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: Is It Still Too Early?
Would you believe there are still questions on where (or whether) AI genuinely adds value in building software, and in general business environments? Sierra co-founder Bret Taylor suggests today’s AI environment may be a bubble, where all the big investments and excitement could lead to both huge successes and painful losses. Except it’s not a bubble, argue our friends in r/ArtificialIntelligence, because look at all the enterprise revenue. Except that, as veteran engineer Mike Judge argues, if every dev were already 10x more productive with AI, we should already be inundated by low-grade shovelware. (Even though new software releases have largely remained flat.) Except that, as veteran engineer and founder Justin Searls argues, the real gains started in Spring 2025 when AI coding tools “finally became worth a damn,” and anecdotally, there still seems to be a significant lack of adoption of AI coding tools among engineering teams anyway. In practice, only 36% of Anthropic users utilize the company’s Claude model for coding. Only 4% of ChatGPT users utilize that model for coding, with the majority using it for writing, tutoring, and how-to questions. Let’s consider that ChatGPT is the dominant LLM by Web traffic, with 10x more Web visits than other models. Yet only 4% use it for coding. The real issue may be that we’re still just in early innings. As AI adoption increases, the speed of shipping software will also likely increase. (Also, it means there are probably still many opportunities to build great tools to make more developers want to adopt AI.)
Next Up: The rest of the week’s top AI news for busy devs.
💻 Development
MCP Registry
MCP has emerged as the standard for orchestrating AI agents with software tools. Here’s a running list of public MCP servers.
Vibe Coding Is Turning Us Into Babysitters
Some senior engineers suggest that vibe coding has fundamentally changed their job into proofreaders of low-grade AI-generated code.
Article: The Rise of Subagents
DeepMind’s Philipp Schmid discusses the use of “subagents,” highly specialized agents operated by an orchestrator.
[GH]: AgentScope - Using Agents to Build LLM Apps
This project offers modular, customizable building for multi-agent workflows.
🤔 Interesting AI Projects, Research, and Updates
What’s the Deal with Non-Determinism in LLMs?
Horace He explores how floating-point numbers and GPU kernels are a fundamental part of why LLMs don’t give the same response twice.
OpenAI + MCP Integration = Your Emails Could Be Leaked
Using OpenAI’s MCP integration? Have we got some bad news for you.
Research: RL Potentially the Path to Better Reasoning
This survey notes how reinforced learning may indeed be the path to turning LLMs into large reasoning models.
PyTorch + vLLM + META Boost Inference with P/D Disaggregation
This collaborative project presents disaggregating first and subsequent tokens as a way to boost throughput and cut latency.
[GH] RustGPT: Rust-Based LLM
This transformer-based LLM was coded entirely in Rust.
Gemini 2.5 DeepThink and OpenAI Both Win ICPC Gold
Gemini 2.5 DeepThink and OpenAI’s “ensemble of general-purpose reasoning models” win at the International Collegiate Programming Contest.
💼 Hiring and Community
Startups Hiring This Week:
- AI/ML Dev → iCodex
- Computer Vision & Robotics Software Engineer → Skyways
Mid-Markets Hiring This Week:
- Product Manager, AI & Data → ELLKAY
- Research Engineer, Robot Character → 1X Technologies
Enterprises Hiring This Week:
- Data Center Hardware Apps Engineer → NVDA
- ML Engineer → LDOS
More High-Profile Departures: AAPL’s Walker and OpenAI’s Yao
As the war for AI talent continues, AAPL’s head of AI and search prepares to walk while Tencent poaches a top OpenAI researcher.
💡 Spotlight: Newly Launched AI Startups
- Conduct: $15M from Creandum for AI-driven ERP modernization
- MetalBear: $12M from TLV Partners for AI-powered testing in the Cloud
- Sphinx: $9M from Lightspeed for AI-powered data science analysis
- Fabrix Security: $8M from Norwest for AI-powered IAM
- Geordie: $6.5M from Ten Eleven Ventures for enterprise agent security
- Geniez: $6M from StageOne Ventures for running AI on mainframes
- Scalekit: $5.5M from Together Fund for agentic authentication
- Hiverge: $5M from Flying Fish Ventures for AI algo generation
- Architect AI: $4.7M from Project A for website agents
- Eve Security: $3M from LiveOak Ventures for agentic security
🚀 New Model Launches
- GPT-5-Codex: Specialized for coding across CLI / IDE / Web / GH
- VaultGemma 1B: DeepMind’s differential privacy model
- Qwen-3-Next: BABA’s new model focused on ultra-long context
- ERNIE-4.5-21B-A3B-Thinking: BIDU’s newest MoE reasoning model
🏭 Industry News & Trends
Massive OpenAI + ORCL Deal Makes Ellison Richer
ORCL is experiencing new life via blockbuster compute deals. The OpenAI deal made CTO Larry Ellison the richest man in the world.
OpenAI Pursuing E-Commerce with ChatGPT Orders?
Would you give your credit card info to an LLM?
MSFT Pivots on OpenAI Ownership, New Anthropic Deals
MSFT will own ~1/3 of OpenAI but change its 20% revshare down to 8%…and explores adding Claude to its Copilot suite.
CoreWeave Launches Venture Fund
The massive GPU cloud compute provider is starting its own VC fund.
Spotify Users Band Together to Sell Their Data, to Spotify’s Dismay
10,000 listeners band together to sell their account insights for AI purposes, a move that didn’t exactly thrill Spotify.
Workday Acquires Sana Labs for $1.1B
HR giant Workday announces the acquisition of the AI learning and development platform as part of a larger push into agentic functionality.
⚖️ Copyright, IP, Licensing, and Regulation
This Week’s Lawsuits:
- Anthropic agrees to $1.5B copyright settlement with three authors
- Penske Media (Rolling Stone, Variety) sues GOOG over AI overviews
- Encyclopedia Britannica & Merriam Webster sue Perplexity over plagiarism
RSS Co-Creator’s New Open Standard for Licensing Content for AI
Eckart Walther unveils Really Simple Licensing (RSL), that uses an OAuth 2.0-inspired protocol to access content for licensing.
CA Bill to Regulate “Companion Chatbots” Heads to Newsom’s Desk
A response to suicides and other tragic incidents, the bill aims to prevent chatbots from engaging minors around self-harm or explicit content.
Oh Good, LLMs Are 2x More Likely to Spread Misinformation Now
Within the last year, most major models have apparently gotten worse when it comes to spreading false info.
Albania Appoints AI Anti-Corruption Minister
What’s the worst that could happen?