ooooh!

Worth a look peeps:

From Amanda Craven. No email or signup; direct download. Can I let you in on a little secret? I’ve been writing books, courses and emails for a long time now and you’d think the words would always flow. But the truth is that sometimes I still sit there staring at the blank subject line box, willing it to magically fill itself in. And when that happens, the whole thing grinds to a halt. Because if no one opens your email, it doesn’t matter how brilliant the rest of it is. So, I did something about it. I put together a brand new freebie for you: 30 Copy & Paste Curiosity Hooks. These are proven subject lines, headlines and openers you can literally grab, tweak a word or two, and use right away. Think of it like a little emergency stash for when your brain goes blank but you still need to send that email, post that update, or finish that sales page. It’s my gift to you - something quick, fun and instantly useful. Because sometimes the difference between silence and sales really is just a few words: https://drive.google.com/file/d/1dZrOHSG67G96f6GOCEKXsZIynbdQPAfT/view Amanda x

VaultGemma is Google’s latest Large Language Model (LLM) trained from scratch with differential privacy (DP). Key features: Sequence-level differential privacy: roughly meaning that any given “sequence” (section of training data) has bounded influence on the model’s output; prevents the model from exposing private data in responses when a single training example is involved. It uses the same training mixture as in Gemma 2, with similar pre-processing (splitting long docs, packing shorter ones) but applies DP techniques in training. Empirical tests: They probed memorization (e.g. giving a prefix of training data and seeing if the model completes with the suffix). VaultGemma at 1B parameters shows no detectable memorization under these tests. So, the basic pitch: high privacy guarantees + a real LLM that’s useful, not just a toy. That is rare, and worth paying attention to. Pros: What looks really good Here are the strengths / why VaultGemma might matter, especially for people like us who care about ethics, practicality, and pushing AI forward: 1. Strong privacy by design Because the model is trained with differential privacy (DP-SGD etc.), it formally limits what the training data can “leak.” If you’re dealing with sensitive data (personal, medical, financial), VaultGemma offers a solution that’s mathematically grounded. Their empirical tests show promise: no detectable memorization in the prefix→suffix test, which addresses a frequent concern (i.e. that the model might “regurgitate” private data). 2. Open and accessible model It’s open, has a model card, etc. That means transparency: researchers, developers can inspect, test, adapt. Size of ~1B parameters — “lightweight” compared to huge finetuned behemoths — meaning easier to deploy, lower cost. Also more feasible to run privately / in constrained environments. 3. Bridging the utility gap Historically, models trained with strict privacy constraints underperform compared with non-private ones. But VaultGemma seems to be narrowing that gap. Google talks about “scaling laws” for DP, meaning they are exploring how performance degrades (or doesn’t) as privacy constraints get tighter.

https://buffer.com/resources/how-often-to-post-on-linkedin/