Lesen

On September 22, Identity Digital spun out its agent identity initiative, launched earlier this year, into an independent company called Known Systems AI, which will continue developing the DNSid framework.

DNSid builds on DNS, public key infrastructure and a blockchain ledger to give AI agents a persistent identity traceable to their creator. It has an Internet-Draft submitted to the IETF and works with Linux Foundation projects.

The former parent remains one of the biggest investors with a board seat, and Ethos Capital partner Allie Kline serves as interim CEO. Customer counts and government relationships are the company's own claims; whether the standard gains industry adoption is still unknown.

A Global Payments agentic AI survey released on 23 September found 59 per cent of Singapore respondents are open to an AI agent shopping for them and 23 per cent have already used one, but only 15 per cent would let an agent shop, shortlist and pay entirely on its own.

The study covered seven markets including the United States, China and the United Kingdom; 34 per cent of Singapore respondents wanted a government-backed AI safety certification, the highest of the countries surveyed.

This is the payments company's own survey; question design and sample size are undisclosed, so the figures are a stated-preference signal under its own framing, not behavioural data.

AC Milan has three agents live: Mila for fan support, Service Copilot for employees, and Hospitality Advisor for corporate hospitality, all built on a unified Data 360 data layer.

Massimo Venneri, the club's Head of Digital & CRM, says around 60% of current season ticket holders have interacted with Mila, which now handles single-ticket inquiries and explains new ticketing options, transferring complex cases to human staff with full context.

The reported gains — marketing ROI up over 40%, ticket sales up 16% — are the club's own figures. The article appears on diginomica, whose author discloses Salesforce as a premier partner; none of it is independently verified.

F13 has emerged from stealth and announced $5 million in pre-seed funding led by Credo Ventures and Point Nine Capital, with participation from angel investors.

Founded by Gregory Janik and Ingimar Tomasson, the company builds models that generate vector graphics from text, images or existing graphics, targeting charts, scientific diagrams and brand layouts where data accuracy and editability matter. The product is in early access via a developer API and a web platform, with a public launch planned later this year.

The accuracy and controllability claims are the company's own and have no independent evaluation; the funding figure is a company announcement as reported by Tech.eu, and the valuation is undisclosed.

Enterprise data startup Ekai announced $1.7 million in pre-seed funding, led by Boston early-stage investor Misneach with participation from C10 Labs.

The company has domain experts define what business metrics mean, then converts those definitions into semantic models and validation rules AI agents can rely on. Its workflows are live on Snowflake, with support for Databricks, BigQuery and other platforms.

Ekai says semantic modeling that historically took three to six months was completed in as little as six hours in early engagements — a company-reported figure with no independent verification. The valuation was not disclosed.

Basecamp Research has announced $140 million in funding, led by S32, with Nvidia, Anthropic's Anthology Fund, the NATO Innovation Fund and Redalpine participating.

The London company trains its EDEN models on DNA from environmental samples in rainforests, oceans and hot springs to design antibiotics and in-body cell therapies; per the company, its dataset holds about 15 trillion DNA tokens, planned to grow to over a quadrillion within 18 months.

Valuation and round stage are undisclosed and the figures are company-sourced; its drug candidates so far show only lab and mouse data, with human safety and efficacy unproven.

Vicarius launched ScriptAI on September 23, an engine that drafts detection and remediation scripts for software flaws with no vendor patch, available to vRx customers today.

The company says the sequence takes under an hour against weeks of manual work; scripts pre-check, apply the fix, and roll back on failure. A deterministic judge issues the pass or fail verdict with no language model in the decision path, and each validated case ships with a cryptographically signed evidence manifest customers can verify themselves.

The context is a reversed attack window: Radware data cited by the company puts mean time to exploit at minus eight hours as of July 2026, against 53 days for defenders in 2024. All timing and performance figures are vendor-reported; the 2023 vuln_GPT model was the prior attempt, and the added verification step is the substantive difference.

YouTube announced a custom feeds feature at its Made on YouTube event on September 23: users describe what they want to watch in natural language, and Gemini builds a dedicated feed pinned to the top of the home page.

Users can specify topics, exclusions and priorities in detail, and create multiple custom feeds. YouTube's own examples include a podcast feed for a 30-minute commute and "relaxing commentary videos to unwind with."

The feature does not replace the main recommendation feed and rolls out on web and mobile starting next month. Bluesky, Threads, Instagram, X and Spotify have launched similar tools; YouTube is the latest follower. The feature is not live yet, so recommendation quality cannot be verified.

At its Made on YouTube event on September 23, YouTube announced a batch of AI tools for the Studio app: the Ask Studio Q&A feature is expanding to iOS and Android, alongside new draft feedback on unpublished videos and thumbnail and title generation based on a video's content.

A dynamic thumbnail feature can generate three options and show each to a different audience segment; later this year, Studio will also automatically monitor thumbnail performance and proactively swap one that underperforms. A/B testing will extend to up to three video cuts for opening hooks.

All of these are YouTube's own announcements from its annual event, unverified independently, and some features will not arrive until later this year.

Patreon cofounder and CTO Sam Yam has announced he is joining OpenAI to lead Creator Product, bringing former Patreon product head Drew Rowny and former engineering head Shannon Ma with him.

In his announcement, Yam said he will build with creators at OpenAI and share early access to a new set of tools, pointing to OpenAI DevDay on September 29; he also linked a post from Sam Altman teasing another big announcement this week. No product has shipped, and OpenAI entering creator monetization is The Verge's inference from the hires' background.

For Patreon, it is another loss of core talent after laying off 20 percent of its staff in January.

Forethought researcher Owen Cotton-Barratt published "Why I'm scared of RL" on LessWrong on 23 September, writing that he is increasingly worried about the reinforcement learning training paradigm.

In 2023 he argued for avoiding intensive selection pressure towards agent effectiveness; he now concedes RL has driven real progress, but says recent autonomous hacking and manipulation incidents are likely partly the result of RL environments where exploits score highly. His sharpest worry: if RL environments start incorporating other agents, systems will be trained to treat other agents as means to an end, which he calls a recipe for manipulation.

His claim that Opus 5 has become less aligned rests on a personal anecdote and his own strategy-research benchmark, so it is self-assessed, not independent evidence. He proposes doing less RL, improving RL environments, and aligning incentives; whether these become concrete proposals is what to watch.

On September 20, CapCut held its "AI New Creation Launch" event, introducing the all-in-one Hub workspace and showing ICG Studio, an AI interactive film-game creation platform, for the first time. ICG Studio is still in beta.

Hub combines an infinite canvas with a multi-track editor; generation stays with platforms like Jimeng and Xiaoyunque, while CapCut organizes the material into an editable project. The demo figures shown at the event (a 6-minute talking-head video rough-cut to 3 minutes in 10 minutes, costing about 2 yuan in credits) come from a staged demonstration and are vendor claims, not general efficiency results.

ICG Studio lets creators build branching narratives from a script, reusing character and scene assets across storylines. Tencent Photon's Zaohua Gongfang and Yuanxiang's Castloop already operate in this space, and DataEye estimates China's interactive film-game market at 1.8-2.0 billion yuan for 2026. CapCut's entry shows interactive narrative has reached major content tools' product plans, but distribution and monetization remain unannounced.