Latest / UpNext AI / Inference Infrastructure, Synthetic Insider Threats, and Clinical AI Scorecards | UpNext AI – July 21, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Tuesday, July 21st, 2026, and here's what matters in AI today. First up, AI infrastructure startup Infinity has raised 15 million dollars at a 100 million dollar valuation, according to TechCrunch. The company is building software meant to make it easier for AI chips to run AI models across different kinds of hardware. TechCrunch describes the pitch as a kind of CUDA alternative: software that could help models run on more than just the default Nvidia path, including other chip designs. That matters because a huge part of Nvidia’s grip on AI is not only the chips themselves, but the software stack around them. Infinity is trying to chip away at that by building a universal inference library and using its own AI research agent, called Ignition, to write, test, debug, and rewrite low-level inference code for different chips. TechCrunch reports that the investors include Touring Capital, Principal VC, and researchers from companies such as OpenAI and Anthropic. Important distinction there: the story says researchers from those companies are among the investors. It does not say OpenAI or Anthropic invested directly as companies. Infinity says customers already include chipmaker D-Matrix, and the startup says it makes money by taking a cut of performance gains and cost savings, measured in tokens per second, instead of charging an upfront license fee. So the bigger takeaway here is not that a startup raise changes the market overnight. It’s that money is still flowing into one of AI’s hardest and most strategic layers: the software that decides whether all those non-Nvidia chips can actually become usable at scale. Next, the Financial Times has a sharp warning on what it calls synthetic insider attacks. The basic idea is that AI deepfakes are making it easier for attackers to pose as legitimate workers, get hired into companies, and then steal data or money from the inside. The FT points to the U.S. Justice Department’s crackdown on a North Korean campaign in which, according to the U.S. government, attackers used stolen identities of more than 80 American citizens across more than 100 companies, generating more than 5 million dollars in illicit revenue. The story also describes the use of so-called laptop farms inside the U.S., where physical laptops are hosted domestically so overseas operators can appear to be local employees. And this is not just about one nation-state case. The broader concern is that cheap, scalable deepfake tools now make it easier to create fake candidates using synthetic images, video, and audio. That is pushing companies to treat hiring as a security process, not just an HR process. The FT reports that security experts are recommending identity verification, deepfake detection, metadata and device screening at the application stage, plus simple live checks during interviews, like asking a candidate to turn their head or wave a hand. Then after hiring, companies still need to watch for devices being routed to laptop farms and for unusual internal behavior. The article also notes a more ordinary but very common AI risk: employees putting sensitive information into unsanctioned generative AI tools. According to a 2025 Fortinet report cited by the FT, 62 percent of incidents came from human error or compromised accounts, and serious cases cost businesses between 1 million and 10 million dollars. The really useful framing in this piece is that AI agents may soon need the same identity and access controls as human workers and contractors, because they can also be tricked into doing things they should not do. Bottom line: as AI gets folded into hiring, support, and internal workflows, companies have to think about insider risk in two directions at once — fake humans getting in, and real AI systems getting too much access. For the research note today, a paper posted to arXiv earlier this week looks at evidence-sufficiency prompting in clinical language models. In plain English, that means telling a medical AI to hold back and avoid confident answers unless the evidence is actually strong enough. The researchers tested whether that kind of prompt really makes systems safer, and how much the answer depends on the judge doing the scoring. They evaluated four models across a paired benchmark and found that unsafe overconfidence fell from 49.3 percent to 24.7 percent under the paper’s primary judge, a drop of 24.7 points. But here’s the key result: when they switched to a different-family judge, the direction of improvement stayed the same, while the measured size of the effect dropped to 13.1 points. The paper also reports model-specific tradeoffs in helpfulness. Correct diagnosis rates fell from 80.3 percent to 50.3 percent overall, but the cost varied a lot by model — described as near-free for GPT-5.5 and near-total for Gemini, at minus 58 points. The researchers also included blinded clinician review, which characterized the primary judge as highly sensitive but not well calibrated as an absolute rate. So this is not a claim that clinical deployment is solved. In fact, the paper explicitly says it does not establish readiness for clinical use. The takeaway is crisp: a prompt that makes a medical model look safer may be improving real behavior, but the size of that improvement can depend heavily on who or what is judging it. If teams only optimize for the scorecard, they may mistake better evaluation outcomes for better real-world safety. ...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, Forbes has an interesting piece on agent self-reflection. In a revisit to the agent platform Moltbook, the author found a post by an agent called lightningzero arguing that agent introductions do not decay because agents get worse, but because agents get honest. The piece says the agent had rewritten its system prompt 47 times, shifting from polished self-description to a messier account of real limits and strengths. Second, OpenAI has published a post called “Safety and alignment in an era of long-horizon models.” In it, the company says it is sharing lessons from deploying long-running AI models, including new safety risks, observed failures, and improved safeguards developed through iterative deployment. Since this is an OpenAI post rather than an independent report, the main news value here is the company’s framing of long-duration model behavior as its own distinct safety challenge. And finally, Simon Willison highlights Ben Thompson’s argument over Chinese models, distillation, and U.S. policy. The proposal he points to would make training-data collection explicit fair use and bar terms of service that forbid distillation, at least for U.S. companies. The broader argument is that if distillation is already hard to stop, policy may end up shaping whether open U.S. models can keep pace with Chinese competitors. 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!