Pinned
I didn't expect that Evaluation is not only the core of training, but also crucial in deployment (harness).
My Bold Prediction:
Data-Centric AI Era: 2022-2024
Evaluation-Centric AI Era: 2025-?
You Jiacheng
14.7K posts
- Based on my investigation, they are using CNC lathe with milling functionality. The first video shows the actual manufacturing process (partial). The second video will show the simulated full process. (1/2)
00:00Replying to @AWeirdPhysicist and @boxcardavidIt's literally impossible, they're being sold for 0.70$ each. If they were 70% manufacturing fee, and you wanted 10$/hr, on a 100k machine(minimum), you still wouldnt get there. - wtf? a user reported a bug at 17:49, and DeepSeek fixed it at 17:59???
- Replying to @francoisfleuretHere is an article about how to train a single model with 4k H100s.
- sorry, but it's not random access memory. it's append-only log.Context is the new RAM
- > LlamaBarn > literally ZERO llama modelsLlamaBarn (sneak peek 👀)
- Bros, this table approximately means a SoTA model can be trained in ONE DAY on xAI's Colossus.
- 3→3.5→3.7→3.8→3.9→3.95→3.97→3.98→3.99→3.995→3.997...
- what a beautiful theory!
00:10Even with full-batch gradients, DL optimizers defy classical optimization theory, as they operate at the *edge of stability.* With @alex_damian_, we introduce "central flows": a theoretical tool to analyze these dynamics that makes accurate quantitative predictions on real NNs. - NVIDIA will be the biggest loser if o-series take over GPT-series. (scaling inference compute instead of model size)
- WHAT??? Microsoft processed 50T tokens in 202503. ByteDance Doubao processed 12.7T tokens *per day* in 202503.
