Oct 15, 2026 · 5 min listen · Last updated October 15, 2026
From storyflo. This is your daily audio brief. Theo, October 15th. I. stories that bear on what's coming next. Let's get into it. First, from The Decoder. 8B but growth slows ahead of planned November IPO. 8 billion, marking a solid 45 percent increase.
Listen · storyflo · A.I.
Daily A.I. Brief · October 15th
0:00-5:21
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's Q3 revenue hits $16.8B but growth slows ahead of planned November IPO
Anthropic reported third-quarter revenue of $16.8 billion, marking a solid 45 percent increase. However, this growth is a significant drop from the previous quarter's impressive 140 percent surge. With an IPO on the horizon in November, there's a sense of caution among investors, who are now anticipating the company's valuation might dip below $1.5 trillion. This slowdown in growth, coupled with a challenging market environment, is leading to more tempered expectations for what many see as a critical moment for AI on Wall Street. It’ll be interesting to see how they navigate this shift as the IPO approaches.
Bank of America puts 4,000 people on fixing vulnerabilities uncovered by Anthropic's Mythos and other AI models
Bank of America is ramping up its efforts to address software vulnerabilities, bringing in a total of 4,000 people for this task. This includes 1,000 outside contractors specifically hired to tackle issues identified by AI models, including Anthropic's Mythos. The bank is clearly taking a proactive approach to cybersecurity, recognizing the importance of these AI tools in revealing weaknesses that might have otherwise gone unnoticed.
This move reflects a broader trend in the financial sector, where institutions are increasingly relying on advanced AI to enhance their security measures. By investing in both internal and external resources, Bank of America is not just responding to current threats but also looking to strengthen its overall infrastructure against future risks. It's interesting to see how AI is shaping the landscape of cybersecurity, isn’t it?
Google wants Gemini 4 Argon to be the only AI model robots need
Google's latest AI model, Gemini 4 Argon, is designed to streamline robotics by handling a range of tasks like controlling robots, generating simulations from images or sketches, and assessing training videos. It’s impressive to see it score 87 percent on Google's own ERQA benchmark, which suggests it has a solid grasp of various functions. However, the examples shared indicate that it still requires considerable human input to operate effectively. This blend of capability and reliance on human guidance highlights the ongoing challenge in AI development, where we’re making strides but still need that human touch to guide the technology. It's fascinating to think about how this will evolve.
GPT-6-Astra solves 49 of 50 physics research problems at $2.40 apiece
So, there’s this new AI called GPT-6-Astra that tackled a benchmark in theoretical physics and nailed 49 out of 50 problems. That’s pretty impressive, right? What’s fascinating is that in a few cases where it stumbled, the experts’ own solutions had errors. It’s like the AI is holding a mirror to the field, showing that even the pros can miss things. But here’s the catch: researchers say this isn’t quite the same as what we’ve seen in math breakthroughs. They feel the AI lacks that intuitive “taste” that humans have. Still, they’re optimistic about the future, especially for training young physicists and creating a more level playing field in the discipline. It’s a big shift in how we might think about physics research going forward.
Nvidia's Nemotron 3.5 Super VL is an open model built to run enterprise AI agents nonstop
Nvidia just rolled out the Nemotron 3.5 Super VL, an open multimodal model boasting 120 billion parameters designed specifically for enterprise AI agents. What’s interesting is that they’ve also introduced two smaller Bolt models, which allow these agents to operate directly on local machines. This could really change how companies deploy AI, making it more accessible and efficient.
A bunch of organizations are already diving in and fine-tuning Super VL to fit their needs, but here’s the catch: to truly maximize its potential, you’ll need Nvidia hardware. It’s a solid move for them, ensuring that users are somewhat tethered to their ecosystem. It’ll be fascinating to see how this plays out in the enterprise space.
I Tested Four Ways to Give an AI Assistant a Memory. The Obvious One Was the Worst Value.
So, here’s the scoop on giving AI assistants a memory. The author tried four different methods to see which one worked best. You’d think just pasting in the entire chat history would be the most effective, right? Surprisingly, it turned out to be the least efficient method. It cost 20 to 31 times more than simply looking up notes, and when it came to completing tasks, it didn’t perform any better.
The other methods, which focused on more streamlined approaches like note retrieval, really shone in comparison. It’s fascinating how what seems like the obvious choice can sometimes lead to the worst outcome. It makes you wonder about the underlying mechanics of memory in AI and how we can optimize it for better performance. It’s a reminder that sometimes, less is more, even in tech.
How Many Guesses Does It Take to Crack an AI's "Random" Number?
So, there's this intriguing exploration of how we can actually quantify the randomness in AI-generated numbers. The author dives into the logits from various open models, both base and instruction-tuned, to analyze how predictable these "random" outputs really are. It turns out, by breaking down the data into bits, we can estimate how many guesses it might take to figure out the number an AI is generating.
What’s fascinating is that instead of just accepting these outputs as random, the piece encourages us to think about the underlying patterns. It’s almost like peeling back layers to reveal a more structured approach to what seems chaotic at first glance. The author’s method of translating these logits into a guess distribution adds a layer of clarity, making it easier to grasp how we might outsmart the AI in a guessing game.
In the end, it’s a reminder that even in the realm of AI, where randomness seems to reign, there’s often a hidden logic waiting to be uncovered. It’s a neat intersection of math and machine learning that really makes you rethink how we interact with these systems.