8 Ways You Can Save on Your 2023 AI/ML Budget
Want to cut the 2023 budget, customize or do even more?
Here are eight battle-tested ways you can save money on the 2023 budget.
This is meant as a buffet. Pick and choose what is relevant to your case! :)
Intro
Are you on the leading edge of using AI/ML?
This is a great time to think about improving efficiency and cutting legacy costs. A medium to large team implementing most of these methods could likely save between half a million to one million a year.
Summary
I expand on each of these eight ways below:
- Reframe existing Quality Assurance work as labeling
- Move to Open Source software
- Move to Enterprise with Unlimited use model
- Use existing tools instead of recreating the data science wheel
- Keep data in one place to reduce data storage and network costs
- Engage more existing staff to do annotation
- Use Random Review over multiple people looking at same file
- Improve Annotation Efficiency
Let’s dive in!
Reframe existing Quality Assurance work as Labeling
Shift existing QA work to be done inside of Diffgram. This saves on labeling costs. We have seen companies do this well since 2019 and a big uptick in people using this approach recently. It also means their value becomes a function of how many uses cases for the data they generate, instead of just the one time use.
Move to Open Source software
Open Source Annotation software like Diffgram can reduce your licensing costs. For smaller companies who meet the license limit this means you can reduce the expense of costly alternatives like legacy Labelbox (if they are still around), V7 Labs etc.
Move to Enterprise with Unlimited Data Use Model
For larger companies Diffgram offers an Enterprise license. The key difference vs alternatives like Labelbox is the usage is unlimited due to better technology. This means you get more value for the same or lower price. It’s super easy to migrate with one click data migration. We offer easy enterprise licensing, install help, training, customization, and more.
Use existing tools instead of recreating the data science wheel
Often you can get the same or even better results by using existing tools, like AutoML. For example Diffgram has one touch integrations with Hugging Face, Azure, and GCP methods. You can create events to call your own methods, scripts, or preferred 3rd party solutions. This existing tech saves you a huge amount vs your team spinning wheels on specifics, or continuing to pay legacy firms like Labelbox. We know historically Diffgram didn’t have these integrations but now that we do you can take advantage of it!
Keep data in one place to reduce data storage & network costs
By using “pass by reference” methods you can keep your data in one place. This reduces storage and network costs.
Engage more existing staff to do annotation
Have existing staff who have low ROI cycles that can be shifted to annotation work.
Use random review over multiple people looking at same file
We have heard lots of excuses for needing multiple people but none of them really hold water. The math is really simple. Do you want 3x the expense or 1.1x the expense? If multiple people review the same file it’s 3x (x being the cost to review one file), if you randomly sample 10% for review (or your desired QA level) then it’s 1.1x. We have seen similar quality levels… for nearly 65% less cost.
Improve annotation efficiency
Of course I realize you are likely already doing this but just a reminder. From little UI search things, to pre-labeling, to interactive automations. There are many many approaches here, and judicious investment of time can yield excellent ROI. Recently one firm using Diffgram went from 1 minute 33 seconds to 8 seconds per annotation using out of the box methods. And we believe there is still more room to improve.
Improve team communication — bonus
By moving to Diffgram your teams can keep on the same page with the same Schema, up to date annotations, and more. We have seen this reduce a lot of overhead in communication in multi-team contexts.
Recap
The majority of these methods are relatively easy to implement. Some of them are just a small shift in thinking and some can be done in a few hours or days by one person or a small team.
If you have any questions or want to learn more about Enterprise options we are here to help.
