1.
Chinese AI company Zhipu claims its new is a better bug-finder than Anthropic, OpenAI
2.
Bipartisan 'Uprising' Against Flock Cameras: a Larger Fight Against Big Tech and Surveillance?
Politico notes that over 20 local jurisdictions in America "either stopped using Flock cameras or began the process of doing so in July, according to a tracker maintained by DeFlock, an activist group that has been mapping the company. It's the highest amount in a single month since they began tracking in 2021." Some local officials said the public safety promises weren't worth the cost. The cameras "didn't help us with anything.
3.
Forward Deployed Engineer: The AI's Hottest Job Paying $280K to $1M+ (Full Course with resources)
You know how there's been a lot of talk about AI jobs lately? Well, it's not just about the headlines. I was looking at job platforms and I saw something surprising - Forward Deployed Engineer jobs popping up everywhere. We're talking about companies like Clera, Salesforce, Databricks, and even startups, with openings in the US, Canada, Europe, Japan, and India. It's becoming a real career category.
OpenAI has over 20 FDE openings, and AWS just announced a $1 billion investment to build a Forward Deployed Engineering organization. The pay is insane - new-grad roles at Palantir can start at $135K-$145K base, while OpenAI offers $162K-$280K base plus equity. Some top FDE packages have already crossed $500K total compensation, and senior roles can reach seven figures.
The interesting part is that these companies are paying top dollar for people who can take a powerful AI model and make it work inside a real business. That skill is becoming scarce, and it's not about training a frontier model - it's about deploying one. I've been thinking about this a lot, and I realized that the job of a Forward Deployed Engineer is not just about AI, but about understanding the company's workflow, data, and rules. It's about owning the path from early discovery to stable production.
If you're interested in this role, I've put together a practical course to help you learn the core FDE stack in just 8 weeks. We'll cover Python, SQL, APIs, cloud, Docker, and more, and you'll get to build a real FDE-style project from scratch. I'll also provide free learning materials, prompts for customer discovery, and tips on how to build evals, get real-world experience, and create your portfolio. The goal is simple: learn the stack, build something useful, and deploy it for a real user.
4.
I Beat Substack’s AI Detector
The interesting part is that I didn't exactly "hack" the system, but rather, I found a way to work within its limitations. I'm not going to go into the nitty-gritty details just yet, but essentially, I was able to create a writing style that mimicked human patterns without triggering the detector. It's not about whether or not this is a good thing, but rather, it's a fascinating technical problem to solve.
I've seen people using these detection tools to virtue signal and "win" arguments, but the irony is that some of these folks are using AI-generated content themselves. I came across a post with a Pangram, and the author had a fully AI-generated YouTube video on their website – it was a pretty funny sight. Anyway, I'll dive deeper into how I did it and what it means for the future of AI writing in my next newsletter.
5.
$12B of US ratepayers' money wasted on a modeling mistake and PJM wants to do it again
PJM’s capacity‑auction model has been quietly miscounting the output of existing gas plants. By assuming winter air is no more efficient than summer air and ignoring the winter‑hardening upgrades many plants completed after Storm Elliott, the model under‑estimates available capacity by roughly four gigawatts.
That mis‑estimate inflates the perceived shortfall, prompting the market to buy far more capacity than needed. The extra purchases translate into about $12 billion of extra costs for the 66 million ratepayers between 2025 and 2027, even though the actual additional power needed is only a few megawatts.
Because PJM’s capacity market treats new and existing plants the same, the over‑procured capacity pays the same premium to both, effectively rewarding idle existing plants while pushing up prices for new builds. The upcoming emergency auction, which will lock in contracts through 2043 without firm demand commitments, risks repeating the same over‑spending if the modeling errors aren’t corrected.
If the model were adjusted to credit winter efficiency gains and the winterization upgrades, the analysis shows savings of $6.7 billion for 2025‑26 and $4.9 billion for 2026‑27, with only a modest reduction in procured capacity. Better modeling could spare ratepayers a large portion of the bill while still maintaining reliability.
6.
Klaviyo’s CEO on Building at $1.5B ARR With Agents: “Dark Factory,” Composer, and Why Every Single Employee Had to Hit L3 by June
Klaviyo’s internal “Dark Factory” is basically a prompt‑driven PM that writes specs, splits a problem into micro‑services, drafts the API contracts, then fires sub‑agents to build each piece. The whole thing runs over a weekend, pausing only to ask the team for clarification when a requirement is fuzzy. By June every employee—engineers, designers, sales, even interns—had to be operating at the L3 level, meaning they constantly run multiple agent sessions and validate the outputs, not just fire a single query.
That framework birthed Composer, a marketing agent that hit 95 k users in its first month and is already seeing weekly repeat usage. Its first prototype emerged from a single Dark Factory run, and credit consumption is climbing about 30 % each week.
Composer leans on two layers: a live data feed that tells it how shoppers across Klaviyo’s network are reacting, and a “coach” agent that scores each campaign idea on predicted engagement and revenue before anything ships. The result is an AI that behaves like a talented high‑school athlete with a dedicated coaching staff, letting even casual users get sophisticated results without a steep onboarding curve.
7.
How to Setup Claude Cowork? The Complete 7-Part Guide
I dug into the latest Claude Cowork setup because the usual “add a bio and you’re good to go” just isn’t enough. The real trick is building a tiny operating system around the agent: decide what it should know, when it should pull in new info, how it signals a task is finished, and where it drops state for the next run. That little scaffolding is what separates a flaky morning brief from a reliable weekly report.
The current version lives in a shared home with Chat, works on desktop, web, and mobile, and can run in Anthropic’s cloud while still reaching local files through the Claude Desktop app. It’s organized into Projects—each with its own instructions, files, memory, and scheduled tasks—plus Connectors for Gmail, Drive, Slack, etc., Skills that teach repeatable procedures, Plugins that bundle those pieces, and the ability to click and type inside permitted apps. Think of it as an agent with a workspace, tools, memory, and an autonomy dial.
A practical pattern that keeps showing up is a single project folder with sub‑folders for CONTEXT, WORK, TEMPLATES, OUTPUTS, and ARCHIVE, plus a tiny working‑log.md that you update each run. Global instructions stay static—who you are, how you want to be treated, permission boundaries—while each project holds the specific facts it needs. Keeping those layers separate lets Claude stay focused and avoids the “read everything” trap that drags performance down.
8.
Java News Roundup: Simple JSON API, GlassFish, Jakarta EE, JNoSQL, Open Liberty, LangChain4j
9.
Microsoft blames AI for delayed Exchange update, can’t say when it will arrive
Microsoft has blamed extra work created by AI bug-finders for the delayed release of a major Cumulative Update to Exchange Server Subscription Edition (SE). Redmond’s Exchange team made that admission last Thursday in a post titled “Where is Exchange SE CU1 anyway?” that reveals the software giant is “getting questions from our customers on when they can expect us to release Exchange SE Cumulative Update 1 (CU1).” “After all, in the past we mentioned that it would be released by the end of the first half of calendar year 2026, later updated to ‘second half of 2026’.
10.
Anthropic Criticized For Adding Watermarks to Text that Claude Generates - or Processes
This week Anthropic announced its Claude chatbot will watermark the text it generates, reports the blog Futurism. "It works by making subtle changes in the AI's word choices across the text it generates, which are supposed to be imperceptible to a human but, in aggregate, form a pattern that is detectable with the tool." Anthropic said it was implementing the watermark system in response to the European Union's landmark AI Act passed in 2024, which requires that AI companies mark content that's been generated or edited by their systems.