Yapay Zeka Modelleri

Google has launched Lyria 3 Pro, an advanced AI music generation model capable of creating original compositions up to three minutes long with structured controls for intros, choruses, and arrangements. The tool prioritizes ethical training and avoids artist imitation.

OpenAI has shut down its Sora video-generation model amid growing concerns over its potential misuse in disaster misinformation campaigns. According to government sources, the decision aligns with national preparedness protocols.

Access to large language models (LLMs) is being democratized through innovative open tools and policy advocacy, according to Nomic AI CEO Brandon Duderstadt. In a recent Gradient Dissent episode, he detailed how fine-tuning and prompt engineering are lowering barriers to AI adoption.

EleutherAI, a nonprofit collective led by scientist Stella Biderman, trains and releases open large language models to promote transparency and public access in AI development. Unlike corporate labs, EleutherAI prioritizes open science and community-driven innovation.

Google has launched Lyria 3 Pro, an AI music generator capable of producing three-minute songs with full structural complexity. Unlike rivals facing copyright lawsuits, Google claims its training data is legally secured.

MolmoWeb, a fully open-source AI agent developed by AI2, navigates websites using only screenshots, outperforming larger proprietary models. With 30,000 human task trajectories and under 8 billion parameters, it redefines accessibility in web automation.

Google has launched Lyria 3 Pro, its most advanced AI music generation model yet, enabling users to create three-minute-long, highly customizable tracks. The upgrade integrates across Gemini and enterprise platforms, signaling a major expansion in AI-driven creative tools.

A groundbreaking benchmark reveals Qwen3.5-397B achieves 20 tokens per second on a single RTX 5090 GPU, setting a new standard for local LLM inference. The test, conducted on an AMD EPYC system, highlights the potential of consumer-grade hardware for enterprise-scale AI.

Google Lyria 3 Pro now generates AI-made songs up to three minutes long, six times longer than its previous limit. The upgrade integrates seamlessly across Google’s ecosystem, marking a major leap in generative music AI.

OpenAI's latest public repository reveals Claude, Anthropic's AI model, as the third top contributor—a surprising sign of cross-company AI collaboration. This development signals a new era of interoperability in generative AI.

MiniMax M2.7 is the world’s first open-weight LLM demonstrating true self-evolution, autonomously analyzing failures and optimizing its own performance. This breakthrough redefines AI agent development and marks a pivotal moment in open AI research.

The Local LLM by Ente introduces a privacy-first AI model that processes data on-device, challenging mainstream cloud-based AI. Experts analyze its implications for user autonomy and corporate surveillance.

OpenAI has abruptly discontinued Sora, its flagship AI video generator, just three months after announcing a partnership with Disney. The move raises questions about AI development timelines and commercial viability.

OpenAI has shut down Sora, its groundbreaking AI video generator, after just 25 months amid ethical concerns and partnership losses. The move signals a shift in AI video development toward regulated, localized innovation.

OpenAI has abruptly shut down its AI video generator Sora, just six months after its public launch, amid mounting concerns over deepfakes and industry pushback. The move has stunned major stakeholders including Disney and raised new questions about the future of generative AI.

Paged Attention in large language models revolutionizes GPU memory usage by eliminating wasteful fixed allocations, dramatically increasing concurrency and reducing costs. This breakthrough addresses a critical bottleneck in AI scaling.

DeepSeek has unveiled a groundbreaking engram-based architecture that dramatically reduces training costs and improves reasoning scalability. This innovation, distinct from traditional MoE models, signals a paradigm shift in large language model design.

GPT-5.4 Mini touts itself as an affordable, lightweight solution for multi-agent development—but internal data reveals a 3x price increase over GPT-5.3. Investigative analysis examines performance, pricing, and community backlash.

Reka AI has unveiled its new Reka Edge model, designed for physical-world applications, during an AMA with its research team. The model prioritizes efficiency, edge deployment, and real-time reasoning — marking a pivotal shift in local AI development.

Google DeepMind’s Gemini 3.1 Flash-Lite can generate complete, functional websites in near real time, marking a seismic shift in web development. Experts debate its practicality versus creative limitations.

Anthropic's Claude AI now mimics human input to control Mac computers, marking a major leap in personal agent technology. This capability blurs the line between assistant and operator, raising new questions about autonomy and security.

Gemini Embedding 2 now allows direct video-to-vector indexing without transcription, enabling sub-second natural language video search. Developers are already building cost-effective tools for security footage and surveillance systems.

NVIDIA's Nemotron 3 agents are redefining AI systems with advanced reasoning, multimodal RAG, voice integration, and safety guardrails. Built on open-source post-training techniques, these agentic models set a new standard for enterprise AI.

A student’s ambitious effort to convert Mistral Nemo into a 16-expert MoE model has ignited interest in the AI community. With budget limits halting further training, developers are weighing how to support open-source MoE advancements.