Google just leaked its own video model days before I/O.
Gemini Omni showed up in live interfaces, complete with remix tools and in-chat editing.
Classic Google can't keep a secret to save their life.
But wait, there's more:
A tabular ML model that handles 1 million rows instantly
Mira Murati's startup just dropped a real-time AI that interrupts you mid-sentence
I'm Alex. Welcome to L8R by Innov8.
Let's dive deep 🐰
In today's post:
Google's Gemini Omni video model leaks before I/O
TabPFN-3 runs ML on 1M rows without training
Mira Murati's lab drops real-time AI that interrupts you
Google's Secret Gemini Omni Video AI Just Leaked
Google's massive new video AI just broke cover right before their biggest event of the year.
Leaked screenshots show a secret model called Gemini Omni hiding inside the live Gemini app.
This means you will soon edit and remix high-quality videos right inside your normal chat window.
The details:
Early testers found hidden menus for Gemini Omni that let you remix videos and use templates directly in your chat.
The model easily passed the famous spaghetti test for video consistency and perfectly rendered a professor writing math on a chalkboard.
Generating just two short video clips drained 86% of a user's daily AI Pro quota.
Industry experts expect Google to officially announce the tool at Google I/O on May 19, 2026.
Why it matters:
Video AI is finally moving away from clunky, separate websites.
Google wants to put text, images, and video into one single brain.
When you can type a prompt and instantly fix a video clip in the same chat, making content gets a hundred times faster.
This forces every other AI company to rethink how their apps actually work.
💡 L8R's Take:
Google is making the right move by putting all these tools in one simple chat window.
But that massive 86% drain on your daily quota is a giant red flag.
If they do not fix those crazy computing costs before launch, this tool will be completely useless for everyday creators.
TabPFN-3 Makes Tabular Machine Learning Instant
Prior Labs just released TabPFN-3, a massive upgrade to their tabular foundation model.
It makes predictions on huge spreadsheets in a single step without any training time or setup.
This completely changes how businesses handle their data.
The details:
The model handles datasets up to 1 million rows, which is 10 times bigger than the previous version.
It spits out predictions for 1 million samples in just 0.2 seconds using a single H100 GPU.
A new API feature called 'Thinking Mode' uses extra test-time compute to boost prediction accuracy.
TabPFN-3 beats classic machine learning methods 93% of the time on the TabArena benchmark.
Prior Labs released this right as they entered a deal to be acquired by SAP for a massive €1 billion investment.
Why it matters:
You usually need a whole team of data scientists to build good predictive models
. They spend days tuning settings and waiting for models to train. TabPFN-3 skips all that boring work.
You just feed it data and get highly accurate answers right away.
💡 L8R's Take:
This is the death of traditional gradient-boosted trees for standard business data.
If you are still wasting days tuning models, you are doing it wrong.
TabPFN-3 is the new standard, and SAP made a brilliant move snatching them up.
AI Writes a Zero-Day Exploit for Cybercriminals
Voice AI is finally learning how to have a real conversation.
Thinking Machines Lab, the new startup from former OpenAI CTO Mira Murati, just dropped a model that listens and speaks at the exact same time.
It means the end of awkward pauses and waiting your turn to talk to a machine.
The details:
The new TML-Interaction-Small model runs on 200-millisecond micro-turns so it can process audio, video, and text instantly.
It uses a fast front-end model for quick chats and a smart 276-billion parameter background model for heavy reasoning.
Response time hits lightning-fast speeds of just 0.40 seconds.
Mira Murati teamed up with AI heavyweights like PyTorch creator Soumith Chintala to build this system.
The research preview launched on May 11, 2026, with a full public release planned for late 2026.
Why it matters:
Current voice bots use clunky external tools to figure out when you stop talking.
This creates a big bottleneck that ruins the flow.
Thinking Machines built the interactivity right into the core of the AI.
Now you can interrupt the bot, change your mind mid-sentence, and actually talk to it like a normal human.
💡 L8R's Take:
OpenAI and Google are going to sweat over this.
Murati knows exactly where her old bosses are weak, and she built a weapon to exploit it.
Voice AI without the awkward lag is the real killer app of 2026.
🚀 Quick L8R Summary
Gemini Omni: Google's video generation model leaked before I/O, showing remix and in-chat editing features.
TabPFN-3: Prior Labs dropped a tabular ML model that handles 1M rows instantlyno training needed.
Interaction Model: Mira Murati's startup built real-time voice AI that thinks in 200ms chunks, beating GPT's turn-based system.
📩 Innathe L8R engane undarunnu 👇?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha L8R il kanaam bie…👋