The Humans in the Loop: When You Get the Money, You Get the Power
This Week in AI for Devs: AI Resources, But at What Cost?
Welcome to The Humans in the Loop, your executive summary of AI news for devs. I’m Andrew, the [human] author, and I’d love your feedback on this newsletter. Please feel free to email me at andrew@heavybit.com with your thoughts. Thanks!
Our top story: Compute + Energy + Silicon Don’t Come Cheap
OpenAI is in the headlines due to some blockbuster deals and reportedly even greater ambitions. The vendor has penned a $100B deal with NVDA to deploy data centers to deliver 10GW of AI infra (that’s more than the power capacity of Portugal). The vendor, fresh from signing a $300B cloud deal with ORCL, is signing more deals worldwide for resources, like a recent semiconductor deal with Samsung and SK Hynix in South Korea to deliver as many as 900,000 chips per month as part of the $500B Stargate project. And the company’s CEO recently suggested the company’s goal is to build out 250GW of capacity by 2033 (for the record, the entire US uses 1,200 GW annually). All told, some estimates put the total investment into planned data centers at a whopping $850B, which will mean America’s AI program will likely be a significant part of its economy, while recent energy challenges, like the lack of natural gas infrastructure in Texas, may require the energy industry to rethink how it produces and transports it. AI is not only changing the face of technology, it’s potentially changing the face of America.
Coming Up: The rest of the week’s top AI news for busy devs.
💻 Development
Willison on Agentic Loops
Veteran researcher Simon Willison discusses how to design agentic processes with testing, security, and potential errors in mind.
Guide: Building a Triage Agent for AWS GuardDuty
Sec engineer Dakota Riley covers how to build an AI agent for sec automation using AWS’ threat detection tools.
Course: Building Data Agents
This course covers how to build a data-analyzing agent, measure its performance, and iteratively improve it.
Shrinking SWE-Bench 50x to Run in the Cloud
This article discusses how it shrank the popular AI software engineering benchmark to run well even on Cloud machines.
🤔 Interesting AI Projects, Research, and Updates
Opinion: Is There Actually Not an AI Bubble?
Anthropic’s Julian Schrittwieser suggests that AI is actually progressing exponentially, and we may see full 8-hour AI workdays by 2026.
AI Agents: Private Data + Untrusted Content + External Comms = Data Theft
Security author Bruce Schneier notes that exciting new agentic features, like Notion’s AI agent, may expose you to data theft risk.
Inside the Architecture of NVDA GPUs
Veteran AI engineer Aleksa Gordić kicks off a deep-dive into the architecture of NVDA GPUs for AI use cases.
Video: Sutton Says LLMs Can’t Learn Continually
Professor Richard Sutton, a founder of modern RL, suggests LLMs are a dead end in terms of genuinely learning.
💼 Hiring and Community
Startups Hiring This Week:
- Sr. Platform Engineer → Allora Labs
- VP Data Science → hackajob
Mid-Markets Hiring This Week:
- Data Scientist → Clay
- AI/ML Engineer → Customized Energy Solutions
Enterprises Hiring This Week:
- Data Review Scientist → MKKGY
- Sr. ML Engineer → RVN
💡 Spotlight: Newly Launched AI Startups
- Upscale AI: $100M from Mayfield for “AI networking”
- Obot AI: $35M from Nexus Venture Partners for enterprise MCP
- Paid: $21M from Lightspeed for AI agent builders
- Rocket: $15M from Salesforce Ventures for “vibe solutioning”
- Datawizz: $12.5M from Human Capital for improved LLM routing
- Yasp: $5M from Capnamic for agentic AI compilation
🚀 New Model Launches
- Claude Sonnet 4.5: Once again the benchmark leader for computer use
- DeepSeek v3.2-exp: Cuts API costs in half
- Qwen3-Max: BABA’s first trillion-parameter model
- Intellect-2 32B: First 32B model trained via distributed RL
- Sora 2: OpenAI’s video model, powering an app to compete with TikTok
🏭 Industry News & Trends
Thinking Machines’ First Product: Tinker
The secretive startup founded by former OpenAI CTO Mira Murati launches an automation tool for model creation.
META Acquires RISC-V Chipmaker Rivos
Perhaps the next step in META’s journey to build in-house model-training silicon.
BABA’s New Frontier Lab Launches Seven Models at Once
Qwen3-Max isn’t the only new launch. The e-commerce giant is investing $50B in new model development.
Yep, OpenAI Is Getting Into Selling Stuff Online
OpenAI is starting with Etsy transactions via ChatGPT’s new Instant Checkout feature, with Shopify support to come.
⚖️ Copyright, IP, Licensing, Regulation, and Safety
Newsom Signs SB 53 for AI Regulation
The “milder” version of the failed SB 1047 bill adds whistleblower protections and a framework for a public compute cluster.
ChatGPT Adds Parental Controls
New controls include quiet hours, toggle on/off for voice and images, and opt-outs for model training.