July 2026.3
This week is about efficiency in different forms: making frontier AI cheaper to run, knowing when a side project has already served its purpose, and focusing your career on the work rather than the ladder.
📖 Story 1: Kimi K3 Tech Blog: Open Frontier Intelligence
kimi.com · Read
Moonshot AI introduced Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with native vision and a 1-million-token context window.
Only 16 of its 896 experts are active at once. Moonshot says this makes K3 more efficient to scale than Kimi K2, although running it still requires serious infrastructure. The company recommends deployments with at least 64 accelerators. Open weights are scheduled for release by July 27.
K3 appears competitive with proprietary frontier models, particularly for coding and long-running agent tasks. But it also produces a lot of reasoning tokens. One test used more than 16,000 output tokens to generate a single SVG and cost roughly 25 cents.
That led to another question: why is DeepSeek still so much cheaper?
The discussion pointed to sparse architectures, compressed attention, custom inference kernels, aggressive caching, and high cache-hit rates. Others suspected strategic pricing or subsidies.
Kimi is pursuing similar optimizations. Its API offers much cheaper cached input and claims cache-hit rates above 90% for coding workloads. But K3 also expects its previous reasoning history to be preserved, and Moonshot warns that performance can become unstable in incompatible agent harnesses.
Open models may be catching up with the proprietary frontier. The harder competition is now efficiency: how much hardware, context, reasoning, and money it takes to complete the same task.
💬 HN Discussion
Developers were impressed by how close K3 appears to the leading proprietary models, but questioned its benchmark claims, high token usage, and the practicality of calling such a large model “open” when most people cannot run it themselves.
💬 Reddit Discussion
The LocalLLaMA thread focused on DeepSeek’s unusually low cost. Commenters debated whether it comes from better architecture and infrastructure, aggressive caching, subsidized pricing, or some combination of all three.
📖 Story 2: It's OK to abandon your side-project - Robb Owen
robbowen.digital · Read
We hear about side projects that become startups or popular open-source tools. Most do not.
Robb Owen built an app to practise Latvian noun endings. But implementing every grammar rule taught him the system so well that the finished app became boring and unnecessary.
As a product, it failed. As a learning project, it succeeded.
That is a useful way to view abandoned work. A project may have completed its real purpose even if it was never launched or maintained.
The missing step is often a retrospective:
What did I learn? Why did I stop? What would I do differently?
A short postmortem turns an abandoned repository into reusable experience rather than unfinished guilt.
💬 HN Discussion
Commenters liked the idea of writing an “end-of-life wrap-up” for old projects.
Several also argued that completion does not have to mean launching or attracting users. A project can be complete once it answers a question or teaches the skill it was created to explore.
Side projects do not always owe the world a product. Sometimes they only owe you a lesson.
📖 Story 3: Note To My Younger Self
yewjin.substack.com · Read
Yew Jin Lim wrote a nice reflection on what he would tell his younger self after two decades in Big Tech.
The core advice is simple: love the work, not the ladder. Find what fascinates you, keep asking when something does not make sense, and build real relationships with the people around you.
A lot of career growth comes from habits that sound boring but compound over time: listen properly, explain with precision, document your impact, and ask for feedback before the official review cycle.
The feedback part is especially practical: ask people to rate your work from 1 to 10. If it is not a 10, ask what is missing.
The principles can collide. Sometimes you optimise for learning, sometimes for impact, sometimes for the team. The real skill is knowing which one matters most in the moment.
💬 Community Moment
Super Dario: One More Week
https://superdario.pawb.de/🛠️ Projects Worth Checking Out
- GitHub - bradAGI/awesome-cli-coding-agents: Curated directory of terminal-native AI coding agents
- GitHub - oguzhaninan/Stacer: Linux System Optimizer and Monitoring
- GitHub - simonw/llm-prices: Prices of various LLMs
- GitHub - markusfisch/PieLauncher: Android home screen launcher that uses a dynamic pie menu instead of tables of icons.
- GitHub - apache/echarts: Apache ECharts is a powerful, interactive charting and data visualization library for browser
