Nov 8, 2026 · 5 min listen · Last updated November 8, 2026
From storyflo. This is your daily audio brief. Theo here. I. desk. Five stories from the last twenty-four hours — here's where I'd start. Let's get into it. First, from Towards Data Science. How to Build Multi-Agent Systems with the OpenAI Agents SDK.
Listen · storyflo · A.I.
Daily A.I. Brief · November 8th
0:00-5:19
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
How to Build Multi-Agent Systems with the OpenAI Agents SDK
So, there’s this fascinating guide on building multi-agent systems using the OpenAI Agents SDK, and it dives into how you can create practical applications like an electric vehicle purchase adviser and a support agent in Python. What’s really interesting is how it breaks down the mechanics of these systems. You can leverage concepts like fan-out, where one agent can distribute tasks to multiple others, and pipelines that streamline processes by chaining agents together.
It also touches on hierarchies, which help organize agents based on their roles, and swarms, where agents work collaboratively to tackle complex problems. The beauty of it is that these different structures reuse the same foundational building blocks, making it easier to scale and adapt your systems. It’s a neat way to think about how agents can interact and support each other, all while keeping the code manageable and efficient. If you’re into coding or just curious about how these systems can work together, it’s definitely worth a look.
EP 41 | Nu Holdings Stock Analysis: Will Nubank Leave Big Banks Behind?
Nubank's just hit a billion dollars in quarterly profit, which is pretty impressive. The stock's currently sitting at about $12.03, and analysts are giving it a strong buy rating for the next few years, suggesting potential returns of up to 170%. Their recent Q2 results show revenue of $5.51 billion and a net income of $1.06 billion, but there are some mixed signals with credit risk and arrears. They’re expanding into Mexico and the U.S., but there are challenges ahead, especially with servicing costs and local capital limits. The upcoming Q3 results on November 12 will be crucial, as will their Investor Day on December 8, where we might get more clarity on their strategy moving forward. It’s a fascinating time for Nubank, and it’ll be interesting to see if they can really disrupt the big banks.
AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
Alibaba has rolled out a new benchmark called Co-Design Bench to evaluate AI coding agents using actual user requests, and the results are pretty eye-opening. Even the highest-performing model, GPT-6 Astra, only fully solves about 26.4 percent of the tasks presented. This is a stark contrast to older benchmarks, where these models often score above 99 percent. It really highlights how, despite the impressive numbers in controlled settings, these AI agents still struggle significantly in real-world applications. So, it seems that human intervention is still crucial, as nearly three out of four tasks require some sort of cleanup. It’s a reminder of the complexities we face in practical coding scenarios.
The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts
Richard Sutton, a Turing Award winner, is making it clear that just throwing more compute power at AI isn’t going to cut it. He argues we need fresh learning methods to truly advance intelligence. It’s not just him, though; former OpenAI researcher Tworek echoes this sentiment, suggesting that the current trend of blindly scaling transformers has led to a kind of intellectual stagnation among scientists. It’s fascinating to think about how many are now recognizing this gap in the field. A wave of startups is emerging, looking to innovate beyond the limitations of raw compute. It’s a pivotal moment for AI development, really.
The brain-as-computer metaphor shaped modern AI but may be misleading the consciousness debate
Anil Seth, Ned Block, and Giulio Tononi recently shared some intriguing thoughts on the nature of consciousness in a commentary for Nature. They argue that the prevalent idea of the brain as a computer has led us to misunderstand what consciousness really is. Their main point is that current AI systems, no matter how advanced, lack true consciousness and are unlikely to achieve it through silicon-based hardware. They highlight that equating consciousness solely with computation is a flawed perspective. This is a significant reminder to be cautious about attributing rights or moral considerations to AI, as it doesn't possess the qualities that define conscious experience. It’s a fascinating discussion that really challenges us to rethink how we view both AI and consciousness.
Andrej Karpathy built a curated video platform in a weekend using Claude and 20 years of Hacker News
Andrej Karpathy, who co-founded OpenAI, recently put together a video platform over just one weekend, using a tool called Claude's Opus 5.5. What’s interesting is how he took a staggering 43 million comments from Hacker News and distilled them down into 6,454 curated videos. It’s like he created a personalized video experience that, according to him, even surpasses the recommendations you’d get from YouTube.
He’s been soaking in insights from Hacker News for two decades, so you can imagine the depth of context he brings to this project. It’s a neat blend of AI and community knowledge, showing how technology can sift through vast amounts of information to create something genuinely useful. It’s a reminder of how quickly things can come together when you have the right tools and a wealth of experience.
Change One Monty Hall Rule and Five LLMs Still Say Switch
So, there’s this interesting experiment where researchers tweaked the classic Monty Hall problem a bit and ran it through five different language models. They wanted to see if these models would still suggest that players should switch their choice after one door is revealed. Despite the models having access to the rules and reasoning through the problem, they ended up with around 33% accuracy.
What’s really curious is that one of the models actually quoted the original rule correctly but then went ahead and ignored it in its final answer. It’s a fascinating look at how these systems process logic and reasoning, especially when faced with a slight change in the scenario. This kind of stuff really makes you think about how we interpret rules and decisions, doesn’t it?