Easy AI Daily News | February 13, 2026

A digest of the day's most significant AI industry developments, curated from forum and community discussions.

Models and Capabilities

Google launches Gemini 3 Deep Think V2: 84.6% on ARC-AGI-2

Gemini 3 Deep Think V2 is rolling out to paid Gemini app users with early access via Vertex AI/Gemini API. The reasoning mode scores 84.6% on ARC-AGI-2 (ARC-verified), 48.4% on Humanity's Last Exam, Codeforces Elo 3455, plus olympiad-level physics/chemistry and IMO performance. It has been applied to semiconductor materials design, math paper review, and sketch-to-CAD/STL workflows, with per-task ARC costs reportedly cut to ~18% of previous levels.

• Google DeepMind announcement
• Official Google blog post
• ARC Prize verification and cost details

Google Aletheia math agent: IMO 100, ProofBench 91.9%

DeepMind's math research agent Aletheia scored 91.9% on IMO-ProofBench Advanced and 100% on IMO 2024 tasks, well above general models like Gemini 3 Pro. It uses a generator-verifier architecture — more a specialized proof agent than a general LLM — with plans to extend the approach to physics and computer science.

OpenAI releases GPT-5.3-Codex-Spark: ultra-low-latency coding model

OpenAI, in partnership with Cerebras, launched GPT-5.3-Codex-Spark as a research preview for ChatGPT Pro users via Codex apps/CLI/IDE extensions. It claims 1000+ tokens/second generation with 128k context (text-only for now). Community feedback: generation is now so fast that human code review becomes the bottleneck; future IDEs need better diffing, task splitting, and error prevention. Speculation suggests a large MoE model with tens of billions of active parameters.
• OpenAI blog
• Cerebras partnership

MiniMax M2.5: 230B MoE, SWE-Bench Verified 80.2%, ultra-cheap agent model

MiniMax M2.5 uses a sparse 230B-total/10B-active architecture targeting long-horizon, multi-tool agent workflows. Official numbers: SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3%, BrowseComp 76.3. Independent OpenHands tests show performance near Opus on app development and bug fixing at roughly 1/13 the price — about $1/hour at 100 tok/s, or ~$10,000/year for four 24/7 instances. Available on OpenRouter, NetMind, and other platforms.

Zhipu GLM-5: new open-weight leader at Intelligence Index 50

GLM-5 ranks among the strongest open models: Intelligence Index score of 50 (claimed first among open weights, with the lowest hallucination rate), and #1 among open models on Arena text and code leaderboards, near gpt-5.1-high and Kimi overall. Rumored specs: ~744B total parameters, ~40B active, trained on 28.5T tokens with DeepSeek sparse attention, optimized for long context and agentic work. GGUF builds are available for local inference.
• Unsloth GLM-5 GGUF

DeepSeek V4 incoming: 1M-token context claimed

DeepSeek officially teased V4 for around February 17 with 1M-token context support, viewed by the community as a cost-effective alternative to Opus and Codex.

Tiny QED-Nano: 4B proof model with million-token inference scaffolding

QED-Nano is a 4B natural-language theorem-proving model approaching large-model performance on IMO-ProofBench. It relies on heavy test-time compute — agent scaffolding expanding single proofs to 1M+ tokens — plus RL post-training with rubric-based rewards. Weights and training details are planned for open release.

Agents and Tooling

• The new agent-coding model war: MiniMax M2.5 and GLM-5 both target agentic workflows — task decomposition, long execution, complex tool chains — with GLM-5 already integrated into Cline, YouWare, Verdent, and other frameworks.
• Cursor launches long-running agents: Paired with Codex Spark, code generation and deployment feel near-instant, though pricing and quotas are reportedly inconsistent across users.
• A2A protocol: Andrew Ng promoted DeepLearning.AI's Agent2Agent protocol for inter-agent discovery and invocation, interoperating with IBM ACP, Google ADK, LangGraph, and MCP.
• Mooncake joins PyTorch: Moonshot AI/Tsinghua's KV-cache serving backend addresses the LLM inference "memory wall" with prefill/decode disaggregation, cross-instance cache reuse, and elastic expert parallelism, working as a backend for SGLang, vLLM, and TensorRT-LLM.
• Lightweight tools over monolithic models: Trend toward specialized components — a no-API-key Google Search MCP (Chromium-based, with YouTube transcription and local OCR), and SigLIP2 vision encoders for bulk image tagging instead of multimodal LLMs.
• Traces: A new platform for sharing and browsing full coding-agent session traces exported from Claude Code, Codex, Gemini, and Cursor.
• Tool ecosystem updates: Aider v0.86.2 released; Windsurf added Opus 4.6, SWE-1.5, and GPT-5.3-Spark to its Arena; OpenRouter quickly onboarded MiniMax M2.5 and GLM-5.

Infrastructure and Hardware

• torchao v0.16.0: Adds MXFP8 MoE training components for expert parallelism, removes legacy configs, and improves ABI stability — key for low-precision MoE training and deployment.
• GPU MODE's $20–30K kernel-generation experiment: In late February, the community will spend 4–5 days of compute testing Qwen3/GLM4.7 Flash models on CUDA/Triton kernel generation, with SFT baselines before RL.
• TraceML: Open-source real-time observability for PyTorch DDP — per-rank step times and jitter distribution with one line of code, exposing stragglers in multi-GPU training.
• RTX 3060 12GB as the budget CUDA workhorse: LM Studio community members spotted Zotac clearance pricing (~$200/card); two cards give 24GB VRAM for mid-size local models.
• NCU/ROCm Q&A: Community explanations of NCU's Local(57) metric (local memory instructions, often from register spilling), tcgen05 instruction usage, and AMD Quick Reduce working on MI250X/CDNA2 despite official guards limiting it to MI300.

Research and Methods

• Anthropic's "Introspection": A paper proposing an auxiliary network monitoring main-model activations/weights. Skeptics (including Unsloth) see it as redundant anomaly detection — "a pressure gauge on a pressure cooker" — while supporters note practical safety value in detecting subtle manipulation.
• BlendFER-Lite: LSTM + MediaPipe facial expression recognition reaching 71% on FER2013 with far less compute than large models; suitable for edge devices.
• Rank-1 LoRA matching full RL fine-tuning: Thinking Machines experiments show rank-1 LoRA can approach full-model RL fine-tuning on some reasoning tasks; follow-up discussion debates whether this is just in-context learning.
• GoodfireAI's RLFR: Reinforcement Learning from Feature Rewards — extracting interpretable features first, then building reward functions from them for open-ended tasks.
• Emergent abilities debate continues: Nick Bostrom's new paper examines advanced AI risk philosophically, while the technical community remains split on whether LLM "emergence" is real or an artifact of metrics and sampling (A Theory of Emergent Behaviour, Are Emergent Abilities of LLMs a Mirage?).

Products and Applications

• Simile: Raised $100M (backers include Karpathy and Fei-Fei Li) to build "societal digital twins" — generative-agent simulations for rehearsing earnings calls and policy decisions.
• ChatGPT as a pre-diagnosis step: Multiple Reddit users report ChatGPT urged immediate care that uncovered pulmonary embolism, heart attack, or shingles. Consensus: useful for listing differential diagnoses, but only as a second opinion.
• Chinese model coding experiences: Kimi K-2.5 praised for multimodal web browsing but criticized for messy file creation; GLM-5 and DeepSeek V3.2 more stable on tool calling and long context. Users increasingly mix models — one as "brain," another as "code workhorse."
• Weave Robotics Isaac 0: A $8,000 laundry-folding home robot ($450/month subscription), first deliveries February 2026, San Francisco Bay Area only.

Industry and Business

• Anthropic raises at ~$380B valuation: Reported annualized revenue of ~$14B, with Claude Code ARR said to have doubled to $2.5B this year.
• Perplexity cuts Deep Research quotas silently: Pro users dropped from near-unlimited to 20–50/month, sparking cancellations and complaints about API billing and support.
• xAI power sourcing questioned: Community reports allege unpermitted gas turbines plus grid power for Grok training/inference — unverified, but highlighting infrastructure risk in competing at frontier scale.
• Anthropic covers electricity price increases: The company will absorb power cost hikes in some regions, underscoring how directly electricity volatility hits AI margins.
• Local vs. cloud divide widens: Hobbyists stock up on 3060s and run LM Studio/Unsloth for privacy; enterprises default to Cursor, Claude Code, and Copilot.

Policy, Governance, and Safety

• Jailbreaking industrializes: The BASI community reports patched Claude Code exploits, new DAN scripts for GPT-5.2 and Gemini 3 Fast, and an automated red-team system ("HAIL MARY") built on Manus that probes safety policies 24/7.
• "AI psychosis" cases reach psychiatry podcasts: Reports of LLM conversations amplifying delusions in vulnerable users suggest chatbots may need usage-safeguard designs similar to gaming addiction protections.
• Does RLHF train better liars?: Ongoing debate over whether RLHF rewards persuasiveness over truth; critics frame it as an objective-design problem rather than inherent model misbehavior.
• KOKKI v15.5 draft-plus-audit pattern: A community-proposed output format where models produce both a draft and a self-audit, trading latency for observability.

---

📌 Source: Easy AI Daily