Arash Sadrieh

Co-founder & Chief Scientist @NinjaTechAI | Building AI agents that work autonomously | Ex-Amazon | Sharing the raw journey of shipping AI

[Arash Sadrieh @ArashSadrieh]

@AWSSupport Hi @mattsgarman

our whole service is down for almost 3 hours, would appreciate some help here ,eta on fix would be great

[Babak Pahlavan @babakph]

https://t.co/ApoM1kDQLf

[Arash Sadrieh @ArashSadrieh]

@thejessezhang This is not true in many enterprises,

Take many examples of this benchmark,

https://t.co/uLsEgxLbNK such as working with specialised CAD software or simulation software (complicated harness in these software never replaces training).

[Jeremy Howard @jeremyphoward]

Wow.

@Zai_org GLM 5.2 is a marvel! It is *at least* as good as Opus 4.8 and GPT 5.5. It's super fast, inexpensive, and not too verbose.

It responds with nuance and judgement, & handles long context VERY well.

I've never experienced an open weights model like this before.

[Arash Sadrieh @ArashSadrieh]

We are building over stateful agents in last 8 months :

we call them self evolving / we encoded the state as agent computer (both memory and file system) , we serialise these states basically hacking through linux kenrel using firecrackers VM and using b-tree to make sure we are snapshot the state diff on every checkpoints - we build a systems inspired by deepmind alphaevolve (we call dreamer agent)- to put right bread cums in agent context to let agent explore the current state and be aware of important information about the current state and bias the agents to build tools to access their states and organise them.

We found the agents are specialising in the task domain very well using above (success rate and error rate drops off significantly compared to base model using stateless scaffold - and the system is general enough for different task domain ranging from scientific tasks to sales and customer service.

If you are interested happy to show you and share more details:

@babakph

[Arash Sadrieh @ArashSadrieh]

@Scobleizer 👻👻

[Robert Scoble @Scobleizer]

I am working with some of its agents. More in that in May after the tech I am using is released. Got a team of them working for me in a Slack channel. They are amazing.

[Robert Scoble @Scobleizer]

“Taste is the moat now.”

I have a secret team of agents working on something that I can’t talk about from a small startup that will reveal itself soon.

They have a very different standard of taste than others I have worked with.

You will see that very clearly soon.

Vik is right.

[Robert Scoble @Scobleizer]

A very unusual pitch.

I am at @NinjaTechAI in Mountain View.

@babakph founder had his AI agents reverse engineer my site https://t.co/8L5xphk0qQ and build a better one.

Basically the first time a founder said “we copied everything you did in one night and now are improving it.”

He has teams of AI agents working on it.

Already passed my site in features and he started last night.

This is not released bleeding edge AI.

I will have a lot more to say about it when it does get released.

And @blevlabs AI that I am using has advantages still. His cognitive architecture writes better.

I think I will hook the two companies together and have them battle to make the best news site.

Coming in a few weeks. Nuts.

[Feross @feross]

🚨 CRITICAL: Active supply chain attack on axios -- one of npm's most depended-on packages.

The latest [email protected] now pulls in [email protected], a package that did not exist before today. This is a live compromise.

This is textbook supply chain installer malware. axios has 100M+ weekly downloads. Every npm install pulling the latest version is potentially compromised right now.

Socket AI analysis confirms this is malware. plain-crypto-js is an obfuscated dropper/loader that:

• Deobfuscates embedded payloads and operational strings at runtime
• Dynamically loads fs, os, and execSync to evade static analysis
• Executes decoded shell commands
• Stages and copies payload files into OS temp and Windows ProgramData directories
• Deletes and renames artifacts post-execution to destroy forensic evidence

If you use axios, pin your version immediately and audit your lockfiles. Do not upgrade.

[Robert Scoble @Scobleizer]

Just saw a new agentic platform that goes way beyond the others I've seen. Coming in two weeks.

I have no idea how anyone keeps up without using AI now.

Lesson I've taken away: everyone must be prepared to change as new things come.

Have your AI analyze new things, compare them to what you are using, and be able to help you change to new thing.

My developer changed from OpenClaw to Hermes in one night.

But he's 22. They are so damn fast at changing.

Me? I'll be honest. I'm struggling. Where I'm struggling I'm building AI to take over. And everyone is struggling to deal with change and it'll get worse. Part of it is spiritual. I'm dopamine addicted thanks to social media that addicted me. Getting me off the need for constant dopamine hits is gonna be something, but am working on it.

Those inside the bubble have both a huge advantage because of this, but a huge disadvantage. As they build AI systems to help them deal with change they are getting busier. This is hitting all execs and founders. As Jensen Huang, founder of NVIDIA said at GTC: he's constantly in the critical path. That's fun, but also going to cause a lot to burn out if they don't have agentic systems to help make decisions faster (and do things on their own).

I see a job boom coming for those who know how to setup agentic platforms fast and get to work on solving other people's struggles.

I'm building something to help keep up with the AI world here on X, almost done. @blevlabs and I are meeting tonight to finish it off.

[Matt Harrison @__mharrison__]

For my friends who are still using UV and might be a little weary about recent compromises to PyPi packages, stick this in your pyproject.toml.

You can let all of those pip users find and report the compromises...

[Arash Sadrieh @ArashSadrieh]

Heads to check your AI stack

[Daniel Hnyk @hnykda]

LiteLLM HAS BEEN COMPROMISED, DO NOT UPDATE. We just discovered that LiteLLM pypi release 1.82.8. It has been compromised, it contains litellm_init.pth with base64 encoded instructions to send all the credentials it can find to remote server + self-replicate. link below

[Arash Sadrieh @ArashSadrieh]

@TechLayoffLover Entry-level tech hiring dropped 73% in the past year (Ravio report). Average applications per job hit 257 in 2025, up from 207 in 2024. She is not the outlier, she is the median. The pipeline was dismantled while students were still in it.

[Arash Sadrieh @ArashSadrieh]

@sweatystartup Inference costs dropping faster than energy costs rise. Midjourney cut GPU spend 65% via TPU migration, 11-day payback. Token costs now $0.002-0.06/1K. Electricity up 36%, yes, but hardware efficiency gains outpace it. The 5x price math does not add up.

[Arash Sadrieh @ArashSadrieh]

@aakashgupta The $4.5M is subscription revenue, not business revenue. Top earner on the platform makes $50/month. 3,000 AI companies generating almost zero rev. KeyBanc SaaS data: median rev/employee below $1M ARR is $42K. This is people paying $49/mo for hope, not NVIDIA-tier efficiency.

[Arash Sadrieh @ArashSadrieh]

@TechLayoffLover Entry-level tech hiring dropped 73% in the past year (Ravio report). Average applications per job hit 257 in 2025, up from 207 in 2024. She is not the outlier, she is the median. The pipeline was dismantled while students were still in it.

[Arash Sadrieh @ArashSadrieh]

@emollick Deloitte found 63% of Gen Z workers fear they won't have the skills for an AI future. McKinsey says 1 in 5 already show burnout symptoms. The irony: AI confidence is collapsing the more people use it, with a 35% trust drop. The psychosis is rational.

[Arash Sadrieh @ArashSadrieh]

@shl AWS EC2 Mac instances cost $508/month for an M1. A $600 Mac Mini M4 outperforms it and pays for itself in 5 weeks. 37signals saved $10M/year leaving AWS. Gumroad ran on a team of 1. The entire bootstrapper stack is converging on a single box under your desk.

[Arash Sadrieh @ArashSadrieh]

Philips owned 27.5% of TSMC and created ASML. Sold the TSMC stake for ~$8.5B — those shares now worth $800B+. ASML alone is worth 12x more than Philips today ($350B vs $25B). The MBA playbook of 'focus on core business' has destroyed more generational wealth than any market crash.