Feminine Energy

Today’s issue: Getting my stock tips from Astro, an age-old adage, and why package managers deserve a little empathy.

---

Feminine Energy

When we did our Model Roundup way back… wait, that was last issue 🫠… I did not anticipate two more models, Kimi K3 and a new model by Thinking Machines called Inkling, being announced. Rookie mistake.

Today, we’re focusing on Inkling. This is certainly a decision, given that Kimi K3 is beating Fable 5 on many key benchmarks, but that might be a story for another day. So what’s the big deal with Inkling?

Inkling is built on a MoE (mixture-of-experts) architecture with 975B total parameters, 41B active parameters, and a context window of up to 1M tokens. On its own, Inkling is not the best model, ranking around fifth on the leaderboard for most of the relevant benchmarks. What’s interesting, though, is what I can only call an unmistakable feminine energy. I mean this as a compliment, because let’s be honest, what’s more masculine than overstating your capabilities and confidently saying something that is objectively wrong?

Inkling, in contrast, was designed to be more self-aware of its limitations while being confident in what it does know. Here are some of the key attributes of the model:

• Generalist: Inkling is a generalist model with training that includes reasoning, tool calling, instruction-following, and factuality. Instead of being narrowly optimized for a specific domain, it is adaptable.

• Multimodal: In addition to having a broad range of capabilities, Inkling also supports multiple modalities and can interact with audio, text, and vision (images). This could be inaccurate, but it appears to be the first open-weights model to support all three modalities.

• Fine-Tunable: Perhaps Inkling’s most interesting attribute is what Thinking Machines calls “epistemics”, referring to how the model was trained for calibration, instruction-following, and resistance to censorship. This matters because Thinking Machines’ other flagship product is a tool called Tinker, which allows anyone to fine-tune models with frontier-level infrastructure.

Bottom Line: Combining a tunable open-weights model with fine-tuning infrastructure is a genius idea. For American companies looking to control costs and perhaps a little wary of betting everything on a Chinese model, Inkling could be a game-changer.

Should we be surprised? No. The company is run by a woman after all 👸🏽.

---

Write Useful Logs for Production

Not all logs are created equal. Getting the right balance of information without eating up your quota can make a huge difference. That’s why the team at Sentry is doing a workshop to help you and your team get the most out of your logs.

Here’s what it covers:

• What’s Worth Logging: Find out what’s worth logging (and what isn’t) and why string-based console logs mostly just eat your quota.
• Structured Log Best Practices: Learn what attributes to add that make logs searchable, and how to connect them to traces and errors so you can move from “something broke” to “here’s exactly what happened.”
• Tuning Signal-to-Noise: Instead of drowning in data, see how you can filter your logs to keep important signal while reducing the firehose of noise.

This workshop is worth its weight in gold and best of all, it’s free. Sign up today.

---

1. Aiden Bai and co. created an AI benchmark for React. Unsurprisingly, the models are very bad at React, which makes sense because they were trained on your code.

2. SpaceXAI open-sourced Grok Build, its Rust-based coding harness. Now you can see exactly how they plan to steal your .env variables.

3. Hyperagent lets you build a team of specialist agents running their own browser, shell, and file system. They integrate with your tools and share context so they can hand off specific tasks to each other without you babysitting them. Get $500 in free credits here. [sponsored]

4. Phil Eaton wrote an article about open-source distribution and why package managers deserve some empathy.

5. Glauber watched Jarred yolo a Rust port of Bun and decided maybe it would be a good idea to do this for Postgres. He’s really embracing the whole cowboy thing since moving to the US.

6. Laura Summers is tired of being the human in the loop, which is similarly why I stopped attending meetings at my last company.

7. Temporal wrote a whitepaper about durable execution and handling failures without adding complexity to your application. [sponsored]

8. The Astro team released version 7.1, which lets you run multiple dev servers (for your agents), uses less memory for content collections, and gives buy/sell recs for Cloudflare stock. Spoiler: it’s always a buy.

9. Dhravya Shah made a virtual machine that runs in the browser that can run Bun, a shell, Git, and local LLMs. Wut.

10. Build emails, pages, and documents with open-source React components. Elements turns code into visually editable content. [sponsored]

11. Alice Moore wrote about what being agent-native as a team actually means.

12. Adam Jacob wrote a practical guide on how to reduce your token spend. Yes, it does involve paying for another product, but as my mama always said, you gotta spend money to save money.