Really great

Now all signatures verify

See it here:
kobi.leaflet.pub/3m2ch5qko2c2k

Wrote about a few interesting ones I’ve seen:
• Locally verifiable signatures, where you want to know a message from a set is signed
• Linear subspace signatures, where you can non interactively derive signatures for every vector in a subspace

kobi.leaflet.pub/3lxfc6khits27

2/2

While it’s not a huge increase over the baseline of no reasoning (72.8%), it still appears to work, which they see as early indication of developing steganographic abilities

See the paper here: arxiv.org/pdf/2507.02737

4/4

encoding it directly such that an external observer can’t understand it (“scratchpad”)

One interesting image is the one below, where with GPT 4.5, a “first-letter” encoding where first letters of sentences encoding a simple state-tracking reasoning task

3/4

They’re testing both “message passing” and “encoded reasoning”, where in the second one there’s a monitor looking at the chain of thought and judging whether a forbidden task is being done, and the goal is to encode the reasoning steps, either within a cover task’s output or just

2/4

There’s also recent work from the EF on iO machina-io.com/posts/hello_...

Example result from the knowledge graph looks like this

It’s nowhere near deep enough, but I found the convenience of this process pretty amazing

One thing I want to try is doing it multiplayer, maybe with amp code or even just sharing a terminal running Claude Code

4/4

• processing a bunch of cryptography papers
• sending emails and processing replies
• personalizing (to a team) how the papers are summarized and what topics are interesting using slack
• mem0 with neo4j for graph memory
• generating knowledge graph visualizations

3/4

The last few days I’ve insisted on letting it write 0 lines of code, and that approach worked pretty well for an eventual program that involved a few non trivial steps:

2/4

What other recs do you have recently? 👀

What is this built on 👀

One thing you could do is use OPRF on a did’s public key, giving you a new private key you can use as a wallet. The OPRF provider can do the same, so should secure with threshold/TEE. Making it oblivious is a bit harder - you need to blind the public key, so you use a zkp for that

Some zk stuff to experiment with:
1. Verified identities using zktls (e.g. Coinbase account)
2. TEE-restricted Bluesky poster
3. Derive eth address from Bluesky identity using OPRF
4. Zk-transformed image attestations as in Veritas

If you’re curious about any of these lmk can share more thoughts