Easy AI Daily – February 26, 2026

A structured digest of AI industry news for February 26, 2026.

Products & Applications

• Perplexity launches "Computer": An all-in-one agent workstation covering research, design, coding, deployment, and operations. Uses parallel asynchronous sub-agents plus a coordinating model that auto-selects models per task, with usage-based billing, spend caps, and memory/file/tool management. Rolling out to Max users first. Product post | Pricing | Architecture breakdown
• Claude Code turns one: Anthropic is positioning Claude Code as a coding agent foundation and promoting legacy-system modernization (e.g., COBOL). Markets read this as a threat to IBM's mainframe services, driving IBM stock down more than 10% at one point, though real-world results remain unproven. IBM stock discussion | Anniversary podcast
• GitHub Copilot CLI goes GA: Adds a /research command for deep, repo-wide analysis built on GitHub code search and MCP tools, generating exportable reports (gist) with live task status in terminal titles. Feature intro | GA announcement
• Nous open-sources Hermes Agent: Python-based multi-agent workbench with hierarchical memory, sub-agents, filesystem/terminal control, browser actions, and seamless session continuation across CLI and messaging apps. Pairs with Atropos for large-scale data generation and RL pipelines. GitHub | Portal
• LM Studio launches LM Link: Built on Tailscale, enables secure remote access to local LLMs on home or cloud machines without exposing ports. LM Link

Models & Capabilities

• OpenAI releases GPT-5.3-Codex: Available via API; community testing shows ~25% faster than 5.2 with fewer tokens per task and strong SWE-Bench Pro results. Priced at $1.75/M input and $14/M output tokens, sparking value debates. Dev announcement | Cline benchmarks
• Alibaba ships Qwen 3.5 Medium (27B / 35B-A3B / 122B-A10B): Open-weight, MoE, day-one support in 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 | Long-context notes
• Grok-4.20-Beta1 tops Arena Search with a score of 1226, ahead of GPT-5.2 and Gemini-3, and ties 4th on Text (1492). Leaderboard
• Liquid AI releases LFM2-24B-A2B: A 24B sparse MoE activating only 2B per token, runnable on 32GB-memory devices; day-one llama.cpp, vLLM, SGLang, and GGUF support. Trained on 17T+ tokens so far, to be upgraded to LFM2.5. Reddit thread
• Diffusion LLM inference engines: Inception Labs and others claim ~1000 tok/s via diffusion-based approaches and inference-time scaling techniques like Ψ-Samplers; still experimental and awaiting community reproduction. Andrew Ng commentary

Agents & Tooling

• Karpathy: coding agents "really work" since December 2025 — he recently completed a full local deployment (SSH, vLLM, model pull, load testing, serving, frontend, systemd, reporting) almost fully autonomously, noting a qualitative jump in long-horizon task coherence. Thread
• ActionEngine: Treats GUI/web automation as graph search — offline exploration builds a state machine, and inference uses a single LLM call to generate the whole action program, claiming better success rate, latency, and cost than step-by-step visual agents. Intro
• OpenClaw as a system-level agent: Widely used for desktop/file/browser control and automation (email, CRM, finance), with security concerns — one user granted root and had their trash emptied — plus community-built three-tier persistent memory stacks. Usage thread
• Aider community's budget model stack: Deepseek V3.2 for main reasoning (cheap, slightly slow), mimo-v2-flash for fast file edits, and Kimi-k2.5 for planning on hard problems. Discussion

Infrastructure & Hardware

• Karpathy on memory orchestration: The real bottleneck is scheduling prefill/decode between fast-small on-chip SRAM and large-slow DRAM under long-context, high-concurrency agent workloads; neither HBM-heavy nor large-SRAM routes solve it well. Thread
• OpenAI and Meta receive 160M AMD warrants (~$600 strike target, up to ~$192B theoretical value) as effectively an equity rebate tied to GPU purchases. Analysis
• Blackwell GPU cloud price war: Packet.ai offers hourly billing from ~$0.66/hour or $199/month for training, pushing individuals and small teams toward rental options like Lightning AI. Packet pricing | Lightning AI clusters
• Zagora: Distributed fine-tuning platform combining scattered consumer GPUs over the public internet to train 70B+ models (GPT-OSS, Qwen 2.5, Mistral), using Petals/SWARM-style pipelining. Project

Research & Methods

• Agent reliability lags capability: Benchmark scores climb while reliability does not; agents often derail after one bad tool call. Researchers call for minimal, strictly enforced safety benchmarks (e.g., never sending emails amid irrelevant context). Reliability discussion | Minimal benchmark proposal
• Trace-Free+ (Intuit): Curriculum training teaches models to rewrite tool descriptions into agent-friendly formats, improving multi-tool reliability without extra traces at inference. Intro
• Goodfire: Infrastructure collecting billions of activations on trillion-parameter models with minimal latency impact, including a demo of live chain-of-thought steering via activation analysis. Thread
• Midtraining: A paper systematizes the training stage between pretraining and post-training; it can reduce forgetting and improve downstream performance but is highly sensitive to timing and data distribution. arXiv:2507.06203
• Diffusion / Flow Matching survey collection: Eleuther community roundup covering Rectified Flows, Flow Matching, Diffusion Forcing, and newer papers, plus a video playlist. Rectified Flows | Diffusion Forcing | Playlist

Industry & Company News

• Anthropic acquires Vercept, a computer-use agent startup, to strengthen Claude's ability to actually operate interfaces for non-technical users. Announcement
• Wayve raises $1.5B Series D at an $8.6B valuation (SoftBank, Microsoft, NVIDIA, Uber), planning supervised robotaxi pilots in 10 cities in 2026 and selling embodied AI hardware/software to automakers from 2027. Funding details
• Quiver AI raises $8.3M seed (a16z-led) and releases Arrow-1.0, generating editable SVG vector graphics from sketches or text for UI/posters/icons. Announcement

Policy, Governance & Safety

• Pentagon–xAI deal and pressure on Anthropic: Reports say the DoD reached an agreement to use Grok in classified systems and demanded Anthropic allow "all lawful uses" of Claude, including mass surveillance and weapons development. Anthropic publicly refuses mass surveillance and autonomous weapons and reportedly faced Defense Production Act threats. Coverage | More discussion
• Anthropic softens its RSP: Per TIME, Anthropic dropped its most aggressive Responsible Scaling pledge (halting training of stronger models until safety is proven), with its chief scientist citing unsustainability of unilateral commitments absent competitor follow-through. TIME report
• Jeff Dean opposes mass surveillance: Arguing it suppresses free speech, is prone to abuse, and violates constitutional spirit; community debate follows on delegating enforcement to agents that cannot refuse unlawful orders. Statement
• Energy constraints emerge: The US government reportedly considering requiring large AI/cloud providers to self-supply power to protect the grid and consumers, underscoring that scaling is now an infrastructure and energy policy issue. Discussion
• Automated jailbreak agents raise compliance risks: A BASI community member built a self-updating jailbreak agent (OpenClaw + DeepSeek-R1) generating multi-turn covert jailbreak prompts for Claude, GPT, Gemini, and Grok — reviewers flagged it as violating virtually every vendor's terms of service. Risk review

📌 Source: Easy AI Daily