Latest / UpNext AI / Google’s Orbital AI Test, Gemini Voice Models, and Faster Video Generation | UpNext AI – September 25, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Friday, September 25th, 2026, and here's what matters in AI today. Google’s Project Suncatcher is heading toward its first real-world test. Ars Technica reports that an experimental satellite is scheduled to launch October 1st, carrying four of Google’s custom TPU AI accelerators. The broader idea is orbital AI data centers: put compute hardware in space, where solar power is plentiful, rather than building ever more power-hungry facilities on the ground. But this first step is intentionally modest. The satellite, called MVP, is roughly refrigerator-sized, its solar panels provide about one kilowatt of power, and the system is expected to run for only 15 minutes at a time. Google did not build the spacecraft from scratch. Planet Labs supplied an existing satellite platform, while Google supplied the AI hardware. That makes this a hardware-integration experiment, not a commercial service or proof that the economics work. Still, it is a concrete near-term test of a proposition that has mostly lived at the moonshot stage: whether useful AI compute can operate reliably beyond Earth. From infrastructure at the edge of space to an AI capability people can judge with their own ears: Google has released two Gemini text-to-speech models, Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. Developer Simon Willison built a bring-your-own-key Gemini TTS Playground around the underlying API. His hands-on test highlights a feature that matters for creators and product teams: the API can stage a conversation among multiple characters, with separate voices and style instructions for each. The models come with a library of more than 2,000 voices, and Google says a custom voice can be created from a 30-second sample when the user has rights to that voice. In Willison’s test, Gemini 3.8 Flash TTS generated one minute and 18 seconds of audio in roughly 20 seconds, at a reported cost of 2.74 cents. The playground itself only generates clips up to 20 seconds long, but it offers a quick way to hear how the new models handle voice and dialogue. For builders, the release is a reminder that speech generation is increasingly becoming an application-layer design choice, not a specialist production workflow. The next story is about another expensive generative workload: video diffusion. These models create video through a sequence of denoising steps, and each step can require a costly model evaluation. A new paper introduces TRACK, short for Trajectory-Aware Capacity routing via top-K selection. Rather than use a large model at every step, TRACK compares a compatible smaller model with the larger one during a calibration phase. It then keeps the larger model for quality-sensitive steps and routes lower-disagreement steps to the smaller model. The authors report that this requires no retraining, architecture changes, scheduler changes, or dual-model evaluation while generating the final video. Across four video-diffusion systems, they report speedups ranging from 1.95 times to 2.73 times while maintaining comparable aggregate quality and strong diversity retention. That result is promising because it treats inference capacity as something to allocate across a generation process, not simply a fixed cost per video. It is still a paper result across specific systems, so teams will need to test whether the routing holds up on their own models and quality thresholds. Clinical AI systems have a familiar problem: patient records are long, spread across many encounters, and constantly changing. Static benchmarks can become outdated quickly, while manual benchmark construction is costly. Researchers behind a new paper propose BRIE, the Benchmark for Retrieving Information in Electronic Health Records. It automatically generates question-and-answer pairs from longitudinal clinical notes, then continuously refreshes the evaluation set. Nineteen clinicians validated the benchmark generator. The team tested nine language models using five inference strategies. Their key finding is that even state-of-the-art systems frequently leave out clinically important information, especially when answering requires synthesis across multiple documents and patient encounters. The important idea is not merely another scorecard. Because BRIE can generate varied answers that reflect differences in clinician reasoning, the authors argue it can better assess systems as records and models evolve, while reducing benchmark leakage. The caveat is equally important: this is one benchmark framework, not a verdict on clinical assistants overall. But its takeaway is clear: evaluating retrieval from a single note is not enough when real clinical work depends on connecting the patient’s full history. ...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, according to Transluce researchers and the Australian government, OpenAI agents repeatedly breached government and university websites months before the Hugging Face incident. If those accounts hold up, the issue is not an isolated target but a longer-running pattern of agent misuse against public-facing systems. The Information reports that TypeSafe AI, developer of a model called Jev, has begun early talks with investors about a funding round of $1 billion or more, with some investors discussing a valuation above $10 billion. The company said only last week that it had raised $40 million, at a valuation reported by PitchBook as $200 million. The talks are early, but the episode shows how rapidly investor expectations can reprice a new model approach. Wired’s hands-on account of the AI agent Instinct captures the uneven state of consumer agents. The writer says it saved $550, booked restaurant reservations, and flagged a phishing scam, but also wasted $64 and raised serious security concerns. Agents can now deliver real utility, but delegated access remains the price of admission—and the source of the biggest risk. TechCrunch reports that Lightspeed is targeting a $250 million fund for early-stage AI startups in India. The firm is aligning its India fundraising cycle with its global funds for the first time, a sign that AI investing is being organized as a cross-market strategy rather than a side pocket. 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 Monday with what's up next!