Subscribe
Share
Share
Embed
Daily AI news and research, distilled. UpNext AI breaks down the most important developments in artificial intelligence—from major industry moves to cutting-edge papers.
Welcome to the UpNext AI podcast. It's Thursday, June 25th, 2026, and here's what matters in AI today.
We’ll start with Google DeepMind and A24. According to WIRED, Google DeepMind has made a 75 million dollar investment in A24, in what A24 described as a research partnership. The companies are working together through A24 Labs to create new filmmaking tools and workflows. Why this matters is bigger than one studio deal. It’s another sign that major AI companies are moving beyond software products and deeper into the cultural industries that shape taste, distribution, and creative workflows. WIRED frames this as part of AI companies deepening their influence in Hollywood, and that’s the real takeaway here. A24 says this is not about handing artists finished tools from the outside, but about helping shape what gets built. And just as important, the reporting makes clear this is an investment and research partnership—not an acquisition, and not evidence that Google now controls the studio. The broader context in the piece is also striking. It mentions several other money figures floating around the AI-and-entertainment conversation, including 300 million dollars, 1 billion dollars, and 18.5 million dollars. The specific details vary across those examples, but the pattern is clear: AI money is showing up in media in a much bigger way. So the headline here is simple. Google DeepMind’s 75 million dollar move into A24 is one more sign that the AI race is no longer just about models and cloud credits. It’s also becoming a fight over creative infrastructure and influence.
Next, a paper on AI security that feels especially practical. Researchers posted a paper called Detect, Unlearn, Restore, focused on defending text summarization models against data poisoning. The problem is training-time poisoning during fine-tuning—when someone manipulates the examples used to tune a model so the system later produces biased, harmful, or distorted summaries, even while standard evaluation still looks mostly normal. The authors say poisoned document-summary pairs can show unusually high training influence in white-box settings, which means if you can inspect the model and training process, you may be able to spot suspect examples. In black-box settings, where you do not have that internal access, they say poisoned models become two to three times more sensitive to meaning-preserving perturbations, which gives you a way to audit behavior from the outside. What makes this stand out is that the paper does not stop at detection. Across nine architectures and six benchmark datasets, the authors report 85 to 92 percent detection precision, and say gradient-ascent unlearning restored up to 96 percent of original behavior with less than 0.6 percent ROUGE degradation. In plain English: the researchers are arguing that poisoned summarization models leave fingerprints, and that in many cases you may be able to detect the damage and roll it back without retraining from scratch. That is useful because summarization is exactly the kind of capability that gets deployed everywhere—internal copilots, document workflows, support tools, compliance systems. If poisoning can hide inside a fine-tuning pipeline, post-hoc defenses like this become much more valuable.
For the research section, let’s look at BenchX, a benchmark for cancer detection and localization in medical imaging. This paper was published yesterday, and its core point is straightforward: medical AI can look strong on average and still fail badly when the hospital context changes. BenchX is designed to test that directly. The benchmark covers 85,355 CT scans and evaluates 12 tumor-detection models across factors like tumor size, location, patient subgroup, and imaging protocol. The paper says performance can shift when demographics differ or when scan setups change—for example across contrast phases, ages, or sexes. One especially important result in the paper is that models optimized for average accuracy performed poorly in rare or underrepresented groups, including young, female African American patients. The researchers also say large language models were used to extract and organize subgroup information from clinical data so the analysis could be scaled and reproduced. The key term here is protocol bias—basically, changes in how scans are acquired or processed that can quietly change how well a model works. Bottom line: BenchX argues that if you only test medical AI on clean average-case data, you can miss exactly where it is least reliable. For clinical AI, subgroup-level and protocol-level evaluation is not a nice-to-have. It is the real test.
...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. ...
First, OpenAI and Broadcom have unveiled Jalapeño, a custom chip built for LLM inference. OpenAI says the chip is designed to improve performance, efficiency, and scale across AI systems. There are no technical specs in the supplied materials, so for now the main point is strategic: another frontier lab is moving deeper into custom infrastructure.
Next, Bloomberg reports that two senior Google AI researchers, Jonas Adler and Alexander Pritzel, are set to leave for Anthropic, according to people familiar with the matter. Bloomberg says both were seen internally as key contributors to Gemini work. Google DeepMind CEO Demis Hassabis, speaking this week, described the market for AI talent as ferociously competitive.
And finally, a nice practitioner note from Simon Willison. Inspired by Mozilla’s new MDN MCP service, he says he tried converting MDN’s browser compatibility data into a SQLite database. The project also used AI-assisted tooling to help build the workflow, and the resulting database is available in a form that can be explored in the browser.
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!