Easy AI Daily | February 26, 2026

A roundup of AI industry news from zhichai.net's Easy AI Daily for February 26, 2026.

Products & Applications

• Perplexity launches "Computer": an all-in-one agent workstation that combines research, design, coding, deployment, and operations in a single interface. Under the hood it uses parallel asynchronous sub-agents plus an orchestrator model that selects models per task, with usage-based billing, spending caps, and memory/file/tool management. Available first to Max users. (Launch post | Pricing | Architecture breakdown)
• Claude Code turns one: Anthropic is positioning Claude Code as a coding-agent platform and pitched a legacy-system modernization use case (COBOL). Markets read this as a threat to IBM mainframe services — IBM stock briefly fell over 10% — though whether it can actually overhaul critical financial systems remains unproven. (IBM reaction 1 | IBM reaction 2 | Anniversary podcast)
• GitHub Copilot CLI goes GA, adding a repo-level /research command built on GitHub code search and MCP tools that produces deep-dive reports exportable as gists; the terminal title shows live task status.
• Nous open-sources Hermes Agent: a Python multi-agent workbench with multi-level memory, sub-agents, filesystem/terminal control, browser actions, and session continuity across CLI and chat apps. Works with Atropos for large-scale data generation and RL pipelines. (GitHub | Portal)
• LM Studio's LM Link: securely access local LLMs on home or cloud machines over Tailscale without exposing ports. The community is asking for a mobile client and a mode that doesn't depend on third-party accounts. (LM Link)

Models & Capabilities

• OpenAI ships GPT-5.3-Codex via API: roughly 25% faster than 5.2, fewer tokens per task, strong on SWE-Bench Pro. Pricing: $1.75/M input tokens, $14/M output tokens, sparking "expensive but strong" debates. (Announcement | Cline benchmarks | Price discussion)
• Alibaba releases Qwen3.5 Medium (27B / 35B-A3B / 122B-A10B, open weights, MoE) with day-one support across vLLM, GGUF, LM Studio, and Ollama; near-lossless at 4-bit + KV quantization and 800K–1M context. Developers report 35B-A3B's local agent tool-calling rivals commercial cloud models while activating only ~3B parameters per token. (Release | Long context & FP8 | Local agent feedback | Arena listing)
• Grok-4.20-Beta1 takes #1 on Arena Search with a score of 1226, beating GPT-5.2 and Gemini-3; it also ties for 4th on the Text leaderboard at 1492. (Leaderboard | Announcement)
• Liquid AI releases LFM2-24B-A2B: a 24B sparse MoE activating only 2B per token, runnable on 32GB-memory devices, with day-one support in llama.cpp, vLLM, SGLang, and multiple GGUF quants. Pretrained on 17T+ tokens (ongoing); a future LFM2.5 upgrade is planned. (Reddit discussion)
• Diffusion LLM inference engines: Inception Labs and others demonstrated diffusion-based LLM inference claiming ~1000 tok/s, further accelerated by inference-time scaling techniques like Ψ-Samplers. Still experimental; real-world performance awaits community reproduction. (Andrew Ng commentary | Speed discussion | Diffusion Duality paper lead)

Agents & Tooling

• Karpathy: coding agents "actually work" since December 2025. He describes deploying a local stack almost entirely by agent — SSH setup, vLLM install, model downloads, load testing, serving, frontend, systemd, and a written report — noting a qualitative jump in long-task coherence. (Thread)
• ActionEngine treats web/GUI automation as graph search: it explores offline into a state machine, then generates an entire action program with a single LLM call at inference time, claiming better success rate, latency, and cost than step-by-step visual agents. (Method intro)
• OpenClaw as a system-level agent: individuals and teams use it to control files, browsers, and full machines for email/CRM/finance automation — but safety concerns abound (one user granted root and had their trash wiped; others built three-layer persistent memory stacks). (Usage thread | Discord discussion)
• Aider community's budget multi-model stack: DeepSeek V3.2 for main reasoning (cheap but slower), mimo-v2-flash for fast file edits, moonshot Kimi-k2.5 for hard-problem planning with mimo for implementation — seen as a good cost/experience balance. (Aider Discord)

Infrastructure & Hardware

• Karpathy on the real bottleneck: memory orchestration, not raw compute. He splits LLM compute into small-fast on-chip SRAM vs. large-slow external DRAM, arguing that scheduling prefill and decode between the two tiers under long-context, high-concurrency agent workloads is the core unsolved problem — for both HBM and big-SRAM approaches. (Full thread)
• OpenAI and Meta receive 160 million AMD warrants (~$600 strike target) tied to large GPU purchases — effectively an equity rebate on GPU spend, with theoretical returns up to $192B in market value. (Deal analysis)
• Blackwell cloud price war: Packet.ai offers Blackwell GPUs from ~$0.66/hour or $199/month; individuals and small teams increasingly rent via Lightning AI rather than buying B200s. (Packet pricing | Lightning AI clusters)
• Zagora builds distributed fine-tuning by pooling scattered consumer GPUs over the public internet to train 70B+ Transformer models (GPT-OSS, Qwen 2.5, Mistral), using a Petals/SWARM-style pipelined approach. (Project)

Research & Methods

• Agent reliability lags capability: multiple works note benchmark scores soar while reliability doesn't — agents often derail after one bad tool call and errors compound. Some propose "extremely simple but strictly enforced" minimal safety benchmarks (e.g., never send emails regardless of irrelevant context). (Reliability survey | Failure modes | Minimal benchmark proposal)
• Trace-Free+ (Intuit): a curriculum-style training that first teaches models to rewrite tool descriptions into agent-friendly formats, improving multi-tool reliability without extra traces at inference. (Intro)
• Goodfire demonstrated interpretability infrastructure collecting billions of activations on trillion-parameter models with minimal latency impact, including live chain-of-thought steering via activation analysis. (Technical thread)
• Midtraining — a short training stage between pretraining and instruction tuning — reduces forgetting and boosts downstream performance, but is highly sensitive to timing and data distribution; done carelessly it hurts. (Paper)
• Diffusion / Flow Matching survey series: the Eleuther community compiled recent work from Rectified Flows and Flow Matching to Diffusion Forcing, plus new papers from ByteDance and Tencent, with a video playlist. (Rectified Flows | Diffusion Forcing | Resources)

Industry & Companies

• Anthropic acquires Vercept, a computer-use agent startup, aiming to make AI actually operate interfaces and complete tasks for users rather than just suggest steps — reinforcing the Claude Code / Claude Tools line. (Announcement | Founder retrospective)
• Wayve raises $1.5B Series D at an $8.6B valuation (investors include SoftBank, Microsoft, NVIDIA, Uber), planning monitored robotaxi pilots in 10 cities in 2026 and selling embodied-AI hardware/software to automakers and robot companies from 2027. (Funding & roadmap)
• Quiver AI raises $8.3M seed led by a16z and launches Arrow-1.0, which converts design sketches or text descriptions into editable SVG vector graphics for UI/poster/icon workflows. (Announcement)

Policy, Governance & Safety

• Pentagon pushes AI military use boundaries: reports say the DoD reached a deal with xAI to use Grok in classified systems and demanded Anthropic permit Claude for "all lawful purposes" including mass surveillance and weapons development. Anthropic publicly refuses mass surveillance and autonomous weapons and was reportedly threatened with the Defense Production Act or supply-chain risk designation. (xAI-Pentagon deal | Anthropic ultimatum | Community discussion)
• Anthropic weakens its RSP: per TIME, Anthropic dropped its flagship pledge not to train stronger models before proving adequate safety, with the chief scientist arguing unilateral commitments are unsustainable when competitors don't follow. Critics call it caving to commercial pressure; others blame the absence of global regulation. (TIME report | Discussion)
• Jeff Dean opposes mass surveillance: he argued publicly that mass surveillance suppresses free speech, is prone to abuse, and violates constitutional spirit. Others worry delegating enforcement to agents that can't refuse unlawful orders makes policing systems harder to constrain. (Statement | Risk discussion)
• Energy constraints surface: the US government is reportedly considering requiring large AI/cloud companies to self-provision power to protect the grid and ratepayers — signaling that scaling is now an infrastructure and energy-policy problem, not just an algorithm/GPU one. (Report)
• Automated jailbreak red-teaming: a BASI community member built a self-updating jailbreak agent using OpenClaw + DeepSeek-R1 to generate multi-turn covert jailbreak prompts against Claude, GPT, Gemini, and Grok. Reviewers noted it violates virtually every vendor's TOS and poses serious compliance risks. (Discussion | Risk review)

📌 Source: Easy AI Daily (zhichai.net)