Sep 23, 2026 · 8 min listen · Last updated September 23, 2026
From storyflo. This is your daily audio brief. Hey, it's Theo. September 23rd. Here are five stories I'd flag if you missed yesterday's end-of-day. Let's get into it. First, from The Decoder. Anthropic engineer explains why Claude's writing got worse although the model got smarter.
Listen · storyflo · A.I.
Daily A.I. Brief · September 23rd
0:00-8:02
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 engineer explains why Claude's writing got worse although the model got smarter
Jackson Kernion from Anthropic sheds light on an interesting quirk with their Claude models. As they’ve focused on improving the model's abilities in math, coding, and technical explanations, the writing style has shifted. It’s become more like these dense info dumps that can feel overwhelming to us humans. So, while the model is technically smarter, it doesn’t always translate to better writing. They’ve made some tweaks with Opus 5.5 to address this issue, but it seems that Opus 4.6 still holds the crown when it comes to clear, engaging writing. It’s a fascinating example of how optimizing for one area can sometimes lead to unexpected trade-offs in another.
OpenAI hires Patreon co-founder Sam Yam to lead a new Creator Product division
OpenAI has brought on Sam Yam, one of the co-founders of Patreon, to head a new division focused on Creator Products. It’s interesting to see how his experience from over a decade at Patreon will shape the tools they’re planning to develop for creators. He’s not going solo, either; he’s bringing along a couple of other executives from Patreon to form this new team. They’re aiming to unveil some of their work at OpenAI's upcoming DevDay, which should give us a glimpse into how they plan to support creators in a fresh way. It’s a fascinating move that could really change how creators interact with AI tools.
Alibaba launches Qwen Audio 3.1 with new models and slashes AI audio prices by up to 95 percent
Alibaba's Qwen team just rolled out Qwen-Audio-3.1, introducing five new models focused on speech recognition and text-to-speech. What’s interesting is the ASR model's ability to recognize multiple languages and dialects while filtering out those pesky filler words. There’s also the ASR-Next, which can identify different speakers and even pick up on emotions and background noise—pretty impressive, right?
On the TTS side, it’s designed for multilingual synthesis, making it versatile for various applications. The kicker? They’ve slashed prices on these AI audio tools by up to 95%. It’s a big shift that could open up a lot of possibilities for developers and businesses.
Inside Basecamp Research, the AI startup turning evolution into training data
Basecamp Research is making some intriguing strides in AI by tapping into the vast world of biology. They’ve secured $140 million in funding from big names like Nvidia and Anthropic, which is a solid vote of confidence. What’s fascinating is their approach: they’re using genetic material from diverse ecosystems—think rainforests, oceans, and hot springs—to train AI models. This isn’t just about crunching numbers; it’s about designing antibiotics and developing tools for cell therapies, which could have a real impact on healthcare.
CTO Philip Lorenz shared some insights, emphasizing that biology presents a more complex challenge for AI than language does. It’s not just about achieving high scores on tests; the real challenge lies in translating that into effective molecules. There’s a lot of potential here, and it’s exciting to think about how this could reshape our understanding of medicine and biology. The way they’re merging evolution with technology feels like a fresh perspective in the AI landscape.
Meta's AI agent Muse draws 500,000 users in a week along with claims it copied OpenClaw
Meta launched its new AI agent, Muse, and it’s already made quite a splash, racking up over 500,000 users in just a week and topping the App Store charts. What’s interesting is that Meta has openly acknowledged that Muse takes a lot of cues from an open-source project called OpenClaw. They even noted that some file names and content are almost identical, which raises questions about originality and inspiration in tech development. Meanwhile, OpenAI is reportedly weighing its options in response to this situation. It’s a fascinating moment in the AI landscape, highlighting how quickly things can evolve and the fine line between inspiration and imitation.
Nvidia-backed Nscale keeps its biggest customer, Bytedance, out of its IPO filing
Nscale, the AI cloud provider that’s got Nvidia in its corner, is gearing up for an IPO but has decided to omit its biggest customer, Bytedance, from the main prospectus. This is interesting because Bytedance, the parent company of TikTok, plays a significant role in Nscale's operations. By not including them, Nscale might be trying to present a more independent image to potential investors, or perhaps they want to avoid any scrutiny that could come from Bytedance's ongoing regulatory challenges. It’s a strategic move that raises questions about how Nscale plans to position itself in the market and what it means for their relationship with Bytedance moving forward. This decision could impact investor perceptions, especially in a landscape where partnerships and customer stability are key indicators of success.
So, there's this guide that walks you through creating a world model in Python, which is pretty cool if you're into AI and machine learning. It starts from the basics, making it accessible for beginners. You’ll learn how to set up a model that can simulate environments like CartPole, which is that balancing game where you try to keep a pole upright on a cart.
What’s interesting is how the model learns to predict the next state of the environment, almost like it’s daydreaming about what happens next. The guide also dives into measuring when the model's predictions start to fall apart, which is crucial for understanding its limits. It’s like teaching a kid to ride a bike and then watching them wobble before they find their balance. Overall, it’s a solid starting point if you’re curious about building and testing your own models.
How GRPO Trains Small Language Models with Verifiable Rewards
So, there’s this interesting approach called GRPO that’s focusing on training small language models, and it really highlights the importance of the reward function in the learning process. It’s not just about having a sophisticated model; the way you set up rewards can significantly influence how well the model learns and performs.
They dive into local reasoning experiments using something called Unsloth, which is a tool for testing these models. What’s fascinating is how they’ve structured the experiments to show that even small changes in the reward system can lead to different outcomes in how the model understands and generates language. It’s like tuning a musical instrument — the right adjustments can make a world of difference.
The researchers emphasize that verifiable rewards are key, meaning the rewards need to be clear and measurable. This clarity helps ensure the model is learning the right things, rather than just chasing after vague or misleading incentives. It’s a reminder that in machine learning, the details really matter, and how you frame the problem can lead to surprising insights and improvements.
Overall, this work sheds light on the intricate dance between model design and reward systems, suggesting that a thoughtful approach to both can lead to more effective and reliable language models. It’s a nice blend of theory and practical application, showing that there’s still a lot to explore in this field.
So, let’s dive into how machines turn our words into numbers, which is pretty fascinating. It all starts with something called TF-IDF, which stands for Term Frequency-Inverse Document Frequency. Essentially, it combines how often a word appears in a document with how common that word is across a bunch of documents. The idea is to give more weight to words that are unique and less common, helping the machine understand what’s important.
First, we tokenize the text, breaking it down into individual words. Then, we create a vocabulary, which is just a list of all unique words across our documents. From there, we calculate the term frequency for each word in a document, which tells us how often a word shows up. But that alone doesn’t tell the whole story, especially for common words like "the."
That’s where the inverse document frequency comes in. It helps us understand how rare or common a word is across all documents, giving us a clearer picture of its significance. Finally, we combine these two metrics to get the TF-IDF score for each word, allowing machines to classify and retrieve information more effectively. It’s a neat process that really highlights the mechanics of language processing!
I Trained a Tiny Network to Compress Data. It Drew a Pentagon.
So, there’s this fascinating project where someone took on the challenge of recreating a model from scratch using NumPy, which is pretty impressive on its own. They focused on a concept called "Toy Models of Superposition" from Anthropic. What’s really cool is that they did everything manually, deriving their own gradients without leaning on any pre-existing numbers. It’s like building a car from raw materials instead of just assembling parts from a factory.
The end result? A tiny neural network that learned to compress data, and the output was a pentagon. It’s a simple shape, but the fact that a network can learn to represent complex data with such a basic geometric form is a neat demonstration of how these systems can work. It’s a reminder of how even small models can yield surprising insights when you dig into the mechanics.
This kind of hands-on approach not only deepens understanding but also highlights the beauty of machine learning—how creativity and technical skill can come together in unexpected ways. It’s a little glimpse into the playful side of AI development that often gets overshadowed by the big names and flashy results.