We do have a platform we'll be launching as well to manage training + serving for you which will require more diligent privacy guarantees.

wmo routes requests between frontier models and open source models that continuously train using Tinker. As the smaller models improve, more traffic gets routed to them.

Calculating cost is just tokens in/out.

It does this by simulating production tool responses through text world modeling (similar to QwenAgentWorld, summary here https://x.com/silennai/status/2073887455884058814).

We can then use this to train a router for frontier, OS, and local models (use defaults or pick which ones to optimize against).

wmo ingests agent traces, builds the simulation, embeds the traces, runs different models you choose against the simulation scenarios, and then uses a KNN for model selection (similar to https://arxiv.org/abs/2505.19797).

- Cache aware: cache is taken into account for the effective price in routing.

- Confidence gated: we don't deviate from the best fit model when paired evidence over retrieved neighbors is below 0.5 standard errors or on queries unlike anything in the fit set.

- Optimize for cost or quality: train a balanced, cost max, or quality max router.

Usage

`wmo build` creates the simulation (or add your own benchmark)

`wmo optimize` tunes the router

`wmo serve` starts the server and can run everything fully locally. The simulation and router can update over time as more agent traces are gathered and new models are added.

Router results vs Fable

- RouterBench: -66.5% cost, -1.7% performance, -24.7% latency p50. 77.5% of traffic to Sonnet 5, 16.1% Fable 5.

- TauBench: -44.5% cost, +6.3% performance, -20% latency. 83% to Opus 5, 17% to Kimi-K2.6 (over K3).

- Terminal Bench 2: -64% cost, +8% performance, -50.6% latency. Sonnet 5 is fully along the pareto front. Training a specialized router per task isn't cheap. In sparse data regimes the value can be "here's the best model".

We're working on sample effiient continual learning for agent specific models at experientiallabs.ai"

Comments URL: https://news.ycombinator.com/item?id=49063454

Points: 59

# Comments: 27

Basically, an LLM pretends to be a virtual machine executing instructions. Based on GEPA and Qwen AgentWorld.

Just - clone, - `uv sync` - `uv run wmh build` and you'll get a wizard that will help you create your own world model from your traces.

http://github.com/experientiallabs/world-model-harness

Comments URL: https://news.ycombinator.com/item?id=48735124

Points: 2

# Comments: 1

"Physics simulation involves discontinuities (contacts, friction regimes), long rollouts, and chaotic dynamics where small parameter changes lead to large outcome differences. Even with simulator internals, differentiating through thousands of unstable timesteps would yield noisy, high-variance gradients. Evolution is simpler and more robust for this regime." "The real tradeoff is sample-efficient but complex (RL) vs compute hungry but simple (GA). DQN extracts learning signal from every timestep and assigns credit to individual actions."

DQN likely would have handled this much better.

Comments URL: https://news.ycombinator.com/item?id=46891654

Points: 5

# Comments: 2

You save vocab_dim*model_dim params (ex. 617m for GPT-3).

But the residual stream means that the weight matrices are roughly connected via a matmul, which means they struggle to encode bigrams (commutative property enforces symmetry).

Attention + MLP adds nonlinearity, but it still means less expressivity.

Which is why they aren't SOTA, but are useful in smaller models.

Comments URL: https://news.ycombinator.com/item?id=46726671

Points: 1

# Comments: 1