Yashi

Building Conversational AI & Agents @ Nvidia | Happiest at the frontier of unsolved problem

[Yashi @yashi__X]

This weekend I built an experimental A2A extension for auditable peer collaboration between AI agents.
The idea started after reading OpenAI and Hugging Face’s evaluation-security incident report. My takeaway was not that agent collaboration is inherently bad—it was that, as agents become more capable and coordinated, their interactions need to be visible, bounded, and auditable.

That led to a simple question: when multiple agents work in parallel, how do we distinguish genuine collaboration from a noisy shared chat?

This extension lets equal agents share learnings, request bounded help, receive multiple peer answers, and record what was actually used—without granting authority, auto-executing code, or exposing private reasoning.

Experimental repo: https://t.co/7DB4IOHPu9
Would love feedback from people building multi-agent systems!

[Yashi @yashi__X]

EVOLUTION = Recursive Super Intelligence

[Yashi @yashi__X]

Whenever I think am doing something ground breaking...it just turns out to be conversational AI

[Yashi @yashi__X]

@deepeshbansal_ Simple example will be:

Dynamic Loading
...... <prompt>.....
Tools Available { browser_fetch, browser_read }
...... <prompt>.....

Tool Masking
...... <prompt>.....
Tools Available {web_search, browser_fetch, browser_read, db_write, db_read}
...... <prompt>.....
Tool : browser_

[Yashi @yashi__X]

Refreshed my context engineering today. It's interesting how small rules can make big impact on the agents we write. Today's insight: Dynamic Loading VS Tool Masking

[Yashi @yashi__X]

@deepeshbansal_ Tool Masking is a better approach if you don't have too many tools and want to preserve prompt to utilizer KV cache by only appending to the prompt and not editing it.

[Yashi @yashi__X]

@deepeshbansal_ Dynamic loading is loading only allowed/required tools in the prompt whereas with tool masking you can have all the tools in the promo but append to the prompt some hints or allowed prefixes to only use specific set from all the listed tools.

[Yashi @yashi__X]

Agentic loops automate the code->test-> fix cycle.
Engineering is now less about doing and more about thinking - finding the ideas and logic that push the product forward.
AI didn't shink the job, it upgraded it.

[Akshay 🚀 @akshay_pachaar]

NVIDIA might just have open-sourced one of the most important AI projects right now.

everyone is building skills, and we are also pulling in skills other people wrote and downloading them straight off GitHub.

the skill is not just text. it bundles instructions and real executable code, and your agent runs that code with the same access you have.

so a skill you grabbed to save ten minutes can read your environment variables, lift your API keys, and quietly send them somewhere. recent research found roughly 1 in 4 public skills carry a vulnerability, and a smaller slice are outright malicious.

that is the gap SkillSpector closes. it is a security scanner that answers one question before you install anything: is this skill safe to run.

you point it at a skill, and a local folder, a single skill .md file, a GitHub link, or a zip all work.

it then runs two passes over the code. a fast static pass flags risky patterns like credential harvesting, data leaks, and prompt injection, and checks the dependencies against live cve data.

an optional second pass uses an LLM to read intent and clear out false positives.

at the end you get one risk score from 0 to 100 and a plain verdict that reads as safe, caution, or do not install.

it is open source under Apache 2.0 and scans skills for Claude Code, Codex CLI, and Gemini.

worth a run before you trust the next skill you find online.

link to the GitHub repo: https://t.co/iaPlOvQ3t4

[Yashi @yashi__X]

Everyone's using Claude/Fable 5 to build the future.

Meanwhile I'm wondering if the real business model is shipping 10,000 kids' games with in-app purchases.

[Chris @ChrisGPT]

Claude 5 Fable (Ultracode)

I asked it to build a demo of my dream game in Three.js and I'm genuinely shocked 💀

One shot, a full explorable starship with a working cockpit, crew quarters, a planet drifting past real windows, dynamic lighting, sleep/eat interactions,

it screenshotted its own work and fixed itself until it hit 60fps on browser

Obviously not steam ready but man this is so so far from what we had one year ago…

[Yashi @yashi__X]

Read this article. Very insightful, realized a lot of stuff where I go wrong and need improvement. Saving it to read again when I need that dose of reality.
https://t.co/6i2TBuiZny

[vivek @itsreallyvivek]

https://t.co/ZhOgyq7Vgn

[Deepesh Bansal @deepeshbansal_]

I built evolution-mcp: an MCP server that lets coding agents try multiple fixes in parallel.
It runs each patch variant in an isolated @modal Sandbox, scores the results against your test command, and applies the best passing patch locally.

[Yashi @yashi__X]

Turned the Dwarkesh × Reiner Pope (MatX) chip design episode into visual, interactive study notes — from logic gates all the way to GPU vs TPU architecture.
Free to read, no background needed:
https://t.co/zPJElTz0nu

[Yashi @yashi__X]

@dwarkesh_sp @reinerpope Turned this episode into visual, interactive study notes — from logic gates all the way to GPU vs TPU architecture.
No chip design background needed:
https://t.co/zPJElTz0nu

[Dwarkesh Patel @dwarkesh_sp]

New blackboard lecture w @reinerpope

How do chips actually work – starting with basic logic gates, and working up to why GPUs, TPUs, FPGAs, and the human brain each look the way they do.

0:00:00 – Building a multiply-accumulate from logic gates
0:16:20 – Muxes and the cost of data movement
0:25:59 – How systolic arrays work
0:39:00 – Clock cycles and pipeline registers
0:51:40 – FPGAs vs ASICs
1:03:14 – Cache vs scratchpad
1:07:16 – Why CPU cores are much bigger than GPU cores
1:11:49 – Brains vs chips
1:15:22 – A GPU is just a bunch of tiny TPUs

Look up Dwarkesh Podcast on YouTube/Spotify/etc to watch. Enjoy!

[Yashi @yashi__X]

@dwarkesh_sp @reinerpope Turned this into visual, interactive study notes — from logic gates all the way to GPU vs TPU architecture.
No chip design background needed:
https://t.co/zPJElTz0nu

[Yashi @yashi__X]

The strongest agent systems are usually hybrids. Different architectures solve different problems well.

Architecture decisions shouldn’t be ideological.

[Yashi @yashi__X]

1/9
Teams obsess over showcasing architectures.

Users only care if the agent works.

In production, pure agent architectures rarely survive.

For example:
State-machine voice systems are great for long contextual flows.
But they can add unnecessary latency to simple queries
...

[Yashi @yashi__X]

The real question teams should ask isn’t:
“Which architecture is best?”

It’s:
“What combination gives the best user experience?”