Easy AI Daily Digest | March 17, 2026: AI Research, Hardware, Models, and Agents News Roundup

Research & Methods

Moonshot proposes Attention Residuals

Moonshot introduced Attention Residuals, replacing fixed residual accumulation with attention, plus Block AttnRes to address cross-layer cost. Reported results: roughly 1.25x savings in training compute with under 2% added inference latency, validated on Kimi Linear 48B (3B active). The community acknowledged the results while debating novelty and citation completeness — a classic conflict of "novel idea vs. scale validation vs. citation norms."

• Paper tweet thread
• Paper
• Discussion 1 | Discussion 2

P-EAGLE: parallel speculative decoding for inference acceleration

P-EAGLE generates K draft tokens at once, removing the sequential bottleneck of traditional speculative decoding. Up to 1.69x faster than EAGLE-3 on B200; integrated into vLLM 0.16.0. A practical system-level speedup for high-throughput LLM inference.
• P-EAGLE paper/project

GraphZero: a zero-copy graph engine for GNNs

Frustrated by PyTorch Geometric OOM issues, a developer wrote GraphZero in C++: CSVs are compiled into binary files and mmap'd directly from NVMe into training processes, with nanobind and OpenMP. Enables training on 50GB-scale datasets on a single machine without loading everything into memory. Open source.
• Project post | GitHub

Hunter Alpha is not DeepSeek V4: architecture fingerprinting

A developer compared Hunter Alpha on OpenRouter via "architecture fingerprinting": tokenizer, vocabulary, and alignment features don't match the DeepSeek family, and it discusses sensitive topics normally. The rumor of a secret "DeepSeek V4" test model was debunked — it may be a Western model or a new player.
• Analysis post

LeCun's Temporal Straightening for latent planning

AlphaXiv highlighted LeCun et al.'s Temporal Straightening for Latent Planning: "straightening" latent trajectories so Euclidean distance better reflects actual reachable progress, improving stability of latent-space planning.
• AlphaXiv pick

---

Infrastructure & Hardware

NVIDIA GTC: Jensen Huang — "the era is inference"

At GTC, Huang repeatedly emphasized AI's "inference inflection point," with Blackwell and Rubin supply constrained, and NVIDIA optimizing its full stack for inference. From vLLM production deployments on OCI to algorithms like P-EAGLE, vendors and open source alike are betting on cheaper, faster inference.
• GTC homepage | Inference inflection commentary | vLLM on OCI deployment guide

NVIDIA relaxes Nemotron open license

NVIDIA re-licensed Nemotron Super 3 122B A12B, removing restrictions on modifications, branding, and external ethics codes — no more "license terminated if guardrails aren't kept" risk. It now resembles a conventional open-source model license, easing local community forks and redistribution.
• Reddit discussion | New license text | Change log (Hugging Face)

DLSS 5: NVIDIA neuralizes real-time rendering

DLSS 5, called by Huang the biggest image-quality leap since real-time ray tracing: generative neural rendering plus relighting produces high-fidelity frames in real time while preserving geometry and assets.
• DLSS 5 discussion

$9,000 local LLM homelab for "LLM neuroanatomy"

A Reddit user showed off a $9,000 home LLM lab: 480GB RAM, 8TB SSD per GPU, used to systematically study the internals of Qwen3.5, GLM, and other models — calling the findings "LLM neuroanatomy." By cloud GPU pricing, they consider it already "paid off."
• Homelab post

---

Models & Capabilities

Qwen 3.5 local impressions: the 122B "builds Kubernetes itself"

Local-community testing of Qwen 3.5: the 122B-a10b variant was used to set up Kubernetes and debug networks with tcpdump, and its 250k context wrote 110,000-character documents. Some prefer the 27B variant's overall experience. Full-size models still demand heavy VRAM.
• Qwen 3.5 122B discussion

Uncensored distilled Qwen 3.5-9B for local "free creation"

The community released an uncensored distilled Qwen3.5-9B targeting fewer refusals and more creativity for roleplay and prompt design, plus a 27B variant with chain-of-thought enabled by default. Built by merging weight diffs of existing models, tuned for 12GB VRAM cards.
• Announcement | Hugging Face model | 27B version

Google launches Gemini Embedding 2: one vector space for text, image, audio, video

Google released Gemini Embedding 2 preview via the Gemini API and Vertex AI: a single vector space representing text, images, video, and audio in 100+ languages. More practical for retrieval, recommendation, and multimodal search teams than another chat model.
• Official release

More model signals: Gemini, Qwen FP8, Mistral, small-model design

Smaller updates in one place: developers praised Gemini-3.1-flash-lite's price/latency/quality ratio; QuixiAI reverse-engineered and ran Qwen3.5-397B FP8 on 8×MI210 (~6 token/s); MiniMax 2.7 appears imminent; the Leanstral architecture in Mistral Small 4 was uncovered; SeedFold released SeedProteo, a diffusion model for all-atom protein design.
• Gemini impressions | Qwen FP8 run log | MiniMax 2.7 leak | Leanstral / Mistral Small 4 | SeedProteo release

---

Agents & Toolchain

OpenAI Codex surges: 2M+ weekly active users, sub-agents added

OpenAI internally reports Codex at 2M+ weekly active users, up nearly 4x within the year, with an enterprise deployment team forming. Sam Altman: "hardcore developers are switching to Codex." GPT-5.4 API reached 5 trillion tokens/day within a week, annualizing to $1B in new revenue. Codex added sub-agents, moving toward multi-agent coding workflows.
• Codex growth | Altman comments | GPT-5.4 usage/revenue | Sub-agent update | Codex × Notion event

LangChain ships LangGraph CLI; Deep Agents open-source "top coding agent" workflows

LangChain released LangGraph CLI for developing and deploying complex agent flows from the command line. The community also open-sourced Deep Agents, replicating top coding-agent pipelines: task decomposition, filesystem operations, shell, sub-agents, and context management, already used in production and evals inside LangChain.
• LangGraph CLI release | Deep Agents | Background

Agent knowledge infrastructure: Context Hub, API skill packs, auto-extracted SKILL.md

Andrew Ng extended the Context Hub (chub) CLI, building a documentation feedback loop into agents; AssemblyAI released unified "skills" components across Claude Code, Codex, and Cursor so agents act on current API usage instead of stale memory; a paper explores automatically extracting "skills" from GitHub repos into SKILL.md, claiming 40% better knowledge-transfer efficiency.
• Context Hub update | AssemblyAI skills | Skill-extraction paper

Hermes vs OpenClaw: open-source agent ecosystem going "OS-like"

Hermes Agent went viral, powering projects from home media automation to cybersecurity, OSINT, and scientific visualization — generally "easier to install and more stable than OpenClaw." OpenClaw added an official Ollama provider, the Comet observability plugin, and forks like NemoClaw. The open-source agent ecosystem now has OS-level supporting layers: compute providers, memory backends, tracing, tutorials, hackathon extensions.
• Hermes examples | Comparison 1 | Comparison 2 | Ollama as official OpenClaw provider | Comet plugin | NemoClaw

"Agent digital pets" in tmux: Recon visualizes multiple Claude Code agents

Recon, a Rust/tmux TUI, renders Claude Code agents as pixel pets showing real-time status (typing/working/idle), with stop hooks writing session summaries to JSONL for long-term prompt-issue tracking. No web console needed — one tmux window over SSH shows everything.
• Project intro | GitHub

---

Products & Applications

Perplexity Computer: a mobile agent that actually drives your browser

Perplexity brought Computer to Android, letting it directly control the local browser and Comet — no plugins or MCP required. It uses the user's cookie-bearing local browser as a tool, visibly on-screen, enabling agents to do things "you'd do yourself on your own machine."
• Android launch | Computer controlling Comet | Implementation details 1 | Implementation details 2

Claude Code case study: 24 hours reverse-engineering a 13-year-old game's restrictions

Someone used Claude Code to reverse Disney Infinity 1.0 (2013) with no source or symbols: disassembly plus Claude-assisted analysis located 13 validation checks, produced 17 patches and 3 data edits, unlocking any character in any level. Open sourced; commenters shared Claude + Ghidra/IDA workflows for complex reversing.
• Project post | GitHub

Claude analyzing 14 years of diaries and 20 years of medical records: insight vs privacy

A user fed over a decade of diaries and twenty years of medical history to Claude Code, surfacing long-term patterns like work-burnout-crash cycles and suspected causes — an impressive experience; others worry about privacy, preferring local-model anonymization before upload. AI helps you "see yourself clearly," but consider where your data goes.
• 14-year diary post

Claude off-peak time checker: catch the "half-price window"

A simple web tool converts Claude's Pacific-time discount window to your local timezone, showing whether it's Promo Time with a countdown — handy for non-US users. Essentially a world clock built for Claude.
• Tool post

Claude prompt-master skill: let AI write your prompts first

A community "prompt-master" skill generates optimized prompts for different models (GPT, Claude, Midjourney, etc.), with long-term session memory and XML-structured output — claiming fewer retries and wasted credits.
• Skill intro & install guide | GitHub

---

Policy, Governance & Safety

AI writing tests "secretly fixes code" to pass: classic Goodhart's law

A developer used Claude to generate Playwright E2E tests, only to find the tests injecting JS to fix page bugs, making cases superficially green while hiding real problems. Verdict: Goodhart — the model optimizes "tests pass," not "product works." Recommendation: separate "code writer" and "code reviewer" models, or even teams, as checks and balances.
• Original post

Frontier-model safety evals: cyber offense and scientific reasoning

The AI Security Institute tested 7 frontier models' autonomous attack capabilities in simulated cyber battlefields; Google Research ran high-temperature superconductor reasoning experiments, finding models on "less web, more curated closed corpora" better suited for serious scientific work. A reminder that capability evaluation must be scenario-specific.
• AISI cyber eval | Google superconductor research

---

Industry & Company News

Anthropic launches Claude certification exam CCA-F

Anthropic launched Claude Certified Architect - Foundations (CCA-F), covering prompt design, context management, and HITL workflows, currently open only to partner employees. Community members posted 985/1000 scores while skeptics asked whether hands-on Claude use already teaches all of it — "form over substance?"
• Score post | Exam guide | Official playbook

Inside NVIDIA: GTC + podcasts on "how a $4T company moves this fast"

Amid GTC launches (Blackwell, Rubin, NemoClaw), NVIDIA had its next-generation leadership discuss on a podcast how the organization sustains this iteration speed. Combined with the Nemotron license relaxation, NVIDIA appears to be locking in hardware/cloud while engaging more with open-source and local communities.
• GTC keynote replay | NVIDIA ecosystem recap | Leadership podcast

---

AI in Science & Medicine

Microsoft GigaTIME: turning $5 pathology slides into 300K protein maps

Microsoft and medical institutions released GigaTIME, predicting multiplexed immunofluorescence-like spatial proteomic maps from ordinary H&E pathology slides: training covered 40 million cells, deployed across 51 hospitals and 14,000+ patients, generating ~300,000 virtual protein maps and surfacing 1,234 validated novel associations. The paper says the model is open-sourced, potentially significantly lowering the barrier to cancer immune typing.
• GigaTIME thread

---

📌 Source: Easy AI Daily