The streaming engine is the most interesting part.
Producer thread synthesizes chunk N while consumer plays chunk N-1.
Chunk 1 plays for 6.36s.
Chunk 2 synthesizes in 4.53s.
4.53 < 6.36 → zero gap between chunks in real playback.
First chunk latency: 5.47s. Overall RTF: 1.25x
Phase 3 had a surprising finding.
FP16 made things SLOWER on RTX 3050.
FP32 baseline: RTF 1.46-1.58x
FP16 autocast: RTF 0.80-1.47x
Why? RTX 3050 has only 2 tensor cores. FP32→FP16 conversion overhead exceeds compute savings on single-sequence autoregressive generation.
So much for quick commerce @letsblinkit
Bought a last minute gifting book, and the product i received was damaged rusted and old book!
Thank you Blinkit, for keeping the trust of your customer!!
Dayum! @Spotify got down for some time and my productivity decreased exponentially 😂.
I was busy fixing why it is getting stopped at 10 Sec mark for every song....
All media channels, digital platforms and individuals are advised to refrain from live coverage or real-time reporting of defence operations and movement of security forces. Disclosure of such sensitive or source-based information may jeopardize operational effectiveness and endanger lives. Past incidents like the #KargilWar, 26/11 attacks, and the #Kandahar hijacking underscore the risks of premature reporting. As per clause 6(1)(p) of the Cable Television Networks (Amendment) Rules, 2021, only periodic briefings by designated officials are permitted during anti-terror operations. All stakeholders are urged to exercise vigilance, sensitivity, and responsibility in coverage, upholding the highest standards in the service of the nation. 🇮🇳
Read more: https://t.co/bHscgUBMEV
#MediaAdvisory #NationalSecurity #MIBIndia #ResponsibleReporting
@rajnathsingh@DefenceMinIndia@SethSanjayMP@HQ_IDS_India@adgpi@indiannavy@IAF_MCC@PIB_India
@imfcapdev Must Answer: Why Fund a State Enabling Financial Terrorism?
22 April 2025 — Pahalgam, J&K: 28 civilians killed. But what funded this horror?
Our latest blog uncovers the shocking truth: a ₹22.5 Cr crypto scam linked to Pakistani operatives was used to finance terror activities behind the Pahalgam massacre.
Read the full investigation that exposes the laundering trail, the terror network, and the global implications—especially for institutions like the IMF that are still considering aid to Pakistan.
Read the full blog:
https://t.co/2kNSoauOfn
The war isn't only happening in the border!
It's also a war of narrative... Now it's the duty of the citizens of India to hold the correct Narrative!
As it was never Us who started
Neither was it about IND vs PAK
It was about the terrorism Sheltering in PAK!
I bought this @zebronics headphone in 10th June 2024 and its not been a year.
Faced issues:-
- Elastic Strap got broken
- The sound is coming out distorted. Only by adjusting it with the help of its volume up and down buttons can I remove the distortion.
And yeah, it's Heavy!
Just wrapped up my first Kaggle competition! 🥳
Built a Grammar Scoring Engine for Spoken English using Whisper, BERT & ML.
🎙️ Whisper (ASR)
✂️ Grammar cleanup
🧠 DistilBERT + LightGBM + Meta-ensemble
📈 Pearson r: 0.625
#AI#NLP#Kaggle#Whisper#BERT#EdTech#SpeechRecognition
Building a Grammar Scoring engine for voice samples rn!
You think ‘Architecture’s set, time to sleep,’ then MAE & RMSE haunt you like night terrors.
#AI#MachineLearning devs, what metrics keep you up at night?
This is "Jack of all trades":
> C, C++, JS/TS, Rust, Python, Java
> React, Next, Angular, Flutter/React Native
> Node, Flask, FastAPI, Go, Redis
> MongoDB, Postgres
> Deep learning and machine learning: Major projects or research work in this
> DevOps: AWS, Kubernetes etc
> Unity or Unreal (not mandatory but a plus point)
I should add knowing all these is not enough, you need to be confident enough in all of them such that you can undertake any project.
We made 5 challenges and if you score 47 points we'll offer you $500K/year + equity to join us at 🦥@UnslothAI!
No experience or PhD needed.
$400K - $500K/yr: Founding Engineer (47 points)
$250K - $300K/yr: ML Engineer (32 points)
Challenges:
1. Convert nf4 / BnB 4bit to Triton
2. Make FSDP2 work with QLoRA
3. Remove graph breaks in torch.compile
4. Help solve Unsloth issues!
5. Memory Efficient Backprop
If you have any questions about the challenges, please feel free to ask! We're looking for people to help push Unsloth forward - so come join us to democratize AI further!
Our past work includes:
1. 1.58bit DeepSeek R1 GGUFs: https://t.co/gALGkUg5Cg
2. GRPO with Llama 3.1 8B in a Colab: https://t.co/LFdkNxwAYg
3. Gemma bug fixes: https://t.co/7kX94PyKQR
4. Gradient accumulation bug fixes: https://t.co/Tq4c5Qwqyw
Details & submission guide: https://t.co/iXxRUTijWV