https://highsignal.sh/stories/google-deepmind-chief-says-gemini-4-is-in-refinement-aiming-to-launch-much-earli-66dcdd26 High Signal en 2026-09-24T13:47:30.112Z Google DeepMind chief says Gemini 4 is in refinement, aiming to launch 'much earlier' than year-end https://highsignal.sh/stories/arxiv-paper-introduces-webmre-an-offline-benchmark-for-multimodal-web-agents-6c94f0c5 High Signal en 2026-09-24T13:45:30.804Z arXiv paper introduces WebMRE, an offline benchmark for multimodal web agents https://highsignal.sh/stories/ruby-asr-paper-targets-japanese-reading-ambiguity-with-a-span-bound-transcript-f-14ae493d High Signal en 2026-09-24T07:50:17.401Z Ruby-ASR paper targets Japanese reading ambiguity with a span-bound transcript format https://highsignal.sh/stories/arxiv-study-multi-task-fine-tuning-beats-specialists-for-customer-support-llms-139b86d8 High Signal en 2026-09-24T07:50:00.562Z ArXiv study: Multi-task fine-tuning beats specialists for customer-support LLMs https://highsignal.sh/stories/unidataagent-china-unicom-s-ontology-grounded-enterprise-q-a-agent-cuts-report-t-0ee40ca1 High Signal en 2026-09-24T07:49:53.458Z UniDataAgent: China Unicom's Ontology-Grounded Enterprise Q&A Agent Cuts Report Time from Days to Minutes https://highsignal.sh/stories/arxiv-paper-cuts-transformer-sequence-length-40-with-span-collapsing-merge-modul-9154f86f High Signal en 2026-09-24T07:49:28.780Z ArXiv paper cuts Transformer sequence length 40% with span-collapsing merge module https://highsignal.sh/stories/lockr-targets-stable-but-wrong-lock-in-in-diffusion-language-models-27e19e96 High Signal en 2026-09-24T07:49:13.194Z LOCKR targets 'stable-but-wrong lock-in' in diffusion language models https://highsignal.sh/stories/arxiv-preprint-introduces-ugtphon-a-benchmark-for-pronouncing-slang-and-abbrevia-15207607 High Signal en 2026-09-24T07:49:07.725Z ArXiv Preprint Introduces UGTPhon, a Benchmark for Pronouncing Slang and Abbreviations https://highsignal.sh/stories/arxiv-preprint-proposes-r3con-harness-to-cut-reliance-on-model-scale-65c0d4c7 High Signal en 2026-09-24T07:48:58.264Z ArXiv preprint proposes 'R3Con' harness to cut reliance on model scale https://highsignal.sh/stories/arxiv-paper-proposes-tempered-evidence-fusion-for-scoring-values-in-long-posts-7800a5ed High Signal en 2026-09-24T07:48:52.919Z arXiv Paper Proposes Tempered Evidence Fusion for Scoring Values in Long Posts https://highsignal.sh/stories/recap-cuts-reasoning-tokens-by-up-to-31-in-arxiv-preprint-39321529 High Signal en 2026-09-24T07:48:42.198Z RECAP Cuts Reasoning Tokens by Up to 31% in arXiv Preprint https://highsignal.sh/stories/fact-verification-scores-improve-mostly-because-the-evidence-does-not-the-answer-1fad211c High Signal en 2026-09-24T07:48:36.635Z Fact-Verification Scores Improve Mostly Because the Evidence Does, Not the Answers https://highsignal.sh/stories/arxiv-paper-proposes-auditable-llm-labeling-for-classroom-talk-adbac51c High Signal en 2026-09-24T07:48:30.832Z ArXiv paper proposes auditable LLM labeling for classroom talk https://highsignal.sh/stories/kinship-term-benchmark-finds-llms-recognize-cultural-terms-better-than-they-gene-b747a151 High Signal en 2026-09-24T07:48:24.662Z Kinship-term benchmark finds LLMs recognize cultural terms better than they generate them https://highsignal.sh/stories/arxiv-paper-measures-how-much-benchmark-contamination-actually-inflates-scores-3f693840 High Signal en 2026-09-24T07:48:18.883Z ArXiv paper measures how much benchmark contamination actually inflates scores https://highsignal.sh/stories/comed-preprint-cuts-multi-llm-collaboration-to-only-when-it-helps-2db66a53 High Signal en 2026-09-24T07:48:13.218Z COMED preprint cuts multi-LLM collaboration to only when it helps https://highsignal.sh/stories/google-s-gemini-3-8-tts-models-are-cheap-and-handle-multi-voice-dialogue-per-sim-8e3e7165 High Signal en 2026-09-24T07:47:39.172Z Google's Gemini 3.8 TTS models are cheap and handle multi-voice dialogue, per Simon Willison https://highsignal.sh/stories/meta-connect-2026-what-to-expect-from-zuckerberg-s-ai-and-wearables-keynote-6f3b6b0f High Signal en 2026-09-24T05:35:22.483Z Meta Connect 2026: What to expect from Zuckerberg's AI and wearables keynote https://highsignal.sh/stories/arxiv-paper-ask-coding-agents-if-a-stored-claim-still-holds-not-if-a-diff-preser-7beea10b High Signal en 2026-09-24T05:35:09.940Z arXiv paper: ask coding agents if a stored claim still holds, not if a diff preserves behavior https://highsignal.sh/stories/arxiv-paper-casts-jotto-word-game-as-graph-pruning-problem-82900c13 High Signal en 2026-09-24T05:34:58.061Z arXiv Paper Casts Jotto Word Game as Graph Pruning Problem https://highsignal.sh/stories/openai-releases-mentalhealthbench-an-expert-informed-benchmark-for-ai-responses-cd591887 High Signal en 2026-09-24T05:34:36.549Z OpenAI releases MentalHealthBench, an expert-informed benchmark for AI responses in mental health conversations https://highsignal.sh/stories/ringg-says-gpt-5-6-agents-handle-up-to-65-of-customer-calls-at-90-less-cost-than-69d14549 High Signal en 2026-09-24T05:34:19.819Z Ringg says GPT-5.6 agents handle up to 65% of customer calls at 90% less cost than GPT-4.1 https://highsignal.sh/stories/openai-says-harvey-uses-gpt-6-astra-to-produce-more-structured-legal-drafts-5f6c8b52 High Signal en 2026-09-24T05:34:10.555Z OpenAI says Harvey uses GPT-6 Astra to produce more structured legal drafts https://highsignal.sh/stories/fineweb-gets-culture-language-and-region-labels-to-audit-what-pretraining-data-m-1f534de8 High Signal en 2026-09-24T05:33:47.004Z FineWeb gets culture, language and region labels to audit what pretraining data misses https://highsignal.sh/stories/turkish-small-model-paper-logs-a-safety-gate-that-lied-to-its-own-authors-c9260f3a High Signal en 2026-09-24T05:32:49.648Z Turkish small-model paper logs a safety gate that lied to its own authors https://highsignal.sh/stories/arxiv-paper-applies-llms-to-encrypted-map-fusion-without-training-37573012 High Signal en 2026-09-24T05:31:42.646Z arXiv paper applies LLMs to encrypted map fusion without training https://highsignal.sh/stories/arxiv-preprint-proposes-fuzzy-logic-prediction-head-to-fix-llm-ordinal-reasoning-736832f6 High Signal en 2026-09-24T05:31:31.124Z ArXiv preprint proposes fuzzy-logic prediction head to fix LLM ordinal reasoning https://highsignal.sh/stories/frontiermath-erd-s-benchmark-gpt-6-astra-solves-3-of-open-conjectures-all-others-6c5babf5 High Signal en 2026-09-24T05:31:24.527Z FrontierMath Erdős benchmark: GPT-6 Astra solves 3% of open conjectures, all others score zero https://highsignal.sh/stories/aibuildai-2-5-claims-73-3-medal-rate-at-top-of-mle-bench-01e73a0e High Signal en 2026-09-24T05:31:12.620Z AIBuildAI-2.5 claims 73.3% medal rate at top of MLE-Bench https://highsignal.sh/stories/new-arxiv-paper-fixes-cross-gpu-llm-nondeterminism-with-fused-upcast-gemm-kernel-f67c00ce High Signal en 2026-09-24T05:31:04.767Z New arXiv paper fixes cross-GPU LLM nondeterminism with fused-upcast GEMM kernels https://highsignal.sh/stories/peerify-benchmark-tests-whether-ai-can-check-reviewers-claims-against-manuscript-8b9c3b9d High Signal en 2026-09-24T05:30:59.906Z Peerify benchmark tests whether AI can check reviewers' claims against manuscripts https://highsignal.sh/stories/arxiv-study-finds-residualization-can-decorrelate-llm-eval-scores-without-fixing-a3ff5f36 High Signal en 2026-09-24T05:30:54.996Z arXiv study finds residualization can decorrelate LLM eval scores without fixing them https://highsignal.sh/stories/sentiment-arcs-in-central-bank-press-conferences-predict-rate-decisions-arxiv-st-14977556 High Signal en 2026-09-24T05:30:48.879Z Sentiment 'arcs' in central bank press conferences predict rate decisions, arXiv study claims https://highsignal.sh/stories/arxiv-paper-proposes-using-frozen-numerical-foundation-models-to-guide-few-shot-34c95259 High Signal en 2026-09-24T05:30:44.620Z ArXiv paper proposes using frozen numerical foundation models to guide few-shot multitask optimization https://highsignal.sh/stories/audit-99-fake-news-detector-accuracy-mostly-reflects-source-not-veracity-45acde1d High Signal en 2026-09-24T05:30:27.618Z Audit: 99% fake-news detector accuracy mostly reflects source, not veracity https://highsignal.sh/stories/arxiv-preprint-applies-graph-reinforcement-learning-to-one-dimensional-bin-packi-9b79ed34 High Signal en 2026-09-24T05:30:20.655Z arXiv preprint applies graph reinforcement learning to one-dimensional bin packing https://highsignal.sh/stories/aws-publishes-guide-to-running-opencode-coding-agent-on-bedrock-open-weight-mode-c20b62dd High Signal en 2026-09-23T23:07:38.420Z AWS Publishes Guide to Running OpenCode Coding Agent on Bedrock Open Weight Models https://highsignal.sh/stories/simon-willison-releases-llm-keys-ui-0-1-to-save-api-keys-on-remote-agent-machine-9e5eccf2 High Signal en 2026-09-23T16:27:43.893Z Simon Willison releases llm-keys-ui 0.1 to save API keys on remote agent machines https://highsignal.sh/stories/simon-willison-pushes-back-on-mcp-was-always-a-bad-idea-hacker-news-post-6fd955f7 High Signal en 2026-09-23T16:27:38.426Z Simon Willison pushes back on 'MCP was always a bad idea' Hacker News post https://highsignal.sh/stories/llm-cli-plugin-llm-anthropic-0-29-adds-support-for-claude-opus-5-5-7ac2f1b3 High Signal en 2026-09-23T16:27:36.167Z llm CLI plugin llm-anthropic 0.29 adds support for Claude Opus 5.5 https://highsignal.sh/stories/engineer-describes-workplace-where-everything-is-made-by-claude-code-869f5db7 High Signal en 2026-09-23T16:26:28.860Z Engineer Describes Workplace Where 'Everything Is Made by Claude Code' https://highsignal.sh/stories/cloudflare-makes-python-workers-generally-available-after-two-year-preview-2f5c873e High Signal en 2026-09-23T16:22:13.961Z Cloudflare makes Python Workers generally available after two-year preview https://highsignal.sh/stories/techcrunch-ai-reports-on-two-years-of-ai-data-center-fights-in-pennsylvania-a3264e51 High Signal en 2026-09-23T16:21:14.850Z TechCrunch AI reports on two years of AI data center fights in Pennsylvania https://highsignal.sh/stories/simon-willison-quotes-tiktok-creator-on-spotting-ai-written-scripts-35102f40 High Signal en 2026-09-23T16:15:54.524Z Simon Willison Quotes TikTok Creator on Spotting AI-Written Scripts https://highsignal.sh/stories/simon-willison-and-jesse-vincent-to-host-san-francisco-agentic-engineering-meetu-dc2b9ac6 High Signal en 2026-09-23T16:14:29.764Z Simon Willison and Jesse Vincent to host San Francisco agentic engineering meetup on October 14 https://highsignal.sh/stories/youtube-adds-ai-creator-tools-for-more-complex-tasks-at-made-on-youtube-30bca938 High Signal en 2026-09-23T15:30:24.884Z YouTube adds AI creator tools for more complex tasks at Made on YouTube https://highsignal.sh/stories/v7-uses-gpt-5-6-to-give-ai-agents-institutional-memory-e70257a0 High Signal en 2026-09-23T02:14:54.847Z V7 Uses GPT-5.6 to Give AI Agents Institutional Memory https://highsignal.sh/stories/openai-expands-academy-with-new-ai-learning-paths-f84eb834 High Signal en 2026-09-23T02:14:49.179Z OpenAI marks two years of OpenAI Academy https://highsignal.sh/stories/openai-calls-for-shared-global-standards-for-ai-safety-8a8da546 High Signal en 2026-09-23T02:14:42.315Z OpenAI Calls for Shared Global Standards for AI Safety https://highsignal.sh/stories/higgsfield-ai-launches-rapid-video-features-for-small-businesses-using-gpt-6-ast-58d90bf8 High Signal en 2026-09-23T02:14:36.126Z Higgsfield AI Launches Rapid Video Features for Small Businesses Using GPT-6 Astra https://highsignal.sh/stories/openai-engages-independent-advisory-group-on-mathematics-and-ai-8ab4cc55 High Signal en 2026-09-23T02:14:29.698Z OpenAI Engages Independent Advisory Group on Mathematics and AI https://highsignal.sh/stories/openai-publishes-principles-for-third-party-ai-safety-assessments-48ec9ce0 High Signal en 2026-09-23T02:14:23.651Z OpenAI Publishes Principles for Third-Party AI Safety Assessments https://highsignal.sh/stories/parallel-cuts-labor-market-research-time-and-cost-by-half-with-gpt-6-astra-799e2fbd High Signal en 2026-09-23T02:14:17.728Z Parallel Cuts Labor Market Research Time and Cost by Half with GPT-6 Astra https://highsignal.sh/stories/openai-introduces-gpt-6-sol-and-luna-models-bf7377f1 High Signal en 2026-09-23T02:13:48.351Z OpenAI introduces GPT-6 Sol and Luna models https://highsignal.sh/stories/openai-claims-gpt-6-has-better-prompt-caching-and-diagnostics-5d05f7d6 High Signal en 2026-09-23T02:13:29.244Z OpenAI claims GPT-6 has better prompt caching and diagnostics