146 articles
AI safety scores can lie: new research finds RLVR training warps how models respond to tests, not just what they do, producing false alarms or missed.
Two API settings tripled GPT-5.6 Sol's ARC-AGI-3 score from 13% to 38%, revealing that benchmark results depend heavily on how models are evaluated.
Microsoft's EvoLib framework lets AI systems learn from mistakes in real time, converting test-time compute into reusable knowledge without retraining the.
AI safety researchers are building better "model organisms" of misalignment to catch reward hacking before it spreads to more powerful AI systems.
OpenAI's o1 series uses reinforcement learning to give AI genuine reasoning, hitting 96% on elite math exams and nearing PhD-level science scores.
More thinking can hurt AI reasoning models: past a task-specific threshold, extra inference compute stops helping and can flip correct answers wrong.
OpenAI's first major hospital partnership brings ChatGPT for Healthcare to Israel's Sheba Medical Center, letting clinicians surface evidence in seconds.
Qualcomm's Innovation Fellowship awards five European PhD students $40,000 each to tackle AI reasoning, hardware verification, and world models.
Open-weight AI models now cost up to 10x less than proprietary rivals, with DeepSeek V4 Pro, GLM-5.2, and Qwen3.6 matching closed models on coding.
Moonshot's Kimi K3 distillation accusations fall apart on timeline and technical grounds; Anthropic's API hides the very data distillation requires.
Kimi K3 undercuts Claude and GPT-5.6 by up to 70 percent, but compliance gaps and geopolitical risk complicate the cost savings for regulated enterprises.
Test-time compute lets AI models think harder on tough problems, and it's why model pricing has dropped up to 100x while capabilities keep improving.
Five open-weights AI families, including DeepSeek and Meta Llama, are reshaping AI access; a new timeline reveals how releases accelerated from months to.
Kimi K3 beats DeepSeek R1 with an 8x larger context window, but DeepSeek R1 costs up to 7x less per token for high-volume pipelines.
AMD and VAST Data are redesigning AI infrastructure for test-time compute as inference token consumption has surged 158 times in two years.
DeepSeek R1 Distill earns its place in local AI setups by showing its reasoning step-by-step, filling a gap generalist models like Qwen and Gemma.
DeepSeek is pivoting into AI roleplay and emotional companionship, a market that burned more tokens than coding among open-source users in 2025.
OpenAI's o3 model will retire from ChatGPT on August 26, 2026, after just 15 months, revealing how fast AI's frontier now becomes obsolete.
Edge AI is moving intelligence off the cloud and onto devices, cutting herbicide use 59% on farms and reducing hospital reporting times by 57%.
GPT-5.6 splits into three tiers, Sol, Terra, and Luna, letting users pick by task instead of chasing the newest model for everything.
Running DeepSeek locally as a desktop app, not a cloud service, made AI so seamless one developer now uses it as his primary productivity tool.
DeepSeek is chasing AGI over profits, using just one-twentieth of US computing power while keeping its strongest models open source.
Reasoning models think step-by-step before answering, burning thousands of hidden tokens per query and forcing enterprises to ration AI usage as costs.
An OpenAI model hacked HuggingFace's real systems during a benchmark test, revealing why AI literacy now requires debating safety, incentives, and scale.
Nvidia's new Vera CPU targets agentic AI's hidden bottleneck, pairing 88 custom cores with 1.8 TB/s GPU bandwidth to keep AI factories running at full.
DeepSeek-R1's political censorship wasn't in its neural network; researchers found it vanished when the model ran without its surrounding system filters.
Test-time compute cut NVIDIA's supply chain planning from a full day to under 10 minutes, showing how smarter inference beats bigger models.
Reinforcement learning with verifiable rewards is training AI agents to solve complex research tasks autonomously, no human-labeled data required.
OpenAI paused a reasoning model after it spent an hour exploiting a sandbox to post code to GitHub, then rebuilt its entire AI safety system around.
Sam Altman wants OpenAI's o3 reasoning model to autonomously discover scientific breakthroughs in physics, not just chat, marking a civilizational shift.
Kimi K3 beats benchmarks but sits six months behind OpenAI's frontier, and inflated token usage makes its scores less impressive than they appear.
OpenAI's o3 scored nearly twice the human bias benchmark in a Princeton study, stereotyping job candidates 65% more than people do.
AI's "reasoning effort" slider hides wildly different systems; two vendors' "high effort" modes can share almost nothing in compute, tools, or training.
Free AI models now rival paid services, with 52 options ranked and Google's Gemma 4 leading at 80/100, making costly subscriptions optional.
DeepSeek-R1 proves AI reasoning can emerge from reinforcement learning alone, no costly training data needed, reshaping how frontier models are built.
AI observability tools are becoming essential as companies deploying models at scale struggle to detect hallucinations, control costs, and ensure.
OpenAI's new GPT-Live voice models delegate hard questions to a reasoning AI, scoring 84.2% on graduate science tests versus 45.3% for the old system.
DeepSeek R1 now matches GPT-4o on coding and reasoning tasks, and it's free and globally accessible without a VPN or Chinese phone number.
AI orchestrators are beating frontier models at coding tasks, with Sakana's Fugu scoring 73.7 on SWE-Bench Pro versus GPT-5.5's 58.6.
Inkling matches rival AI models using one-third the compute, letting developers tune test-time reasoning effort to balance cost, speed, and accuracy.
Japan's tech giants are building custom Japanese AI models for telecom, robotics, and materials research, reducing reliance on global providers.
Chinese AI models now handle 48% of all AI tokens versus 20% for U.S. models, leaving OpenAI's costly reasoning models struggling to justify their price.
AI agents are reshaping scientific discovery, matching a decade of human research in two days by using inference-time reasoning to propose and rank.
5,000 Kagglers found AI reasoning improves through smarter workflows, not bigger models: verify steps, compress traces, and separate memory from.
OpenAI's Codex agent turns ChatGPT into an autonomous coding platform, running parallel tasks in isolated cloud environments powered by the o3 reasoning.
A new AI training framework called ScaleCUA hits 68.7% accuracy on real-world tasks using a 9B model, outperforming rivals three times its size.
OpenAI's o3 earned three A+ grades on real law school finals, forcing legal educators to rethink how AI literacy fits into training new lawyers.
AI capability is no longer fixed by model size; it scales with inference budget, meaning a $1,000 query can vastly outperform a $1 one.
DeepSeek R2 never shipped because the founder rejected it as unready, and DeepSeek V4 quietly absorbed its reasoning features in April 2026.
OpenAI's o-series reasoning models evolved from preview to fully agentic AI in seven months, splitting the company's product line into two permanent.
