Latest / UpNext AI / Post-Transformer Reasoning, AI Materials Agents, and Reliability Checks in Medicine | UpNext AI – September 10, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Thursday, September 10th, 2026, and here's what matters in AI today. Our lead is an AWS post on Pathway’s work developing its brain-inspired BDH architecture on Amazon SageMaker HyperPod. The pitch is a direct challenge to the usual recipe for AI progress: bigger models, more training data, longer context windows, and more inference-time computation. Instead of producing a long chain of thought as text, Pathway’s BDH-CQ system performs iterative reasoning in a latent internal state, then decodes candidate answers. The company argues this can let a model refine a solution without paying for a growing stream of reasoning tokens. It also says the architecture maintains internal memory during inference, without test-time weight updates. Pathway’s system uses sparse, local interactions rather than the dense computation associated with transformers. AWS says only 5 percent of its neurons are typically active at a time. On ARC-AGI-1, a visual-rule benchmark, the company reports that a 150-million-parameter BDH-CQ model achieved a 29.2 percent pass-at-two score at a cost of $0.0007 per task, as of August. Those are company-reported results, not an independent verdict on a new architecture. Still, this is a useful signal: the next efficiency gains may not come solely from scaling the familiar transformer stack. They may come from changing where, and how, a system does its reasoning. That search for different AI system designs also reaches the lab. Nature Machine Intelligence has published work on a collaborative agent for autonomous crystal-materials research. The paper’s premise is a system built from two lightweight, synergistic models rather than one monolithic model. The broader idea is compelling even from this early description: specialized models can divide scientific work and collaborate on a research task. For organizations building AI for discovery, the important question is not simply whether an agent can produce a plausible answer. It is whether a compact, coordinated system can help move a real experimental workflow forward. The paper is one study, so it should be read as evidence of a possible design direction rather than a settled blueprint for autonomous science. A separate paper turns from collaboration to a risk that can emerge when many AI agents think alike. Researchers writing on arXiv tested LLM traders with different levels of general-purpose capability in an agent-based financial-market simulation. Their finding is a capability paradox. More capable frontier models showed more correlated behavior. When agents shared accurate reasoning, adding more of them reduced market-level risk. But when they operated in a common misinformation environment, that same correlation became a liability: similar agents could act on the same bad premise together, creating a risk that diversification does not remove. The takeaway extends beyond markets. Improving a model’s individual performance does not automatically improve the outcome of a system full of those models. Teams deploying fleets of agents should measure correlated failure modes, shared data dependencies, and common-model exposure—not only benchmark scores for each agent. For the research note, consider a problem in real-time colonoscopy: an image-segmentation system may fail to identify a polyp correctly, but there is no ground-truth label available at the moment of use to reveal that failure. A new arXiv paper proposes a check called referee-based quality estimation. A primary segmentation model and an independently trained referee analyze the same image, and their agreement becomes a reliability signal. On a 1,223-image external benchmark drawn from four public datasets, the strongest cross-architecture referee reached a 0.960 ROC-AUC—a measure of how well it distinguished reliable from unreliable predictions. When the researchers excluded easy empty-mask cases, performance fell, but the cross-model approach still outperformed the comparison methods. The system needs one additional deterministic model pass, and it is tested specifically for polyp segmentation rather than across all medical imaging. But its practical lesson is broader: when a high-stakes model cannot know whether it is right, an independently trained second opinion may be a useful trigger for human review. ...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. ... OpenAI has added Paul Christiano to the OpenAI Foundation Board and its Safety and Security Committee. The company says he brings experience in AI alignment, safety, and standards, reinforcing the foundation’s governance focus. OpenAI is also offering $5 million in grants for independent research into how generative AI affects teen development, well-being, and safety. The program is aimed at producing evidence on a group whose AI use is growing faster than the research base. A new Interconnects essay argues that AI’s effects on everyday life may take much longer to become tangible than the industry’s pace suggests. Its central warning is that productivity gains concentrated in knowledge work could deepen public skepticism if broadly visible benefits arrive slowly. WIRED reports that Clearview AI is testing a previously unreported prototype called InquiryIQ. The tool tested an xAI model to surface associates, social accounts, and other online information about people identified through Clearview; the report does not establish that it is publicly deployed or routinely used by police. And Volvo is bringing back the XC40 plug-in hybrid, which it discontinued about three years ago. The Verge reports that the refreshed vehicle is due at dealerships early next year with updated styling, a new safety suite, and Google’s Gemini assistant in the interior. 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!