This week in AI - Maya Arazi's Newsletter #1

Exposition

Welcome to the first "This week in AI" — a weekly digest of the AI news you shouldn't miss. I started it because keeping up has become a part-time job in itself, and I figured the least I could do for the people I work with (and anyone else trying to actually keep up) is take that job off their plate.

Each week you'll get one email and a fresh page on mayaarazi.com/news — covering major news, newly released tools and models, GitHub repos worth a clone, a paper or two that matters, and posts from some underrated underdog I follow on X.

Here's the trick: you don't have to read all of it. Every section is a standalone. Skim "What everyone's buzzing about" for the headlines (so you know what people are talking about over dinner), jump to "Tools to actually use" for the practical stuff, or scroll straight to your sector at the bottom — startups, research teams, government, enterprise, or solo operators. The sector sections re-cut the week through that one specific lens, so they work as a "just give me what's relevant to me" fast lane if you're short on time.

Two asks. Share it. Forward freely to anyone who'd benefit (or send them to mayaarazi.com/en#newsletter to subscribe). And tell me what's missing. If a sector falls flat or there's a topic you'd love to see covered, reply and tell me — every piece of feedback shapes the next edition. ♥️

Alrighty, on to the week —

What everyone's buzzing about

Anthropic says AI "recursive self-improvement" may arrive soon — Claude now writes 80% of Anthropic's merged code, and the company says the world should keep open the option to slow or temporarily pause frontier AI development. Anthropic

Claude Opus 4.8 ships 6 weeks after 4.7 — Stronger across coding, agentic tasks, and professional work, with the consistency and autonomy to keep working on long-running tasks. Anthropic honestly describes the release as "a modest but tangible improvement" rather than hype. Anthropic

White House executive order on AI innovation and cybersecurity — Advanced AI capabilities make the nation stronger, but also introduce new national security considerations; the June 2 order establishes a voluntary framework for frontier model review and an AI cybersecurity clearinghouse. White House

Google ships Gemma 4 QAT models optimized for edge — Quantization-Aware Training checkpoints dramatically reduce memory requirements, reducing Gemma 4 E2B to 1GB footprint while preserving quality. Google

Anthropic becomes most valuable private AI startup — $65B Series H closed late May 2026 at $965B post-money, overtaking OpenAI's $852B private valuation. AI Funding Tracker

Tools to actually use

Headroom — Compresses tool outputs, logs, files, and RAG chunks before they reach the LLM, achieving 60–95% token reduction while maintaining quality and accuracy. Netflix teams report $700K saved and 200 billion tokens freed. → What to do: Drop it in as a proxy with zero code changes to cut API bills immediately. GitHub

LiteParse — Open-source document parsing library that parses text with spatial layout and bounding boxes, written in Rust, runs entirely on your machine with no cloud dependencies, no LLMs, and no API keys. → Why it matters: Preserves table structure and layout for AI agents without routing PDFs through third-party APIs. LlamaIndex

Hermes Desktop (my verdict - worth the hype!) — Native macOS, Windows, and Linux GUI bundled with Hermes Agent v0.15.2 under MIT license; the only agent with a built-in learning loop that creates skills from experience and improves them during use. → What to do: Install if you want persistent memory and multi-platform agent workflows without vendor lock-in. Nous Research

Defending Code Reference Harness — Open-source framework for autonomous vulnerability discovery and remediation using Claude; years of learnings from partnerships with security teams. → What to do: Run the quickstart to scan threat model → find → verify → patch on your own code. GitHub

Worth a clone

chopratejas/headroom — Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 60–95% fewer tokens, same answers. Library, proxy, MCP server. Library, proxy, and MCP modes; 6 compression algorithms including SmartCrusher for JSON, CodeCompressor for ASTs, and reversible CCR. GitHub

run-llama/liteparse — Standalone OSS PDF parsing tool focused on fast and light parsing with high-quality spatial text parsing and bounding boxes, without proprietary LLM features or cloud dependencies. Rust core, Node/Python/WASM bindings, ships with OCR. GitHub

NousResearch/hermes-agent — Self-improving AI agent with built-in learning loop that creates skills from experience, improves them during use, and searches past conversations. 180K+ stars, MIT license, runs on $5 VPS or serverless. GitHub

The paper that matters

Do transformers need three projections? Systematic study of QKV variants — Examines whether the standard query-key-value architecture in attention mechanisms is necessary or if simpler variants perform comparably. Worth reading if you're optimizing inference cost or training smaller models from scratch; incremental but directly applicable. arXiv

From my feed

Programmers will document for Claude, but not for each other — HN discussion (177 pts) on how AI coding assistants are finally getting engineers to write explanatory comments, not because humans asked, but because the model needs context. Hacker News

South Korean online communities must scan all images with AI — Due to recent Telecommunications Business Act changes, forum owners must scan every user-uploaded image and video; hardware not provided by government, putting financial pressure on small sites. Starts July 2026. Privacy Guides

For startups

Cognition raises $1B at $26B valuation — The Devin maker closed late May, the week's second-largest round after Anthropic. Application-layer agentic AI is still pulling eleven-figure capital, but the bar is now "real revenue or a deep moat" — pure-demo rounds are gone. Crunchbase

Defense tech funding hits $14.6B in five months — Already past 2025's full-year record of $9.6B; dual-use autonomy and AI-for-defense are leading the surge. → If you build dual-use AI or autonomy: the window is wide open right now and probably won't be in twelve months. Crunchbase

Gemini 3.5 Flash resets the API floor at $1.50/$9 per million tokens — ~40% under Gemini 3.1 Pro at near-Pro quality (Terminal-Bench 2.1 76.2%, MCP Atlas 83.6%). → What to do: A/B test it against whatever you're paying Opus rates for before your next budget cycle — the cost math shifted this week. Google Cloud

For research teams

Microsoft MAI-Thinking-1 ships with full data lineage — A 35B-parameter reasoning model, 256K context, trained from scratch with zero distillation on clean, commercially licensed data; raters preferred it to Sonnet 4.6 in blind eval. → Why it matters: for regulated domains and IP-audited research, it's the first frontier-tier reasoning model that ships with provable data provenance. Microsoft Build

Claude Code's "dynamic workflows" spawn hundreds of parallel subagents at runtime — Research preview only; an orchestrator dispatches subagents and synthesizes results. → How to use: treat it as a test surface for the next 4–6 weeks, not stable infrastructure — the semantics will shift before GA. Anthropic

OpenAI extends GPT-5.5-Cyber to vetted EU cyber teams — Limited preview to EU businesses, governments, and institutions in cybersecurity. → For security-research groups: an EU institutional affiliation is now a meaningful access vector this week — worth checking eligibility if you're working on offensive/defensive evals. BuildFastWithAI

For government & public sector

Colorado AI Act enforcement begins June 30 — High-risk AI deployers (hiring, lending, housing, healthcare) need impact assessments, security risk-management programs, and anti-discrimination safeguards. The grace window is closed; the State AG is actively enforcing. → What to do: audit your AI-touched consequential-decision pipelines now, not after. VerifyWise

NY and Vermont pass three AI laws in one week — Vermont's H 816 bans therapy chatbots (signed May 29); NY's S 9408A bans AI chatbot toys, plus S 6954 (disclosure) and S 1169 (high-risk audit + algorithm-discrimination). → Pattern to watch: category-by-category bans (toys, therapy) are now the template — broad omnibus bills aren't moving. Transparency Coalition

Anthropic expands Project Glasswing to critical infrastructure — June 2 expansion adds power, water, healthcare, communications, and hardware vendors; partners' codebases touch 100M+ people. → For public-sector buyers: vendor-side AI cybersecurity is now its own procurement category — expect it on the next RFP cycle. Anthropic

For enterprise

OpenAI on AWS Bedrock is GA — Frontier OpenAI models and Codex now run inside AWS-native security, compliance, billing, and governance, and the usage counts toward your existing AWS commit. → What to do: if you've been blocked on vendor-review friction, this is the unblocking path — your AWS account team can scope a pilot this quarter. AWS

Microsoft MAI-Code-1-Flash lands in GitHub Copilot — First Microsoft model trained inside Copilot's production harness; 85.8% on Microsoft adversarial coding, ~51% SWE-Bench Pro, 60% fewer tokens on complex tasks, rolling out to ~10% of users first. → Why it matters: an inference-efficient Copilot default shifts your coding-agent vendor calculus — re-bench it against your internal evals before renewals. Microsoft Build

Multi-step prompt injection is now the dominant attack vector — Production-data analysis this week: direct override attempts are universally blocked, but multi-step hijacks that embed sequences across several inputs evade detection. → What to do: if your security team is still defending against single-turn jailbreaks, your threat model is a year out of date — update the red-team brief this sprint. BuildFastWithAI

For solo operators & small teams

xAI Grok Build launches as a terminal-native coding agent — Early beta for SuperGrok and X Premium Plus subscribers (announced May 27). A fifth credible competitor to Claude Code / Codex / Aider / Cline. → If you're a heavy coder: worth a one-week trial before your next monthly subscription decision. xAI

Opus 4.8 ships effort control + 3× cheaper Fast mode — Pick thinking depth via `/effort` on claude.ai; Fast mode is now three times cheaper than the previous generation, with batch + cache savings on top. → What to do: default to Fast for chat, jump to `/effort xhigh` for the gnarly stuff — the cost math now strongly favors batching non-interactive jobs. Anthropic

GitHub Copilot switches to token billing on June 1 — Flat-rate subscription is gone; per-credit pricing wasn't disclosed at launch. → For solo devs: model your real usage NOW before the cost picture changes mid-cycle — and check whether Cursor or Grok Build comes out cheaper for your workload. BuildFastWithAI