kushyphus

I am glad I regret some of the tweets. Means I have moved on.

most big especially software companies already had enough coders to build out the software they needed. The block was never there.

a foreigner settled in india is doing better coverage of hardware space than usual "startup journos". these "journalists" mostly focus on startup drama and next big controversy. on another note so good to see hardware space growing.

@adii_kris I am from Jayanagar can confirm

this is how jayanagar metro station looks in my head

@championswimmer Mahindra has better looking cars

Absolutely bearish about OpenAI hardware device now. Jony Ive is totally washed 😭😭😭😭

when I started tuning llama-server, I changed flags randomly until something worked (or didn't). ncmoe 30? ncmoe 10? why is it suddenly 5x slower? what even is the KV cache eating my VRAM for? so I measured everything. every flag, every ncmoe value, the exact VRAM cost per layer, the exact point where performance falls off a cliff. this is the reference I built for myself. 16 flags, each explained with the "when to change it" and "what breaks if you get it wrong" enjoy

@tusharnandy08 donate one to power educational software? we are giving ai learning platform to underserved students

hey everyone, I have 2 3090s and 4L4s available. If anyone’s looking for gpus, hmu!

@lochan_twt *host

"api costs are too high, lets create our own LLM"

@kunalvg lucky we have enough opensource that it will become possible to finetune and then retrain them for India use case. In the next couple years the cost of training own LLMs should fall drastically and give an opportunity to catchup from the start.

India should read this twice. Our shot at building globally from here rests entirely on intelligence staying open. If AGI ends up locked inside a handful of labs, our future is bleak. India should be the loudest voice for open intelligence in the world. Build it, back it, defend it. That is the only way the next decade is ours to shape, and the only way we avoid becoming a permanent underclass.

I had access to Japanese design twitter for a brief moment and it was beautiful

the funnel I want to optimize for

@LucasSchrodp4 you are a bot Right?

@kush4122 Hello, I'd like to have you as a friend. Can you please send me a request or a private message so we can connect? 🌹🌹

LLM builders should track swear/curse words as a proxy for User satisfaction

@guigotgit @TheAhmadOsman you can literally ask open code free model to connect with lm studio or ideally build a wrapper around llama.cpp

@TheAhmadOsman To me, buying and installing an RTX 3090 is extremely easy. But how do you go from there to having a local LLM API or a model I can point opencode to?

You should buy an RTX 3090 and learn how to run models locally The elite don’t want you to know this but running local models is hella easy, performant, and cheap nowadays

@vaibhavbetter @babugi28 or mor like we wont need it. Are you telling me an opus level model will need 1000s of examples to do basic task. we might not end up needing use case data. Mainly we need a strong physics model.

@babugi28 👀 how many hours of cleaning and cooking data do you need before robots don’t need more data?

Human Archive (YC W26) was one of the first companies in the world to start collecting human data, and I said this last year, but soon you will be able to get free cleaning and cooking services anywhere in the world by just consenting to being recorded. Most people clearly still don’t know what’s coming, but it’s already in full swing. @abhirajbhal, I can understand the initial pushback, but unfortunately your customers will quickly churn for your competitors, and you will be forced to change your mind soon or @urbancompany_UC will no longer exist. I have proof. During @ycombinator, we reached out to you and all your competitors. Everyone told us to fuck off. So we did it ourselves by basically working with a smaller player and giving users an option to be recorded for a cheaper price. 98% of users decided they would rather have a subsidized service and consent to being recorded (we blur all faces and remove PII). In 2 weeks, we went from 48 to 1,315 bookings per day. Fuck it, we could’ve been profitable and even PAYED HOMEOWNERS to come to your home and cook and clean. When we launch next week, this will be a real thing NOT ONLY in India, but across the world. We did this in: 1. Patna 2. Purnia 3. Darbhanga 4. Guwahati 5. Jhanjharpur 6. Jabalpur 7. Indore 8. Siliguri 9. Mussoorie 10. Delhi 11. Gurgaon 12. Faridabad 13. Dimapur 14. Mumbai 15. Jaipur Today, we’re working with some of the largest ops businesses in Asia. Soon, you’ll see workers not only wearing cameras, but also gloves and full-body mocap. @abhirajbhal I genuinely respect you as an entrepreneur. What you’ve built so far is impressive. However, unless @urbancompany_UC starts to recognize what’s REALLY going on here, the stock will continue to drop, customers will churn, competitors will see massive growth, and I’ll be able to say I told you so when you change your mind later on. More than happy to discuss this live

@sridharfyi my product is literally for tackling learning loops. agentic tutor for curriculums

build your company around customer learning loops, not just product building.

@sugandhanisa moved from sf to build for India in India

If you are building in AI you must leave Bangalore and move to SF where you can mingle with other Indians who’ve moved to SF coz they’re building in AI.

@DerekColley_ @sudoingX @UnslothAI @lmstudio lm studio sucks ask Claude to create a wrapper on llama.cpp for your specific hardware and work. create skill to reuse

@sudoingX I wrote a short how-to for people to get @UnslothAI models running in @lmstudio Incl MTP. Select the Q-value you need from the available files in the download. x.com/compose/articl…

i will keep saying it. dgx spark with hermes agent /goal is the most underemployed combo in local ai right now. just switched to qwen 3.6 35B-A3B at Q8 with MTP. 262K context. the agent ran overnight autonomously and is now self updating its memories from last night's session. i woke up and it's already caught up on what changed. hermes agent just works on local models. no hacks. no workarounds. you set a /goal, you sleep, you wake up to progress. the memory system means it picks up where it left off. this is what agentic local ai is supposed to feel like. if you have a spark and you're not running hermes agent on it, you're sitting on the most capable local ai setup money can buy and using it like a desktop.

@samdotb they are copying its deciding what to 'copy' and how where it burns

llms should really learn to copy code with tool instead of using tokens it's insane

@ChinuKabi @system_monarch Bro I checked EU rules. its too wild. As @system_monarch is trying to point out its almost impossible to maintain same tech stack in EU vs World. I would guess the solution is to use some EU frame work which does this..

@system_monarch @kush4122 Why is it Meta's liability to ensure recording doesn't happen there? They can also carry a weapon into those places, it's not on the weapon manufacturer right?

One of my close friends interviewing for Meta E6 was asked this during his system design round: Meta Ray-Ban glasses can record anyone without them knowing. The EU just banned recording in hospitals and courts. Southeast Asia has no such law. Same hardware, same firmware, two completely different legal realities with no GPS restriction and no geofencing built in. How do you enforce location-based recording restrictions globally without breaking the product where recording is perfectly legal?

@tszzl @seconds_0 Term 1 of the contract says they need to build him a full fledged JARVIS @elonmusk

@seconds_0 I’m sure they’re letting him use Claude code with mythos access now

watching elons pro anthropic pivot has been very enlightening