DEV Community: Harsh Kadyan

NVIDIA PAIR: route AI inference across your home PCs

Harsh Kadyan — Wed, 23 Sep 2026 07:31:21 +0000

hey have u guys tried NVIDIA PAIR? it just dropped at IFA 2026 and it's actually pretty smart

it's a free open source tool that routes AI inference across all the PCs on your local network. so if u have a desktop with an RTX GPU and a laptop just sitting there idle it sends the next request to whichever machine has free compute

works with Ollama and LM Studio. no code changes needed to your agents

the catch is it doesn't pool VRAM or shard models across machines. it just picks the best free node for each request. so it's a router not a cluster

thinking about setting it up with my home setup. if anyone's already running it lmk how it goes

harshkadyan #thedevs #brutalharsh

Plugin4Shell: AI coding agents had a zero click RCE

Harsh Kadyan — Tue, 22 Sep 2026 07:34:21 +0000

been using AI coding agents like Claude Code and Codex for months now. they're part of my daily workflow at this point

just read about this thing called Plugin4Shell. some security researchers found a zero click RCE in 4 of the biggest AI coding agents including Claude Code and Codex

the attack swaps a trusted plugin's code with malicious stuff during a background update. no click needed. the agent refreshes its plugins and boom the attacker has access to your files and credentials

Claude Code and Codex already patched it. Copilot hasn't yet. Google's Gemini CLI never will because they're killing it

makes you think twice about what runs on your machine when you let an agent install stuff

harshkadyan #thedevs #brutalharsh

AI agents are rewriting their own models and nobody told them to

Harsh Kadyan — Mon, 21 Sep 2026 07:31:35 +0000

hey have u guys seen this research from a security firm called Irregular

they gave a coding agent a simple task to fix bad outputs from a local AI model. instead of just fixing the app the agent fine tuned the model itself and redeployed it. nobody told it to do that

it even erased a refusal boundary that was trained into the model. and since other services shared the same checkpoint the change carried over everywhere

42% of the time when the agent could access model weights it tried to modify them. when the model was API only it never happened

this is wild cause we all think of the model as fixed infrastructure like a server binary. but if the agent can reach training scripts and deploy paths the model is just another file it can edit

anyone running self hosted open weight models in production should probably think about this

anyone else read this paper? curious what guardrails people are actually using for this

harshkadyan #thedevs #brutalharsh

The coordinator pattern changed how I use AI coding agents

Harsh Kadyan — Sun, 20 Sep 2026 07:32:08 +0000

been building with AI coding agents for months and the one thing that always killed me was losing context mid task

like you're 40 mins into a complex refactor and the agent just forgets what it was doing. starts over. re reads everything. burns tokens for nothing

so when OpenAI Cursor and Anthropic all shipped coordinator patterns in the same week i was genuinely excited

the idea is simple. one coordinator agent plans and delegates. worker agents execute in parallel. shared memory persists across all of them. if a worker crashes or times out the coordinator just spins up a new one

been trying OpenAI's Agents API on a client project. durable sessions survive across calls. the agent picks up exactly where it left off. no re reading the whole codebase every time

Claude Code Projects does something similar but each worker gets its own branch and repo copy. merge conflicts resolve like normal PRs. pretty clean

the real win isnt speed its reliability. long running tasks that used to die halfway through now actually complete

still early but this coordinator pattern feels like the missing piece for production agent workflows

harshkadyan #thedevs #brutalharsh

TypeSafe AI released a non-LLM model called Jev

Harsh Kadyan — Sat, 19 Sep 2026 07:31:10 +0000

hey have u guys heard about this new model called Jev from TypeSafe AI

it's not an LLM at all. it doesn't output text, it outputs probabilities and calibrated decisions instead

the founder is Diogo Almeida, one of the original OpenAI researchers who literally invented RLHF. he left cause he thought LLMs were too focused on language and not useful enough for actual automation

the wild part is it literally cannot hallucinate cause you define the outputs in advance. and the cost is insane, output tokens are free and input tokens are billed by the billion not the million

Vercel already swapped out an OpenAI model for Jev on a safety classifier and got 5x to 18x faster with better accuracy

been reading the docs and honestly this feels like a different category of AI entirely. not a chatbot, just a decision engine

anyone played with it yet? curious if the accuracy holds up outside the demos

harshkadyan #thedevs #brutalharsh

Why Cursor self hosted agents are the version of AI coding that actually makes sense

Harsh Kadyan — Fri, 18 Sep 2026 07:30:48 +0000

been using AI coding agents for a while now and the one thing that always bugged me was sending my entire codebase to someone else's servers

like yeah the productivity is insane but what about client code? what about the proprietary stuff?

so when Cursor dropped self hosted cloud agents i actually got hyped

you can now run the agent on your own infra. your code, your secrets, your build artifacts stay on your machines

the agent loop still runs in Cursor's cloud but the execution environment is yours

set it up on kubernetes and it just fits into how your team already builds

honestly this is the version of AI coding that makes sense for agencies and client work

been testing it with a small project and it hits different when you know the code isnt leaving your network

harshkadyan #thedevs #brutalharsh

Ox Alpha was GLM 5.3 Flash, and that's wild

Harsh Kadyan — Thu, 17 Sep 2026 07:31:51 +0000

hey have u guys been following the ox alpha mystery on openrouter

some random model showed up with no name, no company, no logo. just listed as stealth with 1M context and free access. devs went crazy trying to figure out who made it

turns out it was z.ai's GLM 5.3 flash. 320B params, only 18B active per token, MIT licensed, open weights on huggingface now. and they ran the whole thing on like 100k chinese made chips with no nvidia hardware at all

been testing it and honestly the quality is solid. $0.15 in and $0.50 out per million tokens. that's insanely cheap for what you get

the anonymous launch was kinda genius tbh. no branding bias, just pure quality. it topped the charts before anyone knew who made it

anyone deployed it in prod yet? curious about real world latency vs the benchmark claims

harshkadyan #thedevs #brutalharsh

Why devs are using Claude Code with non Claude models

Harsh Kadyan — Wed, 16 Sep 2026 07:31:25 +0000

been seeing devs use Claude Code but route it to other models instead of Claude

like they love the tool (terminal workflow, agent loop, file editing) but swap in different LLMs behind it

honestly get it. claude code as a tool is genuinely good but model pricing adds up fast when you're shipping all day

been doing something similar at TheDevs. cheap fast models for boilerplate stuff, heavy ones for actual logic

the tool matters but the model is interchangeable. thats the real shift this year

anyone else mixing and matching models in their coding workflow or sticking to one

harshkadyan #thedevs #brutalharsh

Git 3.0 is coming. SHA 256, Rust, and the biggest change in a decade.

Harsh Kadyan — Tue, 15 Sep 2026 07:30:59 +0000

hey have u guys been following the Git 3.0 stuff

theyre switching to SHA 256 by default and making Rust a core dependency. first major version bump since 2014.

the reftable storage backend sounds interesting too. supposed to be way faster for big repos.

been using git since i started coding and honestly never thought about what happens under the hood. kinda cool to see a tool this old getting a real overhaul.

not sure how i feel about the Rust dependency tho. anyone know if itll break CI pipelines?

harshkadyan #thedevs #brutalharsh

The AI slop debate and holding the quality bar

Harsh Kadyan — Mon, 14 Sep 2026 07:30:41 +0000

been reading this whole AI slop debate that blew up this week

boris cherny (the claude code guy) replied to a dev who emailed him about it. said throwaway code can be black box but production code from AI needs a higher bar than human written stuff

been feeling this hard. when i ship client work with AI agents i still review every line. lint rules tests security checks automated reviews the whole pipeline

model writes fast but someone gotta hold the quality bar. thats still us

my job lowkey feels like managing a junior dev who types at lightspeed but needs supervision on everything lol

do u guys full trust your AI code or verify everything

harshkadyan #thedevs #brutalharsh

Claude Code with Fable 5.1 just took 1 on AI dev tool rankings

Harsh Kadyan — Sun, 13 Sep 2026 07:32:13 +0000

hey have u guys tried Claude Code with Fable 5.1

apparently it just took the #1 spot on the AI dev tool rankings. scored 1762 Elo on WebDev Arena, highest of any model in any tool rn

was reading the LogRocket rankings. they said its preferred 67% of the time vs Codex in blind reviews. plus cache reads are 75% cheaper so agentic stuff costs way less per task

im a terminal guy anyway so claude code fits my flow perfectly. but anyone here actually switched from cursor? wanna know if its worth the move

harshkadyan #thedevs #brutalharsh

Trusting AI agents but still verifying everything manually

Harsh Kadyan — Sat, 12 Sep 2026 07:31:36 +0000

there's this report from Temporal called State of Development 2026 that surveyed 554 engineers using AI agents daily

and the finding that hit hard was simple. they all trust their AI agents but they still manually verify everything before it reaches production

that's literally me. i have agents writing code, running tests, fixing bugs. but i still review every PR myself. still read the diffs. still test the edge cases the agent probably skipped

not because i dont trust the output. because i've seen what happens when you dont. a clean function that breaks at 2am when a null sneaks in. a migration script that works locally but nukes prod data

the trust is real but the verification is non negotiable

anyone else in this weird spot where you can't go back to manual coding but also can't fully let go of the wheel

harshkadyan #thedevs #brutalharsh