Releases, papers, SOTA benchmarks
Shanghai AI Lab beats DeepSeek in math via RL, not R1 distillation. I read the filing: ops trade-off is compute cost vs. raw reasoning gains for production models.
I watched DeepSeek’s R1 drop, shaking up markets with open-source reasoning that rivals o1 at a fraction of the cost.
CMU researchers use large models to optimize vision-language prompts without parameter access, a method I find promising for black-box scenarios.
Tencent's 389B MoE is out. Free for business. Beats Llama 3.1? I'll believe it when the unit economics work in the field.
I read EMNLP '24 findings on training-free knowledge editing for large models. It promises efficient new data absorption without retraining. This forward-looking AI topic sits outside our Jan 2025–May 2026 timeline.
I watched three models team up to challenge o1, proving that collaborating with 360+ agents eliminates the need for manual prompt engineering in real-world scenarios.
Tsinghua's new 3D scaling law pushes AI generation boundaries, addressing forward-looking topics beyond current timelines.
I read SJTU's study on LLM contradiction detection. It targets prompt integrity, a niche ops concern for now. This work falls outside our Batch 3 foresight timeline (Jan 2025–May 2026), so it won't impact immediate deployment strategies.
I read that LLMs still fail basic counting tasks like spelling "Strawberry." This highlights persistent reasoning gaps, forcing enterprises to verify model capabilities beyond marketing claims before deployment.
I note LLMs still confuse version numbers with decimals, a basic reasoning flaw that suggests token-level processing overrides numerical logic.