Aaron Wang

My hot take is that most AI bills are paying flagship rates for workflows a 10x-cheaper model does just as well. Through @EragonAI's model router, we benchmarked routed multi-model configs against Anthropic models on common enterprise workflows we see across customers: 97% of frontier quality at 56% of the cost — and every point on the cost-quality frontier is a routed config, not a single model. One model for everything was always a transition phase. The future lies in routing the right tasks to the right model at the right price.

What you keep vs. what you pay, per config: - Best Performance — 101% of Opus quality · 94% of cost; - Balanced — 97% · 56% - Cost-Optimized — 92% · 19% We took the average of 5 runs per config judged by an LLM, across 3 real operational workflows: deep research, email triage, vendor analysis. The mechanism: an optimizer runs avg-of-5 judged trials per candidate route, holds a statistical guardrail against the all-Opus baseline, and rolls back any step substitution that degrades quality. For example: the Slack-fetch step tolerates a 55x cheaper model. The decide step doesn't. Routing knows the difference. Eragon routes every step of your agent's workflow to the model that's earned it — so every dollar of your AI spend is actually doing work.

Frontier models handle most operational workflows no better than models 10x cheaper, which is why model routing is crucial: each step goes to the cheapest model that holds output quality, with an optimizer picking the assignment. Our benchmark on Eragon's model routing config generates the following result: 97% of all-Opus quality at 56% of the cost on real assistant workflows for the Balanced routed config — and the 3 routed configs anchor the cost-quality Pareto frontier.

@joshua_sirota Office warming party incoming!

I guess what I’m trying to say is that if technology = technical objects and the externalization of human reasoning, it’s kind of counterintuitive to even adopt the position of technopessimism?

When people say they are a techno-optimist, what do they mean by technology? 🤔

Kinda crazy that @EragonAI just navigated through the entire DoorDash page via browser use and ordered us our favorite Oren's Hummus 😎🤑

AI won’t be a feature. It will be the operating layer that company workflows embed itself in. That means more than calling a model API, but owning the model that understands a company's unique ways of being. This is Eragon's manifesto.

SF is great for Vietnamese, Japanese, Nepalese, Thai, and "New American" food. It is NOT great for Italian or Chinese food, which it's traditionally famous for. This is one reason SF gets unfairly maligned as a food city.

When I have to raise a point in family discussions:

agents will take care of all details for your company in the future @EragonAI just filled out a contract for us, no human input required insane

the role of Walter Benjamin as the sacred text of university spirituality is really overdue for some pushback. so many papers ending with a prayerful Benjamin citation, it's really cringe