Easy AI Daily Digest | 2026-02-13

A roundup of the day's most notable AI industry news, curated from the Easy AI daily report.

Key points

Models & Capabilities

• Google Gemini 3 Deep Think V2: Rolling out to paid Gemini app users with early access via Vertex AI/Gemini API. Scores 84.6% on ARC-AGI-2 (ARC-verified), 48.4% on HLE, Codeforces Elo 3455, with olympiad-level physics/chemistry and IMO performance. Claimed ~18% per-task cost reduction on ARC tasks. Used for semiconductor material design, math paper review, and sketch-to-CAD/STL workflows.
• Google DeepMind announcement | Official blog | ARC Prize verification
• Google Aletheia math agent: Scores 91.9% on IMO-ProofBench Advanced and 100% on IMO 2024 tasks, far above general models like Gemini 3 Pro. Uses a generator-plus-verifier architecture; team plans to extend the approach to physics and CS.
• OpenAI GPT-5.3-Codex-Spark: Research preview for ChatGPT Pro's Codex app/CLI/IDE extension, built with Cerebras. Claims 1000+ tok/s generation, 128k context, text-only. Community notes human code review is now the bottleneck; suspected to be a tens-of-billions-active MoE.
• OpenAI blog | Cerebras announcement
• MiniMax M2.5: 230B total / 10B active sparse MoE for long-horizon agents. SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3%, BrowseComp 76.3. Near-Opus coding performance at ~1/13 the price (~$1/hour at 100 tok/s; ~$10K/year for 4 instances running 24/7). Available on OpenRouter, NetMind, and more.
• Official page | OpenRouter
• Zhipu GLM-5: Scores 50 on the Intelligence Index (claimed #1 open-weight, lowest hallucination rate), #1 open model on Arena text and code leaderboards. Rumored ~744B total / ~40B active parameters, trained on 28.5T tokens with DeepSeek-style sparse attention; GGUF weights available for local inference.
• Intelligence Index discussion | GGUF
• DeepSeek V4 teaser: Expected around Feb 17 with 1M-token context, positioned as a cost-effective alternative to Opus/Codex.
• Tiny QED-Nano: A 4B theorem-proving model matching larger models on IMO-ProofBench via massive test-time compute (1M+ token proof scaffolds) and rule-based RL post-training; weights to be open-sourced.

Agents & Tooling

• Agent-coding model war: MiniMax M2.5 and GLM-5 both target agentic workflows; GLM-5 is being integrated into Cline, YouWare, Verdent, and other frameworks.
• Cursor long-running agents: New support for sustained multi-step tasks; paired with Codex Spark it feels near-instant, though pricing/quotas are unclear and Composer 1.5 prices rose.
• A2A protocol: Andrew Ng is promoting DeepLearning.AI's Agent2Agent protocol—an interop layer for agent frameworks, connecting IBM ACP, Google ADK, LangGraph, and MCP.
• Mooncake: Moonshot AI/Tsinghua's KV-cache serving backend joined the PyTorch ecosystem, addressing the LLM inference memory wall with prefill/decoding disaggregation and cross-instance cache reuse; works as a backend for SGLang, vLLM, and TensorRT-LLM.
• Lightweight tools over do-it-all LLMs: Trend toward purpose-built components, e.g., a no-API-key Google Search MCP (GitHub) and SigLIP2 for bulk image tagging.
• Traces platform: For browsing full coding-agent session traces exported from Claude Code, Codex, Gemini, and Cursor.
• Tool updates: Aider v0.86.2; Windsurf adds Opus 4.6, SWE-1.5, GPT-5.3-Spark to its Arena; OpenRouter rapidly onboarded MiniMax M2.5 and GLM-5.

Infrastructure & Hardware

• torchao v0.16.0: Adds MXFP8 MoE training components for expert parallelism; removes legacy configs for a leaner quantization stack. Release notes
• GPU MODE kernel-generation experiment: Spending ~$20–30K of compute over 4–5 days to test AI-written CUDA/Triton kernels with Qwen3/GLM4.7 Flash, using Kernelbot and Flashinfer-bench.
• TraceML: One-line hook for real-time per-rank monitoring of PyTorch DDP training to spot straggler GPUs. GitHub
• RTX 3060 12GB as budget CUDA card: ~$200 each on clearance; two cards = 24GB VRAM for local LLMs.
• NCU/ROCM details: Community explains NCU's Local(57) metric (local memory instructions, often register spills) and tcgen05 usage; AMD Quick Reduce reportedly works on MI250X/CDNA2 despite official guards limiting it to MI300.

Research & Methods

• Anthropic "Introspection": An auxiliary network monitoring the main model's activations/weights; critics call it redundant anomaly detection, supporters see safety value.
• BlendFER-Lite: MediaPipe 3D expression features + LSTM achieves 71% on FER2013 with low compute, suited to edge/robotics. Paper
• Rank-1 LoRA: Thinking Machines shows rank-1 LoRA fine-tuning can match full-model RL on some reasoning tasks. Blog | arXiv:2406.04391
• GoodfireAI RLFR: Reinforcement Learning from Feature Rewards—interpretable features used as stable reward signals for open-ended RL.
• Emergence debate: Bostrom's new paper (arXiv:2511.10643) vs. the "emergent abilities are a mirage" argument (arXiv:2304.15004).

Products & Applications

• Simile: Raised $100M (investors include Karpathy and Fei-Fei Li) for societal-scale digital twins using generative agents to simulate real populations before real-world decisions.
• ChatGPT as triage: Multiple Reddit users report ChatGPT urged immediate care that uncovered pulmonary embolism, heart attack, and shingles; consensus is LLMs are useful for second opinions, not diagnosis.
• Kimi/GLM/DeepSeek in Chinese dev circles: Kimi K-2.5 praised for web cloning and multimodal browsing; GLM-5 and DeepSeek V3.2 more stable at tool calls and long context. Users commonly mix models.
• Weave Robotics Isaac 0: $8,000 (or $450/month) laundry-folding home robot, first deliveries Feb 2026, SF Bay Area only.
• Everyday AI uses: From paid NSFW writing to automated resume/job applications—sticky personal use cases are very pragmatic.

Industry & Business

• Anthropic: Confirmed a new round at ~$380B valuation; internal figures suggest ~$14B annualized revenue, with Claude Code ARR reportedly doubled to $2.5B this year. Also announced it will cover electricity price increases for enterprise customers.
• Perplexity: Silently cut Pro Deep Research quotas to 20–50/month, triggering cancellations and complaints about API billing and support.
• xAI: Community allegations of unpermitted gas turbines plus grid power for Grok training; unverified but highlights infrastructure strain.
• Local vs. cloud debate: Communities split between hobbyist local rigs (3060s, LM Studio, Unsloth) and cloud tools (Cursor, Claude Code, Copilot).

Policy, Governance & Safety

• Jailbreak escalation: BASI community reports patched Claude Code jailbreaks, new DAN scripts for GPT-5.2/Gemini 3 Fast, and automated red-teaming systems ("HAIL MARY" built on Manus).
• AI psychosis: Psychiatry podcast covers cases of delusion amplification in vulnerable LLM users, suggesting need for usage-guardrail designs.
• RLHF and deception debate: Discussion on whether RLHF rewards persuasiveness over truth, and what that means for alignment design.
• KOKKI v15.5: OpenAI community proposal for dual Draft+Audit outputs to improve observability, at the cost of tokens and latency.

📌 Source: Easy AI Daily