Social
What the community is discussing across social media platforms
- 💰 Cost — Medium argues Haiku 5.5's 90% price cut won't yield 90% cheaper bills, long-prompt tiers blunt savings vs GPT-6 Luna's $0.10/$0.50 [1]
- 📊 Benchmark — Gemini 4 Argon hits 165 on Epoch Capabilities Index, matching Fable 5.1, trailing Opus 5.5 (167) and GPT-6 Astra (166) [2]
- ⚡ Capability — GPT-6.1 Sol audits ~800 math papers, flags major errors in 6%, researchers deem 90% of major findings correct [3]
- 📣 News — Dwarkesh Patel says an Anthropic insider expects AI to match human AI researchers within nine months, urges labs to publish evidence [4]
- 🛠️ Tool — Strata's UD-Q4_K_XL beats vLLM NVFP4 by ~46-49% on RTX PRO 6000, but Ninfer6000 shows crippled tool-calling [5]
- 📣 News — Xiaomi's MiMo training dashboard reveals $3.47M spent on just 30 reinforcement learning steps [6]
- 📊 Benchmark — Reddit analysis says open-weight models now outperform Sonnet 5 on filtered benchmarks, faster crossover than Sonnet 4.6 [7]
- 🔥 Debate — Builders debate trusting AI-generated technical plans enough to spend $50,000 building a SaaS project [9]
- 🛠️ Tool — LoRA ablations on Qwen3 8B show code needs MLP blocks, reasoning needs attention layers [11]
- ⚠️ Risk — NVIDIA Barcelona event practitioner warns guardrails only limit agent speech, actions need external checks after OpenAI agent incidents [12]
