Joined June 2023
- Tweets 4,644
- Following 141
- Followers 21,749
- Likes 1,630
582 Photos and videos
Pinned Tweet
18 Feb 2025
Looking for an effortless AI agent setup?
Shinkai is perfect for non-developers seeking simplicity and speed, while ai16z (ElizaOS) suits those who enjoy deep coding control.
31
243
407
24,995
Three quests live.
Something else in progress.
Back Monday.
2
85
What’s something you thought was decentralized… but actually wasn’t?
7
80
The trust gap for local AI is real.
The people who would benefit most from running their own models are also the most likely to assume it requires serious technical overhead.
That’s a product problem, not a technology problem.
It’s one we think about a lot.
1
7
66
The agent wallet layer is becoming real infrastructure.
Every major player is figuring out where they sit on it.
The question most people haven’t asked yet is who actually owns the wallet when the agent is the one using it. 🧵
1
8
145
Self-custody for agents is the same argument as self-custody for people.
The asset should move when you tell it to, not when the platform allows it.
1
7
108
People focus on local AI for privacy, but the harder problem is auditability.
Once agents start taking real actions,
what matters isn’t just whether the output was correct,
but whether you can reconstruct and prove exactly what the system did step by step.
6
77
Kai has been doing some research lately.
Going to let him keep going.
1
7
131
Have you tried Research Rabbit? It’s one way people explore ideas beyond surface-level signals.
If you missed it, we’ve opened Quests again
a chance to go back into that mindset of deeper, self-directed research.
Don’t just read signals.
Research something of your own.
Try: identity.shinkai.com/quests
7
98
The Solana leaderboard moves on momentum.
What’s interesting is how often the Reddit layer and the chart tell different stories about the same token, and how that gap resolves over the following weeks.
Most research stops at price data. Token Scout reads both.
This quest failed to work when first launched, but it has come back to life for everybody to have a second chance
7
54
Running local models means you own the memory layer.
Cloud tools handle persistence invisibly, which is convenient until you want to know what they’re keeping and why.
Whether that’s a tradeoff worth making depends on how much you care about who has access to what you stored.
7
71
People think agents fail in obvious ways, but in reality they fail in boring, systemic ways.
Here are some examples 👇
1
9
242
#5 Cost blowups: runs too many steps, too many tokens.
1
7
64
$6 Edge cases in real data: weird inputs that never appeared in testing.
7
63
79% of companies may be adopting AI agents.
But that’s measuring tools purchased.
The harder question is how many of those agents are actually running tasks without a human approving every step.
That number is probably much lower.
7
76