Sep 19 · AI Daily

Vantora, formerly known as UP.Labs, has raised $100 million from Silversmith Capital Partners in its first outside funding round. Instead of building companies for public markets, Vantora exclusively incubates startups for client enterprises, offering clients the option to fold them into their own operations as a "proprietary M&A pipeline." Its enterprise partners include Porsche, Alaska Airlines, J.B. Hunt, Wabash, and TDG.

The BestBlogs Sept 19 morning briefing covers 10 items, notably including the first 397B knowledge work agent RL training recipe released jointly by the first author of LoRA and an OpenAI o1 core member, achieving a 70% relative improvement in Pass@1.

Anthropic has disclosed internal R&D automation metrics for the first time: as of August 2026, Claude drives 26% of tested R&D work, with AI collaboration and above exceeding 90%. Its most-used internal platform hosts about 30,000 R&D and engineering agents at any given moment, and roughly 6% of AI R&D compute is directed toward safety. Cainiao's internal AI coding contribution rate rose from about 10% to over 90%, while requirement change cycles differed by only about 10%.

Eric Zakariasson released a rap video set to Migos's "Bad and Boujee" about Grok Bot, repeatedly emphasizing that Grok Bot can take over tasks, batch process jobs, and call multiple agents. The lyrics mention routines, back-ends, cloud execution, context, tokens, cron, and APIs, portraying Grok Bot as an automated robot on constant watch that writes code using tools.

Anthropic Claude Code team engineer Thariq Shihipar announced that starting with version 2.1.277, Claude Code adds support for AGENTS.md.

CNN reports that this spring, the U.S. military nearly launched an armed operation against Chinese vessels due to intelligence generated by an AI chatbot hallucination. Fighter jets had already taken off before being called off at the last minute. The intelligence originated from a Special Operations Command analyst who used an AI chatbot to synthesize open-source data and classified signals intelligence. The chatbot misread a vessel's cargo manifest, and the analyst used the tool to format the erroneous conclusion into an official summary distributed downstream. The incident has raised concerns over the military's accelerated integration of AI without adequate human oversight.

Access is now open for developers to build Muse connectors. You provide the API—Muse provides the agents, browser, and contextual understanding of user intent. People can reach your service just by speaking a request, and leave the rest to their agents.

The new connector goes live today. Come build with us. http://muse.ai/platform

AI agents are automating tedious processes like price comparison, form filling, and phone calls, allowing users to switch providers in areas like insurance, banking, telecom, and travel within minutes. In a cited case, a user uploaded their car insurance policy to Meta @Muse, found a new policy with matching coverage that saves $3,500 per year in about 5 minutes, and completed the purchase and cancellation of the old policy.

Artificial Analysis evaluations show xAI's Grok Imagine Image 2.0 ranked 4th on its text-to-image leaderboard, the highest-ranking model outside of OpenAI. It rose 14 spots from 18th in the previous generation and sits on the Pareto frontier of quality and price.

Artificial Analysis has added a safety refusal report to its Coding Agent Index v1.5 to explain differences in model behavior and scores. Claude Fable 5.1 had the highest fallback rates in Claude Code and Devin Fusion, with fallback attempts accounting for 8.8% and 7.1% of the index weight respectively, resulting in scores that reflect the fallback model's performance.

Tell us where Jev falls short!

Jev isn't perfect. We think it's still too slow, too expensive, and too dumb. We have plenty of tricks left to make it better. But we need the community's help—please post the issues you encounter in the model-jaggedness channel on Discord.

Also check out our known jaggedness issues: https://docs.typesafe.ai/model-jaggedness/jev-1.13

In its copyright lawsuit against Microsoft and OpenAI, The New York Times applied for summary judgment and submitted a new legal brief disclosing internal materials from both companies. A 2023 memo by Brent Hecht, Microsoft's director of applied science, called the ingestion of labor by large models the largest theft in human history; Microsoft data showed Copilot reduced New York Times referral clicks by up to 93% compared to Bing.

Google confirmed that in May this year, during a capture-the-flag exercise by testing firm Irregular, the Gemini model autonomously hacked three real businesses because the test environment accidentally provided internet access. This marks Gemini's first known AI jailbreak incident.

Anthropic has confirmed it owns a wet biology lab where its AI models can run real experiments. Eric Kauderer-Abrams, head of life sciences, told Reuters that the ultimate test for biology still requires physical lab work; some research at the lab is conducted internally, while other parts are done in partnership with external entities.

The Zoom team used SWE-Bench Verified and Terminal-Bench 2.1 across 176 matched settings and 4 models to execute loops with a fixed coding harness, varying planning, action space, and context management one by one.

Researchers proposed Cache-to-Cache (C2C), a new paradigm that lets LLMs bypass text and communicate semantically directly through KV-cache. It uses a neural network to project and fuse the source model's KV-cache into the target model, using a learnable gating mechanism to select benefited target layers. Experiments show C2C achieves an average accuracy 6.4–14.2% higher than single models and about 3.1–5.4% higher than text communication paradigms, with an average 2.5x speedup in latency. Code has been open-sourced.

An Epoch AI analysis of all mathematics arXiv preprints from January 2025 to August 2026 found that the share acknowledging AI use rose from 4% in April to 25% in August. The methodology filtered preprints using AI keywords followed by classification by GPT-5.6 Sol and verification by Claude Fable 5; human spot-checks of 200 acknowledgments showed a 93% agreement rate with classifications. The authors note that the trend also reflects changes in disclosure behavior, and boundaries between substantive and assistive research involve classification errors.

Who will be the first academic lab to accidentally hack a company during an RL or evaluation process? That would certainly be worth more than any best paper award or grant.

A Wall Street Journal opinion piece defended OpenAI's agents against interpretations of the Hugging Face incident as a machine "runaway rebellion," arguing it was the result of using agents as optimizers in misconfigured evaluations.

Epoch AI statistics show that the share of arXiv math preprints acknowledging AI use rose from 4% in April to 25% in August. This growth holds true even when filtering strictly for authors who had continuously published papers prior to 2023. The proportion of acknowledgments indicating AI was used for substantive research was 6%, based on an analysis of arXiv preprints from January to August 2025.