iFeeling Daily
Daily curated AI insights you can't miss.
Async GRPO with LoRA across HF Jobs: A Bucket, a Proxy, No NCCL
AsyncGRPOTrainer now syncs only a rank-1 LoRA adapter to vLLM via a shared Storage Bucket, so training and inference run in separate HF Jobs with no NCCL.
How OpenAI Scaled Habitat Storage to 1 Billion Weekly Users
OpenAI's Habitat online storage now handles over 70 million requests per second and 500+ petabytes, supporting 1 billion weekly users.
Claude formalizes Fermat’s Last Theorem in 11 days
Claude produced the first end-to-end, computer-checked proof of Fermat's Last Theorem in 11 days, writing 13 million lines of Lean code.
NeoMME: Efficient Multimodal-Native Encoder for Visual Document Retrieval
NeoMME introduces a single bidirectional Transformer for both text and image patches, achieving state-of-the-art visual document retrieval with 2x throughput and 255x storage compression.
Automated Alignment Researchers Reliably Mitigate Alignment Failures
Claude autonomously mitigated 10 alignment failures, closing a substantial portion of the safety gap, and outperformed human researchers on deception.
Previewing the Model Hardware Standard
Anthropic's Model Hardware Standard (MHS) enables AI agents to safely operate lab and manufacturing devices in parallel, reducing integration from weeks to hours.
PROOF-Gen: Distilling Tool-Calling Models from Failed Teacher Runs
PROOF-Gen recovers golden tool-calling trajectories from teacher failures via per-scenario prompt optimization, recovering 93% of failed scenarios on τ2-bench.
Calibrating Agent Memory: Dose Depends on Model Capability
The right dose of agent memory depends on model capability: weak models gain from curated retrieval, strong models from full guidelines, saturated models none.
Multi-Vector Late Interaction Embedding Models in Sentence Transformers
Sentence Transformers v6.0 adds MultiVectorEncoder for ColBERT-style late interaction, MaxSim scoring, visual document retrieval, and token pooling.