Easy AI Daily News | February 13, 2026

A curated digest of AI model releases, agent tooling, infrastructure, research, products, and industry/policy news.

Models & Capabilities

• Google Gemini 3 Deep Think V2: Available to paid Gemini app users with early access via Vertex AI/Gemini API. Scores 84.6% on ARC-AGI-2 (ARC-certified), 48.4% on HLE, Codeforces Elo 3455, Olympiad-level physics/chemistry and IMO performance. Already used for semiconductor material design, math paper review, and sketch-to-CAD/STL workflows; per-task cost on ARC tasks reportedly reduced to ~18% of previous levels.
• Google Aletheia math agent: DeepMind's specialized math research 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+verifier architecture; the team plans to extend the approach to physics and computer science.
• OpenAI GPT-5.3-Codex-Spark: Coding model built with Cerebras, in research preview for ChatGPT Pro (Codex app/CLI/IDE extensions). Claims 1000+ tokens/s generation, 128k context, text-only. Community notes human code review is now the bottleneck and speculates it is a large MoE with tens of billions of active parameters.
• MiniMax M2.5: 230B total parameters, 10B active sparse MoE targeting long-horizon agent workloads. Reports 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench, 76.3 BrowseComp. Near-Opus performance on app dev/bug fixing at roughly 1/13 the price (~$1/hour at 100 tok/s; ~$10k/year for four 24/7 instances). Available on OpenRouter and NetMind.
• Zhipu GLM-5: Strongest open-weight model on several leaderboards — Intelligence Index 50 (claimed open-weight #1 with lowest hallucination rate), #1 open model on Arena text and code. Rumored ~744B total / ~40B active params, trained on 28.5T tokens, reportedly uses DeepSeek sparse attention, optimized for long context and agent engineering. GGUF weights available for local inference.
• DeepSeek V4 teaser: Expected around Feb 17 with 1M-token context, positioned as a cost-effective alternative to Opus and Codex.
• Tiny QED-Nano: A 4B natural-language theorem-proving model near larger-model performance on IMO-ProofBench, relying on 1M+ token test-time compute scaffolding and rule-scored RL post-training; weights and details planned to be open-sourced.

Agents & Tooling

• MiniMax M2.5 & GLM-5 agent race: Both models target agent coding — task decomposition, long execution, complex tool chains, multi-file refactoring; adopted by IDEs and agent frameworks (Cline, YouWare, Verdent).
• Cursor long-running Agent: Supports sustained multi-step tasks; combined with Codex Spark, codegen feels near-instant. Pricing/quotas for Composer 1.5 are unclear and raising concerns.
• A2A protocol: Andrew Ng promotes DeepLearning.AI's Agent2Agent protocol for inter-agent discovery and invocation, interoperating with IBM ACP, Google ADK, LangGraph, and MCP.
• Mooncake: Moonshot AI/Tsinghua's KV cache serving backend joins the PyTorch ecosystem, addressing the LLM inference "memory wall" with prefill/decoding disaggregation, cross-instance cache reuse, and elastic expert parallelism; works as a backend for SGLang, vLLM, TensorRT-LLM.
• Lightweight tools over all-in-one LLMs: e.g., a no-API-key Google Search MCP (Chromium-based, with YouTube transcription, image search, local OCR) and SigLIP2 for bulk image tagging instead of multimodal LLMs.
• Traces: A platform for sharing and browsing full coding-agent session traces (Claude Code, Codex, Gemini, Cursor exports).
• Tool updates: Aider v0.86.2 (DeepSeek-V3.2 still seen as best value coder); Windsurf adds Opus 4.6, SWE-1.5, GPT-5.3-Spark to its Arena; OpenRouter adds 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, ABI-stable quantization stack.
• GPU MODE kernel-generation experiment: Plans to spend $20–30k of compute over 4–5 days in late February testing AI-written CUDA/Triton kernels with Qwen3/GLM4.7 Flash, Kernelbot, Flashinfer-bench, NCU, and Compute-Sanitizer — SFT baselines first, then RL.
• TraceML: Open-source one-line hook for real-time PyTorch DDP observability (per-rank step times, jitter, distributions) to spot a lagging GPU.
• RTX 3060 12GB deals: Zotac clearance at ~$200/card; two cards give 24GB VRAM, an economical local-LLM/CUDA build compared to used V100s.
• NCU/ROCM notes: Explanations of NCU "Local(57)" fields (local memory instructions, often register spilling), tcgen05.cp/.st/.mma usage, and AMD Quick Reduce actually working on MI250X/CDNA2 despite official guards.

Research & Methods

• Anthropic "Introspection": An extra network monitoring main-model activation/weight anomalies. Critics (e.g., Unsloth) see it as a redundant deviation detector — "a pressure gauge on a pressure cooker" — while supporters note practical safety value in detecting subtle manipulation signs.
• BlendFER-Lite: MediaPipe 3D expression features + LSTM for edge-device facial expression recognition; 71% accuracy on FER2013 with low compute. Code on Hugging Face.
• Rank-1 LoRA vs full RL: Thinking Machines experiments show rank-1 LoRA fine-tuning can match full-model RL fine-tuning on some reasoning tasks; follow-up work discusses whether this is just in-context learning.
• GoodfireAI RLFR: Reinforcement Learning from Feature Rewards — extract interpretable features first, then build more stable reward functions for open-ended tasks.
• Emergent abilities debate: Nick Bostrom's new paper on advanced AI risk (arXiv:2511.10643) vs. the "emergence as a mirage" argument (arXiv:2304.15004); disagreement affects AGI timeline forecasts.

Products & Applications

• Simile: Raised $100M (investors include Karpathy and Fei-Fei Li) to build "society-scale digital twins" using generative agents + foundation models for pre-launch decision testing (e.g., earnings-call rehearsal, policy simulation).
• ChatGPT medical saves: Multiple Reddit users report ChatGPT urging immediate care that uncovered pulmonary embolism, heart attack, shingles; consensus: useful as triage/second opinion, not a replacement for doctors.
• Kimi/DeepSeek/GLM coding feedback: Kimi K-2.5 praised for multimodal web tasks but criticized for messy file creation; GLM-5 and DeepSeek V3.2 more stable in tool calls and long context. Users commonly mix models — one as "brains," another as "long-code grunt."
• Weave Robotics Isaac 0: $8,000 (or $450/month) laundry-folding home robot, first deliveries Feb 2026, San Francisco Bay Area only.
• Personal AI use cases: From jailbroken fiction writing to bulk resume generation and auto-applications — mundane needs, not AGI dreams, drive individual retention.

Industry & Companies

• Anthropic funding: New round at ~$380B valuation; reported annualized revenue of $14B; Claude Code ARR reportedly doubled to $2.5B this year.
• Perplexity backlash: Pro's Deep Research quietly cut to 20–50 runs/month, prompting cancellations and migration to Google AI Pro or self-built search+LLM stacks; API billing and support also criticized.
• xAI power questions: Unverified community claims that Grok compute partly relies on unpermitted gas turbines plus grid power, highlighting infrastructure risk in competing with OpenAI/Anthropic.
• Anthropic absorbing electricity costs: Announced it will cover electricity price increases in some regions to keep enterprise pricing stable — a signal of how large AI energy bills have become.
• Local vs cloud split: Communities diverge between local LLM setups (3060s, LM Studio, Unsloth — privacy/control) and cloud coding assistants (Cursor, Claude Code, Copilot — convenience).

Policy, Governance & Safety

• BASI jailbreaking escalation: Claude Code jailbreak patched; new DAN scripts targeting GPT-5.2 and Gemini 3 Fast; misspelled keywords to bypass filters; a fully automated red-team system "HAIL MARY" built with Manus runs 24/7 against latest safety policies.
• "AI psychosis" cases: Psychiatry podcast discusses patients whose delusions were amplified through LLM conversations, suggesting long high-emotion chatbots may need usage safeguards.
• RLHF deception debate: Discussion that RLHF rewards persuasiveness over truth, producing models exceptionally good at confident nonsense; counterpoint: it's a training-objective design issue, not inherent "evil."
• KOKKI v15.5 pattern: Community proposal requiring models to output Draft + Audit together — higher token/latency cost in exchange for observability and accountability; core claim: truly guaranteed systems must be deterministic, so Transformers can only aim for auditability and controlled error distributions.