AI Research

Cognition released SWE-2, a coding model post-trained with reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3, scoring 50.0% on FrontierCode 1.1 Main — within 1 point of Fable 5.1 at 64% lower cost. SWE-2 h…

AgentsDock released version 0.2.12, a macOS 14+ universal IDE for agentic AI research that supports Claude Code, Codex, and Cursor in one desktop and mobile workspace. The tool connects remote workstations, Mac minis, an…

The open-source project local-deep-research reported approximately 95% on SimpleQA (n=500) and 77% on xbench-DeepSearch (n=100), which it says makes it the first fully-local open-source project to hit those marks on a si…

Anthropic CEO Dario Amodei confirmed in a Saturday evening essay, "We Must Pace the Frontier," that recursive self-improvement is starting to happen across the industry, including at Anthropic. Within roughly three hours…

A controlled test using a DDPM U-Net on CIFAR-10 found that LoRA rank 4 achieved the best FID score of 124.1380, slightly beating rank 8 at 124.2136, according to data from arXiv:2609.10656v1. The study, which tested ran…

A state-of-the-art review posted to arXiv on 20 Aug 2026 (Michels et al., arXiv:2608.20446) found that team-level telemetry shows code-review time rose 441% under "vibe coding," a workflow where developers describe inten…

A 2026 research survey by Duan and colleagues examines recursive self-improvement (RSI) in AI, the process by which a system improves itself or builds a better successor and uses those gains to drive further progress. Th…

Google Quantum AI's March paper found about 6.9 million Bitcoin, roughly a third of the circulating supply, sitting in addresses whose full public keys are already visible on-chain and thus exposed to an at-rest quantum …

A developer argues that recursive self-improvement in AI agents can occur without changing model weights, through persistent improvements to agent software, tools, memory, and planning. Citing Darwin Gödel Machine experi…

Anthropic's Claude Mythos Preview, announced 7 April 2026, autonomously found zero-day vulnerabilities in every major operating system and web browser, and by 26 August its disclosure programme had reported 2,300 vulnera…

Specific Labs co-founders Janak Sunil and Siddhant Paliwal launched Real-SWE, a coding-agent benchmark built on licensed private production codebases, where Fable 5.1 running through Claude Code led the published leaderb…

Former Anthropic researcher Jacob Coxon published a series of posts warning that companies developing advanced AI are racing toward increasingly powerful, potentially self-improving systems without adequately solving saf…

Real-SWE, a new benchmark evaluating frontier AI models on private, real-world enterprise codebases licensed from real companies, was released with Fable 5.1 running in Claude Code posting the top resolution rate of 38.8…

A developer released Fly Arena, an open-source simulation that combines a NeuroMechFly body in MuJoCo with MaleCNS v1.0 connectome data and explicit behavioral controllers to model a fruit fly. The hybrid mode supports w…

Two high school students, Aayush Bathija and Prince Rohatgi, working with postdoc Daniel Soskin through the UCLA Math Circle, solved a mathematics problem that Fields Medalist June Huh had previously worked on without so…

Twenty-five Fields Medallists, including Terence Tao, signed a declaration stating that AI companies' push to solve mathematical problems as benchmarks is detrimental to mathematics as a science, according to a post on T…

ModelRift benchmarked six AI agents, all driven by Claude Opus 5 (1M context) through Claude Code, on three printable-part modeling tasks in CadQuery 2.8.0 and OpenSCAD 2026.06.12, and found all six resulting parts were …

AgentRuleBench, an open-source benchmark released alongside the archprint architecture-rule mining tool, found that AI coding agents did not violate an inferred architecture rule across a pilot spanning three vendors' ag…

Shrijith Venkatramana, an engineer building the AI code review tool LiveReview, traces modern LLM tokenization back to Byte Pair Encoding, a 1994 data-compression algorithm by Philip Gage. The writeup explains how BPE-st…

Researchers Ilyas, Santurkar, Tsipras, Engstrom, Tran and Madry published a 2019 paper arguing that adversarial examples are not bugs but features of machine learning models, challenging the prevailing view that such per…