Releases, papers, SOTA benchmarks
ByteDance Seed open-sources a code model hitting SOTA benchmarks while introducing a new data management paradigm for smaller models.
I read that GPT-4o hits 6.2% accuracy on a new Chinese search test. Meanwhile, AI Mode expands in the US, mixing links with chat.
I read OpenAI’s new tech report detailing why GPT-4o exhibits sycophancy, a key issue shaping AI industry extensions through 2025–2026.
I read the release notes. The hype is real, but I worry about latency costs for routine tasks.
I see capital betting heavily on AI labs via chips and credits. Yet, ByteDance's new 200B model outperforms DeepSeek's 671B version, challenging scale assumptions.
I read the new academician-led review on multimodal LLM alignment. It maps critical industry extensions through 2026. The timeline is tight, but necessary for 2025 deployment.
I read about a 14B model beating Claude in finance/paper translation, challenging market odds.
I'm tracking CVPR 2025's 'Foundation Models + X' workshop call for papers. Academia is tackling robustness and real-world gaps. I think we need better evals before trusting these models in production.
I watched Musk debut Grok 3, touting its reasoning prowess and Arena dominance. The launch coincides with shifts in X subscription pricing.
I read Renmin Univ. and Ant Group's Gilead paper on LLMs as diffusion models. It proposes a 2025-2026 industry roadmap replacing next-token prediction with diffusion processes.