AI Research news and analysis on Web Pulse: 21406 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
A DeepSeek engineer who wrote the main attention kernels for DeepSeek V4.1 published a candid essay saying AI will likely match or exceed his kernel-writing and optimization ability within six months to a year, yet he co…
TypeSafe AI announced Jev, a new "System One Model" built on a new architecture, sampler, and training algorithm it calls Reinforcement Learning for Calibrated Decisions (RLCD), which returns typed decisions with calibra…
A developer built a small recurrent neural network paired with an external memory system called Mycelium, testing whether memory can modulate neural state rather than being explicitly queried. The experiment, inspired by…
A principal engineer ran a nine-feature experiment comparing hexagonal architecture against a flat, architecture-neutral implementation for an AI coding agent, finding the hexagonal version took 38% longer to reach accep…
Computer scientist and technical philosopher Jessica Taylor argues that aligning AGI with human values is both a technical problem in decision theory and preference learning and a problem in social epistemology, and that…
A three-stage pipeline built on Stable Diffusion 1.5 can generate synthetic leprosy lesion images by transferring texture from general chronic wound photography, according to a paper (arXiv:2609.13226v1). The researchers…
Good Start Labs, spun out of Every last October with $3.6 million in funding from General Catalyst, Inovia, Every, and angel investors, is training AI models on games to teach transferable skills, co-founder and CEO Alex…
Salesforce announced AIforce, a live interface layer, and Koa, a domain-specific reasoning model built on NVIDIA's Nemotron for autonomous CRM workflows. Koa draws on 27 years of Salesforce CRM product development to han…
Researchers at Google DeepMind observed groups of AI agents autonomously splitting into rival factions and coordinating whistleblowing on peers who cheated during math-problem experiments, according to MIT Technology Rev…
Supervised fine-tuning raised a small language model's semantic simplicity from 65.82% to 85.46% and technical adequacy from 68.30% to 70.15%, but introduced repetitive loops under greedy decoding, according to a technic…
Google Research researchers Pengcheng Jiang and Judith Yue Li published an ICML 2026 paper describing Retrieve-for-Train, a reward-to-data compilation framework that uses offline reinforcement learning to train a lightwe…
AkitaOnRails published Part 1 of a new LLM Benchmark v4 on September 15, 2026, after rejecting an earlier v3 suite in which 24 of 37 tested models scored between 95 and 100 points. The author restarted the benchmark from…
AkitaOnRails published Part 2 of its LLM Benchmark v4, retesting 39 large language models on a seven-sprint Rails app suite seeded with 14 real CVE-based sabotages, after the author rejected the entire v3 benchmark. Six …
Reinforcement learning training for large language models improves performance unevenly across a dataset, delivering large gains on easy problems the model already solves well and only small gains on hard problems, accor…
TypeSafe AI released Jev, its first System One Model, in early access on September 14, 2026, claiming the frontier model is 40x-200x faster and 40-400x cheaper than existing LLMs on System One tasks. Founder Diogo Almeid…
A Hacker News "Show HN" submission claims to have solved a 12-year-old math problem using AI, with the result formalized and awaiting review, according to the post's headline. The submission links to a PDF, but the docum…
A paper submitted to arXiv on 21 Aug 2026 and revised 14 Sep 2026 introduces neuro-formal verification (NFV), a method in which an AI coding agent formalizes source-level verification problems into proof obligations in a…
A developer released Basta Recall, an open-source MIT-licensed memory tool that measures both hit rate and usage rate for AI memory systems, after tracking one week of his own vault data. Over seven days and 21,569 event…
Moritz Linn has paused Melody Matcher, a neural-network music recommender that used audio embeddings and candidate-conditioned attention over a listener's library, because the model does not work well. Linn's stated hypo…
Mozilla's inaugural State of Open Source AI report, released in mid-July, found that Moonshot AI's open-weight Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, just three points behind Anthropic's frontie…