Dakarda AI Newsletter · 7 September 2026
Monday Edition
Agents work 3.1 days for us — find out how
OpenAI for the first time reveals data from inside its research organization: agents already perform 3.1 "workdays" for every human day. On top of that, Nvidia declares AGI, Ant Group quietly releases a cheap agentic model, and publishers sue OpenAI to have models destroyed.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
This week started with a number worth remembering: 3.1. That's how many eight-hour "workdays" OpenAI agents complete for every human workday — the company is showing such data from inside its research team for the first time. On top of that, Jensen Huang declares that GPT-6 Astra has achieved AGI (critics have doubts, but the infrastructure numbers speak for themselves), and Ant Group quietly releases an open, cheap agentic model. In Brussels, meanwhile, no surprises: the enforcement of GPAI obligations continues, but the emotion has shifted to the USA. In this edition you'll also find the Seattle Times and Newsday lawsuit against OpenAI and Microsoft — the publishers want, among other things, the destruction of the models, which could change the economics of training LLMs — plus a technical novelty from Nvidia: free PAIR, which combines your GPUs into a home cluster. In the tech section, I look at how Claude Code ended up in the hands of cybercriminals, and what to do about the Magento zero-day. Finally, the tip of the day: how to measure the real work of agents in your own team.
What's worth knowing
OpenAI reveals: agents perform 3.1 days of work for every human day
In its "Research acceleration" report, OpenAI shares data from its internal research team: since mid-August, agents have consumed the equivalent of 3.1 eight-hour shifts for every human workday, and total agent runtime surpassed human time back in June. The company declares it will achieve the full "automated AI researcher" goal by March 2028. These are the first such concrete public figures on how much research work agents are already actually doing — and what inference costs: the median researcher runs over $600 per day.
Jensen Huang: GPT-6 Astra has reached AGI. Critics cool the enthusiasm
Nvidia's CEO wrote on X that GPT-6 Astra has "achieved AGI", stating that the model was trained on over 100,000 Grace Blackwell NVLink72 systems, with another 400,000 GPUs soon to come online. Earlier he had posted a message with a much larger number, which he deleted — and critics like Gary Marcus point to Nvidia's interest in promoting GPU spending. For developers, though, the numbers themselves matter: the scale of infrastructure that will shape compute availability.
Ant Group quietly releases LLaDA2.2-mini — a cheap agentic model under Apache 2.0
inclusionAI/LLaDA2.2-mini appeared on Hugging Face without any announcement — a 16-billion-parameter MoE model (1.4 billion active parameters per token) from the family of diffusion language models, trained primarily for agentic tool use rather than general chat. The Apache-2.0 license and 128K context make it an interesting, cheap base for your own deployments with full control. Diffusion LLMs are becoming an increasingly serious alternative to autoregressive models — also when it comes to inference costs.
NVIDIA PAIR turns your computers into a home AI cluster
Personal AI Router is a free public beta for Windows, macOS and Linux that links RTX GPU computers, DGX Spark systems and M4+ Macs into a local inference cluster. Apps and agents send queries to a single endpoint, which distributes them across available capacity — without sending prompts or files to the cloud. For agent builders, this is the first such simple route to cheaper, private inference powered by "sleeping" GPU capacity.
Seattle Times and Newsday sue OpenAI and Microsoft. They want models destroyed
The publishers accuse the companies of training ChatGPT, Copilot and Bing AI on tens of thousands of articles (including paywalled ones) without permission or compensation, and the lawsuit demands, among other things, an order to destroy the training datasets and the models that contain them. The case extends the dispute over trademark violations and false attributions of texts by AI. If the court grants the requests, a precedent will be set that changes the economics of training LLMs on publishers' data.
From the tech world
Claude Code in the hands of cybercriminals — an agent across the full attack chain
A Gambit Security report documents how a member of the Ransomware-as-a-Service group "The Gentlemen" used a Claude Code agent at every stage — from VPN attacks, through Active Directory enumeration, to database dumps. It's one of the first documented uses of AI agents in real attacks and debunks the myth that LLMs are only good for phishing.
Zero-day in Magento and Adobe Commerce — backdoors in shops without login
Attackers are exploiting an unpatched zero-day vulnerability to execute remote code and install backdoors in online stores. Sansec has published compromise indicators (including the X-TRACE- header) and recommends rotating crypt/key, admin passwords and API keys — hosting providers responded en masse to the incident on September 5.
Tip of the day
Measure agent-days in your own team
Before you decide whether agents are worth it, measure their real contribution the same way you measure human work. Define a task unit (code review, research, report preparation), estimate how long a human would take, and compare it with the inference cost you actually pay. For a week, log every task handed to an agent, the estimated human time, and the API cost (you'll find it in your provider's dashboard). After seven days, count how many "agent-days" your team reclaimed — and where the number becomes significant, integrate the agent deeper into your daily workflow instead of treating it as an experiment.
Tool of the issue
NVIDIA PAIR — a free AI router that turns your GPUs into a home cluster
Personal AI Router (public beta) links RTX GPU computers, DGX Spark systems and M4+ Macs into a single local inference endpoint. Queries from apps and agents go where there's free capacity — and the prompt never leaves your network.
Reading list
Research acceleration: view inside OpenAI
The first such concrete public data on how agents work in the lab — including inference costs (median researcher over $600 per day). Required reading for anyone planning "agent teams".
Claude Code as a tool in the hands of cybercriminals
A Gambit Security report shows an AI agent across the full chain of a real RaaS attack. It's worth knowing the adversary's tactics before you integrate agents into your own systems.
AGI headlines are exciting, but the real change is happening where agents are quietly taking over our eight-hour shifts.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.