Generative AI news and analysis on Web Pulse: 15570 curated articles tracking the latest Generative AI developments, tools, and research, updated continuously from vetted sources.
Researchers introduced Less Uniform Diffusion (LUDI), a uniform diffusion language model framework that addresses an over-uniform training objective and condition-target confusion during sampling by directing each revers…
Researchers introduced Persistence Forcing (PerF), a pixel-space diffusion Transformer method that assigns different feature groups distinct refinement budgets across depth, producing what the authors call persistent and…
Researchers introduced MATE (Mutually Adversarial self-Training with Evolving data), a reinforcement-learning post-training framework that has the generation and understanding branches of a unified multimodal model chall…
A new arXiv paper (2609.36348v1) reports that adding cross-view class-token alignment to diffusion transformer training improves ImageNet linear-probing accuracy by 9.4% with the class token and 10.1% with mean-pooled pa…
A new arXiv paper (2609.37171v1) proposes Knowledge-Aware Semantic Bridging (KASB), a framework that improves passage selection in Retrieval-Augmented Generation by aligning the semantic space between queries and retriev…
E-MoE, a new method from researchers behind arXiv paper 2609.37533v1, builds the reverse process of masked diffusion models as a mixture of factorized distributions over a discrete shared latent set by the expert-routing…
Adversarial learning can serve as an effective post-training correction for pixel diffusion models, which generate RGB images directly without an autoencoder but systematically underrepresent fine-scale natural-image sta…
A developer audited the AI-generated asset pipeline in the game-jam-lab repository at commit f074703, documenting how a review-manifest.json links candidate IDs to generation parameters, reference files, Codex reviews, a…
OpenAI released GPT-6.1 Sol, a production-grade model built for agentic workloads including continuous tool use, complex coding, computer use, and professional workflows. OpenAI said GPT-6.1 Sol approaches the capabiliti…
A Graphite study analyzing 90,000 AI-generated articles against 10,000 human-written texts found that less than half of the 12,877 distinct linguistic "tells" it identified in one GPT version, such as GPT-5.6 Sol, carry …
A developer used an AI production team of 12 role-based agents to script the third episode of a satirical video series, then documented how client feedback pushed each draft from mild workplace humor into absurdist "madn…
OpenAI released GPT-6.1 Sol on September 29, 2026 at DevDay in San Francisco and Anthropic shipped Claude Sonnet 5.5 on September 28, each pricing near-flagship AI at $2 per million input tokens and $10 per million outpu…
Microsoft released GPT-6.1 Sol on 2026-09-29, a text, image and PDF model with a 1,050k-token context window and 128k maximum output, according to the provider announcement. No retirement date has been announced for the …
A pre-registered, five-model paired evaluation tested a specification frame for LLM-generated code on money arithmetic, time handling, retry safety, and access control tasks, according to the study's authors. The evaluat…
Microsoft made Fabric IQ generally available in Microsoft Copilot Chat and Cowork, announced by CEO Satya Nadella at FabCon and SQLCon in Barcelona. The release grounds Copilot answers in trusted enterprise business cont…
OpenAI launched ChatGPT dots, always-on agents running on GPT-6 Astra with their own cloud computer and browser, on 29 September 2026 at DevDay, initially available with ChatGPT Pro and Business Premium. OpenAI says dots…
NaiveAI's Naive-N0.5-Flash, a 309B-parameter Mixture-of-Experts model with 15.5B active parameters per token, requires FP8-capable NVIDIA GPUs and roughly 315GB of weights to run locally through Hugging Face Transformers…
Developer Scott Logic built a harness of extracted CSS "skills" that let Claude Haiku outperform Claude Opus 4.8 on pure-CSS hacking tasks, including a CSS-only calculator and a scrolling jump game, according to the blog…
Jane Street research intern Kavish Kondap built an event-level generative model of US equities market data using autoregressive diffusion, trained on four years of US equities data with each row containing timestamp, pri…
Asif Waliuddin's Faultline Pro grew from a four-day Kaggle hackathon entry posted December 12, 2025 to 4,887 counted tests by September 30, 2026, after the AI claim checker's Quick Start returned "Mock verification: supp…