• 5 Posts
  • 208 Comments
Joined 1 year ago
cake
Cake day: June 4th, 2023

help-circle





  • When this whole ‘training’ trend started a few years ago, there were companies offering image and video labelling services.

    It turned out they were mostly sweatshops in low-income countries, where people sat in front of monitors and just dragged boumding boxes around sections of images and picked from an icon menu. Here’s a car, here’s a person, here’s an apple. That sort of thing. You didn’t even need to know how to read or write.

    Of course, the quality was questionable, so they needed a second layer of supervisors verifying the choices. But even with that, the cost was way lower than having an engineer or QA person do it. IIRC, there was a bit of hue and cry when stories came out of big tech companies supporting sweatshop conditions.

    Sounds like it’s still ongoing.




  • I’ve been using ChatGPT, specialized ones on Huggingface, and a bunch of local ones using ollama. A colleague who is into this deep says Claude is giving him best results.

    Thing is, depends on the task. For coding, I’ve found all suck. ChatGPT gets you up to a point, then puts out completely wrong stuff. Gemini, Microsoft, and CodeWhisperer put out half-baked rubbish. If you don’t already know the domain, it will be frustrating finding the bugs.

    For images, I’ve tried DALL-E for placeholder graphics. Problem is, if you change a single prompt element to refine the output, it will generate completely different images with no way to go back. Same with Adobe generators. Folks have recommended Stability for related images. Will be trying that next.

    Most LLMs are just barely acceptable. Good for casual messing around, but I wouldn’t bet the business on any of them. Once the novelty wears off, and the CFOs tally up the costs, my prediction is a lot of these are going away.


  • Installed RabbitMQ for use in Python Celery (for task queue and crontab). Was pleasantly surprised it also offered MQTT support.

    Was originally planning on using a third-party, commercial combo websocket/push notification service. But between RabbitMQ/MQTT with websockets and Firebase Cloud Messaging, I’m getting all of it: queuing, MQTT pubsub, and cross-platform push, all for free. 🎉

    It all runs nicely in Docker and when time to deploy and scale, trust RabbitMQ more since it has solid cluster support.









  • I’ve always kept a strict separation between work and personal projects, including a personal laptop, accounts, and yes, paying for AI services. For a while, a few years ago, while commuting on the company shuttle, I even had my own MiFi cell access point and a laptop battery booster so I could work on my own projects on the bus and not be accused of using company resources.

    Most employment contracts spell out that anything you create using company resources is the property of the company. Legally, they own everything that passes though their computers, software, and networks.

    Also, many corporations run system monitoring services on their laptops and MDM mobile data management on mobile phones (for example JAMF on Apple devices). These monitor things like file access, copying, communications, and web access. This data is sent to central servers for processing and looking for anomalies based on pre-set rules. This might sound tin-foily, but it’s mandated by legal in a lot of companies, including small and medium sized ones.

    If you want to use non-company data to do AI work, or develop a service or idea on your own, or even keep your text messages and email private, you’ll want to use your own equipment, accounts, and services.

    Edit: also, if you get laid-off or fired, you’ll want to have a decent personal rig so you can continue working on your own projects while looking for work. Even if working on a novel on the side, suggest keeping everything off company systems.



  • We used to have the same problem. Years ago, a relative recommended a Miele canister-style. They were pretty pricey, but took a chance. It could practically pull the floorboards up (fortunately, the power level is adjustable). It lasted forever.

    Replaced it with the same brand. Apparently, some models are now made overseas and use cheaper components, but the higher-end models are still made in Germany. Totally worth it.