SilenN

Congrats!!

"Anonymous aggregate PostHog product telemetry is enabled by default. It never includes prompts, traces, actions, observations, paths, model names, credentials, or raw customer content."

We had this exact problem so we solved it for ourselves. Happy to help if you run into any issues.

I have a /hmmm command I use for the second part that works reasonably well: "Stop using jargon and speak coherently. State it more simply and concisely, like one human talking to another. Make it like google dev docs style. More dead prose. No aphorisms, no flourishes. Simple."

Thank you!

Thanks for the positivity tyre! If you look at our git history, we pivoted and only started building the gateway recently. Before that we were building research infrastructure that now powers the intelligence features we provide.

Telemetry is off by default. PostHog is for usage analytics on the open source repo. Audit it if you're skeptical.

We make money off enterprise licenses and hosting models.

I have strong reason to suspect you either can't read or are a bad actor.

Yep! One interesting example is often Opus 5 on low reasoning ~= Opus 5 on high reasoning.

See you soon

Ans: we rarely switch, often times it's just a "switch to using this model for your agent"

Also a hosted marketplace, not just BYOK

That is another way to do. Or we can automatically figure out which models the subagents should be using for you. And update them as new models come out and the work your subagents do changes. More than one way to skin a cat.

Yep! If you look at the commit history that's exactly what happened.

The trick is to rarely switch, or switch at task boundaries. Often the conclusion of routing is actually "this one model is actually at the pareto front for this task, just use it always".

^this

There's two ways to functionally measure this, reconstruction fidelity (which we're able to get to 0.7 - 0.95), and downstream performance (which agrees on the best and worst models, but can sometimes disagrees on the middle of the pack routing).

Exactly

Expensive, in the thousands. We have our own infra in house and are working on bringing these costs down

Technically 0 because a) it ingests your already existing traces and does an initial training run b) in the app we'll have pre-trained routers you can start with that will then learn over time

Fixed formatting which will help with readability. We do routing, distillation, and token compaction.

Let me know if you have any questions!

Thanks :)

Thanks for the heads up, removed mention!

Happy to answer any qs.

Valid criticism. Happy to answer any qs. We're still working on solidfying results.

It's open source!

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

Open source models.

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.

That's cool, thanks for sharing!

world-model-harness makes it easy to go from agent traces to faithful replication of your production environment where your agents run.

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

Thanks! I do have a section on this in the article "Why genetic algorithms aren't state of the art"

"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.

Simply, it's when your output embedding matrix = input.

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.