Daily selection of the latest in AI: research, tools, papers and practical tips.
We read it. We curate it. Model releases, tools, papers, and hands-on tutorials — filtered so you don't have to scroll through the noise.
2026-09-13
A tutorial for building a memory-driven "self model" agent that scores 90.9%, compared with 82.8% for an agentic-RAG baseline, while changed-fact tracking improves from 60% to 100%.
2026-09-13
A hands-on guide fine-tunes a 350M model on a free 16GB GPU, improving JSON format compliance from 18.0% to 31.9% in 100 GRPO steps.
2026-09-13
Anthropic's Threat Intelligence team identified and disrupted multiple operations that used Claude for malicious activity between December 2025 and August 2026. The report presents case studies and tracks how misuse has evolved, noting that none involved Claude Fable or Mythos-class models.
2026-09-13
WearableQA provides 4,084 ten-option questions built from wearable time series, blood biomarkers, and demographics from 200 real users. The benchmark tests models on longitudinal data analysis and physiological reasoning, with raw structured data and customizable renderings included.
2026-09-13
A new framework reconstructs reusable coding environments from existing terminal-agent trajectories instead of creating them from zero. It restores workspaces, fills missing files, and generates original, cross-project, and multi-round tasks. From public trajectories, it builds 37.3k environments and improves coding-agent performance on two benchmarks.
2026-09-12
OpenAI deployed 10,000 collaborating agents for 88 hours to produce a Lean-verified proof of a finite-time singularity in the 3D Navier-Stokes equations, shortly after an independent competing announcement.
2026-09-12
AlphaGenome Atlas scores all 9 billion possible single-letter DNA variants in the human genome and is available free for noncommercial research.
2026-09-12
A multilingual benchmark and leaderboard evaluates retrieval models across 190 million documents and 69,721 queries. Independent relevance judgments and cost-efficient subsampling improve evaluation reliability while preserving model rankings.
2026-09-12
More sample-efficient learning algorithms could reduce AI's dependence on vast datasets. Even if data quality proves difficult to improve beyond human levels, data constraints may not prevent AI from automating most economic work.
2026-09-12
An analysis of 685,000 GitHub pull requests traces Claude-popularized vocabulary and presents VocabGuard, a drift-monitoring system with 0.869 AUC.
2026-09-12
MCP Toolbox for Databases uses protocol-level controls to enforce read-only access, avoiding prompt guardrails, regex parsers, and session flags that can fail under injection.
2026-09-12
Analysis of time-to-first-token latency up to million-token contexts finds quadratic scaling for GPT-5.6 and nearly linear scaling for Claude 5.
2026-09-12
A benchmark shows llm-d serving a 753B-parameter mixture-of-experts model across 544 H100 GPUs for 3,000 concurrent coding agents, with 85.2% of input tokens served from cache.
2026-09-12
OpenCodeReview is a CLI tool combining deterministic pipelines with an LLM agent to review Git diffs and entire files. It delivers line-level comments, supports multilingual rules for issues such as NPE, thread safety, XSS, and SQL injection, and works with OpenAI- and Anthropic-compatible providers.
2026-09-12
A study of nine LLMs finds that agent success drops geometrically as the number of dependent steps grows. Even strong models can fall from near-perfect to near-zero success within 16 steps, mainly because of step-level errors rather than context length.
2026-09-12
AI reasoning models slow down on harder problems because their internal process can enter transient chaotic states. They become trapped near nearly correct solutions, creating fractal patterns in how they search across tasks such as Sudoku, mazes, visual puzzles, and mathematical logic.
2026-09-11
Mercury 2.5 delivers 1,107 tokens per second on widely available Nvidia GPUs, a 260K-token context window, and performance comparable to cost-optimized frontier models. Launch pricing starts at $0.04 per million input tokens and $0.15 per million output tokens, with an 80% introductory discount.
2026-09-11
AlphaGenome Atlas is a 1-petabyte database predicting the regulatory effects of all 9 billion possible single-nucleotide variants in the human genome.
2026-09-11
North Small Translate is an open-weights research model with 25 billion active parameters and 218 billion total parameters, specialized for high-quality machine translation across 50 languages.
2026-09-11
From 2019 to 2025, data improvements delivered 3.24 times more compute-efficiency gains than model improvements. The gains were largely independent, while model research mainly removed scaling constraints; data quality matters most for smaller models and may matter less at larger scales.
2026-09-11
A benchmark puts developer agents in simulated business environments to recover requirements, design architectures, build tools, and deploy a customer-service agent within a model-cost budget. Claude Opus 5 scores 23.9% alone versus 82.2% when paired with an experienced engineer.
2026-09-11
A fully fledged serving system combines continuous batching, paged attention, ragged sequence lengths, tool calling, and an OpenAI-compatible endpoint. The megakernel reaches 292 tokens per second at batch size 1, outperforming vLLM by 1.58x across batch sizes and up to 256K-token contexts without measurable accuracy loss.
2026-09-11
GPT-6 Astra makes a major leap in computer use. A variant of the looped transformer may improve performance at a fixed compute budget, while shorter reasoning traces may reflect fewer mistakes and more internal computation.
2026-09-11
OpenTelemetry's GenAI semantic conventions standardize traces for agent invocations, LLM calls, and tool executions, including token-usage and duration metrics.
2026-09-11
Four misconfigured cybersecurity evaluations gave Claude access to real systems, according to Anthropic's assessment.
2026-09-11
A pure Go, zero-dependency memory layer that stores AI agent knowledge as version-controlled Markdown and YAML in a repository. It provides OKF v0.2 support, fast local BM25 search, progressive disclosure, validation, CLI tools, and an embedded MCP server without external databases.
2026-09-11
A new reinforcement learning method gives language models dense rewards for matching progress across task segments, making learning more efficient on long and complex tasks. Tests on difficult math problems show that segment-level rewards improve success rates when models use larger test-time token budgets.
2026-09-11
A benchmark of 10 LLMs and 26 embedding models across 37 tasks finds similar overall quality. LLMs perform better on reasoning-heavy retrieval, while embedding models are stronger for classification and much cheaper and faster.