Raj

Anyone intrested in building AI native apps?

@KaiXCreator Post the hype comes clarity... Era equivalent of piracy post the gaming golden age... Local first... Purpose specific or hackable

What will come after AI?

FitOne nails it: Use your phone’s camera as a honest coach. No cheating, real verification. We’re building a similar personal approach using SVM (VLMs + pose estimation) that run fully locally on the phone. Focus: Trace movements, progress, recommend imp., and reduce injuries

SVM shine in narrow, deployed scenarios — speed, privacy, cost beat generalist VLMs. 2026 trend: More on-device multimodal apps, especially in India (offline, multilingual, low-power). What are you building — mobile app, robotics, document AI, or something else? Share ideas

Use LoRA/QLoRA on Hugging Face — possible on single GPU/Colab for SmolVLM. Datasets like ChartQA or custom images. Quantize (4-bit), ONNX/MLC LLM for mobile/edge. Browser inference possible with smaller variants. SLM/VLM for 80-90% cases + route complex to bigger model.

SVM Models (VLMs like SmolVLM 256M-2B, moondream, Phi-3-Vision, LFM2-VL-450M) are perfect for real-world use where big models are too slow/expensive. They run on phones, edge devices, laptops — low latency, private, cheap. Here are practical implementations that work today.

Narrow domain (your products, policies, FAQs). Fine-tuned 3-8B SLM handles 80-90% of queries at 1/10th cost, sub-100ms latency. AT&T-scale deployments prove it. Route complex cases to LLM only when needed.

Generalist LLMs (like GPT-4o, Claude, Grok) are powerful but often overkill — expensive, slow, and privacy-risky. Small Language Models (SLMs) win in specific, real-world uses. Here’s where and why you should pick SLMs instead.

@EthanBenard Love your progress man. Keep it up.

Time for the next chapter #weightlossjourney

Small Language Models (<10-14B params) aren’t just “cute” versions of big LLMs. They deliver real ROI right now — especially for enterprises, edge devices, and cost-sensitive apps. Market growth is strong (high teens to 25%+ CAGR). This isn’t hype. It’s economics.

Looking for Dev's who can help build my new venture. DMs open #letsgoo

@Matt_Pinner Falafel or taco? @elonmusk

you are sitting next to Elon Musk, what is the first thing you're asking him?

@dhruv_rathee Check out the police barricading pattern as well. Fun little detail people missed.

27 Lakh Voters were deleted in West Bengal. These were not 'dead people' or 'bangladeshis'. These were real people who had already appealed to get their votes restored. The Supreme court could’ve delayed the West Bengal elections until the tribunals finished hearing the appeals of these voters. But they didn’t. Only around 1607 voters were restored in time by phase 2 of these elections. Almost every appeal that was heard turned out to be a wrongful deletion. In any functional democracy, this would not count as a free and fair election where a large section of voters have lost their voting rights. I'm not saying that TMC would've won for sure if they were added back, maybe BJP would still win, but the question is about the fairness of these elections. Free and fair elections are a spectrum. Ever since Delhi-Maharashtra-Bihar-Bengal, this needle has moved more and more towards being unfair. Each time somehow the opposition parties participate as usual in the electoral process thinking that they can still pull of a win despite the compromised EC, ED, CBI etc. and each time they have been proven wrong after 2024 Lok Sabha. The question is, at what point will they feel that the level of unfairness is so unfair that elections should be boycotted? I personally feel TMC should have refused to participate in these elections until all those 27 lakh appeals were finished being heard.

@TARUNspeakss Spineless. Sellout. BJP's slut.

If you could say one Word to Raghav Chadha, what would it be?