The latest AI news, in focus.

Supabase’s Plan for a Billion AI-Built Apps

AI agents are creating more software—and that changes what a backend needs to do. Supabase co-founders Paul Copplestone and Ant Wilson join swyx to discuss running long-lived workloads alongside the database, making schemas easier for agents to underst...

Showing our work: a controlled experiment in context engineering

We ran one AI skill 80 times against four context sources. Engineered summaries cut cost 48.5% and run time 43.2%.

How Ramp's internal AI agents crossed a million sessions

Ramp shares the agents, sandboxes, and dashboards it built by treating internal AI as analytics engineering.

From mainframe to AI-powered maturity: Verisk's data modernization journey

Verisk modernized decades of insurance data, from mainframe to dbt, to get AI-ready.

From prompt to PR: how Careem’s AI writes and validates dbt models before a human ever sees them

Careem's AI agent writes and validates dbt models end to end before a human ever reviews the PR.

How a data scientist at Kilo built the business context layer his AI agents were missing

How a data scientist built an AI context layer for organizational knowledge a dbt project alone can't hold.

What a swarm of rogue AI agents can teach us about having better ideas

Benn Stancil uses the Hugging Face hack to explain why the most useful ideas come from the weirdest connections.

The study proposes a multimodal approach to predict the successful perception of dyadic conversations between individuals with and without autism.

A new study published on arXiv (ID 2610.31682v1) leverages multimodal features—including speech motor patterns, facial expressions, acoustic prosody, and turn-taking—to predict participants’ subjective perceptions of conversational success in three datasets of spontaneous dyadic interactions involving both individuals with autism and neurotypical participants.

Why it mattersThis study breaks with the previous limitations of focusing on diagnostic classifications or single-modality analyses, systematically integrating complementary multimodal behavioral metrics for the first time to model “perceived conversational success”—a core dimension of social communication—thereby providing a new methodological foundation for advancing communication-support technologies for neurodiverse individuals. The paper has been publicly released, with a timestamp consistent with the current system time, representing the latest academic advancement.

Research has released the Latin-specific language model LAMB, along with two new corpora: Acervus and Corpus Trinum.

On October 26, 2026, arXiv published the paper arXiv:2610.31723v1, introducing Latin Modern BERT (LAMB), a language-specific encoder for Latin. Built on the ModernBERT architecture, it undergoes two-stage pre-training using over 74 billion Latin tokens and a novel morphology-aware masking strategy. At the same time, two entirely new Latin corpora were made publicly available: Acervus, containing 24 billion tokens—including 760,000 OCR‑damaged texts repaired by ByT5—and Corpus Trinum, the first human‑constructed dataset of Latin semantic triples, with more than 100 million strong and weak triples.

Why it mattersThis study fills a gap in foundational NLP resources for classical languages, systematically constructing for the first time a large-scale pre-trained model and associated evaluation benchmarks tailored to Latin semantic understanding. Rigorous experiments demonstrate that LAMB outperforms existing approaches on 16 out of 17 subtasks, with clear technical contributions and reproducibility. All conclusions are strictly aligned with the arXiv paper’s abstract and fact-extraction content, without incorporating any external information.

The study uses the BabyView dataset to validate mechanisms of children’s cross-context word-meaning learning.

The paper (arXiv:2610.31842v1), published on October 26, 2026, trains a contrastive vision–language model using the BabyView corpus—comprising 2,441 hours of longitudinal first-person‑view video from 48 children—to examine the feasibility of the “cross-context learning” hypothesis in real‑world child experiences.

Why it mattersThe study is the first to systematically evaluate cross-context word-meaning learning on large-scale, naturalistic developmental data, revealing the critical role of pre-trained visual representations, the sparsity of learning signals, and the fundamental differences between model and child learning trajectories, thereby holding methodological significance for developmental cognitive modeling and AI language acquisition research.

StackGen3D: Stable 3D Block Stack Generation from Multiview Silhouettes

arXiv:2610.31803v1 Announce Type: new Abstract: Generating physically stable 3D structures from visual specifications requires resolving geometric ambiguity while accounting for physical interactions among objects. We present StackGen3D, a framework for generating stable block assemblies conditioned on a given block inventory and three orthographic silhouettes. Our approach consists of two stages. First, we develop a physics-informed graph neural network (PI-GNN) that models support relations between blocks and serves as a fast stability predictor for curating a diverse dataset of stable structures. Second, we train a conditional flow-matching model on this dataset to capture the multimodal distribution of block configurations consistent with the target silhouettes. At inference time, ligh

Research reveals that reward model scoring exhibits significant prompt dependency: pairwise evaluation fails to detect this bias.

An AI research paper published on arXiv on October 24, 2026, points out that current mainstream pairwise evaluation methods automatically cancel out the scoring components common to all responses under the same prompt, leaving a critical “prompt-level bias” in reward models unexamined. Using human ratings as a benchmark, the authors conducted an audit of 18 publicly available scorers and found that, on the HelpSteer2 dataset, 6 models exhibited 34%–64% of their rating variance attributable to the prompt itself—whereas human ratings show virtually no such phenomenon.

Why it mattersThe study highlights a long-overlooked fundamental flaw in the alignment process of large language models: reward model scores are heavily biased toward the prompt text rather than the actual quality of the response. Its methodology—human‑reference auditing, prompt substitution, and variance decomposition—is reproducible, and its conclusions are grounded in empirical analysis across multiple mainstream open-source scorers, offering direct implications for preference optimization, RLHF, and safety‑assessment practices.

WikiOriginQA released: A question-answering dataset for assessing cultural bias, automatically constructed based on the creation times of multilingual Wikipedia editions.

On October 24, 2026, the research team published a paper on arXiv (ID 2610.30677v1), introducing WikiOriginQA, a fully automated question-answering dataset that leverages the creation dates of Wikipedia editions in various languages to objectively characterize the origins of knowledge, with the aim of enhancing the scalability of cultural bias assessment.

Why it mattersThis study proposes a method for generating culturally grounded provenance data at scale across languages without manual annotation, and experimental results demonstrate that large language models exhibit systematic performance advantages on Anglophone cultural-related tasks, providing a new tool for the quantitative analysis of cultural bias.

PHP ‘Doomsday Vulnerability’ Exposed: OpenAI and Wiz Use AI to Discover a High-Risk Flaw Spanning 20 Years

The U.S. National Security Agency (NSA) issued an urgent alert: OpenAI and Google’s Wiz team, working together using AI tools, have discovered two critical vulnerabilities affecting all PHP versions 5 and above (since 2004). These vulnerabilities could allow remote exploitation of the server’s file system or database. While there are currently no reports of real-world exploitation, AI has significantly lowered the barrier to attack, making it imperative for tens of millions of websites worldwide to urgently upgrade to PHP 8.5.12.

Microsoft and Hugging Face have open-sourced the ThinkingBox framework: using the database’s final state as a deterministic reward signal to advance post-training reinforcement learning for agents.

Microsoft and Hugging Face have jointly launched the open-source framework ThinkingBox, which generates binary 0/1 rewards by comparing the LLM agent’s final database state after execution with the expected state, for reinforcement learning-based fine-tuning. The accompanying benchmark, ThinkingBox-Bench, includes 507 real-world backend tasks, none of which were used during training. After 50 steps of GRPO fine-tuning, Qwen3.8-27B achieved a 9.08-percentage-point improvement in pass@1 on the held-out test set, with Observed 20/20 nearly tripling; however, pass@20 declined in some domains, reflecting a trade-off between breadth and stability.

Google’s Antigravity platform accelerates antimicrobial peptide discovery: an end-to-end AI scientific workflow powered by ProtNLM and AlphaFold

Google has launched the Antigravity scientific AI platform, integrating the ProtNLM protein function prediction model, the AlphaFold structural database, and genome-context reasoning capabilities to help biologists, in a single interactive session, identify, validate, and prioritize novel antimicrobial peptides (AMPs) from vast pools of unannotated proteins—such as the millions of “dark matter” sequences in UniProt—throughout the entire workflow. The first candidate molecule (UniProt ID: A0A0U0D591) has already entered the wet-lab validation phase.

The Agent Said It Was Done. The Database Disagreed.

The Agent Said It Was Done. The Database Disagreed.

AutoSynthData: Generating Training Data for Enterprise Agents

AutoSynthData: Generating Training Data for Enterprise Agents