Cold Open — Samsung hands ChatGPT and Codex to its whole workforce

Monday, June 22, 2026. We scanned more than 2,100 fresh items off the wire overnight. Three made the front page, a handful more made the radar, and one of them owns the morning before your coffee cools.

🎧 This is the print twin of today's Cold Open episode. Listen to today's episode.

The lead · Samsung puts ChatGPT and Codex in front of its entire workforce

OpenAI announced that Samsung Electronics is deploying ChatGPT Enterprise and Codex to its employees — to all Samsung Electronics employees in Korea, and to all employees worldwide in its Device eXperience (DX) division. OpenAI calls it one of its largest enterprise deployments to date, and Samsung says it will use the tools across the whole company: software development, marketing, product development, and manufacturing.

The detail builders should notice is what Codex is now being sold to do. It started as a tool for writing, reviewing, and debugging code — but OpenAI is explicitly pitching it for non-technical teams too, "to turn ideas into working software, internal tools, websites, and automated workflows." The coding agent is being framed as a company-wide platform, not a power-user toy.

"Samsung Electronics, a global leader in technology and manufacturing, is embracing AI not as a tool limited to certain teams or functions, but as a core platform for improving how employees around the world work and innovate." — Harrison Kim, General Manager, OpenAI Korea

Why it matters

The numbers under the announcement are the real story. OpenAI says more than 5 million people now use Codex every week, and that Codex weekly active users in Korea have grown nearly 800% since February 1, 2026. When a coding agent goes from a developer's side tool to standard-issue across a company the size of Samsung Electronics, two things follow for the rest of us: the agent stops being optional in the way teams plan work, and the definition of "who codes" gets a lot blurrier when non-engineers can ask Codex to build the internal tool they used to file a ticket for.

The fine print

This is an adoption announcement, not a capability benchmark. The productivity gains are claimed by the vendor and the customer, not independently measured — and a company-wide license is the easy part. The hard part, the part these rollouts still have to prove, is governance, security, and whether work actually changes once the seats are handed out. Read it as a signal of where the market is heading, not as evidence of the outcome.

Sources: openai.com

02 · Cloudflare gives AI agents (and everyone) throwaway deploys

Cloudflare shipped temporary accounts for Workers. You can now run npx wrangler deploy --temporary and Cloudflare will deploy your application to a fresh, ephemeral project that stays live for 60 minutes — without you creating a Cloudflare account at all. If you want to keep it, the deploy hands you a claim link to take ownership before the timer runs out. Simon Willison tested it end to end and it worked as advertised.

Why it matters. The framing is "for AI agents," and that fits — an autonomous agent that can stand up a live URL with zero setup and let it expire on its own is genuinely useful. But as Willison points out, the AI hook undersells it: accountless, self-expiring deploys lower the friction of shipping for everyone, agent or human.

Sources: blog.cloudflare.com · simonwillison.net

03 · sqlite-utils 4.0rc1 brings migrations and nested transactions

Simon Willison released sqlite-utils 4.0rc1, the first release candidate for v4 of his widely used Python library and CLI for working with SQLite. Two features lead it: bundled database migrations (a port of his battle-tested sqlite-migrate package, now built in) and a new db.atomic() API for nested transactions. The major version bump flags some minor backwards-incompatible changes, and he is asking people to test the RC before he commits to a stable release.

Why it matters. Migrations are the unglamorous backbone of shipping data changes without breaking things. Baking them — plus safe atomic transactions — into a tool that already underpins countless small scripts and projects lowers the bar for doing schema changes the right way.

Sources: simonwillison.net

Also on the radar

• Open OCR — PP-OCRv6: PaddlePaddle's new 50-language OCR family on Hugging Face, scaling from 1.5M to 34.5M parameters so you can trade accuracy for footprint.
• Physical AI safety — NVIDIA Halos for Robotics: a full-stack functional-safety system aimed at robots and physical AI — the unsexy layer that has to exist before agents touch the real world.
• Agentic science — NVIDIA Vera CPU at Los Alamos: infrastructure pitched specifically for agentic scientific AI at a national lab.
• Small + local — fine-tuning Qwen 3 0.6B: a clear writeup on getting good results from a tiny local model for question categorization — a reminder that not every job needs a frontier model.

Trends in dev tools

What moved in the tools engineers actually ship with.

• Coding agents are going company-wide. Codex passing 5 million weekly users — and growing nearly 800% in Korea since February — is the clearest sign yet that the coding agent is becoming default infrastructure rather than a power-user habit. (openai.com)
• The deploy loop is being redesigned for agents. Cloudflare's wrangler deploy --temporary lets an agent (or a human in a hurry) push to a live URL with no account and no setup, then expire on its own. (blog.cloudflare.com)
• Boring-but-critical plumbing is moving into everyday tools. sqlite-utils 4.0rc1 bakes schema migrations and atomic transactions into a CLI millions of small scripts already lean on. (simonwillison.net)
• The contest is moving "beyond the model." Industry roundups of June's release blitz note the action shifting from raw model launches toward the agent runtime — orchestration and governance — as the thing vendors compete on. (drvikramsingh.ai)

Popular skills

The agent-skills wave — portable SKILL.md folders an agent loads on demand — kept compounding.

• One format, every agent. Anthropic's open Agent Skills format now runs across 30+ platforms — Claude Code, Codex, Gemini CLI, GitHub Copilot, Cursor — so a skill written once works in all of them.
• The catalog is exploding. Researchers at Bosch Research and Carnegie Mellon counted the public skills ecosystem growing 18.5× in 20 days — from ~2,179 listings to over 40,000 — with npx skills add becoming the npm install of agent capabilities. (firecrawl.dev)
• Vendors are shipping official skills. Supabase now ships agent skills you install with npx skills add supabase/agent-skills, handing your agent safe database and security guidance instead of hoping it guesses right. (supabase.com)

AI fun fact

The three pillars under today's models can be traced to a single lab in 1991. By Jürgen Schmidhuber's own account in "Munich 1991: The Roots of the Current AI Boom", his Munich group that year prototyped unsupervised pre-training for very deep networks (the "Neural History Compressor"), an attention mechanism that scaled linearly (an ancestor of the Transformer he calls "fast weight programmers"), and artificial curiosity built from two neural nets competing — a precursor to GANs. Pretraining, attention, and adversarial generation, all sketched out thirty-five years ago.

Sources: openai.com · blog.cloudflare.com · simonwillison.net · huggingface.co · firecrawl.dev · people.idsia.ch