11 ways ChatGPT saves me hours of work every day, and why you'll never outcompete those who use AI effectively. A list for those who write code: 1 of 16

Interesting article on deepfakes used for social engineering attacks. e.g. fake videos of a manager asking you to send a payment or (yikes!) fakes of someone holding loved ones for ransom. Plus some ideas on how to protect from these attacks.

Join us in reimagining drawing for the digital age ✍️ With $100K in cash prizes and 18 winners, we’re calling on creatives to take drawing to a place it’s never been. Download Mental Canvas Draw on @surface and join the #ReimagineDrawingChallenge at https://t.co/0tO3UrYaYU 🏆

This makes it more feasible to do “only works in the laboratory” kinds of attacks since you can run them on your replica

Then you can do adversarial attacks against your replica — and those attacks often work on the original model!

One particularly interesting attack is making a replica by probing a black box model for its responses. You then train a model to produce the same response.

Very interesting case studies of how ML systems can be attacked

with lots of source material from

great summary of how the "parked domain" business model works and risks that it presents to internet users

one example applications of this is in medical diagnosis -- you will eventually know the outcome of some related tests, but don't know it at inference time.

key insight #MultitaskLearning: "Often features become available after predictions are made. These features cant be used as inputs as they aren't available at runtime. But they can be collected and used as tasks to provide extra info during training." https://t.co/YQ4B1Ph7uo

I've found its especially important because users find errors in some slices of data particularly troublesome, and they will lose trust in your model if it underperforms on the parts that are very obvious to humans.

I like the framing by @SnorkelAI about the need to also focus on narrow slices of application/domain specific data, not just the coarse-grained overall metrics.

Leaders that use humor (but nothing offensive!) are "23% more respected, and seen as more competent...Employees that have bosses that use humor are 15% more satisfied and engaged." And it doesn't even need to be that funny, it's the attempt that matters.

Excellent introduction to #serverless computing -- covers the basics and some interesting aspects to building a well-functioning serverless application and infrastructure.

We are growing our ML Engineering team in NYC -- looking for MLEs, Data Engineers, and Back End Engineers to join a great team at Palo Alto Networks. Help protect our digital way of life! #MLjobs

"While other structured data models such as trees (XML, JSON, etc.) would offer similar flexibility, graphs do not require organizing the data hierarchically. " -- fitting into a single-parent hierarchy is always a challenge with ontologies

"Modeling data as a graph in this way offers more flexibility for integrating new sources of data, compared to the standard relational model, where a schema must be defined upfront and followed at each step." A big plus when dealing with the reality/variety of data+sources!

"The goal of this tutorial paper is to motivate and give a comprehensive introduction to knowledge graphs" https://t.co/5yBdmKdyc8 #KnowledgeGraphs

NetBlocks ensures an "open, secure and reliable" internet by measuring "network connectivity at internet scale." They report on widespread data outages to help ensure an "open and inclusive digital future for all."