MACHINE LEARNING

→ View original post on X — @simonw

Shutting Down Embedding Models Wastes Stored Vector Investments

By

@simonw

–

22 April 2026 23h43

Shutting down embedding models like this is a huge pain for people who have spent time and money running large amounts of text through them and and storing the vectors – that investment essentially becomes useless if you can no longer calculate fresh vectors to compare

→ View original post on X — @dair_ai

Agents That Self-Generate World Knowledge via Outcome-Based Rewards

By

@dair_ai

–

22 April 2026 23h15

How far are we from agents that can self-generate world knowledge? The work proposes an outcome-based reward that measures how much an agent's self-generated world knowledge actually improves its task success rate. The external guidance is then removed at inference. Result: A

→ View original post on X — @nvidiaai

NVIDIA NeMo RL Accelerates Agentic Performance with FP8

By

@nvidiaai

–

22 April 2026 23h00

Improve agentic performance with accurate RL post-training on low-precision FP8. NVIDIA NeMo RL, an open-source library within NVIDIA NeMo, supports FP8 to speed up RL workloads by 1.48x on Qwen3-8B-Base—enabling faster iterations for agentic tool use and multi-step

→ View original post on X — @clementdelangue

ML Intern Learning Through Mistakes and Persistence

By

@clementdelangue

–

22 April 2026 22h39

persistence always pays off, even for ml interns (btw, I haven't done anything for the past 30 mins, just observing the intern making mistakes and fixing them haha)

AWQ Quantization Optimization for Agentic AI Workloads

By

–

22 April 2026 22h38

For real agentic workloads (North), short-context calibration wasn't enough. We calibrated AWQ on long internal agentic traces (up to 64k tokens) and added token masking in llm-compressor to exclude repetitive chat templates/tool descriptions from calibration stats. Plus QAD

Cohere Hiring ML Systems and Audio Inference Engineers

By

–

22 April 2026 22h38

Enjoyed the read? If you have deep experience in ML frameworks (training or inference) and love working on problems like these, our team is hiring! ML Systems Engineer, Frameworks & Tooling: https://
jobs.ashbyhq.com/cohere/c99e61c
9-ed92-426d-9711-188dfc0f729f?departmentId=7130c75e-15b8-493c-959f-e9b8ea5c1c09
… Audio Inference Engineer, Model Efficiency:

BF16 to FP8 Quantization: Per-Channel Scaling for LLM Accuracy

By

–

22 April 2026 22h38

The tricky part: naïvely casting BF16 group scales to FP8 dropped the quality. Our fix: quantize scales per-channel (outer vector scaling) + rescale by 1/8 to avoid FP8 clipping. Result: >99.5% of W4A16 accuracy recovered on Command A & Cohere MoE. Paired with a CUTLASS

W4A8 Inference Production-Ready Integration in vLLM

By

–

22 April 2026 22h38

Excited to share our work on production-ready W4A8 inference, now integrated in vLLM! By combining 4-bit weights (low memory) with 8-bit activations (high compute), we hit the sweet spot for both decoding and prefill — up to 58% faster TTFT and 45% faster TPOT vs W4A16 on Hopper.