Latest / UpNext AI / Meta’s AI Image Opt-Out, Cross-Chip Inference, and Biomedical Agent Collaboration | UpNext AI – July 8, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Wednesday, July 8th, 2026, and here's what matters in AI today. First up, Meta is rolling out its new Muse Image model, and the big listener-relevant detail is this: according to Wired, Instagram users with public accounts now need to opt out if they want to block AI generations made from their content. Wired reports that Muse Image is deeply integrated into Instagram, and that public photos can be used when someone tags an account in a prompt to generate an image using that person’s likeness. If your account stays public and those settings stay at their defaults, people may be able to create AI content using your Instagram posts and reels. The other important detail here is that this is not opt-in. It’s opt-out. Wired says users who want to stop future generations without making their whole account private have to change settings inside Instagram. The report also says existing AI images already created with that content will not be deleted, and users may not be notified when someone makes those generations. So the practical takeaway is simple: if you keep a public Instagram account and you do not want your photos used in Meta’s AI image features, this is a settings check worth doing now. Next, a more infrastructure-heavy story. TechCrunch reports that French startup ZML has released ZML slash LLMD, a new inference server designed to speed up AI inference across a broad mix of chips. The pitch is straightforward and potentially important. Instead of tuning around one hardware silo, ZML says its software can help open-source large language models run across several chip platforms, including Nvidia, AMD, Google TPU, Apple Metal, and Intel Arc. The company’s stated goal is to break some of the software and architecture barriers that create vendor lock-in and make it easier for enterprises and cloud providers to use a mix of hardware. That matters because inference, the work of actually processing prompts, is where a lot of AI cost now shows up. TechCrunch says ZML is positioning this as a way to make running AI less costly and potentially more energy efficient, especially if customers can mix in cheaper or lower-power hardware. A couple of caveats are important. This is a release, not proof of broad adoption. And while the company is described as a hot French AI startup endorsed by Yann LeCun, the bigger market impact is still something to watch. TechCrunch also reports that the Paris-based company has a team of about 20 people and previously raised $20 million. Still, the broader theme is clear: more of the AI stack is now about inference efficiency, and startups think there is room to compete not just on models, but on the software layer that decides where and how those models run. For today’s research section, a paper on biomedical question answering called From Voting to Agent Collaboration: Answer-Type-Aware LLM Pipelines for BioASQ 14b. The core idea is practical. Instead of using one prompting strategy for every question, the researchers built different workflows for different question types. Yes-or-no questions used snippet shuffling and self-reflection to make answers less sensitive to evidence order. Factoid questions used full snippets plus chain-of-thought style in-context learning to improve biomedical entity identification. And list questions used a multi-agent setup, where evidence extraction, candidate generation, verification, and final aggregation were handled collaboratively. The paper says the system was evaluated in the official BioASQ 14b Task B challenge and achieved competitive results across multiple batches, including first place in the factoid subtask of Batch 4. Bottom line: in specialized domains like biomedicine, matching the reasoning pipeline to the question type appears to improve reliability. ...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. ... The Decoder reports that OpenAI’s GPT-5.6 is launching Thursday after an earlier delay tied to the U.S. government and additional testing. Ars Technica reports that hackers can use nine of the most popular AI tools to assemble massive botnets, using a tactic it calls HalluSquatting that exploits large language models’ inability to say, “I don’t know.” And OpenAI says Australian Payments Plus is using ChatGPT Enterprise and Codex to move faster through payments complexity, while keeping human judgment central. 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!