kalomaze rethinks 'cribbing SSL backbone features' after peers hit multimodal walls
kalomaze · x · 2026-09-21
kalomaze says he used to think it was lame to crib SSL backbone features for your model instead of learning end-to-end from pure data — until two of his mutuals independently hit a multimodal wall doing it the e2e way. A data point in the open-source debate over reusing pretrained backbones vs. pure end-to-end training.
Related event: Developer Says Multimodal Training Still Needs SSL Backbones(4 posts)→
More from Models
- Two years after 'intelligence too cheap to meter', $10/$50 models are the new norm — teortaxesTex · 2026-09-21
- Gary Marcus: "General" LLMs Fall Apart the Moment They Leave Their Training Distribution — GaryMarcus · 2026-09-21
- AI hallucinated a passenger's middle name, almost costing him his flight — menhguin · 2026-09-21
- Distilled Qwen3.8-35B-A3B GGUF quantized model trends on Hugging Face — empero-ai · 2026-09-21
- Codex users blast OpenAI over opaque usage resets: 'tell us the limit and refill rate' — StewartalsopIII · 2026-09-21
- Open models hold just 10% of global AI inference spending, and bets are on that changing fast — sarahookr · 2026-09-21