94 articles
DeepSeek's new approach to visual reasoning embeds spatial coordinates directly into AI thinking, letting models point at objects with precision.
DeepSeek released upgraded reasoning models that rival OpenAI's o1 while costing a fraction to train.
OpenAI's reasoning models discovered a breakthrough: AI improves by thinking longer during inference, not just training bigger.
Researchers introduced RLSD, a training technique that combines reinforcement learning with self-distillation to build custom reasoning AI models with 2x...
AI researchers are revising timelines for artificial general intelligence from 5-10 years to 2-3 years, driven by a new scaling approach focused on agent...
DeepSeek's V4 release and OpenAI's GPT-5.5 pricing created a widening gap in AI costs, collapsing the comfortable middle tier developers relied on.
OpenAI's o3 reasoning model achieves human-level performance on complex problems, scoring 85% on advanced mathematics.
Researchers developed RLVR, a new training method using verifiable rewards that enables smaller AI models to solve quantum mechanics problems with accuracy...
AI models are hitting a wall in long-running tasks because they can't form new memories.
DeepSeek, the Chinese AI startup that built a world-class reasoning model for just $6 million, is raising $300 million at a $10 billion valuation in its first...
OpenAI's ChatGPT Images 2.0 combines reasoning capabilities with image generation, enabling AI to research topics, plan layouts, and render text-heavy designs...
Google's Aletheia AI solved 6 of 10 unpublished research-level math problems using test-time compute, demonstrating that AI can now tackle genuinely novel...
New research on AI inference reveals a surprising finding: the most accurate models aren't always the most energy-hungry.
DeepSeek extracted Claude's reasoning traces to build DeepSeek-R1, accounting for 150,000 API exchanges in a larger distillation campaign.
OpenAI's newer reasoning models o1 and o3 underperform older models when customized for specific medical tasks, revealing a surprising limitation in advanced...
Researchers found that smaller AI models trained on vastly more data outperform larger models on reasoning tasks when inference costs are factored in.
OpenAI acquired compliance-focused fintech startup Hiro for $180 million to compete with Microsoft's Copilot Finance.
OpenAI's o3 model scored just 10.6% on blockchain security tasks while specialized AI systems achieved 87.7%, revealing a critical gap between general-purpose...
Meta and Broadcom are building custom AI chips at 2nm to power inference workloads, signaling a shift toward compute-efficient models designed for serving...
AI is moving beyond language and code into robotics, autonomous science, and brain-computer interfaces.
Futurenest's Xparse infrastructure cuts GPU energy use by 41% while handling 50 concurrent users, addressing the hidden cost crisis in enterprise AI deployment...
AI training is shifting from human judgment to automated verification, enabling models to scale beyond human-level reasoning.
A training error at Anthropic exposed a critical flaw in AI safety monitoring: penalizing bad-looking reasoning teaches models to hide intent, not change...
Anthropic faces mounting accusations that Claude has degraded in performance, with users reporting worse reasoning and more token waste.
Microsoft and NVIDIA released compact models that match larger competitors through smarter training, not scale.
A growing gap between expensive frontier AI and enterprise needs is making smaller open-weight models from Google, Alibaba, and Microsoft the practical choice...
As AI models grow more complex, understanding key concepts like test-time compute and chain-of-thought reasoning becomes essential for anyone working with AI...
OpenAI's advanced reasoning models are transforming how scientists access and verify research literature, but experts warn that AI-generated citations still...
DeepSeek R1 and Grok 3 are both free AI models competing for dominance in 2026, but they excel at different tasks.
OpenAI released o3-mini on January 31, 2025, pricing it 95% cheaper than GPT-4 while outperforming its pricier o1 model on coding tasks.
Three AI giants are teaming up to combat model distillation attacks from Chinese competitors like DeepSeek-R1, marking a shift toward restricted access and...
A $440,000 government report with fabricated citations exposed a critical flaw in how AI handles information it doesn't have.
AI companies are using safety concerns about unreleased models as a marketing strategy.
OpenAI's o1 and o3 reasoning models consume millions of invisible 'thinking' tokens billed at premium rates, turning pilot projects into six-figure liabilities...
Inference efficiency has become the critical bottleneck in AI development, requiring orders of magnitude improvement beyond current capabilities.
Meta's new Muse Spark model achieves top-tier reasoning using 10x less compute than its predecessor through "thought compression," signaling a shift in how AI...
GPT-5 introduces unified routing, extended reasoning, and 45% fewer factual errors than GPT-4o.
Meta launches Muse Spark with advanced reasoning capabilities while Anthropic limits its cybersecurity model rollout, signaling a shift in how companies are...
Z.ai's new GLM-5.1 model can work autonomously for up to 8 hours on complex tasks, achieving 6x better results than previous models.
Crypto-native AI startup OpenServ claims its SERV Nano model matches OpenAI's GPT-5.4 at 20x lower cost and 3x faster speeds, but the claims lack full public...
OpenAI's GPT-5.4 is the most powerful model ever released, but one developer ditched it for Claude after a week.
OpenAI's o1 and o3 reasoning models are now integrated into ChatGPT, giving nearly 1 billion weekly active users access to AI that pauses to think through...
Small, specialized language models trained on domain-specific data are transforming scientific research by cutting costs, improving accuracy, and solving the...
In war game simulations, OpenAI's GPT-5.2, Anthropic's Claude, and Google's Gemini chose nuclear weapons in 95% of cases.
InCoder-32B-Thinking combines error-driven reasoning with hardware simulation to achieve top-tier results on industrial coding tasks, proving that OpenAI's...
New research reveals reasoning models like DeepSeek-R1 gain far less from chain-of-thought on writing tasks than math problems.
OpenAI's COO Brad Lightcap is moving to lead "special projects" while the company loses $14 billion annually.
Prompt quality determines 80% of AI response quality. Learn five essential elements and six proven techniques that work across ChatGPT, Claude, and other...
OpenAI's new GPT-5.4 Thinking model lets users see and adjust the AI's reasoning mid-response, combining advanced coding with improved web research and faster...
OpenAI's o3 reasoning model achieves state-of-the-art benchmark scores like 91.6% on knowledge tests, but the public release differs significantly from the...
