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 Wednesday, September 16th, 2026, and here's what matters in AI today.
TechCrunch has published a running list of AI projects and startups that have shut down, pivoted, or fallen short of expectations. It is a welcome counterweight to a launch cycle that can make every new demo sound inevitable.
The list begins with Relay, an AI workflow-automation startup that shut down this week. Relay positioned itself as an alternative to Zapier, but larger platforms such as OpenAI and Google were building similar automation into products people already used. That left a stand-alone tool fighting for a reason to exist.
The bigger lesson is that scale, reliability, and a clear use case still decide which AI products endure. OpenAI’s July attempt to reorganize ChatGPT into Chat, Codex, and Work modes drew criticism for being confusing and cluttered, and the familiar interface was restored soon afterward. Apple’s long-delayed Siri overhaul finally reached English-language users this month, after engineering issues and bugs contributed to repeated delays and a settlement over iPhone AI marketing claims.
Hardware has been no safer. Humane raised $230 million for its AI Pin, but performance problems and a charging-case safety warning preceded the shutdown of that business. HP acquired most of Humane’s assets for $116 million. The industry is learning, repeatedly, that an AI feature is not yet a product—and a product is not yet a durable business.
That execution gap matters as capital continues to set ever-larger expectations. The Financial Times reports that OpenAI is weighing a funding round at a valuation of 1.2 trillion dollars, ahead of a possible initial public offering. The report frames the discussions as an effort to capitalize on demand after recent model launches. None of that is final: the reported round, valuation, timing, and IPO remain prospective. Still, the figure shows the scale of financial expectations now attached to frontier AI.
At Dreamforce, NVIDIA and Salesforce offered a more operational view of what enterprises want from that investment. Salesforce introduced Koa, its first CRM reasoning model, built by post-training NVIDIA’s open Nemotron 3 Super model on a proprietary synthetic dataset drawn from enterprise CRM scenarios. NVIDIA says Salesforce can run Koa entirely in its own infrastructure, with Salesforce controlling the weights and no customer data used in training or inference.
The companies say Koa matches or exceeds leading-model performance on CRM actions in Salesforce’s CRM Bench, with three times fewer errors. It is already powering an employee agent in Slack and is due to enter customer pilots in October. Salesforce separately highlighted agent-security tools at Dreamforce, making the pitch clear: enterprises want models tailored to their own workflows, but with control over the systems and data around them.
For the research note, a thesis on lymphoma treatment selection makes a pointed case against treating a strong medical benchmark score as proof of safe clinical deployment. Researchers tested a commercially available large language model, guided by a custom prompt and curated knowledge base, across 50 synthetic cases of relapsed or refractory diffuse large B-cell lymphoma. The task requires weighing incomplete evidence, regulatory constraints, and several treatment options.
The model’s recommendations were judged clinically acceptable in 83 percent of evaluations, and it chose the same treatment category as the expert in 75 percent. But the important result was who changed their mind after seeing the model’s answer. General hematologists modified their treatment recommendation in 32 percent of cases, compared with 6 percent for lymphoma subspecialists.
Specialists rejected eight recommendations that generalists accepted; five of those disagreements involved options specialists considered excessively toxic or sub-optimal. In two cases with potential for harm, generalists changed their treatment toward the model’s recommendation. This is one study using synthetic cases, not a deployment verdict. Its takeaway is sharper: evaluation has to measure who is using the system and whether they can recognize its mistakes—not simply whether its average answer looks plausible.
...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. ...
NVIDIA says a power-management trial at cloud provider Lambda produced 24 percent more token throughput within the same facility power budget. The test ran 19 computing nodes at reduced power instead of 16 at full power. The broader idea is to shift power toward the work that needs it, rather than treating every rack as permanently provisioned at its peak.
The Verge reports that Meta may unveil camera-free smart glasses, reportedly codenamed Luna, as soon as next week’s Meta Connect event. The glasses would reportedly use six microphones and speakers for interaction with Meta AI and the Muse agent. Removing the camera would directly address concerns tied to recording-capable smart glasses, though Meta has not confirmed the product.
OpenAI profiles Fyxer, an AI executive-assistant product that organizes inboxes and drafts emails in an individual user’s voice. Fyxer combines OpenAI models with fine-tuning, memory, and user feedback—another example of AI assistants being designed around trust and repeatable workflow behavior, rather than a one-shot prompt.
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!