venividi

ML Research | Document AI | AI Safety | Mech Interp. | Published @ WACV, AAAI, ICLR, ACL

[Shubham Sharma | Video Editor @editxshub]

@MTSlive

[lil c @seclilc]

AI;DR

(AI; didn’t read)

[Suraj Sharma @suraj_sharma14]

I think India accidentally built the internet’s greatest nostalgia archive.

It started with one barbershop.

Now there are websites for truck journeys, chai tapris, regional songs, 90s cartoons, pan shops, bus rides and basically every piece of desi life you forgot you missed.

> https://t.co/GH9QuwYunn → ₹20 haircut + 90s barbershop bangers

> https://t.co/1MG24W55lW → Horn OK Please at 2 AM

> https://t.co/KbvIwPBU3H → Classic truck radio

> https://t.co/HdD5jubQgW → UP Parivahan nostalgia

> https://t.co/6hjRNCAirz → Missing home + Marathi classics

> https://t.co/BANAbuZPyz → Soft Tamil night drive

> https://t.co/xwxRyQN2Fj → 2000s Kannada childhood unlocked

> https://t.co/crdYmXGUWM → Pure Kannada nostalgia

> https://t.co/a3gWUaO7px → Sindhi lada, full volume

> https://t.co/zaX4ov0ZpE → Old Punjabi Bluetooth-memory-card songs

> https://t.co/X9AmyYpCv1 → Bihar bus journey core

> https://t.co/w676HEEG27 → Gujarati chachar chok

> https://t.co/kJcffs7BFQ → Mazdoor anthem energy

> https://t.co/WWbE0PjGVN → Cutting chai + plastic stool

> https://t.co/0lzs44brad → Chai ki tapri classic

> https://t.co/ttUYsY5dy5 → Long-distance bus uncle playlist

> https://t.co/Ij0ge45Onr → Roadways bus vibes

> https://t.co/VwExVWy1HB → Mistri / construction site radio

> https://t.co/b6PmEPGNaf → Auto wala, full volume

> https://t.co/7YvIImffu1 → Pan shop corner energy

> https://t.co/DnF8JMQhzx → Deluxe saloon upgrade

> https://t.co/l7lllxIK8q → Dhaba + full thali mood

> https://t.co/o4sgCk4NxY → Qawwali mehfil, late night

> https://t.co/eKvjd2LypA → Gali walk + old Bollywood

> https://t.co/qy4hneF2Dc → Interactive truck + 90s bangers

> https://t.co/3lMURmJPet → 90s cartoon theme songs

> https://t.co/Mrewbn0slC → 80s–90s gali nostalgia

> https://t.co/yt4Ao1OTqM → Another cutting chai classic

> https://t.co/gGyBf2nkUu → Cutting chai special

> https://t.co/7mL6WU4z4a → Another truck-radio gem

> https://t.co/X9AmyYpCv1 → Bihar Parivahan Nigam bus

> https://t.co/w676HEEG27 → Gujarati nostalgia

> https://t.co/kJcffs7BFQ → Mazdoor radio

And somehow, every single one feels like a tiny time machine.

The internet is weird.

One day it's another AI wrapper.

The next day someone builds a website that makes you remember sitting at a chai tapri, travelling in a UP bus, listening to songs from a Bluetooth memory card or getting a ₹20 haircut in 2007.

Internet is actually beautiful sometimes.

Drop the desi nostalgia website I missed in the replies.

Let's see how deep this rabbit hole goes. 👇

[Raffi @raffichill]

Please write up a brief handoff document describing everything we've tried and what we are attempting to achieve so I can pass it along to another agent

[venividi @robustlywrong]

@jrosseruk @SPARexec Hey J, interesting project! I’ve applied. I’ve worked on a PoC testing whether SLMs behave differently when they know they’re in an eval vs. deployment setting, and under different incentives like throughput vs. accuracy. (https://t.co/CSH75SNzT5)

[venividi @robustlywrong]

@agupta I use Sonnet 5 simply because it is cheaper than Opus and it can get the task done (with a more detailed prompt). I usually ask Opus to define the task and set constraints for Sonnet and then it is able to do pretty well at a lesser cost.

[Intelligent Internet @ii_posts]

Preprint

[Douglas A. Boneparth @dougboneparth]

Remember to treat your body like a data center.

Consume as much literature and water as possible.

[venividi @robustlywrong]

@exploding_grad @MATSprogram Hey, great work! Can you pls tell me if you had prior ai safety/research experience before you applied?

[venividi @robustlywrong]

@jiegaosophia @MonashUni I have shared my CV over email! Thank you.

[Jun Song @jun_song]

@AnthropicAI Can you guys try to be more creative?

[venividi @robustlywrong]

@wdymarnav hey @wdymarnav ! could you add me too - thanks!

[venividi @robustlywrong]

@quasicoh i thought i was the only one 😆😆

[Aritra 🤗 @ariG23498]

Meta open sources a 30B dense Vision Langage Model, distilled from Muse.

Muse Glimmer, comes with Day 0 support from:
> transformers
> llama.cpp
> vLLM
> SGLang
> fine tune with TRL
> and so much more

[venividi @robustlywrong]

@atagade19 @Shawn__Zhou @ihsgnef Hey @atagade19 , I’ve applied to both the distillation and subliminal learning ones! They’re really interesting and I’m looking to apply to PhD roles starting next year!

[venividi @robustlywrong]

@Sree_Sharvesh @SPARexec Hey @Sree_Sharvesh ! Really interesting project, I just applied. I was thinking about if having multiple sub-agents discover and combine small vulnerabilities over time could help make attacks emerge naturally (especially over long-horizon interactions). Curious what you think!

[venividi @robustlywrong]

@EzraJNewman Thanks Ezra! I just applied. I was going through some literature around CoT faithfulness and wondered whether it might be better to first check if VEA could end up measuring willingness to disclose suspicion rather than awareness itself. Curious to hear what you think!

[venividi @robustlywrong]

@EzraJNewman Hey @EzraJNewman! I'll def apply. My research is in FinAI, so I've seen similar cases in prod. I had run a PoC with <5B LM on eval vs deployment and throughput vs accuracy. Qwen flagged ambiguous fields more reliably under accuracy framing; LFM varied with batch composition.

[venividi @robustlywrong]

@iarthsingh Thanks!! Any ongoing research projects I can contribute to?