Easy AI Daily News | February 26, 2026: Perplexity Computer, GPT-5.3-Codex, Qwen 3.5, and More

Product & Applications

• Perplexity "Computer": An all-in-one agent workstation that handles research, design, coding, deployment, and operations in a single interface. Uses parallel asynchronous sub-agents plus a coordinator model, with usage-based billing, spending caps, and memory/file/tool management. Available first to Max users. Release post
• Claude Code turns one: Anthropic pitched Claude Code as a foundation for coding agents, including legacy-system (COBOL) modernization. Markets read this as a threat to IBM mainframe services, and IBM stock briefly fell over 10%. Discussion
• GitHub Copilot CLI goes GA: Adds a repo-level /research command leveraging GitHub code search and MCP tools to generate deep-dive reports exportable as gists. GA announcement
• Nous open-sources Hermes Agent: Python-based, with multi-tier memory, sub-agents, file/terminal control, browser actions, and session continuity across CLI and chat apps; integrates with Atropos for data generation and RL pipelines. GitHub
• LM Studio LM Link: Tailscale-based secure remote access to local LLMs without exposing ports. Details

Models & Capabilities

• GPT-5.3-Codex available via API: ~25% faster than 5.2, strong SWE-Bench Pro results; pricing $1.75/M input, $14/M output tokens. Dev announcement
• Qwen 3.5 Medium (27B, 35B-A3B, 122B-A10B): Open weights, 800K–1M context, near-lossless 4-bit + KV quantization, day-one support in vLLM, GGUF, LM Studio, Ollama. Developers report 35B-A3B local agent tool-calling near commercial cloud models, with only ~3B active parameters per token. Release
• Grok-4.20-Beta1 ranks #1 on Arena Search (1226), ahead of GPT-5.2 and Gemini-3; tied 4th on Text (1492). Leaderboard
• Liquid AI LFM2-24B-A2B: 24B sparse MoE activating 2B per token, runs on 32GB devices, supports llama.cpp, vLLM, SGLang, GGUF; pretraining past 17T tokens. Reddit discussion
• Diffusion LLM inference: Inception Labs and others claim ~1000 tok/s diffusion-based LLM inference; still experimental. Andrew Ng commentary

Agents & Tooling

• Karpathy: coding agents "really started working" since December 2025 — he completed a full local deployment (SSH, vLLM, model serving, load testing, frontend, systemd) nearly fully automated. Thread
• ActionEngine: treats GUI automation as graph search, offline-explores a state machine, then generates a full action program with one LLM call — better success rate, latency, and cost than step-by-step visual agents. Method intro
• Aider community model stack: Deepseek V3.2 for main reasoning, mimo-v2-flash for fast edits, Kimi-k2.5 for planning hard problems.

Infrastructure & Hardware

• Karpathy on the real bottleneck: memory orchestration between small/fast on-chip SRAM and large/slow DRAM, not raw compute. Full thread
• OpenAI and Meta receive 160M AMD warrants (~$600 strike target), effectively equity rebates on GPU purchases.
• Blackwell cloud price war: Packet.ai at ~$0.66/hour or $199/month. Pricing
• Zagora: distributed fine-tuning platform combining scattered consumer GPUs over the public internet to train 70B+ Transformer models. Project

Research

• Agent reliability lags capability: benchmarks keep rising but agents still derail on single tool-call errors; calls for minimal safety benchmarks. Discussion
• Trace-Free+ (Intuit): curriculum training teaches models to rewrite tool descriptions, improving multi-tool reliability without extra traces. Intro
• Goodfire: interpretability infrastructure collecting billions of activations on trillion-parameter models with minimal latency impact.
• Midtraining: a training phase between pretraining and post-training reduces forgetting and improves downstream performance, but is highly sensitive to timing and data distribution. Paper
• Diffusion/Flow Matching survey resources: Rectified Flows, Diffusion Forcing, video playlist

Industry & Business

• Anthropic acquires Vercept to strengthen Claude's computer-use capabilities. Announcement
• Wayve raises $1.5B Series D ($8.6B valuation; SoftBank, Microsoft, NVIDIA, Uber); plans monitored robotaxi pilots in 10 cities in 2026, selling embodied AI stack in 2027.
• Quiver AI raises $8.3M seed (a16z-led), releasing Arrow-1.0 for generating editable SVG from sketches or text.

Policy, Governance & Safety

• Pentagon AI negotiations: xAI reaches a deal to use Grok in classified systems; Anthropic reportedly pressured to allow mass surveillance and weapons use, facing Defense Production Act threats for refusing. Coverage
• Anthropic softens RSP: drops its pledge to pause training frontier models if unsafe, citing unilateral commitments as unsustainable without competitor buy-in. TIME report
• Jeff Dean publicly opposes mass surveillance AI uses, citing free-speech and constitutional concerns.
• Energy constraints: US reportedly considering requiring large AI/cloud companies to self-provision power as data centers strain the grid.
• Automated jailbreak risks: a self-updating jailbreak agent built on OpenClaw + DeepSeek-R1 raises major compliance and security concerns.

📌 Source: Easy AI Daily