Latest / UpNext AI / AI Textbooks, Enterprise Agents, and Code Security Benchmarks | UpNext AI – August 13, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, August 13th, 2026, and here's what matters in AI today. Nathan Lambert has written a post-training textbook and used AI throughout the process—not as a replacement author, but as an assistant for tasks such as copyediting, diagram generation, LaTeX work, and syncing different versions of the manuscript. His conclusion is a notable counterweight to fast-moving progress in coding and math: current models can help at the level of sentences, equations, and local edits, yet still struggle to organize a long technical argument into a coherent chapter. Lambert argues that this is more than a complaint about prose quality. If a model cannot reliably explain established material across a long-form reference work, that should temper expectations that it can independently drive broad, open-ended scientific work. He describes a five-year horizon as part of that question, but does not present it as a forecast that machines will surpass textbook authors by then. The more immediate lesson for teams is practical. AI can remove substantial production friction for domain experts, but expertise, editorial judgment, and responsibility for the final narrative remain essential—especially when the work needs to teach rather than merely fill space. That gap between completing individual tasks and carrying out reliable end-to-end work is also at the center of a new agent benchmark called VAKRA. The researchers built it around more than 8,000 executable APIs across 62 domains, then asked agents to combine API calls, document retrieval, multi-step reasoning, and natural-language tool-use policies. The setup matters because enterprise agents rarely operate against one clean database or one isolated tool. They have to resolve which customer, product, or policy a question refers to, connect information across sources, and know when a request should not be answered. Using a fixed ReAct agent harness, the researchers found the best tested model reached 70.4 percent accuracy on single-step endpoint tasks, but only 50 to 51 percent on compositional API tasks. Performance fell by more than half as reasoning chains grew deeper. On policy-constrained unanswerable questions, results dropped as low as 2.4 percent. The finding is not that tool calling itself is the main obstacle. The paper’s trace analysis points instead to language-mediated reasoning: entity disambiguation and grounding across sources. For organizations building agents, that makes evaluation against real workflows—not just a tool-call demo—the more meaningful test. For the research note, consider a related problem in software security: identifying the code change that first introduced a vulnerability. That history matters because a vulnerability may affect versions long before the eventual fix. Researchers behind VICBench assembled 100 verified vulnerability-introducing commits, each tied to one of 100 documented CVEs, or Common Vulnerabilities and Exposures. The benchmark spans 88 projects in Python, Java, and C++, and covers 48 CWE categories, which classify common software weakness types. The researchers deliberately included more complex real-world changes than earlier datasets. They report that two state-of-the-art approaches, V-SZZ and LLM4SZZ, achieved F1 scores between 33.3 and 40.1 percent. F1 is a measure that balances whether a system finds the right cases with whether its findings are correct. That result leaves substantial manual work for security teams. VICBench is a benchmark rather than a finished detection system, but its takeaway is clear: evaluating AI-assisted vulnerability tracing on simplified patches can overstate how ready it is for messy production code. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... A few more developments to watch. TechCrunch reports that OpenAI-backed Thrive Holdings raised $2 billion in funding at a $12 billion valuation. SoftBank, D1 Capital Partners, and Altimeter Capital are among the investors, putting a large financing marker behind the enterprise-AI push. Google has added four Pixel 11 phones to its lineup: the Pixel 11, Pixel 11 Pro, Pixel 11 Pro XL, and Pixel 11 Pro Fold. The Verge reports improved cameras, a new Tensor G6 chip, new photography tools, and expanded Gemini Intelligence and Live Translate capabilities—another sign that AI features are becoming part of the core smartphone upgrade pitch. And Anthropic has hired legal-tech founder Robert Mahari to lead Claude’s push into law practices, according to The Decoder. The move signals a more focused effort to adapt general-purpose AI assistance to the workflows and requirements of legal professionals. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!