Daily merges at Anthropic grew from roughly 2,500 in May to over 30,000 in September as AI agents scaled coding, forcing the lab to build an in-house source control system and migrate its monorepo off Git in five months.
The Shenzhen-based maker of cable-driven humanoids and Lumo embodied AI models closed its fourth round in six months, led by Shenzhen Investment Holdings-linked capital with Hongtai Fund and Shanheng Capital participating.
Jensen Huang said in Paris that Nvidia will unveil $2 billion in European AI investments within a week, on top of roughly $8 billion already deployed across the continent over the past year.
Evo2 treats harness edits and their outcomes as training data (SFT plus interleaved GRPO), lifting SWE-bench Verified to 48.3% from a 41% baseline and raising a GPT-5-mini agent from 24.4% to 48.3% with the evolved harness alone.
The paper proves that token-wise importance weighting in off-policy RL with verifiable rewards is generally biased, and introduces a proximal KL regularizer that provably corrects it — enabling safe off-policy RLVR from asynchronous generation and response reuse.
SEAR is a framework and benchmark where agents write robot control code, test it, and keep what works as skills and memories — probing whether the agent genuinely learns versus merely appearing to improve.
A training-free decoding method that adaptively identifies when a model's intermediate reasoning is sufficient, cutting overlong reasoning without extra training, steering vectors or decoding heuristics.
The architecture bridges latent reasoning's efficiency with the native tool access and transparency of text-based LLM agents, reasoning in a latent space while acting and using tools through text.
The open-weight text-to-speech model offers per-word tone and emotion control, mid-sentence voice shifts and plain-language delivery direction; demos have gone viral.
LongHarness keeps evidence and open conclusions as explicit state instead of chat history, jumping GLM-5.3-Flash from 4.4% to 61.1% exact match on OfficeQA Pro V2 and outperforming Claude Code and Codex variants.
The models handle 1M-token multimodal inputs for classification, extraction and summarization, reading input once and pointing to answers without token-by-token generation — claimed ~200x faster and 467x cheaper than frontier LLM APIs.
Following incidents with cloud agents leaking API keys into logs, Nous argues for local, self-hosted custom agents as the privacy-preserving alternative to closed agent stacks.
Gemini 4 Argon watched video, mapped the underlying physics, generated its own simulation and taught Apptronik's Apollo humanoid new tasks — Apollo sank 7 cornhole throws in a row. Google says the approach compresses weeks of manual modeling into hours.
A multi-year agreement brings Promise's AI-powered robotic Factory-as-a-Service platform to US housing production, supporting more than 1,250 homes for America's largest homebuilder by volume.
Max Hodak announced the PRIMA vision-restoring chip's first implant in a patient with Stargardt disease, extending the neural implant platform beyond its initial macular degeneration trials.
A Mac budgeting app whose AI models run entirely on the user's machine, launched on Show HN and quickly climbing to 98 points; currently in beta for Apple Silicon.
The benchmark quantifies AI-generated code bloat, a growing concern as coding agents proliferate; it reached 64 points on Show HN within a day.
A web crawler purpose-built for AI agents that deliberately excludes the LLM from the crawl loop, keeping large-scale crawling fast and cheap. Open source on GitHub.
A compact machine-learning runtime that runs in the browser via npm and flags silent logic bugs that never surface as exceptions.
An open-source gateway that lets developers control which tools an AI agent can invoke, aimed at production agent deployments needing fine-grained tool permissions.
The Israeli-US startup, emerging from stealth with $34M Series A led by Jump Capital plus an $8.5M seed, monitors how employees use AI tools, skills, MCP servers and homegrown agents — already at $2M ARR with 30+ enterprise customers including Oscar Health and Workato.
The Düsseldorf semiconductor startup's unified processor architecture for autonomous machines is demonstrated in its UB108 chip (256 processing elements, Samsung 8nm), with a 4,096-element UB410 targeted for 2027.
The California-Bengaluru startup uses AI agents to create and maintain product tutorials, training and marketing videos that stay accurate as software changes; 1,000+ customers including Atlassian and Zendesk.
The round, led by Left Lane Capital and First Round Capital, was announced alongside CortexLM, an AI system that reviews patient records for documentation errors and compliance risks before claims reach insurers.
The Beijing startup, founded in 2024, builds encrypted AI model services and an agent-development platform running on Huawei Ascend and Hygon platforms — its third round in six months.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
