Latest / UpNext AI / Realtime-Bench’s Agent Test, Google’s Genie 4, and OpenAI’s Codex Acquisition
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, November 5th, 2026, and here's what matters in AI today. An AI agent may know what to do and still act too late. A benchmark featured in Jack Clark’s Import AI newsletter tests that problem across interactive web and game tasks, including situations where delays change the outcome. Its researchers compared performance with the clock running against a setting where the clock could be paused. GPT-6 Astra, in its high-reasoning setting, had the highest pass rate with the clock running: 36 percent of tasks. With the clock paused, high-reasoning models approached 70 or even 80 percent success. Those results point to a real constraint for agents handling time-sensitive work: reasoning capability alone does not ensure a timely action. A ticket purchase, for example, depends on knowing whether availability is still current and when to stop deliberating. The same newsletter covers Google DeepMind’s Genie 4, an interactive world model now rolling out as a research preview to some trusted testers. Google says users can change a simulation through natural language while it runs, such as changing the weather or introducing an object. The company plans to bring it to Google AI Ultra subscribers later. That brings us to the tools surrounding agents once they start producing real work. OpenAI says it will acquire Pierre Computer Company, whose developer libraries already power parts of the Codex app. If you’ve navigated project files or reviewed a change there, you’ve used Pierre’s Trees and Diffs libraries. Pierre also makes Code Storage, a service for storing code and tracking its history with Git. OpenAI wants the Pierre team to help adapt Git workflows to teams working alongside coding agents. The company says its engineering systems need to handle more work happening in parallel, and it wants Codex to connect code with the decisions behind it. That makes review and source control central to the acquisition: developers need to find changes, understand them, and manage their history as agents contribute code. OpenAI says Pierre’s open-source projects will remain open-source, and customers can keep using Code Storage after the deal closes while its features are brought into Codex. Closing remains subject to customary conditions, including regulatory approvals; the companies remain separate until then. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. OpenAI also has a more specialized deployment update. Its Daybreak Red offering now gives approved customers access to GPT-5.6 Cyber through Amazon Bedrock in AWS GovCloud’s US-East region. The company positions it for authorized vulnerability research, exploit validation, and security testing by government teams. It says the model can help defenders investigate software weaknesses, reproduce vulnerabilities in controlled settings, and develop mitigations, with qualified practitioners reviewing findings and proposed fixes before implementation. OpenAI says Daybreak Red gives verified defenders greater latitude for authorized cybersecurity work, paired with identity verification, monitoring, and access controls. For this Bedrock offering, it says inference data is not used to train models, and customers do not need to share their Daybreak Red data with OpenAI. Teams need both OpenAI approval and AWS provisioning; approval alone does not enable access. For the research note, consider the storage cost of searching or comparing large collections of text. Systems often turn text into numerical representations called embeddings. In an arXiv preprint posted Wednesday, researchers tested two ways to make those representations smaller: shorten them, and store their numbers at lower precision. They evaluated four embedding models across 38 tasks from the Massive Text Embedding Benchmark. Their strongest combined result: reducing embedding length eightfold and using eight-bit precision cut storage per representation by a factor of 32 while retaining 94.4 percent of baseline effectiveness. Eight-bit precision alone cost at most one and a half percentage points overall; binary precision performed significantly worse across every task category. Retrieval degraded much faster than the other categories, and the researchers measured evaluation times only once. Teams sizing a text-search system should test compression against their own retrieval workload, rather than treating the storage saving as a universal performance result. ...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. Visit up next dot fm slash eleven to check out their latest offerings. ... A few other developments. The Information reports that cybersecurity company Cyera found AI agents inside a financial company’s sensitive data that had gone unnoticed for more than six months. The finding illustrates a basic operational problem: organizations need to know which agents can access sensitive information before they can manage that access. The Information also reports that Palo Alto Networks is in talks to buy Clutch Security for around $500 million, citing people familiar with the matter. Clutch makes software for tracking passwords and authentication codes used by AI agents. The deal has not been finalized. And TechCrunch reports that DuckDuckGo is courting schools frustrated by Google’s AI restrictions, pointing educators to its search engine and options to disable AI features. 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. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!