The Humans in the Loop: The World According to GOOG
This Week in AI for Devs: GOOG Cloud Next Redux
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: Google Cloud Next Show-and-Tell
The annual Google Cloud Next conference has come and gone, leaving behind a host of new AI updates. On the front end, GOOG has unveiled Firebase Studio, which integrates the publisher’s shared cloud dev environment Project IDX, the open-source AI app development framework Genkit, and the publisher’s Gemini model all in one place. And with Agent2Agent, its own multi-agent orchestration framework, GOOG is making its entry into the world of model context protocol (MCP), which is increasingly seeming like a new community standard for managing AI agents. The publisher claims its offering is “complementary” to Anthropic’s offering, rather than competitive. The event also saw the unveiling of GOOG’s next AI hardware, the seventh-generation tensor processing unit (TPU) Ironwood, which the publisher claims is “the first designed specifically for inference.” It also rounded out its event with multimodal (video, image, speech) updates to its unified AI development platform Vertex. While GOOG has been criticized in the past for not being on the AI cutting edge, it’s clearly doing everything in its power to keep up with the latest trends in the space.
Next Up: The bottom line in AI news for busy devs.
💻 Development
Codex CLI: OpenAI’s Answer to Claude Code?
The publisher’s new offering is a command-line coding agent possibly intended to go toe-to-toe with Claude Code.
MSFT: Learn to Build Agents and Make Them Debug
MSFT’s GH repo has 10 lessons to start building agents, and its researchers unveil debug-gym, where you can train agents to debug code.
[GH] any-agent: Python-Based Unified Agent Library
Mozilla’s new library offers a unified interface for agents built with GOOG ADK, LangChain, LlamaIndex, OpenAI Agents SDK, or smolagents.
Critical Thinking In The Age Of AI-Generated Code
Veteran dev Stelios Manioudakis: As low-quality AI-generated code increases, devs need to cultivate the skills of detecting and testing AI-created code.
Thinking About MCP? DIY
Several new GH repos offer the ability to generate-your-own for OpenAPI, stand up a simple proxy server, or create a server for messages/contacts.
🤔 Interesting AI Projects, Research, and Updates
Research: Better MoE Via Better Pruning, Test-Time Optimization
As teams use mixture-of-experts to efficiently run jobs across multiple trained models, researchers boost performance by selectively pruning parameter footprints and utilizing test-time optimization.
Research: Improving LLM Continual Learning With Subspaces
Continual learning adds new training/tuning to models but causes catastrophic forgetting. Constraining updates to parametric subspaces can help.
Research: Good at Managing People? Good at Managing AI Agents
Study suggests that using GPT-4o to simulate successful AI agent orchestration reveals a close mapping to successfully managing human teams.
💼 Hiring and Community
Startups Hiring This Week:
- MLOps/Infrastructure Engineer → 10a Labs
- VP AI Engineering & Research Ops → Multiverse
Mid-Markets Hiring This Week:
- Senior ML Scientist → Viator
- Senior Manager, Responsible AI Solutions → Ford Foundation
Enterprises Hiring This Week:
- Data Scientist - Recommendation Algorithms → Stitch Fix
- Staff AI Platform Engineer → SentinelOne
💡 Spotlight: Newly-Launched AI Startups
- Corvic: $12M from M Ventures for enterprise AI tooling
- Pillar Security: $9M from Shield Capital for AI app protection
- Portia AI: $5.8M from General Catalyst for AI agent compliance
- Unosecur: $5M from DFF Ventures for AI-powered IAM
- Inephany Ltd: $2.2M from Amadeus Capital for model training optimization
🚀 New Model Launches
- META Llama 4 Family (Behemoth 288B, Maverick 17B, Scout 17B)
- GPT 4.1 Family (4.1, 4.1 mini, 4.1 nano)
- GPT o3 and o4-mini (newest reasoning models)
- Grok 3 API
- Together.AI’s DeepCoder 14B
- ServiceNow’s Apriel 5B
🏭 Industry: M&A, Launches, Trends
OpenAI to Acquire Windsurf (Formerly Codeium) for $3B?
Not content with releasing its own command-line offering, OpenAI may have 3 billion reasons why developers are the most important customers for AI.
DeepSeek to Open-Source Its Inference
The creator of the bombshell open-source reasoning model R1 will now be opening up its inference.
NVDA Opens Its First US Factories
In response to demand for its Blackwell chips, NVDA is opening up manufacturing stateside.
Former OpenAI Leaders Pursuing Monster Funding
Apparently, former CTO Murati is seeking $2B for Thinking Machines while former co-founder Sutskever nabs $2B for SSI at a $32B valuation.
⚖️ Copyright, IP, Licensing, and Regulation
Tariffs on Everything Except AI Tech?
In the US, tariffs seem to be coming for just about everything (except tech products from NVDA and AAPL).
Screw EU: META Is Using Your Data to Train AI Again
Despite privacy and regulatory concerns, META is once again scooping up public posts and comments to train its models.
YouTube Backs NO FAKES Act
The video publisher publicly expresses support for the proposed legislation that attempts to address deepfakes.
About the Author
Hi there. My name is Andrew, and I work at Heavybit, the leading VC for developer-first startups. As the Editorial Lead, my goal is to find the most valuable and important AI news for developers and founders. The idea is to curate and bottom-line emerging trends from our perspective in 10+ years of coaching and funding developer-facing companies. Email me and let me know what you think.
That’s all for this edition. In the meantime, please feel free to follow The Humans in the Loop on LinkedIn, BlueSky, and Twitter.