Easy AI Daily — February 13, 2026

A full roundup of the day's AI news from the Easy AI Daily, covering new models, agent tooling, infrastructure, research, products, industry moves, and safety.

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, IMO-level physics/chemistry/math. Already used for semiconductor materials design, math paper review, and sketch-to-CAD/STL workflows. Claims per-task ARC cost reduced to ~18% of previous levels.
• Google DeepMind announcement | Google blog | ARC Prize verification
• Google Aletheia: DeepMind's 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 CS.
• Aletheia introduction
• OpenAI GPT-5.3-Codex-Spark: Built with Cerebras, research preview for ChatGPT Pro Codex users. Claims 1000+ tok/s generation, 128k context, text-only for now. Community notes that human code review is now the bottleneck; speculation points to a tens-of-B-active MoE.
• OpenAI blog | Cerebras announcement
• MiniMax M2.5: 230B total / 10B active MoE targeting long-horizon agent workloads. SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3%, BrowseComp 76.3. Near-Opus performance at ~1/13 the price (~$1/hour at 100 tok/s); available on OpenRouter and NetMind.
• Official page | OpenRouter
• Zhipu GLM-5: Scores 50 on the Intelligence Index (claimed #1 open-weights, lowest hallucination), #1 open model on Arena text and code leaderboards. Rumored ~744B total / ~40B active params, 28.5T training tokens, DeepSeek-style sparse attention, GGUF available for local inference.
• Intelligence Index discussion | Unsloth GGUF
• DeepSeek V4: Teased for ~Feb 17 with 1M-token context; community sees it as a cost-effective rival to Opus and Codex.
• Discussion
• Tiny QED-Nano: A 4B theorem-proving model approaching larger models on IMO-ProofBench, using million+ token test-time agent scaffolds and RL post-training with rubric-based rewards; weights to be open-sourced.
• Introduction thread

Agents & Tooling

• Agent-model race: M2.5 and GLM-5 both target agentic coding — task decomposition, multi-file refactoring, long tool chains — and are being integrated into Cline, YouWare, Verdent, and other frameworks.
• Cursor long-running agents: New feature supports sustained multi-step refactors; pairing with Codex Spark yields near-instant generation, though pricing/quotas are inconsistent and under discussion. (announcement)
• A2A protocol: Andrew Ng promotes DeepLearning.AI's Agent2Agent protocol for inter-framework agent discovery and calls, bridging IBM ACP, Google ADK, LangGraph, and MCP. (post)
• Mooncake: Moonshot/Tsinghua's KV-cache serving backend joined the PyTorch ecosystem; supports prefill/decode disaggregation, cross-instance cache reuse, and elastic expert parallelism as a backend for SGLang, vLLM, and TensorRT-LLM. (PyTorch announcement)
• Lightweight tools: Community favors specialized components — e.g., a no-API-key Google Search MCP based on Chromium, and SigLIP2 for bulk image tagging instead of multimodal LLMs. (noapi-google-search-mcp)
• Traces: A new platform for browsing full coding-agent session traces exported from Claude Code, Codex, Gemini, and Cursor. (traces.com)
• Tool updates: Aider v0.86.2 (DeepSeek-V3.2 still seen as best price/performance coder), Windsurf Arena adds Opus 4.6 / SWE-1.5 / GPT-5.3-Spark, OpenRouter rapidly onboards M2.5 and GLM-5.

Infrastructure & Hardware

• torchao v0.16.0: Adds MXFP8 MoE training components for expert parallelism; removes legacy quantization options for a leaner, ABI-stable stack. (release notes)
• GPU MODE kernel experiment: ~$20–30K of compute over 4–5 days in late February to test AI-generated CUDA/Triton kernels with Qwen3/GLM4.7 Flash, using Kernelbot and Flashinfer-bench evaluation. (FlashInfer Bench docs)
• TraceML: Open-source real-time observability for PyTorch DDP — per-rank step times and jitter with one line of code, exposing straggler GPUs. (GitHub)
• RTX 3060 12GB clearance: ~$200/card at Zotac makes 24GB dual-card setups attractive for local LLM inference; LM Studio community is bulk-buying. (example listing)
• NCU/ROCm notes: Community explains NCU's Local(57) field (local memory instructions, often register spilling) and tcgen05 instructions; AMD Quick Reduce reportedly works on MI250X/CDNA2 despite official guards limiting it to MI300. (Quick Reduce blog)

Research & Methods

• Anthropic Introspection: A monitoring network watches the main model's activations/weights for anomalies. Critics (e.g., Unsloth) call it a redundant pressure gauge; supporters note practical safety value in detecting subtle manipulation. (paper)
• BlendFER-Lite: LSTM + MediaPipe 3D expression features hit 71% on FER2013 with low compute, suited to edge/robotics; code and models released. (paper)
• Rank-1 LoRA vs full RL: Thinking Machines shows rank-1 LoRA fine-tuning can match full-model RL on some reasoning tasks, possibly via in-context learning effects. (blog | arXiv:2406.04391)
• GoodfireAI RLFR: Reinforcement Learning from Feature Rewards — extract interpretable features first, then build more stable reward functions from them for open-ended tasks.
• Emergence debate continues: Nick Bostrom's new paper (arXiv:2511.10643) and the ongoing "are emergent abilities a mirage?" argument (arXiv:2304.15004) still split the community.

Products & Applications

• Simile: Raised $100M (investors include Karpathy and Fei-Fei Li) to build society-scale digital twins using generative agents for pre-launch decision testing.
• ChatGPT medical catches: Multiple Reddit users report ChatGPT urged immediate care that uncovered pulmonary embolism, heart attack, and shingles; consensus is LLMs are useful for triage/second opinions, not diagnosis.
• Chinese-model coding feedback: Kimi K-2.5 praised for web browsing and site cloning; GLM-5 and DeepSeek V3.2 more stable for tool calls and long context. Common pattern: one model as "brain," another as "long-code workhorse."
• Weave Robotics Isaac 0: $8,000 (or $450/month) laundry-folding home robot, first deliveries February 2026, SF Bay Area only. (weaverobotics.com)
• Everyday AI use cases: From creative writing jailbreaks to batch-generating resumes and auto-applying to jobs — mundane needs, not AGI dreams, drive individual retention.

Industry & Business

• Anthropic raises at ~$380B valuation; internal figures suggest ~$14B annualized revenue, with Claude Code ARR reportedly doubled to $2.5B this year. (announcement)
• Perplexity quietly cuts Deep Research quotas for Pro users to 20–50/month, triggering cancellations and migration to Google AI Pro or self-built stacks.
• xAI power controversy: Unverified claims that Grok compute partly relies on unpermitted gas turbines plus grid power — emblematic of the infrastructure race with OpenAI/Anthropic.
• Anthropic covers electricity price increases for enterprise customers, underscoring how directly power costs now hit AI margins. (announcement)
• Local vs cloud split deepens: Hobbyists buy 3060s and run LM Studio/Unsloth for privacy; enterprises standardize on Cursor, Claude Code, and Copilot.

Policy, Governance & Safety

• Jailbreak tooling industrializes: The BASI community reports patched Claude Code exploits, new DAN scripts for GPT-5.2/Gemini 3 Fast, filter bypasses via misspellings, and an automated red-team system "HAIL MARY" built on Manus that probes safety policies 24/7.
• "AI psychosis" cases: A psychiatry podcast documents patients whose delusions were amplified by LLM chat, prompting calls for engagement-limiting design for vulnerable users. (episode)
• Does RLHF train better liars? Debate in the Yannick Kilcher community: RLHF may reward persuasiveness over truth, though others argue it's an objective-design issue rather than inherent "evil."
• KOKKI v15.5: A proposed dual-output pattern (Draft + Audit) where models self-review; trades tokens/latency for observability, on the thesis that transformers can only offer auditability and controlled error distribution, not guarantees.

📌 Source: Easy AI Daily