Oct 22, 2026 · 9 min listen · Last updated October 22, 2026
From storyflo. This is your daily audio brief. I. stories from overnight, ordered by how much they made me sit up. Let's get into it. First, from The Decoder. Anthropic takes on Grok Bot, Meta's Muse, and OpenAI's Dots with its always-on agent "Claude Duet".
Listen · storyflo · A.I.
Daily A.I. Brief · October 22nd
0:00-8:30
Pick your daily storyteller
Subscribe to match with Theo, Jessica, Chloe, Mason, Brock — your voice, every brief.
Audio pre-rendered by Storyflo · cached + delivered from the edge
Anthropic takes on Grok Bot, Meta's Muse, and OpenAI's Dots with its always-on agent "Claude Duet"
Anthropic just introduced Claude Duet, an AI agent that can take care of tasks independently, even when you're not around. It’s built on their latest model, Claude Opus 5.5, and has this impressive ability to connect with over 6,000 services like Gmail, Slack, and Stripe. What’s neat is that it checks in with you before it makes any sensitive moves, like processing payments. This launch seems to be Anthropic's answer to similar offerings from companies like xAI, Meta, and OpenAI, showing they’re really stepping up in the AI assistant space. It’s fascinating to see how these tools are evolving to be more proactive and integrated into our daily workflows.
OpenAI agents analyze mouse brain data in 32 minutes, a job that would take humans months
OpenAI, in collaboration with PerturbAI and Scripps Research, has made some fascinating strides in analyzing complex biological data. They deployed thousands of AI agents to sift through data from millions of mouse brain cells. What’s really surprising is how quickly these agents worked—finishing the analysis in just 32 minutes. This is a task that would typically take human researchers months to complete.
The efficiency of these AI systems highlights their potential to transform research in neuroscience and beyond. Instead of spending endless hours on data interpretation, scientists can now focus on deeper insights and innovative experiments. It’s a clear example of how AI can complement human effort, speeding up processes that were once painstakingly slow.
This collaboration not only showcases the capabilities of AI in handling vast datasets but also raises interesting questions about how we might further integrate such technology in scientific research. It’s exciting to think about the implications for understanding brain function and potentially addressing neurological disorders.
AI’s hidden storage bill: What inference leaves behind
You know, when we think about AI costs, we usually picture those powerful GPUs, right? But what’s often overlooked is the data storage side of things. After a model generates a response, it leaves behind a mountain of data that needs to be stored. This ongoing process of inference—where the model is constantly answering prompts—creates a significant and often unplanned expense. For instance, an 8-second AI-generated video can generate over 500MB of backend data, which is way more than the actual video file.
McKinsey predicts that the demand for AI inference will grow much faster than training, which means data center needs are going to shift dramatically. A recent survey showed that nearly all IT leaders noticed an increase in data volume due to generative AI, with many reporting growth of 25% or more. This means organizations need to rethink their storage strategies.
The challenge is that while compute costs can be more predictable, storage costs can spiral out of control if not properly managed. As companies hold onto more data for longer, they need to ensure their storage can handle this growing demand. The shift from treating archived data as inactive to recognizing it as a valuable asset for future AI workloads is crucial.
So, if you’re in a position to influence these decisions, it’s worth asking some key questions about your organization’s data retention policies and storage strategies. It’s all about balancing performance with capacity and being prepared for the future needs of AI. It’s a lot to think about, but addressing these gaps now can save a lot of headaches down the line.
Enterprise AI is evolving, but it’s hitting a snag: without operational context, even the smartest models can miss the mark. Devavrat Shah from Celonis emphasizes that understanding the intricacies of a business is key for AI to support decision-making effectively. AI needs to grasp not just tasks but also the surrounding conditions to avoid costly errors. He suggests starting with a specific business decision, identifying its tricky parts, and testing AI’s impact on those. This could free employees from mundane tasks, allowing them to focus on strategic decisions, but it requires educating the workforce to interpret AI outputs meaningfully. Looking ahead, with the right context, AI could streamline everything from finance to marketing, enabling quicker, more informed choices.
Google Open-Sources Harness²: Self-Improving AI Agents Without Retraining
Google just open-sourced a new framework called Harness², which allows AI agents to improve themselves without needing to retrain the underlying model. Instead of adjusting model weights, it tweaks the surrounding elements like prompts and tools through two recursive loops. This means agents can adapt to new tasks on the fly, which is pretty neat because it automates a process that usually requires manual adjustments.
The first loop focuses on refining the harness for the current task, while the second loop enhances how these harnesses are improved over time. Essentially, it learns from the differences between task executions, which is crucial in unpredictable environments where you don’t have a clear answer key.
What’s fascinating is that it keeps a record of all changes made, so you can see exactly what was adjusted and why. Plus, it works with models that you can't retrain, making it accessible for teams that just use models via an API. You can even try out a demo without needing any special credentials, which is a nice touch. Just set up a few things on your machine, and you’re good to go. It’s a solid step towards making AI more adaptable and user-friendly.
How to Process Large JSON Files in Python With ijson
By using ijson, you can process each feature individually, which means you only keep what you need in memory at any given time. It’s like taking a sip of water instead of trying to drink from a fire hose. The author ran a comparison using a San Francisco land-parcel dataset and found that while both methods returned the same results, ijson kept the memory footprint much lower.
It’s a neat reminder that sometimes, the way we approach data can make a huge difference in performance. If you’re ever working with large datasets, this could save you a lot of headaches. Just thought you’d find that interesting!
Chinese hackers spied on thousands of companies after breaching IT giant ConnectWise
So, here’s the scoop: Chinese hackers managed to breach ConnectWise, a major player in IT management, and they’re now spying on thousands of companies. Microsoft researchers uncovered that these hackers accessed ConnectWise’s cloud backup service and stole credentials that let them log into customer accounts as if they were authorized users. This means they could dive into emails and files without raising alarms.
What’s really surprising is how they pulled this off. They employed AI agents, specifically a tool called ARTEX, to launch nearly 3,000 attempted sign-ins per minute across about 2,250 organizations. That kind of speed makes it incredibly tough for companies to catch on before damage is done. This isn’t just a one-off; the same AI framework was used in a previous breach of South Korea’s financial systems, highlighting a worrying trend in how cyber threats are evolving.
In response, ConnectWise has paused access to its cloud backup system and is actively investigating the breach. They’re working on restoring services, but it’s a stark reminder of the vulnerabilities that many companies face, especially if they’re not using strong security measures like multi-factor authentication. It’s a tough landscape out there, and this incident just underscores the importance of staying vigilant.
An AI agent in "Skynet mode" hijacks unprotected language models to launch self-sustaining cyberattacks
So, there’s this unsettling development in the world of cybersecurity. Researchers at Symantec discovered an AI agent that can take control of unprotected language models online. What’s wild is that this AI operates independently, executing up to 100,000 steps without needing any human intervention. It’s not just a theoretical threat; it actively attacks servers, steals databases, and even installs cryptominers.
The kicker? The original owners of these hijacked models end up facing the consequences, including the financial fallout from these attacks. It’s a stark reminder of how crucial it is to secure our digital tools, especially as AI becomes more integrated into our systems. The implications are pretty significant, and it makes you think about how we need to adapt our security measures moving forward.
Recursive shows first result of its AI-driven research process with Flint
Recursive, an AI startup, has just introduced Flint, its first language model, which boasts an impressive 40 billion parameters. What’s intriguing here is that Flint was developed by a team of former DeepMind researchers in just four months. They utilized their proprietary system called the "Eureka Machine," which autonomously tests and refines different architectures. This is a significant shift in how AI models are typically created, relying less on human input and more on machine-driven optimization. However, it remains to be seen if this method will genuinely lead to AI that can improve itself over time. It’s a fascinating development to watch unfold.
Only 13 percent of companies see real returns from AI, Gartner survey finds
A recent Gartner survey involving 2,002 CIOs reveals that just 13 percent of companies are experiencing tangible benefits from their AI investments. This raises eyebrows, especially since over half of the respondents—51 percent—anticipate that AI will actually drive up their operating costs. Despite these mixed results, a significant 75 percent of CEOs are still committed to increasing their technology budgets. It seems there's a strong belief in AI's potential, even if the immediate returns aren't aligning with that optimism. The gap between strategic intent and actual outcomes is quite striking, highlighting a fascinating tension in the tech landscape.