lovesdogsnsnow

This is interesting! Sort of a super mixture of experts model. What's the latency penalty paid with your router in the middle?

The pattern I often see is companies prototyping on the most expensive models, then testing smaller/faster/cheaper models to determine what is actually required for production. For which contexts and products do you foresee your approach being superior?

Given you're just passing along inference costs from backend providers and aren't taking margin, what's your long-term plan for profitability?

Aren't interconnect bandwidth and latency important for training workloads?

I had a conversation with Matt Rogers at a tradeshow a long time back. The thing that stuck with me was that despite his experience at Apple, experience building the iPod, iPad and iPhone and direct relationships with the same factories as he'd worked with at Apple, he wasn't able to get the same results at Nest.

Alernate title: No, you can't manufacture that like Apple does, even if you literally worked with Steve Jobs to build the iPod, iPad and iPhone, hire your best colleagues from Apple and work with the exact same factories as Apple uses

https://www.linkedin.com/pulse/what-founder-nest-thermostat-...

Would love to talk shop on the space.

Absolutely - this challenge is especially acute for all ecommerce retailers with broad catalogs. Target, macy’s, walmart, jet, etc all face this challenge.

Systems like Solr, elastic search and endeca (out of the box) all assume relevance means keyword frequency in a product page, with some weighting depending of title, description, tag, etc. Delivering relevant results that users might want to purchase requires taking these systems, adding or customizing their NLP techniques, operationalizing historical user search & purchase data to determine intent, personalizing by shopper history, etc.

The challenges of massive heterogenous catalog affect other areas... Chief among them search result personalization… an individual’s gaming purchase history might cause ‘button down’ to return gaming keyboards, rather than oxford shirts, while a pet products purchase history could lead to a search for turkey returning turkey dog food.

The fact that Amazon fails to personalize search results is evidence of the difficulty & opportunity here. The sort of pervasive personalization found in AirBnb, facebook, google are simply out of reach of most ecommerce retailers…

Does this mean all prefixes have to be regenerated for every addition of a new item? Or can you invalidate the relevant prefixes?