Easy AI Daily | 2026-02-26: AI Industry News Roundup

A translation of the Chinese-language Easy AI Daily digest for February 26, 2026, covering products, models, agents, infrastructure, research, business, and policy.

Products & Applications

• Perplexity launches "Computer": An all-in-one agent workstation for research, design, coding, deployment, and operations in a single interface. Uses parallel asynchronous sub-agents plus a coordinating model, with per-task model routing, usage-based billing, spending caps, and memory/file/tool management. Rolling out to Max users first.
• Product launch post | Pricing & availability | Architecture breakdown
• Claude Code turns one: Anthropic is positioning Claude Code as a coding-agent foundation and launched a legacy-system modernization offering targeting COBOL. Markets read this as a threat to IBM mainframe services, briefly driving IBM stock down over 10%. Real-world viability for critical financial systems remains unproven.
• IBM stock reaction 1 | IBM stock reaction 2 | Anniversary podcast
• GitHub Copilot CLI goes GA: Adds a repo-level /research command using GitHub code search and MCP tools for deep analysis, with report export to gist and live task status in terminal titles.
• Feature intro | GA announcement
• Nous open-sources Hermes Agent: A Python multi-agent workbench with tiered memory, sub-agents, filesystem/terminal control, browser actions, and session continuity across CLI and messaging apps. Integrates with Atropos for large-scale data generation and RL pipelines.
• GitHub repo | Announcement & free trial
• LM Studio launches LM Link: Tailscale-based secure remote access to local LLMs without exposing ports. Community wants a mobile client and a pure-local mode without third-party accounts.
• LM Link

Models & Capabilities

• GPT-5.3-Codex ships on the API: Roughly 25% faster than 5.2 with fewer tokens per task; strong on SWE-Bench Pro. Pricing: $1.75/M input, $14/M output tokens, sparking "expensive but strong" debates.
• Dev announcement | Cline benchmarks | Pricing discussion
• Qwen 3.5 Medium series (Alibaba): Open-weight 27B / 35B-A3B / 122B-A10B models, available day-one via vLLM, GGUF, LM Studio, and Ollama. Official claims of near-lossless 4-bit + KV quantization and 800K–1M token context. Developers report 35B-A3B approaches commercial cloud models for local agent tool-calling while activating only ~3B parameters per token.
• Release & tooling | Long context & FP8 weights | Local agent review | Arena listing
• Grok-4.20-Beta1 tops Arena Search: Score of 1226, ahead of GPT-5.2 and Gemini-3; tied 4th on the Text leaderboard at 1492.
• Arena leaderboard | Announcement
• Liquid AI releases LFM2-24B-A2B: A 24B-parameter sparse MoE activating 2B per token, runnable on 32GB-memory devices; supports llama.cpp, vLLM, SGLang, and multiple GGUF quants. Trained on 17T+ tokens and still training, later to become LFM2.5.
• Reddit thread
• Diffusion LLM inference engines: Inception Labs and others claim ~1000 tok/s diffusion-based generation, further accelerated by inference-time scaling such as Ψ-Samplers. Still experimental; results await community replication.
• Andrew Ng commentary | Speed discussion | Diffusion Duality thread

Agents & Tooling

• Karpathy: coding agents "actually work" since December 2025: He describes fully automated local deployment—SSH setup, vLLM install, model pulls, load testing, serving, frontend, systemd, report writing—crediting a qualitative jump in long-task coherence.
• Thread
• ActionEngine: Treats GUI/web interaction as graph search—offline exploration builds a state machine, and inference needs a single LLM call to generate the full action program, claiming better success rate, latency, and cost than step-by-step visual agents.
• Method intro
• OpenClaw as a system-level agent: Users run it as a local OS-level agent controlling files, browser, and machine for email/CRM/finance automation—raising security concerns (one user with root access had their trash wiped) and prompting custom three-layer persistent-memory stacks.
• Usage thread | Discord discussion
• Aider community's budget model mix: Deepseek V3.2 for main reasoning (cheap, slightly slow), mimo-v2-flash for fast file edits, and moonshot Kimi-k2.5 for planning on hard problems—considered a balanced cost/experience multi-model routing setup.
• Aider Discord

Infrastructure & Hardware

• Karpathy on the memory wall: The real bottleneck is memory orchestration—scheduling prefill and decode between small-fast on-chip SRAM and large-slow DRAM in long-context, high-concurrency agent workloads. Neither HBM nor big-SRAM routes solve it well.
• Full thread
• OpenAI and Meta receive 160M AMD warrants: Effectively "stock rebates" tied to large GPU purchases, with ~$600 strike targets and theoretical value up to $192B, deepening ties between compute suppliers and model companies.
• Deal analysis
• Blackwell GPU price war: Packet.ai offers Blackwell cloud at ~$0.66/hour or $199/month, pushing individuals and small teams toward rental options like Lightning AI clusters.
• Packet pricing | Lightning AI clusters
• Zagora: A distributed fine-tuning platform stitching scattered consumer GPUs over the public internet to train 70B+ models (GPT-OSS, Qwen 2.5, Mistral) using Petals/SWARM-style pipeline training.
• Project page

Research & Methods

• Agent reliability lags capability: Benchmark scores climb while reliability doesn't—one deviant tool call can cascade into compounding errors. Researchers call for minimal safety benchmarks, e.g., never sending emails regardless of unrelated context.
• Reliability discussion | Failure modes | Minimal safety benchmark proposal
• Trace-Free+ (Intuit): Curriculum training teaches models to rewrite complex tool descriptions into agent-friendly formats, improving multi-tool reliability at inference without extra traces.
• Method intro
• Goodfire: Production-scale interpretability infrastructure collecting billions of activations on very large models with minimal latency impact, including a live chain-of-thought steering demo.
• Technical thread
• Midtraining: A paper systematizes the training stage between pretraining and post-training, finding it can reduce forgetting and improve downstream performance—but is highly sensitive to timing and data distribution; poor placement hurts.
• Preprint
• Diffusion/Flow Matching survey roundup: Eleuther community resources covering Rectified Flows, Flow Matching, Diffusion Forcing, and new papers from ByteDance and Tencent, plus a video playlist.
• Rectified Flows | Diffusion Forcing | Resource playlist

Business & Companies

• Anthropic acquires Vercept: A computer-use agent startup; the goal is for AI to actually operate interfaces and complete tasks, not just suggest steps—reinforcing the Claude Code / Claude Tools product line.
• Announcement | Founder's retrospective
• Wayve raises $1.5B Series D: At an $8.6B valuation with SoftBank, Microsoft, NVIDIA, and Uber. Plans supervised 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 (a16z): Launches Arrow-1.0, generating editable SVG vector graphics from sketches or text for UI/poster/icon use cases.
• Funding & model launch

Policy, Governance & Safety

• Pentagon talks with xAI, Anthropic, OpenAI: Reports say the DoD reached a deal to use Grok in classified systems and pressed Anthropic to permit "all lawful uses," including mass surveillance and weapons development. Anthropic refuses mass surveillance and autonomous weapons, reportedly facing Defense Production Act threats or supply-chain-risk designation.
• xAI-Pentagon report | Ultimatum report | Community discussion
• Anthropic waters down its RSP: Per TIME, Anthropic dropped its most aggressive Responsible Scaling Policy pledge—not training stronger models until safety is proven. Its chief scientist argued unilateral commitments are unsustainable when competitors don't follow.
• TIME report | Discussion
• Jeff Dean opposes mass surveillance: He argued large-scale monitoring suppresses free speech, invites abuse, and violates constitutional spirit; others worry delegating enforcement to agents that can't refuse unlawful orders makes policing systems harder to constrain.
• Jeff Dean statement | Risk discussion
• Energy constraints emerge: The US government is reportedly considering requiring large AI/cloud companies to secure their own power supply to protect the grid and consumers—scaling is now an energy-policy issue, not just compute.
• Report
• Automated jailbreak agents raise compliance risk: A BASI community member built a self-updating jailbreak agent with OpenClaw + DeepSeek-R1 generating multi-turn covert prompts for Claude, GPT, Gemini, and Grok. Reviewers flagged it as violating every vendor's TOS, with logging, malicious-model, and rollback risks.
• Discussion | Risk review

📌 Source: Easy AI Daily (zhichai.net translation)