Building a fab is pretty much impossible. You can give TSMC a ton of money to build more capacity, but so can everyone else in the space.

I've mostly had non-graded homework in my studies because cheating was always easy. In highschool they might have told your parents if you don't do homework. In university you do what you want. It's never been an issue overall.

Was that an actual SVG?

No that's GPT-4o native image output.

I wonder how far away we are from models which, given this prompt, generate that image in the first step in their chain-of-thought and then use it as a reference to generate SVG code.

It could be useful for much more than just silly benchmarks, there's a reason why physics students are taught to draw a diagram before attempting a problem.

Someone managed to get ChatGPT to render the image using GPT-4o, then save that image to a Code Interpreter container and run Python code with OpenCV to trace the edges and produce an SVG: https://bsky.app/profile/btucker.net/post/3lla7extk5c2u

Does this match the rules of your test, or is it cheating? :)

It possible with crypto shredding. You store everything encrypted with one key by customer. When you want to delete you erase the key. The data becomes unusable everywhere (backups included). Then a job periodically garbage collects data without a key, but that's more for cost saving.

Big companies do this but it requires some technical maturity. If you operate in Europe you have to implement proper data deletion. I would be more worried about small companies that large ones tombe honest.

You still have to backup those keys somewhere... and if you don't do it the same way as for the data then your backups are effectively worthless.

Much less data to back up so it can be stored in a way that is replicated for redundancy but still mutable. Separating the key and data is what allows for sending data to tape backup etc

If your (backup-via-redundancy) keys are mutable, you do not have a backup. What happens in the case of a ransomware attack, for example?

You've also added (possibly substantial) latency to every single operation that operates on user data.

The specifics of how the keys are backed against different failure modes/attacks is orthogonal to the splitting of data/key.

Yes you would need to carefully design the system that allows deletion of keys while minimizing chances of data loss, but it can be done, and it's going to be cheaper and less complex to do so on a tiny subset of the data.

Latency considerations are also down to design, it's not a given that there will be significant overhead imposed.

One simple way is to keep only a few days / weeks of (immutable) keys backups. You can always stop the deletion of you have a big issue. If the law says you have 14 days to delete all data, you keep only that much backups.

Yea was going to say this. I would be surprised if any major company does this for normal uses.

That's how GAFAMs size companies do it. They need to comply with European regulations for data deletion, and there are very few other options.

Loki was much cheaper to run in my experience, using S3 storage. And you could scale the parts you needed dynamically in K8s.

Elastic was kind of a resource hog and much more expensive for the same amount of data.

That might be dependent on your use case though.

Both models mentionned in the article are available, Gemini robotics for partners only, and Gemini robotics ER in private preview.

It can be very useful for grid stabilization at the scale of seconds to minutes (both with charge and discharge).You don't need that much capacity to do that.

They do have to be wide open and close pretty fast once the end on the booster had passed them.

The rocket is not filled until the last minute, by fueling arms on the tower. And the weight is like 90% fuel, so it makes a pretty big difference.

They said in the event that they will first deliver 3s and Ys. They just add a 2 seater and 20 seater van later to fill other use cases.