The Archive

AI labs have staked their reputations on model benchmarking — the supposedly neutral and accurate system of evaluations that are designed to measure whether one AI product is better than another. However, the methods behind these evaluations are complicated and still maturing, said Irene Solaiman, Hugging Face’s chief policy officer, during a discussion at TechCrunch […]

Microsoft wants enterprises to closely watch the agents employees use with systems not controlled by AI labs like OpenAI and Anthropic.

The US is building more fossil fuel infrastructure than it has in a decade while renewable energy growth slows down, a new report shows. Electricity demand is also climbing faster than it has in more than a decade with more data centers, factories, and electric vehicles plugging into the grid. Despite those growing demands, the Trump administration has sought to hobble the wind and solar industries in favor of fossil fuels. America's energy landscape is already shifting as a result, according to the report published last week by the nonprofit Environmental Defense Fund and research firm Atlas...

CoreWeave had a bad week last week. Its bonds slipped into distressed territory by a common definition - when investors demand 1,000 basis points more in yield than they'd get from government bonds. Last Tuesady Chamath "SPAC King" Palihapitiya noticed the distressed debt, and also asked whether the AI buildout was a private credit bubble. Its shares are down more than 20 percent since Tuesday's close, which is kind of a lot! There even seemed to be some collateral damage in the market on other AI companies, too. Now, I have also asked if the AI industry is a bubble, but when the SPAC King ec...

US fears Saudi government will help ex-OpenAI engineer flee spying charges.

Amazon wants to break Nvidia’s hold on PyTorch, the free toolkit used to build most AI apps, so that more developers run their apps on alternative chips like its own.

OpenCode is an open-source AI coding agent that's become a popular alternative to Claude Code and Cursor. But hackers got into part of its online platform and stole some users’ API keys, and the AI tools it turned to for help kept refusing to look at the malicious code.

AI companion app Tolan has raised another $30 million after attracting more than 250,000 active users.

Eugenia Kuyda's newest startup Wabi is positioned to take advantage of an idea she had more than a decade ago. Wabi offers a personal AI assistant with a twist, free and downloadable to anyone today.

After drawing acquisition interest earlier this year, the Nvidia challenger raised $1.5 billion to remain independent.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. Until recently, you could divide internet users into just two categories: bots and people. Websites have long allowed bots to perform automated tasks, like capturing the content of web pages, but rules…

Halcyon achieves state-of-art text-to-image generation without language pretraining, using 18-275× less compute than LlamaGen and JiT-T2I on GPIC dataset.

OpenVLHarness framework enables multimodal LLMs to use external tools and retain evidence across reasoning steps for visual agentic tasks.

FunCAD-Harness framework evaluates LLM agents on functional CAD mechanism design by verifying physical requirements in simulation.

Study on privileged-correction contamination in multimodal RL with self-distillation, identifying how non-visual factors corrupt token-level feedback.

CoopUQ framework enables safe multi-robot navigation via uncertainty-aware cooperative forecasting in human environments and OOD scenarios.

PhysRSI-eval benchmark quantifies physical recursive self-improvement in VLMs across 18 manipulation tasks; identifies bottlenecks via Physical Barrel Hypothesis.

WeAct enables long-horizon agents to avoid catastrophic forgetting by treating adapter selection as learned actions, managing heterogeneous capabilities.

Agentic Imagination framework trains agents to imagine action outcomes before execution, improving planning under irreversible or sparse-feedback environments.

AnyGoal studies world model generalization beyond expert rollouts via compositional dynamics, linking data coverage to downstream planning performance.

LatComm introduces reusable latent context compression between LLMs, enabling compact encoding that is readable across multiple models and questions.

Newton-based bilevel optimization methods for hyperparameter tuning and meta-learning without lower-level strong convexity assumptions.

GPT-5.1 and Claude agents lose accuracy retrieving tool outputs from external files rather than context windows, measured on USGS PHREEQC geochemistry tasks.

KronLSTM extends xLSTM with Kronecker Delta Rule memory-revision for long-context language modeling with reduced interference.

Quantum Hamiltonian embeddings for supervised learning encode datasets into parameterized Hamiltonians with trainability guarantees.

Hybrid architectures mixing linear recurrence and softmax attention exhibit competing learning where layers suppress each other's signals.

ProteinGeoBench reveals frontier agents retrieve protein metadata reliably but fail at geometric reasoning on 131 structural-biology tasks.

WildAWM synthesizes executable GUI/web/code environments for training computer-use agents without real-world security risks.

KEEP-Bench evaluates whether context compression preserves evidence for plausible resolutions and minority perspectives across 4,102 cases.

RefCraft benchmarks agentic web generation from visual references via 300 realistic tasks, moving beyond text-only or reconstruction objectives.