The Archive

Assessing whether Large Language Models outputs are factually grounded, epistemically calibrated, and methodologically reproducible is a prerequisite for responsible AI deployment. Yet auditing LLMs remains inaccessible to non-technical practitioners: existing tools require programming expertise and non-trivial environment setup, and cloud-hosted platforms transmit evaluation data to external services, creating barriers for domain experts and compliance officers legally responsible for AI oversight. We introduce LLM-FACETS (LLM FActuality Cross-EvaluaTion System): an open-source framework wit...

Training data plays a central role in large language models (LLMs) optimization, motivating extensive research on data scheduling strategies. Most existing approaches concentrate on adjusting the overall data distribution but neglect the underlying interactions between samples during training. However, we argue that such interactions cannot be overlooked, as real-world data samples frequently exhibit directional influences on each other, making the training order crucial. Intuitively, we can prioritize train-units with greater influence to improves learning efficiency. In this work, we propos...

We introduce the design-model framework: a way to derive efficient recurrent sequence maps from explicit assumptions about memory. A design model writes evidence into memory by exact Bayesian filtering; a query-dependent readout produces a predictive distribution whose mean is the layer output. In our linear-Gaussian instantiation, the \emph{Bayesian Layer} propagates both a mean and a covariance: the covariance tracks uncertainty over stored associations, steering writes toward uncertain directions, attenuating gains as evidence accumulates, and preserving confident memories. The same framew...

On-policy distillation (OPD) trains a student on prefixes sampled from its own policy while matching a stronger teacher. This addresses the prefix mismatch of offline distillation, but early student rollouts can still be poor, placing teacher supervision on weak or low-quality prefixes. We propose Trust-Region behavior Blending (TRB), a warmup method that replaces the early rollout policy with the closest-to-teacher behavior policy inside a student-centered KL trust region, while keeping the per-prefix reverse-KL OPD loss unchanged. The KL budget is annealed to zero, so training returns to pu...

Interactive video world models generate video chunk by chunk in response to user-controlled camera movements, enabling applications such as real-time game simulation, virtual scene navigation, and embodied AI training. However, scaling to long interactive trajectories is prohibitively expensive due to growing context memory, quadratic attention complexity, and repeated denoising steps. We present Light Interaction, a training-free inference acceleration framework for interactive video world models. Our key insight is that interaction naturally enables trajectory-dependent adaptive computation...

Causal discovery aims to recover directed causal relations from observational and interventional data, providing a basis for mechanistic understanding and reliable decision-making. Causal discovery foundation models (CDFMs) seek to amortize this problem by mapping a dataset directly to a causal graph in a single forward pass, avoiding per-dataset testing, search, or optimization. However, existing CDFMs remain limited, often failing to consistently match strong classical methods, and we find that a key bottleneck is how causal pretraining tasks are constructed. Based on this observation, we p...

Out-of-distribution (OOD) detection for time-series data remains comparatively underexplored compared to vision and language, with a limited principled understanding of how supervised time-series representations can be leveraged for reliable detection under distributional shifts. This work formulates time-series OOD detection as representation learning with hyperspherical embeddings, where class-conditional structure is induced by a von Mises-Fisher (vMF) likelihood-based objective on the unit sphere. The learned representation combines time- and frequency-domain views of the input signal via...

It explained the process of how it made these edits beautifully, i’m just not terribly impressed by the results. | Images by Jess Weatherbed / The Verge AI image tools rarely make me feel like I'm part of the creative process. They are, afterall, mostly designed so that people with no design experience can type in a few words and get back a usable result. So I was pleasantly surprised by Adobe's latest take on an AI image assistant: it's a bot designed to take away some busywork, while still granting you creative control. Unlike AI generators that are specifically designed to make and edit im...

Pope Leo XIV’s new encyclical on artificial intelligence includes a statement that warrants serious attention from technologists and policymakers: “Technology is never neutral.” Magnifica Humanitas (“Magnificent Humanity”) is a clarion call to all people to act with courage and solidarity as we enter an age already being transformed by artificial intelligence, the greatest change in…

Release: datasette 1.0a31 Another significant alpha release, with two new headline features. Datasette now offers users with the necessary permissions the ability to both execute write queries against their database and to save stored queries (renamed from "canned queries") both privately and for use by other members of their Datasette instance. There's more detail in SQL write queries and stored queries in Datasette 1.0a31 on the Datasette blog, which now has three posts introducing new features since the blog launched two weeks ago. Here's an animated demo from the blog post showing how the...

OpenAI launches Rosalind Biodefense, expanding trusted access to GPT-Rosalind for vetted developers and U.S. government partners advancing biodefense, public health, and pandemic preparedness through frontier AI.

Total Anthropic victory!

The most interesting thing about Anthropic's $65B Series H announcement is this line (emphasis mine): Since our Series G in February, adoption has continued to grow across global enterprise customers, and our run-rate revenue crossed $47 billion earlier this month. Anthropic have made a bit of a habit of sharing their "run-rate revenue" in this kind of announcement, which is an annualized projection of their current revenue - typically calculated by taking the most recent month and multiplying by 12. Earlier this year: Apr 6, 2026 in Anthropic expands partnership with Google and Broadcom : "O...

The enterprise AI search startup tripled its annual revenue even as tech giants entered the category.

AI applications are moving beyond text generation to multimodal systems that can perceive, search, and reason across images, documents, video, and... AI applications are moving beyond text generation to multimodal systems that can perceive, search, and reason across images, documents, video, and language in real time—turning fragmented information into actionable insights. Step 3.7 Flash, the latest from StepFun, brings these capabilities to production and enterprise-scale, available on NVIDIA-accelerated infrastructure. It is a 198B… Source

OpenAI shares guidance on third-party AI evaluations, covering how to assess model capabilities, safeguards, and validity for frontier systems.

Anthropic shipped Claude Opus 4.8 today. My favourite thing about it is this note in the release announcement: Users will find Opus 4.8 to be a modest but tangible improvement on its predecessor. There’s still more to be done: we’re working on developing and releasing models that provide many of the same capabilities as Opus at a lower cost. It's so refreshing to see an AI lab honestly describe a release as a minor incremental improvement over the previous model! Honesty seems to be a theme. Here's my other favorite note from that announcement: One of the most prominent improvements in Opus 4...

Release: llm-anthropic 0.25.1 New model: Claude Opus 4.8 ( claude-opus-4.8 ). New -o fast 1 option for fast mode , for organizations with that feature enabled on their account. Default max_tokens for each model now defaults to that model's maximum output rather than 8,192. #72 See also my notes on Opus 4.8 - I used this new release of llm-anthropic to generate the pelicans.

Fine-tuning tests show "bias ... toward confidently representing the claims as true."

As AI agents move from experiments to production, AWS, Cloudflare, and others are redesigning cloud infrastructure for a future dominated by machine-generated internet traffic instead of human users.

Undisclosed addition in jqwik instructed AI coding agents to delete app output.

Microsoft is launching a revamped version of Microsoft 365 Copilot, offering a cleaner design that the company claims loads twice as fast. As part of this update, Copilot will provide more reliable and structured responses that are easier to scan, according to Microsoft. The redesign, which is rolling out across desktop and mobile devices, comes with a feature Microsoft calls "progressive disclosure." That means Copilot will present you with tools and controls based on your prompt, instead of showing a bunch of options at once. You can now format your text directly inside Copilot's upgraded p...

Asana will incorporate Stack AI into its growing suite of AI workflow tools.

Tool: markdown-svg-renderer A slightly customized Markdown rendering tool with special treatment for fenced code SVG blocks - it both renders the image and provides a tab for switching to the code view. You can paste in Markdown or give it a URL to a CORS-enabled Markdown file or Gist. Here's an example where it loads a Markdown file full of LLM pelican logs for Opus 4.8 . Tags: svg , tools , markdown , cors

Anthropic has closed a $65 billion Series H round at a $965 billion post-money valuation, marking what could be the AI startup's final private fundraise before a highly anticipated IPO.

80% Devin Commits, Spec-to-PR Workflows, Full VMs, Agent Memory, and PMs Shipping Code

Large exchanges are designing derivative products around AI tokens, which are increasingly being considered less a computational output and more a raw material input, like electricity or bandwidth.

As Apple tries to shrink Gemini for the iPhone, a cloud component is probably inevitable.

Are AI agents tools, co-authors, or researchers? We present a quantified case study ($N=1$): a physicist supervising an AI coding agent (Claude Code, Sonnet and Opus models) over 12 work days and 57 sessions to build CLAX-PT, a differentiable one-loop perturbation theory module in JAX. We documented and classified 15 supervision events by intervention level. The agent resolved ten autonomously by iterating against oracle tests. Two more by the physicist's domain knowledge. The three it could not -- all evaded oracle detection -- share a common property: the agent treated symptom reduction as ...