Nenad Mancevic
@manca1
Principal Engineer @HBOMax, architecting massive-scale Video AI & empowering engineers. Roots in hard real-time video streaming & Computer Vision. Views my own.
You might like
[1/6] Tech leadership isn’t about being the smartest in the room. It’s about making the room smarter. What I’ve learned (sometimes the hard way) as a Principal Engineer 🧵👇
Today’s AWS outage is a reminder that resilience matters more than savings. Companies without a multi-cloud strategy are bound to disappoint their customers when outages happen. Multi-cloud may seem expensive, but a few hours of downtime can erase those “savings” for years.
Can we start using tokens/watt for inference throughput, instead of raw FLOPS? This aligns incentives with humanity's benefit: maximizing tokens, while minimizing power consumption. "Smarter AI, Cleaner Planet." (sounds like a commercial)
Now we just need to see AMD's investment in OpenAI and we'll go a full circle.
Excited to partner with AMD to use their chips to serve our users! This is all incremental to our work with NVIDIA (and we plan to increase our NVIDIA purchasing over time). The world needs much more compute...
With the release of Tinker by @thinkymachines, we are slowly entering the post-training era which will dominate model development moving forward. What’s common for all post-training techniques? The need for high-quality data.
Evals and RL environments are the new hot topic in Silicon Valley, but the core ideas aren’t really new. In supervised ML, we’ve always needed a way to measure model quality. That means defining metrics (precision, recall, F1, BLEU, etc.), curating a dataset with high-quality…
The X algo changes seem to limit reach for all account types: large, medium, and small. A post from a 15K-follower account got only 100 views in five hours. The algo seems to prioritize retweets, likes, and comments within the first hour, which invites more low-quality rage bait…
It’s funny that they charge per input/output tokens, yet top off-the-shelf models heavily throttle your requests. There’s a legitimate use case for making more than 1 QPS on the API without having to deal with bunch of 403s and 429s all over the place.
Replies like "can this be solved through better tooling? better models? better pre-human PR automation?" No. Jesus Christ, No. The solution is school. An education. Learning something. School. Its school.
When is @github reverting back to master as the default branch?
Yes, Omarchy runs on master. So will every project I start from here-on out. I got fooled by the good-faith exploit once, but won't repeat that mistake. world.hey.com/dhh/forcing-ma…
Teenagers have likely replaced reading with social media. I’m hearing more and more middle schoolers are constantly on YouTube, TikTok and Roblox. These trends are very concerning.
High-quality data created by domain experts is the new oil. There are two major benefits to having ultra-high-quality data in the AI-first world: 1. LLM post-training - aligning model outputs with human-level expertise. 2. Evaluation and benchmarking - measuring LLM…
Looks familiar? AI Bubble vs the Dotcom Bubble Maybe history does repeat itself
With the recent buzz about hundreds of billions of dollars being poured into AI infrastructure projects, it’s hard not to draw a parallel to the early 20th century, when the world was being wired for electricity. We’re living through a truly unique moment in history.
Of course, the infinite loop goes like this: 1. NVIDIA invests $100B in OpenAI 2. OpenAI buys more NVIDIA GPUs 3. NVIDIA market cap increases by $100B 4. Repeat until singularity is reached.
United States Trends
- 1. Lakers 90.2K posts
- 2. Luka 69.4K posts
- 3. Ayton 14.9K posts
- 4. Marcus Smart 8,443 posts
- 5. Warriors 95.4K posts
- 6. Steph 32.4K posts
- 7. Sengun 25.3K posts
- 8. #DWTS 48.6K posts
- 9. Shai 33.7K posts
- 10. #DubNation 13.1K posts
- 11. Rockets 58.9K posts
- 12. Draymond 8,393 posts
- 13. Reaves 9,132 posts
- 14. Double OT 6,632 posts
- 15. Jimmy Butler 9,201 posts
- 16. #LakeShow 6,569 posts
- 17. #TusksUp 1,089 posts
- 18. LeBron 36.7K posts
- 19. Vando 3,000 posts
- 20. Kuminga 10.2K posts
Something went wrong.
Something went wrong.