Microsoft Lobe

» see lobe.ai as it lived, on the wayback machine

~ the obit ~

i work on edge ai for uav computer vision now. the drag-and-drop machine-learning tool category never made it into anyone's production pipeline. it was always demo-stage tooling, the kind of thing you put on stage at build or google i/o to convince a non-developer that ml was approachable. lobe was the prettiest of those tools. it still died.

lobe started in san francisco around 2015. mike matas had come out of apple, where he worked on the iphone and ipad, then facebook paper. adam menges and markus beissinger filled in the ml and product sides. the app was a single mac and windows desktop binary that ran tensorflow underneath, gave you a clean ui for tagging images, and let you watch the loss curve come down in real time. you could export to coreml or onnx and move the model into whatever stack you actually shipped on.

microsoft acquired lobe in september 2018, terms not disclosed. it was the same season microsoft bought github, around the time satya nadella's ai narrative was still about cognitive services rather than copilot. lobe got rebranded as microsoft lobe, the desktop app stayed free, the website moved to lobe.ai under microsoft hosting. the build 2020 keynote demoed it. there was a second of momentum.

then nothing for three years. the github page noted in early 2021 that the app was not in active development. the desktop release stayed pinned at version 0.10, you could still download it, it kept training image classifiers, microsoft never said it was dead. official deprecation arrived in august 2024 when microsoft shut down the lobe.ai website and open-sourced the codebase on github, the standard way large companies bury a product without holding a funeral.

~ the rap sheet ~

~ what it actually did ~

image classification, mostly. you opened the app, made a project, dragged in folders of labeled images, and the app trained a mobilenet-class network on your laptop. training ran in seconds for small datasets. there was a webcam preview mode where you could point your camera at the object and watch the prediction confidence shift live. it was the cleanest ui anyone had built for the boring 90% of ml work, which is collecting and labeling data.

the export options were the part that made it useful. tensorflow lite for android, coreml for ios, onnx for the cross-platform case, and tensorflow.js for the browser. you could prototype a classifier in lobe in an afternoon and have a working mobile demo by the end of the day.

~ how it ended ~

microsoft never built a paid tier. there was no premium lobe and no azure-hosted lobe. the entire product was a free download with no revenue stream attached. when the copilot wave hit in 2023 and every microsoft ai dollar moved toward openai integration, lobe had no internal champion left. the desktop app category microsoft was building around lobe (a real local-first ml workflow) was not where the strategy was going.

the github archive is the legacy. the lobe-python source is up. the desktop binary still runs on windows 11 if you can find a backup of the installer. it just no longer has a website or a future.

~ what survived ~

the no-code ml app category lost its cleanest entrant and never really recovered. teachable machine from google still exists as a web demo. apple has create ml inside xcode. neither one has the lobe ui quality. the segment got absorbed into the broader autoML feature inside azure ml studio and sagemaker, which is what enterprise wanted anyway.

the lobe founders left microsoft. mike matas works on his own thing now. the company they sold for an undisclosed eight-figure sum in 2018 lasted six years inside its acquirer before getting quietly shelved, which is faster than the average microsoft acquisition arc but slower than most.

@}-->--

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

« previous  ·  random  ·  next »