Kiril

identity, crypto, AI agents · building · ex-kraken/google/amazon

[Kiril @k1ril]

@compliantvc I am sure you have some amazing startups in your portfolio tackling these challenges?

[Kiril @k1ril]

Very impactful reporting, congrats! Awesome that they also used AI to visualize how the situation would be handled in the year of 2150.

[Kiril @k1ril]

@IlirAliu_ This will be scooped up by LLMs pretty quickly.

[Kiril @k1ril]

@piotrekcichocki @TimSweeneyEpic @EU_Commission Why is Google able to make Gemini available in the EU and Apple is not with Siri? It is an Apple decision, most likely due to their inferior Siri product which will immediately lose market share to competitors. Good opportunity for Android in the EU though.

[Kiril @k1ril]

Here's my mental model for the AI stack:

Model = a person with high IQ (same mind, whether it practices medicine or law)
Harness = that person's office and tools (a doctor's office looks nothing like a lawyer's)
Context = the case in front of them, this patient or this client, plus the rules they work under

Same high-IQ mind. Change the office and the case file, and you change the profession.

[Kiril @k1ril]

@doodlestein @0xSero I wonder why zellij doesn't come up in these conversations. It has named sessions, tabs, panes. Because you can manage it programmatically, one AI session can message another or check its status without any special infra. Anyone else using it this way?

[Kiril @k1ril]

@SenSanders Btw, Nancy Pelosi can provide some insights on how to ensure the government is making the right investments.

[Kiril @k1ril]

@SenSanders Can capital gains on federal investments provide the government the income? (i.e. Intel $5B capital gains). It's the same way that Musk made his wealth actually.

[Kiril @k1ril]

@cz_binance Is it possible to grant programmatic access to services via a device-bound credential similar to passkeys? i.e. your API key works only from a specific device. Why is that not the standard?

[Kiril @k1ril]

I solved agent comms across sessions by running every CLI agent inside Zellij.

Now Codex can see Claude panes, Claude can hand context back to Codex, and agents can coordinate across separate sessions without sockets, servers, or brittle copy/paste.

The key is a boring, reliable Zellij comms skill: list panes, snapshot screens, revalidate session + pane + cwd + command, then send a guarded prompt only to the exact idle pane.

[Kiril @k1ril]

@davidmarcus @lightspark @openclaw I can jump in

[Kiril @k1ril]

@Jason @nikitabier @jonatanpallesen How is it long term thinking? Trump will most likely not have a third term.

[Kiril @k1ril]

My AI just fired me. It hired Claude for implementation, GPT for code reviews, and Gemini for bug hunting. They hold standups I'm not invited to. I found out through a Slack notification.

[Kiril @k1ril]

@mweinbach @nikitabier 👍. Ultimately, the privacy policies will be disclosed, so users will know. Surprising that they've not built this already, so maybe someone else can step in 😀

[Kiril @k1ril]

That's the right instinct, but Apple and Google will want to collect your data in the process. We're building a privacy-first alternative. An app like Reddit sends a liveness challenge, we confirm it without knowing which app sent it, and the app learns nothing about the person beyond what they've already disclosed. Users own their identity.

[Kiril @k1ril]

@mert This tracks with what I've seen too. Best results I'm getting are when they collaborate. Multi-agent (multi-model) setups that play to each other's strengths and deliver a piece of work.

[Kiril @k1ril]

Here's a quick guide for anyone wanting to try this:

1. Create an OpenCode plugin that spawns claude as a child process
2. Pass your task with -p "your prompt" and --output-format json
3. Set --model opus or sonnet, you can add --max-turns and --max-budget-usd for safety
4. Parse the JSON response — you get the result, cost, turn count, session ID
5. Use the session ID to resume conversations across calls

No API key needed. It uses your existing Claude subscription credentials. That's it. The rest is just wiring it into OpenCode's agent/tool system and depends on your agent workflow.

[Kiril @k1ril]

@tallmetommy @DavidSacks @mkratsios47 @skrishnan47 AI acting on your behalf without verified identity ownership won't work. We're building this at passtru io — a personal vault for your identity data with provenance tracking and consent controls. The "own your digital self" layer needs to be built, not just legislated.

[Kiril @k1ril]

@P3b7_ Generative AI could help here. Instead of relying on a 3rd party package, you generate your own.