Welcome back. After taking on the AI chip market, Google is making up for lost time in the frontier model race. Its latest model goes toe-to-toe with one of Anthropic’s flagships at a fraction of the cost. Meanwhile, a new model factory wants to change how devs train, deploy, and improve AI models.
Also: A dev recreated the entire world in Minecraft (you can even find your home), why prompt injection defenses fall short, and why an OpenCode dev banned agents from writing tests.
Today’s Brief
How a SpaceXAI engineer ships code while she sleeps
The coolest way to AirDrop your iPhone photos
Turn your coding agent into a software factory (tutorial)
Build self-improving agents (cookbook)

TODAY IN PROGRAMMING
Google’s new model codes like Anthropic’s flagship for less: The search giant just dropped Gemini 4 Neon, and it’s giving Claude Opus 5.5 a run for its money. Per the company, Neon nearly matches Opus on two coding benchmarks for 25% less ($3/$15 per million tokens). It writes about 110 tokens per second. In Google’s Antigravity coding tool, Neon can hand tasks to smaller parallel agents. Google says Antigravity tripled its score on one coding test. Test it in your own setup before trusting benchmark hype.
This AI lab lets engineering teams train and fine-tune models in-house: Prime Intellect just launched Model Factory, letting you own your model instead of renting one. Frontier labs train a model, ship it, and keep improving it based on how it performs. Now your engineering team can do the same. You can train an open-weights model on your own tasks, deploy it with one click, and improve it using real-world results. You pay per GPU-hour. Get started here.
AI digs up a 10-year-old hole in NGINX: A memory bug has been hiding in the web server’s HTTP/2 code since 2015, and it could let attackers run code on your server without logging in. Cloud security firm Wiz spotted it using its AI code scanner, while Xint and other researchers found it independently. Xint even managed to chain it with a Linux kernel bug to gain root access. No attacks have been reported yet, but you’ll want to patch this one; here’s what to do next.

PRESENTED BY TIGERDATA
Sensors keep generating data, and historical records add up fast. As storage costs climb, teams face a tough choice: delete data they might need later or pay more to keep it.
TimescaleDB eliminates that tradeoff with high-performance Postgres built for time-series data.
Up to 95% compression reduces storage costs while keeping historical data queryable.
Automatic partitioning with hypertables keeps ingestion fast as data volumes grow.
Familiar SQL and tools let your team work with the drivers and workflows they already know.
Keep your data. Cut your storage costs. All in Postgres.

INSIGHT
How a SpaceXAI engineer ships code while she sleeps
Agents are writing code faster than you can review it, and they’re picking up every bad habit in your codebase along the way. Lauren Tan, an engineer at SpaceXAI, went from reviewing around three pull requests a day from colleagues in 2024 to roughly 300 a day now. Instead of trying to keep up, she spent six months fixing the environment her agents work in. The result was 3,000 pull requests shipped to production last month, many while she slept. In a recent talk, she breaks down how she pulled it off. Here’s what you can steal from her playbook:
Fix the code before the prompt. When an agent messes up, don’t rush to add another line to AGENTS.md. Write a lint rule or type that stops it from making the same mistake again.
Weed your codebase. Agents love copying whatever’s already there, including that ancient workaround nobody bothered to delete. Clean up outdated patterns before they multiply.
Ban the footguns. Some tools are just too easy for agents to misuse. Tan’s team banned React’s “useEffect” in product code, and the problem disappeared.
Let folders do the policing. Keep Electron main and renderer code in separate directories and block imports between them. That’s one less mistake you’ll have to catch in review.
Cut the comments. Agents can use them to justify bad shortcuts. Grok Bot bans 99% of comments and relies on lint errors to point agents toward the actual fix.
Code review is shrinking. Tan now puts rules files and AGENTS.md at the bottom of her priority list, saving them for decisions that genuinely need judgment. With fewer things left to catch, most product code now merges without deep human review.
P.S. If you're ready to dive deeper, Lauren Tan's pstack (the methodology she uses to ship 2,000 PRs in a month) is a great place to start.

PRESENTED BY MOZILLA FOUNDATION
That’s produced real breakthroughs, but here are the trade-offs:
You don’t control the model
Your data isn’t just yours anymore
Costs and capacity are out of your hands
Some of the most useful AI is made for a specific community, problem, or place. Mozilla is backing projects that do just that. Sound like yours? Apply by November 20 for up to $500K and keep full ownership.

IN THE KNOW
What’s trending on socials and headlines

Meme of the day.
Block Earth: A dev recreated the entire planet in Minecraft at 1:1 scale, furnished interiors included. Type your real address and walk through your own neighborhood (20M views).
Magical AirDrop: Devs are obsessed with this demo that an engineer vibe-coded with GPT 6.2 Astra. Pinch a photo off your iPhone, open your hand, and see where you land (2.6K likes).
One-Way Door: A senior engineer spent days talking to devs in London and posted his field notes on agent adoption. His sharpest claim is that those who truly switch never go back (1.4K likes).
Weak Boundary: An AI agent startup’s engineering head argues prompt-injection filters alone can’t secure agents. He posted the three-layer defense he uses instead (1.8K bookmarks).
Hot Take: An OpenCode developer deleted his entire test suite and banned his AI agents from writing more. His justification set off a brawl in the replies (1.2K likes).
Astra Review: A senior dev ran GPT 6.2 Astra and Sol on real projects for a full day. He made two changes to his setup and says one feels “like pure magic” (1.2K likes).
Launch Audit: A backend engineer spent an hour on a friend’s vibe-coded app before launch. What he found made the owner rotate every credential that same night (2M views).

TOP & TRENDING RESOURCES
Top Tutorial
How to turn your coding agent into a self-running software factory: This tutorial shows how to use CodeDen to move from a single coding agent to a full “dark factory.” You’ll learn how to orchestrate multiple agents, turn GitHub issues into reviewed pull requests, add automated code review, and eventually let the whole workflow find, build, verify, and merge work on its own.
Top Tool
Devin Bots: Give one bot a goal, and it spins up a team of Devin agents to plan, build, test, review, and push the work toward a merge-ready PR. It can split jobs across parallel cloud machines, watch Slack or Linear for new work, run on schedules, and only pull you in when a decision actually needs you.
Trending Cookbook
How to build proactive agents that get better over time (by LangChain): This guide breaks down what changes when agents stop waiting for prompts and start working toward goals on their own. You’ll learn how to design triggers, memory, and review UX so agents know when to act, what to remember, and when something is actually worth interrupting you for.

AI CODING HACK
How to retune your Codex setup for GPT-6.2 Astra
GPT-6.2 just hit Codex, and your config is still tuned for GPT-6. A dev shared a one-shot audit prompt to fix it.
Step 1: Start Codex on Astra with search on (it needs OpenAI’s new guides):
codex -m gpt-6.2-astra --searchStep 2: Paste the full audit prompt. It starts:
Read these two pages. If you can't get the actual text, stop and tell me...Step 3: It checks “~/.codex/config.toml” and AGENTS.md for wrong default models, effort levels above the guides, and instructions Codex can work out itself.
Astra quotes each bad setting against the guides and returns config to paste. It changes nothing until you approve.

Whenever you're ready to dive deeper
If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:
We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.
Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.
What did you think of today's newsletter?
You can also reply directly to this email if you have suggestions, feedback, or questions.
Until next time — The Code team






