MetaMind

» see metamind.io as it lived, on the wayback machine

~ the obituary ~

In 2014 there was no vision API on every cloud. If you wanted a model that could tell one kind of leaf from another, you hired somebody who could train one. MetaMind's answer was a web page: upload a folder of labeled images, wait, and get back a classifier you could call over HTTP. The same site did sentiment on blocks of text. The pitch was that the hard part had already been done by somebody else and you were renting the result.

The company came out of stealth on 5 December 2014, four months old, with eight million dollars from Khosla Ventures and from Marc Benioff personally. Richard Socher was the name that mattered. He had come out of a Stanford PhD on deep learning for language, with work on ImageNet behind him, and his Stanford course material on neural networks for natural language processing was already how a generation of people in this field learned the subject. Sven Strohband, Khosla's CTO, co-founded it with him and took the chief executive seat.

The product was never the whole story, and the people who paid attention knew it. MetaMind was a research group with a billing page attached. The thing being sold to businesses was a demo of what the lab could do, and the lab turned out to be the asset.

~ the case file ~

~ what came out of the lab ~

In June 2015 the MetaMind researchers published "Ask Me Anything: Dynamic Memory Networks for Natural Language Processing", with Socher's name at the end of the author list. The dynamic memory network read a passage, took a question about it, and ran an iterative attention process that could condition each pass on what the previous pass had found, building up something the paper calls episodic memory before answering.

Eight authors out of one small Palo Alto company, and it became one of the more cited pieces of question-answering work of its period. Nothing on metamind.io's price list was worth as much as that paper, and nobody who bought an image classifier in 2015 could have told you it existed.

~ the acquihire math ~

Salesforce announced the purchase on 4 April 2016, reportedly for around $32.8 million, and would not confirm a figure. The schedule that followed was published immediately and it is unusually blunt. Uploaded data was marked for deletion on 11 April, one week after the announcement, before most customers had finished reading the news. Free web accounts stopped working on 4 May and paying monthly customers got one extra month, until 4 June. Two months from deal to dark.

Five months after that, Salesforce launched Einstein, its own branded AI layer running across the sales and service products. This is how the arithmetic works when a platform company with everyone's data buys a model company with none of it. You are not buying a product, you are buying eight people, and the product is the thing you switch off to stop them being distracted by it.

~ the next chapter ~

Socher became chief scientist at Salesforce and stayed four years, then left in 2020 to build You.com. The MetaMind research group largely became Salesforce Research, which kept publishing, which is more than most acquired labs manage.

I work on vision models that run on a drone, which is about as far from a browser-upload classifier as this field reaches, and the MetaMind pitch is still the one everyone in the business has made at some point. The hard part is not the model. The hard part is persuading somebody to hand you their data. Salesforce already had everybody's data, which is precisely why Salesforce was the buyer rather than the competitor.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »