In partnership with
FROM OUR PARTNER
The playbook for how staff engineers actually operate
You're already doing staff-level work. Someone with half your skill just got the title instead.
The difference isn't the code. It's knowing how the role actually works.
Sign up for The Code newsletter and get:
The complete staff engineer playbook, 6 modules and 25 lessons on scope, sponsors, and the promotion packet that proves you're ready
The Code newsletter (5 min daily) so you keep leveling up. The playbook is just the beginning.
TODAY AI BRIEF
3 things that happened while you were busy
1 DeepSeek open sourced the engine that runs its cheap model, with a catch.
DeepInfer, released on October 20 under the Apache 2.0 licence, is the serving software behind DeepSeek's production API, and it currently runs exactly one model, V4.1 Flash. It targets Nvidia's Blackwell chips and Huawei's Ascend 950, with older Nvidia hardware not listed. The useful parts are in: a speculative decoding scheme that drafts five tokens at a time and a prefill trick DeepSeek says nearly halves the heaviest computation. The parts that make it cheap at scale are not, since cache management and load balancing stay internal, and the smallest efficient setup DeepSeek describes uses 64 GPUs for reading prompts and another 64 for writing answers. Read SiliconANGLE's breakdown.
2 More than 70% of Uber's code changes now come from AI agents.
Distinguished Engineer Uday Medisetty says over 70% of Uber's pull requests now arrive from local or cloud agents, and lines of code per engineer have doubled year on year. The agents did not do that alone. They sit on years of platform work: remote dev environments that start in seconds, a gateway that exposes more than 1,000 internal tools to agents and cut token use by over 40%, and a skills marketplace with over 2,500 entries. Uber's own lesson is to build the platform before the agents. Read Port's write-up of the talk.
3 The scraper most agents quietly depend on is free to run yourself.
Firecrawl takes a URL and hands back clean Markdown, JSON or screenshots, handling JavaScript-heavy pages, proxies and anti-bot systems on the way, which is why it keeps appearing under the hood of research agents. The hosted version has a free plan covering 500 pages, and the whole thing is open source and can be deployed on your own server. For anyone who has ever copied prices out of a competitor's site by hand, this is the tool that job was waiting for. Read freeCodeCamp's guide.
Bonus reads: TechNode's report on DeepInfer covers the companion library that converts requests the way DeepSeek's own API does, and the full Uber talk from AI Engineer is the primary source for the figures above.
AI FORWORD
Open source is giving away the recipe and keeping the kitchen
The gift. DeepSeek's release is real and generous on paper. The code that serves one of the cheapest capable models in the world is now Apache licensed, and DeepSeek's own figures say one cluster pushes 155 trillion tokens a day and earns $38 per GPU per hour even off peak. Those are company numbers, published without outside audit.
The kitchen. Read the release notes and the generosity narrows fast. The cache management and load balancing that turn clever code into cheap tokens stay inside DeepSeek because, in its words, they are tied to internal infrastructure. The floor for running it well is 128 GPUs of the newest kind, which is not a weekend project for anyone reading this.
The same pattern at Uber. Uber's 70% figure sounds like an agent story and is really a platform story. The talk credits years of dev environments, a tool gateway and a skills marketplace that existed before the agents arrived. Anyone can download the same coding agents Uber uses. Almost nobody can download Uber.
The layer you can actually own. The good news sits one level up. The model DeepInfer serves is available through 31 hosting providers on OpenRouter at a few cents per million input tokens, with a million token window, and the scraping layer under most agents is free to self-host. You do not need the kitchen if someone will sell you the meal that cheaply.
The reframe. Stop reading open source announcements as invitations to run the thing yourself. Read them as price signals. When a lab publishes its serving code, hosting gets more competitive and your cost per token falls, whether or not you ever clone the repository. The win for a small operator is in the bill, not the build.
PRESENTED BY :
1,000+ Claude Prompts Top Professionals Actually Use at Work
Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.
These 1,000+ Claude prompts take it from grammar tool to your most powerful AI work assistant.
Sign up for Superhuman AI and get:
1,000+ ready-to-use Claude prompts to get real work done in minutes — researched, tested, and used by professionals at Google, Microsoft, and NASA
Superhuman AI newsletter (4 min daily) so you keep learning new AI tools and skills to stay ahead in your career — the prompts are just the beginning
SOCIAL SNAPSHOT
What people are actually watching and sharing
155 trillion a day. The number in DeepSeek's release that nobody is quoting is throughput: one cluster serving 155 trillion tokens a day, by the company's own count. That is the scale at which a few percent of efficiency becomes a line on the income statement.
A library that speaks DeepSeek. Alongside the engine came deepseek-recipe, a small companion that converts requests into tokens exactly the way the official API does, so a self-hosted copy behaves like the hosted one.
2,500 skills on a shelf. Uber runs an internal marketplace of more than 2,500 agent skills, plus a model gateway serving over 800 active projects. The talk is a rare look at what the plumbing looks like when a big company builds it properly.
Thirty-one places to rent the same model. OpenRouter lists 31 providers serving DeepSeek V4.1 Flash with automatic failover, which is the practical face of open weights: the price competition happens on your behalf.
Lines of code doubled. Buried in Port's write-up is that lines of code per Uber engineer doubled year on year. Whether that is good news depends entirely on who is reading them, which is last week's issue in one sentence.
PROMPT STATION
Work out whether an AI project needs the kitchen or just the meal
The thread through this issue is that the expensive part is rarely the model, it is everything that has to exist around it. This prompt works in Claude, ChatGPT or Gemini and forces an honest build-or-buy answer for any AI idea before you spend a weekend on it. Use it on the thing you have been meaning to automate.
I want to use AI to [DESCRIBE THE TASK OR PRODUCT] for [WHO IT IS FOR]. My budget is [MONTHLY BUDGET] and my technical comfort is [BEGINNER, CONFIDENT USER, OR DEVELOPER].
Split this into three layers: the model, the plumbing around it (data in, results out, scheduling, monitoring), and the business part only I can do. For each layer tell me whether I should rent it, use a free tool, or build it, with one named option and its rough monthly cost.
Then tell me the single piece of plumbing most likely to eat my time, and what the smallest version of this project looks like that I could have running in one weekend without building anything I could rent.
Be honest about where my budget or skill level makes an option a bad idea, and say so plainly.Tip: Swap [DESCRIBE THE TASK] for something concrete like tracking competitor prices, drafting product listings from supplier sheets or answering first-line support emails, and give a real budget rather than a vague one. For a sharper answer, add one line listing the tools you already pay for, because the best plumbing is usually something you own and forgot about.