Humyn Labs

In our multi-speaker audio recordings, every conversation produces three files. Host track. Guest track. Combined file. The combined file is what ends up in the training data. It's also the easiest one to fake. Here's how: submit two genuine individual tracks, but replace the combined file with a completely different recording. To most systems, it looks legitimate. Three files, all present, nothing flagged. We found exactly this. Combined files that couldn't have been produced by mixing the individual tracks submitted alongside them. The math didn't add up. Our Authenticity Verifier catches it by reconstructing what the combined file should sound like from the two individual tracks, then comparing that against what was actually submitted. Files that don't match: rejected. Files that are close but not right: flagged for human review. If the parts don't make the whole, something has gone wrong. This tool helps us find out exactly what. More on this soon. Want to talk to us? DMs open.

🟩 Intelligence needs direction. Join us in tomorrow's ThoroughScan sesh as we talk about “Human in the Loop.” Guest Speakers: @AIFlow_ML @hypetris_ @Saksham98329954 @pukerrainbrow @pukecast @RugRadio 📅 28th April ⏰ 2:00 PM UTC / 10:00 PM HKT Set your reminders 👇

🚨🤖Indian startups are building new businesses around egocentric data collection for robotics, with firms such as Humyn AI, FPV Labs, Neo Cambrian and Objectways rapidly expanding into the space.

Using AI to write a brief is allowed. But what looks right and is right are two different things. The difference is a human-in-the-loop who checks.

This isn't an isolated incident. 1,200+ documented AI hallucination cases in legal proceedings globally. 800 from US courts alone. Oregon: $109,700 in sanctions against a single attorney. Tennessee: $30,000 fine. Largest federal appellate sanction yet. Nebraska: first indefinite suspension over AI errors in the US. $145,000 in sanctions across US courts. In one quarter.

A lawyer submitted a brief with 63 citations to the Nebraska Supreme Court. The court read them carefully. The lawyer clearly didn't. This is what happens when AI gets it confidently wrong. And it's happening more than you think. 🧵

@nikitabier 10k likes on this comment and we'll send Nikita a goat

To celebrate reaching a million followers, my mentally ill friends had an actual boar delivered to the office today.

@NASAKennedy the humyn mind is beautiful

The planet can spell your name – literally. 🔤🌍 This Earth Day, see your name written in landscapes captured by Landsat: go.nasa.gov/4ak4Cdu

@agilityrobotics @sciam @denibechard Agreed. It’s the classic performance vs. competence trap. It’s easy to build a greyhound for a track but incredibly hard to build a colleague for a warehouse. Until then, we're just looking at very impressive, very expensive kinetic art.

Running a half-marathon faster than a human is impressive. Moving safely through a facility filled with people is harder. @Sciam’s @denibechard challenges the difference and asked @agilityrobotics’ Jonathan Hurst to explain it. scientificamerican.com/article/a-huma…

@SoundHound If an agent works in a demo but lags in production, did it ever really work? 💀 Great to see someone highlighting the technical debt that usually hides behind the 'wow' factor. Overall a great read!

Most AI agents look great in demos. Very few survive production. In this new TechRadar byline, our Chief Product Officer James Hom breaks down what *actually* matters when moving from proof-of-concept to real-world deployment, from reliability and latency to measurable outcomes. Check it out 👇 hubs.li/Q04d58jv0

@RSoricut @GoogleDeepMind The ability to decode vision outputs directly from generative visualizations is a game-changer. We're particularly interested in how this unified architecture handles edge cases in semantic segmentation compared to traditional specialist models. Great milestone!

Meet Vision Banana 🍌 from @GoogleDeepMind! We provide strong evidence that image generators are generalist vision learners. Traditional computer vision tasks (segmentation, depth estimation, normal prediction) can now be performed at/near SOTA with a single generalist model derived from an image generation model. 🖼️ Explore the results: vision-banana.github.io 📄 See details at: arxiv.org/abs/2604.20329

Okay... wow! Looks like most companies are trying to automate a broken experience rather than fix it first. An AI agent is only as good as the system it sits inside. Pretty appalling to see that most of those systems were never designed with the customer in mind to begin with. Worth a read!

For years, operationally focused automation metrics looked strong while customer outcomes lagged behind. That gap is now visible, measurable, and worse for the enterprise, it’s unforgivable. We just published a first-of-its-kind State of Agentic CX study. stateofcx.parloa.com It is an agent-led, human-augmented benchmark of enterprise CX across web, chat, and voice. The findings were truly alarming. What shocked us most was that the breakdown starts before a conversation even begins. Parloa’s AI agents tested global enterprises on how easy it was to get support. The pattern was clear across the board: getting help is harder than it should be. - More than half of websites required scrolling and clicks to locate support. - A third of phone numbers were buried in footers. - 37.8% presented multiple phone numbers with no clear guidance or alignment. Chat is even more constrained. Fewer than 30% of companies offer it at all. And when they do, it is often just for show, placed on the homepage but missing from the pages where customers actually need help. Ready to discover some uncomfortable truths about CX in 2026? Check out the full report.

what's your most unpopular opinion on AI that would put you in this position?

One voice can no longer hide behind many names. Introducing our proprietary Voice Fingerprinting Engine. The most damaging fraud pattern in audio data collection: one person, five email addresses, five supposedly different files. To most systems, that looks like five unique speakers. Our engine converts every verified human's audio into a mathematical representation of their unique vocal characteristics. That fingerprint is then compared against every other verified human in the dataset. Above 0.80 cosine similarity: Flagged and removed. Between 0.65 and 0.80: Sent to human review. <0.65: Likely different speakers. Thanks to this, 6.1% of submissions were caught and removed in our last major project. Before annotation. Before quality metrics. Before they became a problem. Every human voice is now a fingerprint. No disguise will be good enough to fool the math. Want to know more? DM us!

The human intelligence layer for physical AI just got a $20M commitment. Great insights from our co-founder @manishdiesel on what it actually takes to build at this scale. Real human behaviour, real environments, voice and multimodal data across 20+ countries and 33+ languages. This is exactly what we're building for. Exciting times ahead.