Articles & Guides

StartupBench builds an end-to-end benchmark from AI startup products with demonstrated market demand rather than researcher-designed tasks. Even the strongest evaluated model reliably completes only about 30% of the benchmark.

Personalized Auto-Research argues that a genuine AI co-scientist must understand a researcher’s prior work, methods, collaborators, and communities instead of producing a generic plan for a shared goal.

PTXBench evaluates GPU-kernel generation beyond compilation by checking correctness, runtime execution of target PTX instructions, and speed against frontier libraries. Its results expose a persistent gap between using architecture-specific instructions and delivering competitive performance.

LEGO-RL addresses a practical gap in training coding agents: native harnesses provide realistic tools and execution feedback, but their behavior can be difficult to align with policy-gradient optimization. The framework improves Qwen3.5-35B-A3B on SWE-bench Verified across three agent harnesses.

MathForm combines Mathlib retrieval, compiler diagnostics, and semantic-consistency feedback to refine mathematical formalizations iteratively. Its 8B model is trained on roughly 367K verified Lean 4 examples and outperforms several specialized 32B systems on the reported benchmarks.

EditBridge introduces a diffusion bridge for ultra-high-resolution image editing. Instead of independently super-resolving a low-resolution edit, it uses the original high-resolution image to guide detail recovery while reducing cross-image attention costs.

ASI-Bench evaluates a capability that conventional benchmarks rarely capture: discovering promising directions, choosing research methods, and producing verifiable results with limited human guidance. Its results show a steep performance decline as methodological support is removed.

Stripe has confirmed its acquisition of AI gateway startup OpenRouter, reportedly valuing the deal at $7.5 billion. Behind the playful talk of the “singularity” is a more practical push into AI usage and cost management.

OpenAI is previewing Private Safety Processing, a system designed to detect misuse across multiple conversations without retaining customer data. The move sharpens the enterprise privacy and safety contest with Anthropic.

Google won a $10 million auction for a large Spirit Airlines enterprise dataset containing decades of worker records. Flight attendants argue that removing names does not adequately protect confidential employment information.

AI is spreading across search, email, televisions, and other everyday products, yet broader adoption has not translated into public acceptance. Concerns about jobs, intellectual property, infrastructure costs, and corporate accountability are turning an image problem into a business risk.

Google is adding interactive visuals, simulations, quizzes, document summaries, and research workflows to Search and Gemini. The updates position both products as a more complete toolkit for students.

Several cybersecurity researchers reported losing access to OpenAI’s Trusted Access for Cyber program without warning. OpenAI said a technical issue affected a limited number of users and asked them to re-verify their identities.

As spending on GPUs and data centers continues to rise, compute is becoming a major cost that businesses may need to hedge. Silicon Data wants to create a reference price for GPU rentals and bring compute futures to Wall Street.

Meta platforms carried ads for Kromix, an AI image and video tool that appeared to promote nonconsensual sexual deepfakes. One video used a face closely resembling a prominent female US politician before cutting to pornography.

StateM improves long-horizon agents by redesigning the execution layer rather than changing model weights. On Terminal-Bench 2.1, the system reaches 95.3% raw accuracy with GPT-5.6 Sol xhigh and transfers across several models.

Large Discovery Model combines a generative model with a Bayesian non-parametric reward surrogate, allowing AI systems to propose candidates while using experimental feedback and uncertainty to guide the next search.

MOSS-VL is designed for real-time interaction rather than treating streaming vision as an afterthought. Its gated cross-attention, interaction-focused data, and staged training help the model keep processing incoming frames while generating responses.

HarmProfile treats harmful generation as an object of study rather than merely the result of a successful jailbreak. Its findings suggest that frontier models have distinct risk profiles, while harmfulness and output diversity rise with capability.

A new study proposes “agentic transactions,” reinterpreting the ACID principles of databases for LLM agents that reason, use tools, generate code, and modify persistent workspaces. Its data-agent implementation reports a 10.6% gain over existing agents on relevant benchmarks.

AutoResearchEval examines 100 real frontier-science tasks across the full research lifecycle. Its central finding is that agents do not mainly fail from a lack of raw capability, but from the absence of a reliable metacognitive loop for checking and correcting their own work.