The New Gorky Boy

Today’s issue: TanStack’s growing family, becoming a geriatric React Native developer, and late-stage capitalism strikes again.

Welcome to #512.

---

The New Gorky Boy

It’s been 2 months since SpaceXAI agreed to acquire Cursor (a deal that was finalized today), and with the release of Grok 4.6, we’re starting to see the fruit of our labor data.

According to the “trust me bro” benchmarks, Grok 4.6 is at the current frontier, costs 1/10th as much as Fable 5, and is significantly faster. This is quite an accomplishment even if the benchmarks don’t reflect real-world usage. So the question is: did they cook, and if so, what’s the recipe? Let’s find out.

The recipe

Ingredients:

• Curated synthetic data
• Cursor data (thank you for your contribution, comrade)
• An improved optimizer

Instructions:

• Cook for longer than previous models
• Do some supervised fine-tuning for STEM
• Add a dash of reinforcement learning for agentic tasks

The result

The real question is how this model performs in real life. Thankfully, Eric Zakariasson from the Cursor team wrote down some observations about how it performs.

• Short prompts: Apparently telling Gorky boy to “work harder” doesn’t work with this model. Instead, writing short prompts with clear preferences yields good results. Takeaway: the model has “good taste” (allegedly).

• Strict verification: Instead of giving it open-ended tasks, a prompt like keep on iterating and verifying until it's production ready will force the model to check its own work as it goes, leading to better outcomes.

• Synchronous work: Because of the speed and intelligence of the model, Eric was able to stay in more of a flow state while working on a specific task, rather than following the add-context-and-wait approach most of us are used to.

Bottom Line: Ant and Ope are going to release models that push the frontier far beyond where Grok 4.6 is, but in the long run, fast + cheap + “smart enough” is probably what wins. Good job, Gorky.

---

Agents write the code. Someone still has to fix CI.

[code]smith is an AI agent that optimizes your CI for you — fixing failing builds, addressing reviews, and rewriting your workflows for speed or cost optimization.

• Autofix failing PRs — with [code]smith, you never see a broken PR again.

• Optimize your CI — flag over-provisioned jobs and rewrite workflows for speed and cost savings.

• Build any feature — kick off [code]smith from a Slack message or delegate a ticket straight from Linear.

One click to get started, and you’ll never push a “fix lint” commit again.

---

1. The TanStack family grew again with their latest child package, Hotkeys, which provides a type-safe, framework-agnostic way to add keyboard shortcuts to your app.

2. React Native released version 0.87, which features a “strict” TypeScript API, a Metro update, and the realization that I’ll be in my 80s when they hit 1.0.0.

3. The new Software Stewardship Lab is raising funds to help sustain the open-source maintainers and software supply chains the industry depends on. [sponsored]

4. Ben Vinegar created Hunk, a terminal diff viewer built to make reviewing agent code more manageable.

5. Vicki Boykis wrote an essay about AI-generated PR descriptions and why you, sir / ma’am, should just write it yourself.

6. Nathan Sobo wrote about software craftsmanship in the era of vibes.

7. OpenRouter gives you one API to access 400+ models. You get better prices, better uptime, and no subscriptions to worry about. [sponsored]

8. DeepSeek got a lot more expensive. I thought inference was free in China wtf.

9. Florian Herrengt wrote about the disappearing middle class in software engineering. Late-stage capitalism is coming for everything smh.

10. Lea Verou explains why you can skip the tri-state toggle for “dark mode” on your personal site. And no, it’s not because no one visits it, but that’s also true.