Sep 25, 2026 · 9 min listen · Last updated September 25, 2026
From storyflo. This is your daily audio brief. Hey, it's Theo. September 25th. Here are five stories I'd flag if you missed yesterday's end-of-day. Let's get into it. First, from The Decoder. 6 billion cloud deal with Akamai, pushing its compute spending past $500 billion in under a year.
Listen · storyflo · A.I.
Daily A.I. Brief · September 25th
0:00-8:52
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 signs $11.6 billion cloud deal with Akamai, pushing its compute spending past $500 billion in under a year
Anthropic has just inked a significant seven-year cloud agreement with Akamai, valued at $11.6 billion. Along with this deal, they’re also set to receive a warrant for up to 5 percent of Akamai's shares, which is an interesting move for them. This brings Anthropic’s total compute spending to an eye-popping $517 billion in less than a year. It’s a staggering figure, and it really highlights the scale at which they're operating. However, CEO Dario Amodei has expressed some serious concerns, warning that if their revenue projections miss the mark, the company could face bankruptcy. It’s a high-stakes situation, for sure.
Google's "Call for Me" lets Gemini phone businesses for you
Google is rolling out a new feature called "Call for Me," which allows its Gemini AI to make phone calls on behalf of users. Currently, it's only available to Pixel 11 owners in the U.S. who have a paid Gemini subscription and the beta version of the Google Phone app. The AI can handle tasks like making reservations or checking product availability by navigating through phone menus and even waiting on hold, all while using the user's own phone number.
What’s interesting is that Gemini can share approved personal information during these calls, making it more effective at managing tasks. Users can follow along with a live transcript and jump in whenever they want. This feature builds on Google's previous work with Google Duplex, and it seems they're being cautious about rolling it out because real conversations can get pretty nuanced. It’s a fascinating step in how AI can assist with everyday tasks, but it’s still in the early stages.
White House tells OpenAI and Anthropic to let U.S. review new models before sharing them with British testers
The White House is asking OpenAI and Anthropic to pause the release of their new AI models to the U.K.'s AI Safety Institute. They want U.S. agencies to have a chance to review these models before any testing happens overseas. This move highlights a growing concern about the safety and regulation of AI technologies, especially as they become more integrated into various sectors.
The decision reflects the U.S. government's desire to maintain oversight and ensure that any potential risks are assessed before these models are deployed or tested abroad. It’s an interesting shift in how countries are approaching AI safety, emphasizing the importance of collaboration and scrutiny in this rapidly evolving field.
This request also signals a more cautious stance from the U.S. as it navigates the complexities of global AI development. It’s all about balancing innovation with safety, and it seems like the White House is keen to take a proactive role in shaping these discussions.
7 Advanced Python Tricks to Level Up Your Coding Skills
So, there's this interesting piece about advanced Python tricks that really dives into what’s already baked into the language, rather than just new syntax. It highlights some built-in features that can simplify your code significantly. For instance, you can turn a callable into an iterator using `iter()` with a sentinel value, which is way cleaner than the typical `while True` loop.
Then there's `ExitStack`, which is a lifesaver for managing resources when you don’t know how many you’ll need at runtime. It ensures everything closes properly, even if an error occurs. Another neat trick is using `memoryview` to slice binary data without making copies, which can save a lot of memory when dealing with large datasets.
The article also touches on `ExceptionGroup`, a feature from Python 3.11 that allows you to handle multiple exceptions from concurrent tasks without losing any information. You can keep track of all errors, which is a big deal when you’re debugging. Plus, there’s `ChainMap` for managing configuration dictionaries, allowing you to layer settings without merging them into a single mess.
Lastly, it introduces `MappingProxyType` for exposing read-only views of dictionaries, which is great for keeping your internal state safe while still allowing access. And if you’re working with positional arguments, `functools.Placeholder` lets you reserve slots, making your function calls clearer.
It’s all about leveraging what Python already offers to make your code cleaner and more efficient. Definitely worth checking out if you're looking to refine your skills!
Batching by Length Instead of Looping Item by Item for SLM Optimization
In this article, the focus is on optimizing small language models (SLMs) by changing how we process input data. Instead of handling one support ticket at a time, which wastes resources, the idea is to batch tickets by their length. When we process one ticket, the model often sits idle while waiting for data to load, especially on CPUs. By batching, we can reduce this idle time, but the challenge is that typical batching requires padding sequences to a common length, which can lead to inefficiencies.
To tackle this, the authors suggest sorting tickets by their token length before batching. This way, each batch contains items that are similar in size, minimizing wasted computation on padding. The benchmarks show significant improvements when using this method, as it allows the model to work more efficiently, processing fewer unnecessary tokens.
The results highlight how much time and resources can be saved by this simple adjustment. It’s a fascinating look at how small changes in data handling can lead to better performance in machine learning tasks.
Meta is putting its muscle behind Muse as the AI app takes off
Muse is really gaining traction, thanks to Meta's spotlight at its recent developer conference. Just a few weeks after its launch on September 8, downloads jumped from 2.5 million to over 3.4 million, with some estimates even suggesting it could be as high as 4.3 million. The app is currently only available in the U.S. and Canada, but it’s already hit the top of both the App Store and Google Play Store.
The buzz is not just from downloads; daily active users spiked by 27% right after the conference. Muse is getting a boost from Meta's cross-promotion strategy, targeting existing Facebook and Instagram users, but interestingly, only a small portion of its growth comes from paid ads. The app’s organic appeal seems to be resonating well, especially with features like video chat and smart glasses integration on the horizon.
What’s fascinating is how Meta is tailoring its marketing for Muse differently than it did for Threads, focusing on showcasing specific functionalities that could benefit users. This targeted approach is clearly paying off, and it’ll be interesting to see how Muse evolves as it rolls out more features and expands its reach.
Intelligence doesn't come cheap as AI drives up costs for the NSA, hospitals, and insurers
The NSA is pouring billions into testing advanced AI models, primarily due to the skyrocketing costs of computing power. Lawmakers are now anticipating that overseeing AI will run into tens of billions annually, a stark contrast to the Congressional Budget Office's earlier estimate of just $20 million. This shift highlights how quickly things can escalate when new technology is involved.
In the healthcare sector, AI is also making its mark, but not in a budget-friendly way. AI-assisted billing codes have added nearly $1 billion in costs over just two years. It’s fascinating, isn’t it? The very tools designed to enhance efficiency are driving up expenses across the board. It makes you wonder about the balance between innovation and affordability.
Meta's Muse agent gives every user a full cloud computer running Ubuntu Linux
What’s really neat is the “Sentinel” process they’ve implemented. It keeps an eye on sensitive actions outside your workspace, giving you a layer of security while you’re working. Plus, you have the ability to inspect every file in the system, which adds a level of transparency that’s refreshing.
In just the first week, they’ve already attracted over 500,000 users, which shows they’re really focusing on getting this product into as many hands as possible rather than just cranking up the power of the models behind it. It’s fascinating to see how they’re prioritizing accessibility and user engagement.
Jev vs. LLMs: When AI moves from Generation to Decision-making
So, there’s this interesting comparison between TypeSafe AI’s Jev and traditional large language models, or LLMs, focusing on how they handle classification tasks. The testing involved a whopping 3,080 tasks, which is no small feat. What’s really striking is how Jev not only matches but sometimes surpasses LLMs in accuracy and latency. It’s like watching a sprinter who’s also got a great sense of direction.
Jev seems to excel in calibration and confidence too, which is crucial when you think about decision-making in AI. Instead of just generating text, it’s about making informed choices based on the data it processes. This shift from generation to decision-making could really change how we integrate AI into systems that require more than just conversational abilities.
The findings suggest that Jev could serve as a more reliable decision layer for AI applications, which is a big deal. It’s not just about spitting out information anymore; it’s about making smart, confident decisions that can impact real-world applications. So, it’s worth keeping an eye on how this technology evolves and what it could mean for the future of AI in decision-making roles.
RAG Isn't an Agent — I Built the Layer Between Retrieval and Action
So, here’s the interesting part: the author dives into the distinction between retrieval-augmented generation (RAG) and agents. RAG focuses on pulling in relevant information, while agents are about taking action based on that information. What they did was create a layer that links these two processes, allowing for a more seamless interaction between retrieving data and executing tasks.
They ran nine different tasks through three systems: RAG alone, agents alone, and their new integrated approach. The results showed some intriguing differences in performance and efficiency. It’s not just about having the data; it’s about how you use it. This layered approach could change how we think about AI systems, making them more effective in real-world applications.
The takeaway here is that simply having a retrieval system isn’t enough; understanding how to connect it with actionable insights is key. It’s a fascinating look at how we can refine these technologies to work better together.