Platform. The path is real — but it
Alibaba's Qwen3.6-27B is freely available for self-hosting, but its eight-GPU reference setup and a deteriorating open-source supply chain mean the barrier to production is higher than the free-access headline suggests. The model's
Alibaba's Qwen3.6-35B-A3B activates only ~3B of its 35 billion parameters at once, cutting inference costs without shrinking the architecture. Early assessments highlight strong agentic coding performance and a context window that stretches