Live webinar: how to train voice AI that sounds natural

Courtesy: Voices

Full-duplex voice models listen and speak at the same time, but only if they’re trained on data that sounds like real conversation. Open-source and synthetic datasets rarely do. Join Voices and AlphaSignal on Tuesday, October 27 at 1 PM ET as Sully Goolamnobee, Product Manager, AI Voice, covers the three requirements for full-duplex training data and how custom conversational datasets are built to meet them. Live Q&A included.

_{*This is sponsored}

THROUGH THE VALLEY

1/ Microsoft thinks it's getting close to medical superintelligence. Its AI diagnostic system, MAI-DxO, solved 83% of complex medical cases, compared with 36% for generalist doctors, according to a Science study. Impressive, although these were simulated cases, and you'd probably have specialists handling cases this difficult in a real hospital.

The numbers get more interesting when you look at the underlying models. GPT-5 already scored 79% without Microsoft's multi-agent setup, which added just four points. GPT-5.5 improved more, from 72% to nearly 87%. Makes me wonder how much of the breakthrough is clever orchestration versus simply having better models.

Mustafa Suleyman is now talking about AI predicting patients' future health from medical records as early as 2027. If that works reliably, catching diseases before symptoms appear could be a much bigger deal than winning diagnostic benchmarks.

2/ AI found a 14-year-old PHP bug. Hackers got there before the patch. Two critical vulnerabilities, called php2shell, allow attackers to execute code on vulnerable servers without logging in. Palo Alto Networks identified 22.6 million potentially vulnerable instances, and Cloudflare reportedly saw millions of exploitation attempts within hours of the news leaking.

The bugs had been sitting in PHP for roughly 14 years. A weaker version even surfaced in 2024 without much concern. OpenAI says GPT-5.6 Sol helped turn that suspicious behavior into a working exploit.

I suspect we'll see plenty more of this. There's decades of old code nobody has properly examined, and AI is getting pretty good at finding things humans missed. OpenAI has announced $1B for Patch the Planet to help organizations audit their systems, but finding bugs faster doesn't mean they'll get fixed faster. Getting millions of servers updated is still a painfully human problem.

PHP has released fixes. No complete workaround, so update if you're running it.

TRENDING TOOLS

Sentience - your digital twin that understands your world, remembers what matters, and gets things done for you _^(sponsored)
Canva Craft - Canva’s first image generation model trained on human-designed templates
Aside - AI browser remembers everything on your screen. No need to explain your context every time
Hallmark v2 - the design skill to give your agents taste. Works in Claude Code, Codex, and Opencode
PHP2Shell - Perplexity Computer's skill checks PHP sites and GitHub deployment configs for exposure
1Password for Agents - let agents handle logins without seeing passwords
Higgsfield Content Factory - turn any idea into a viral reel, from “What if” explainers to animated hooks for your ads