OpenAI's ex CTO finally drops her first model

@ai.samaritan

12.4K followers · usually 3.8K views a post

Views

141.4K

Likes

6.9K

Comments

74

Shares

1.1K

SoundTikTok sound · Vivaldi-Four Season-Summer-3

Promoted product

Inklingthinkingmachines.ai

Connectionism: Research Blog by Thinking Machines Lab

Why it worked

The video highlights a novel AI model that can self-fine-tune and offers significant customization, appealing to developers and AI enthusiasts. The demonstration of the model learning to avoid a specific letter is a compelling and easily understandable example of its capabilities.

Summary

This video introduces Inkling, a new AI model from Thinking Machines, founded by OpenAI's ex-CTO. The model is unique because it's designed for customization and can be fine-tuned by users on their own data, with a demonstration showing it learning to avoid using the letter 'E'.

Structure

1Introduction of Inkling and its founder
2Explanation of Inkling's open-weights and customization focus
3Demonstration of self-fine-tuning to avoid the letter 'E'
4

Original caption

This new new AI model fine-tunes itself! Thinking labs finally drops their first model #aiupdates #ainews

More from @ai.samaritan

Product placement

Thinking Machines (ACTS - develops and releases AI models like Inkling, provides a platform called Tinker for fine-tuning, and an Inkling Playground. Removing it would change the video's focus on AI model development and fine-tuning.) Inkling (ACTS - an AI model developed by Thinking Machines that can reason across text, image, and audio, and can be fine-tuned. Removing it would change the video's subject.) Tinker (ACTS - a platform by Thinking Machines for fine-tuning AI models. Removing it would change the video's demonstration of the fine-tuning process.) OpenAI (APPEARS - mentioned as the former employer of Mira Murati. Removing it would not change the video's core narrative about Thinking Machines and Inkling.)

On-screen text

OpenAI's ex CTO finally drops her first model MIRA MURATI'S NEW LAB JUST RELEASED THINKING MACHINES @thinkymachines Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. thinkingmachines.ai/news/introduci... Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. THINKING MACHINES IS THE WILDEST PART. TUNED ITSELF. LET ME EXPLAIN. THINKING MACHINES STARTED BY OPEN AI'S FORMER CTO IS CALLED INKLING. IT'S FULLY OPEN WEIGHTS. IS SITTING ON A PUBLIC PLATFORM RIGHT NOW. THINKING MACHINES users need models that can adapt to very different workflows, not just excel on benchmarks. Inkling GLM 5.2 Nemotron 3 Ultra Claude Fable 5 Kimi K2.6 GPT-OSS (high) THEY OPENLY ADMIT IT'S NOT THE THE WORLD. THAT'S ON PURPOSE. INSTEAD OF COMPETING THEY BUILT IT BEST MODEL TO CUSTOMISE, MEANING YOU TAKE AND TRAIN IT ON YOUR OWN DATA, YOUR OWN WORKFLOWS, YOUR OWN PRODUCT. THE POINT, THEY HAD INKLING CUSTOMISE ITSELF. TO BECOME A MODEL THAT NEVER USES THE LETTER E. ITS OWN TRAINING JOB, RAN IT, EVALUATED AND THEN LOADED ITS OWN NEW BRAIN AND STARTED SINGLE E. THE FULL LOOP, NO HUMAN IN IT. DETAIL BUILDERS SHOULD KNOW. IT HAS A CONTROLLABLE THINKING EFFORT. TESTS, IT MATCHED A LEADING OPEN MODELS CODING PERFORMANCE USING ROUGHLY 1/3 WHICH MEANS A COST FOR THE SAME RESULT. THE FRONTIER LABS SELL YOU INTELLIGENCE. BETTING THE REAL MARKET IS OWNERSHIP MODELS YOU SHAPE AND CONTROL.

Transcript

Mira Murati's new lab just released Thinking Machines started by OpenAI's former CTO is called Inkling. It's fully open weights. The entire thing is sitting on a public platform to tinker with right now. But here's what's interesting. They openly admit it's not the strongest model in the world. That's on purpose. Instead of competing on benchmarks, they built it the best model to customize, meaning you take the base model and train it on your own data, your own workflows, your own product. And to prove the point, they had Inkling customize itself to become a model that never uses the letter E. It wrote its own training job, ran it, evaluated and then loaded its own new brain and started answering without a single E. The full loop, no human in it. There's one more detail builders should know. It has a controllable thinking effort. Dial in their tests, it matched a leading open models coding performance using roughly 1/3, which means a cost for the same result. The frontier labs sell you intelligence. Betting the real market is ownership, models you shape and control.