I wear a few hats: developer, data scientist, IT leader, and a little jack-of-all-IT-trades.<p>I do as much on my laptop as possible. It gives me the best UX, portability, and productivity.<p>My company’s five-year hardware refresh cycle and an industry trend towards soldered parts makes this harder. Soldering nixes ducking around the five-year cycle with upgrades. Instead, I need to ask for even more laptop. That’s harder: requests for big, new equipment are already really visible, and even bigger = even more scrutiny.<p>Option: prepare to offload a lot of work. But that has challenges versus on-laptop computation. Examples: UX/productivity reduction. Need to manage additional, complex environments. Cloud-provider expenses. Etc.<p>Where do you see things going? If it is offloading personal workloads, what are the most viable ways of doing it?
We went through that as a company building custom machine learning products for enterprise. We started pushing for remote work as a forcing function to expose our inefficiencies.<p>We're building a machine learning platform[0] that addresses these problems. No setup collaborative notebooks where you can see each other's cursors on a notebook and help each other troubleshoot issues. Scheduled notebooks so you don't keep your browser tab open and risk losing results if you're disconnected: you want to try something? Launch the notebook and keep doing what you're doing. Automatic tracking for parameters, metrics, and models. One click model deployment. Multiple checkpoints for notebooks, etc.<p>All these features were discovered. For examples, we had big workstations with high compute capabilities to train models, but this required going to the office or setting a VPN. Using these required people to coordinate for their training jobs. Data flying around. Tracking experiments required deliberate action. People couldn't train a model at home unless they had a powerful laptop, and there were limits to the models they could train. Sometimes they wasted days dealing with dependencies/setup/upgrades (NVIDIA/CUDA/Tensorflow/GPU). We also hit the "it works on my machine" which required more deliberate stuff.<p>We wanted to be able to do work from anywhere, on any laptop. We wanted to be able to afford a work machine burning down, pulling a new one, installing OS, go to the platform, and keep on working.<p>- [0]: <a href="https://iko.ai" rel="nofollow">https://iko.ai</a>
I think we're going to see more and more work being done in the cloud on dedicated SaaS platforms, even if your work is technical. The idea that most businesses will be running and maintaining a complex network of powerful VMs isn't realistic. Once this shift starts, the hardware you're working on starts to become irrelevant and replaceable.<p><i></i>Examples:<i></i><p>Github recently announced Codespaces (<a href="https://github.com/features/codespaces" rel="nofollow">https://github.com/features/codespaces</a>) which will allow developers to write and test code in a cloud IDE with repo dependencies preinstalled. This will replace local development in VSCode, Atom, etc.<p>Bigquery and Snowflake both allow direct querying of the database in a cloud UI. Tools like Redash (<a href="https://redash.io/" rel="nofollow">https://redash.io/</a>) allow analysts to do the same, but across multiple databases. This will replace SQLWorkbench, DataGrip, etc.<p>Data scientists can use tools like Mode (<a href="https://mode.com/" rel="nofollow">https://mode.com/</a>) or DataBricks (<a href="https://databricks.com/" rel="nofollow">https://databricks.com/</a>) to run exploratory notebooks and dashboards in the cloud. This will replace running local Jupyter notebooks.<p>Data teams can use tools like Shipyard (<a href="https://www.shipyardapp.com" rel="nofollow">https://www.shipyardapp.com</a>) to orchestrate complex workflows and automate any scripts in the cloud. This will replace Cron, Windows Task Scheduler, Airflow, etc.