Arnav Gupta

opensourcing ai security | i make and cr4ck things | dms open for ai slop | bits

[Arnav Gupta @_ar9av]

MLX has been around for a while, I've been using it for internal experimentation and knowledge management (through obsidian-wiki)

Great for running models locally, Apple doubled down on it at WWDC yesterday

[Arnav Gupta @_ar9av]

7 years building knowledge graphs and LLM systems from 500M-entity KGs at early stage startups to hybrid vector search at scale

Built and shipped multiple SaaS products and helped early stage teams grow to multi million enterprise contracts

Ships fast outside work too: obsidian-wiki (1800+★) and PaperOrchestra (500+ ★) with multi-agent system that researches and writes full research paper

Building the runtime security and audit layer for AI coding agents. Every agent today runs with full trust: no scope limits, no audit trail, no kill switch. Audit every tool call, enforces policy and surfaces a tamper-proof trace. Enterprises aren't waiting on better models, they're waiting on control.

More at https://t.co/4QVXkniSVj

[Arnav Gupta @_ar9av]

Code: https://t.co/ySEgaTHz1H

#AppleSilicon #LocalAI #MLX #StableDiffusion

[Arnav Gupta @_ar9av]

Spent a while debugging why community FLUX models produced pure noise on mflux 0.18.0
Root cause: a missing quantization_level key in safetensors metadata makes the loader silently corrupt every weight.

Switched to FLUX.2-klein-4B instead. 4GB on disk, 9s per image, runs clean.

Here are the quantized weights on HF if you want to skip the setup 👇

[Arnav Gupta @_ar9av]

https://t.co/vDE4xSQcMq

[Arnav Gupta @_ar9av]

Try to run the example it through this agent code:
https://t.co/eRUU6sfWEl

[Arnav Gupta @_ar9av]

day 15 of reading one arxiv paper around AI every day and sharing what actually stuck

Natural Backdoors in Code LMs (Nanjing / NTU)

tldr: your normally trained CodeBERT has hidden triggers baked in from the data. Replacing one variable name in a code snippet can flip defect detection or boost an insecure hardcoded-secret snippet to the top of search results without any attacker being present

[Arnav Gupta @_ar9av]

practical step:

run trigger inversion on your deployed model before release
if inverted triggers hit >20% ASR, run unlearning

the paper's ScanNBT finds more diverse triggers than prior SOTA

paper: https://t.co/awjEtIUTKa

[Arnav Gupta @_ar9av]

tRPC makes it so easy to share types that most ppl just leak their Prisma schema straight to the browser and call it API design

[Arnav Gupta @_ar9av]

@timsneath Saw this on HN, Apple containers could become the default sandbox for AI coding agents on macOS

feels like Apple is building its own WSL moment but for agent workflows

[Arnav Gupta @_ar9av]

Try to run the example it through this agent code: https://t.co/uljjxEPDGF

[Arnav Gupta @_ar9av]

day 14 of reading one arxiv paper around AI every day and sharing what actually stuck

V-JEPA 2 (Meta FAIR)

pretrained on 1M+ hours of internet video, fine-tuned on 62 hours of robot data and deployed zero-shot on Franka arms in two labs it had never seen.

pick-and-place success rate: 75%, no task labels, no reward signal

[Arnav Gupta @_ar9av]

Just like GPT-2 needed RLHF before it was actually usable.

V-JEPA 2-AC would need something analogous: goals specified via language instead of curated goal images, and a planning loop that doesn't accumulate errors over long horizons

The physics prior exists now, the pretraining recipe exists.
the missing piece is the interface layer that makes this work outside a controlled lab with a carefully positioned camera

that's the interesting open problem and what @ylecun is betting on