We had a recent discussion in a group about how many cards PSA grades a month etc. If you run the numbers, they must either have increased the number of graders massively over the last 2-3 years, spend way less time per card (maybe even both) or use some form of tech to automatize the grading process.
Does anyone have info on how exactly they increased the numbers? Obviously they will have more graders than 3 years ago, but is this the only factor that is responsible for more cards being graded per month?
They were quite transparent about the massive increase in staff in all departments, as well as their investment in automation at different stages of the process.
Adding to that the fact that the backlog is now gone, thereās little to speculate about imo. Plenty of staff, not enough product.
The way the market flushed out flippers, and the way the economy is going, wouldnāt be surprised to see specials back to $12 just so they can keep their employees busy and justify the operational costs.
I had discussion on this topic on a couple of discords and apparently the following requirement ā¦:
āIt will also provide unique card identification ā or ācard fingerprintingā ā by identifying the exact card in order to track provenance, resubmissions, condition changes and other attributes over time.ā
ā¦ is not (yet) in place as some people resubmitted cards within a couple of months apart and got their grade upgradedā¦ or maybe they only look at specific āfingerprintā such as "holo patterns but not all cards have holosā¦
āIt will also provide unique card identification ā or ācard fingerprintingā ā by identifying the exact card in order to track provenance, resubmissions, condition changes and other attributes over time.ā
People keep sending the same cards by getting different grades.
But I m pretty sure they use some form of AI as they can spot sometimes things that are hard to spot unless you look carefullyā¦ but I think the opposite is true, they do make mistake (AI) that maybe the grader follow as isā¦
Grading might actually be one of the easiest steps. I know back in 2020 they had a difficult time actually processing the volume of packages they received. The āresearchā step is probably also time consuming since you have to verify every single card. Then QA checks, encapsulation and processing the order to be shipped back. I remember having discussions in 2020 whether plastic supply was actually the limiting step
Anyway, my point is that thereās an entire logistics pipeline at PSA. You canāt just inject AI at step 4 and increase productivity of the whole process (unless that step is heavily rate limiting)
Most likely they just optimized some inefficiencies in some of the most time-consuming steps and scaled their workforce. I would be shocked if they werenāt working on AI right now. But I would be more shocked if they have a working AI solution and didnāt make some kind of announcement about it
I think the AI is not about speeding up the process but rather to increase reliability and consistency of the grade. I think PSA is all about grading ā¦ grading is key and more reliable , greater is the reputation and worthiness of the graded card.
I heard a couple of days ago that in the ācoinā collectible there was a company (CAC) that graded the grade provided by the grading company ā¦ it was like putting a grade on how reliable PSA grade is ā¦ Hope we donāt arrive at this level.
CAC coins:
" Certified coins of the same grade can be of varying quality. Many of todayās collectors want coins that are solid or premium quality for their assigned grade ."
In my opinion if in 10 seconds per side you can not notice flaws you can consider it GEM MT 10. It is useless to waste too much time on a card, especially in the BULK level. Even if it took them 1 minute per card, there are already a lot of cards that donāt deserve 10 and they are 10.
Grading is more to value the card in any future trades / sales. And perhaps for protection, but on this point I have some doubts. You can just use a cardsaver if you want to protect.
BGS and CGC giving sub-grades arenāt consistent either, so why spend too much time on a card? Those few seconds are enough.
We probably spend more time checking the card if it is perfect or not before sending it to grading than the PSA grader
If they use AI to grade cards, once they improve the algorithm, the grading standard will change, making it inconsistent from the past standard.
Most of the times, itās just doing puzzle math, like 5 scratches goes to a 9 automatically. 3 whitening drops another 0.25. surface roughness -0.15 ~ -0.25. Can hardly say its actual AI.
the only difficult part is how to detect all types of damage, and this is solvable by using those surface detectors (like the machines for jewelry). So, the tech is already available (this is just one of the solutions), would be easy to implement.
Once they have this, they can use the algorithm to assist human graders, but I would not use these AI to grade cards, if the AI goes wrong, it is totally possible to go very wrong.
I can understand how they would fingerprint holos, but it would be much harder to fingerprint non-holos. Would love to see how they do it, if they even do it for non-holosā¦
I personally agree with the points you made. I highly doubt they swapped out human graders with machines/AI. Also, IF they did that, it would lead to a high level of inconsistency of grading because they would have to replace the whole grading department at once. Otherwise, parts of your grading would be automatized and others would still be the same as before.
As you mentioned, thereās more to it than just grading, i.e. logistics, slabbing the cards etc. I could imagine though that somewhere in the grading process machines/AI is used to āpregradeā, maybe to help the grader in some way such that instead of letās say 30 seconds per card, he needs 20 instead, enabling him to grade more efficiently.
I think this is a great point. The way I see it working best would be if the AI would count as an additional grader. Say each card is reviewed independently by 3 humans and the AI and the AI is more of a check to make sure the all graders are close to each other. If the AI gives something a 6 and everyone else is at a 10 maybe that triggers a review or something along those lines.
Maybe not to grade cards, but the logistic around grading cards is probably a boon for AI and other tech improvements. I would guess that the grading of each card is pretty cheap, <$1 per card. The rest of the cost is getting it to the grader, encapsulation, and back to the customer.
There have probably been some substantial upgrades in these other areas. I work in AI and a lot of the value comes from really the boring process areas people donāt think about. Card identification and QA checks I would imagine could be entirely automated at this point, if the ROI is justifiable. Also think about the supporting processes, with all the new data available they are probably using some new techniques to predict demand, competition, finances, materials, costs, shippingā¦a company can really make some margin by improving those areas.
For grading cards, AI could probably be used for whitening, centering, and edges. As far as the rest of the card there just arenāt enough samples for each card coming through to build a reliable model. The unique art, the myriad of imperfections, and the amount/angle/resolution of photos/scans that would be needed for each card also complicates the model, in addition the hardware costs. Even if the model was affordable, the input of that model would still be the card and the graderās opinion. In the end you still would have a subjective model, just automated. People are probably just cheaper and better for the foreseeable future.
Agreed. I think card identification would be a great place for automation since I imagine it would be a huge pain in the ass for all graders to know 30,000 cards by heart
I took some ML courses during my degree but have limited practical experience, plus the field moves so fast. My intuition was that the size of of the input data would be very limiting because it would need to be both high-resolution and multi-dimensional (front, back, angles, depth for dent detection, etc) as you mention. Another limitation I can imagine is that even something as simple as swapping the scanner model could require retraining unless you have a really diverse dataset to begin with. And also, as you mention, the generalization problem. I can see a model being able to learn features like āwhiteningā is or a āscratchā or āocā or ācorner damageā on square cards. But I do wonder how well that will generalize with different arts, or like e-series borders (or no borders) or textured cards or cards that havenāt even been made yet. Pokemon backs are probably the ideal case for ML because of how consistent they are. But with fronts it is infeasible to train a model for literally every art so generalizability is very important.
Almost certainly AI will be limited to a āsuggestionā for human grader rather than a replacement - at least today.
Not sure where Iām going with any of this, just thinking out loud
I was actually implementing this stuff in the new coming app, different lighting can easily destroy the whole AI model, strong sun lights are a huge killer. Especially those strong wavy texture ultra-rare modern cards, nearly impossible, I have yet seen a good model that works under the sun. but these models are fairly consistent with a very dark color background, so itās good in certain cases.
On the other hand, If they really implement AI into grading, how would they handle the card into the machine would be another challenge, could easily damage the card. this is one reason I said AI grading is very risky.
Yeah honestly when I think about it it would even be hard to train centering on modern full arts because itās not like the good ol days when you could just count the number of yellow pixels on one side compared to the other