Jonathan Li

building @induction_labs | prev. reasoning @ cohere

[Jonathan Li @jonat_li]

what should a model learn next?

eventually, we think the model should decide for itself. intrinsic discovery is our first step toward that.

[Induction Labs @induction_labs]

We’re introducing intrinsic discovery, a new learning method that allows models to explore and learn from open-ended environments without human supervision.

Using intrinsic discovery, we trained a world model that surpasses GPT-5.6 Sol at simulating terminal environments.

[Jonathan Li @jonat_li]

had the privilege of testing this model before it was released. it's good for its size, and completely insane given the weights are ∈ {-1, 0, 1}

[DeepGrove @deepgrove_ai]

We introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM, SOTA in its weight class.

It solves IMO-level problems and runs at 200+ tokens/s on a Mac Mini M4, 5–16× faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss.

https://t.co/l2y5eZTVS4

[Jonathan Li @jonat_li]

@ShaamilKarim1 congrats, i'm so bullish on @ShaamilKarim1 and the team

[Ivan Zhang @1vnzh]

always cool to see past @cohere teammates build great technology

many such cases

[Harj Taggar @harjtaggar]

Induction’s imagination models are a bet that internet-scale video can teach models how computers actually behave. It's a hugely ambitious vision and @jonat_li is exactly the kind of brilliant 19 year old who can pull it off.

[Induction Labs @induction_labs]

We’re introducing imagination models: a new foundation model architecture that unlocks learning from internet-scale video.

Our first imagination model, Photon-1, learned to use a computer by watching 18 years of screen recording video without action labels.

[Jonathan Li @jonat_li]

@eliebakouch why not just save the last hidden state and compute the full logits on the fly at train time?