Zach Wentz

Thrilled to share my latest article: "The Mixed-Initiative Interface: Designing Control Handoffs Between Humans and AI." In a world where AI handles more decisions, the real UX frontier is the seamless transfer of control—ensuring humans can step in without losing context, state, or accountability. We're moving beyond rigid autonomy to dynamic systems where handoffs are graceful and empowering. Key insights include: - The Handoff Problem: Mixed-initiative systems allow bidirectional control shifts; the challenge is preserving state during transitions to avoid confusion. - Decision Ledger: An append-only log of actions, confidence scores, and reasoning to brief humans on AI's prior steps. - Non-Destructive Overrides: Use proposal queues for AI suggestions that can be reviewed or modified without halting workflows. - Calibrating Confidence: Dynamic thresholds based on consequence, reversibility, and confidence—err toward escalation to build trust. - Reversibility and Intervention Windows: Leverage event sourcing for undos and strategic delays to enable human oversight in high-stakes scenarios. This is essential for anyone building AI-powered products: think pricing optimization, fraud detection, or any domain blending human judgment with machine efficiency. Does this align with challenges you're facing in your work? Read the full piece here: https://lnkd.in/dX_y7YMa

5

"The issue is basically that 'negative reward for lying and stealing' looks the same as 'negative reward for getting caught lying and stealing'. ... [so] AI will wind up with the latter motivation. The reward function will miss sufficiently sneaky misaligned behavior, and so the AI will come to feel like that kind of behavior is good, and this tendency will generalize in a very bad way." https://lnkd.in/eKSXkt7y

2

The voice AI industry has fundamentally misunderstood its own product-market fit. We're obsessing over conversational sophistication while failing at basic speech synthesis. These "performance barriers" Rishav calls out aren't technical limitations; they're the predictable result of prioritizing flashy demos over robust engineering. Companies won't adopt voice agents not because the technology isn't ready, but because vendors have consistently oversold capabilities while underdelivering on reliability. It’s about time someone said it out loud: the barriers to building RELIABLE voice AI (with stats). Here it out 👇

14

1 Comment

A lot shifted in 2025. AI made some things easier and other things more complicated, and it pushed teams to focus more on resilience. The 2025 State of the Internet Year in Review breaks down what stood out most this year. https://lnkd.in/eh2eZXfA

14

Some themes I'm hearing from people having success AI coding agents are explicit context, documentation through time, and automated feedback mechanisms. Explicit context includes the practices and assumptions that had previously been shared via IC senior/junior mentorships. Explicitly articulating architectural patterns, constraints, domain boundaries, language-specific design patterns, the definition of done, refactoring signals, error handling policies, optimization opportunities, etc. What would have been incrementally, lazily, and only verbally surfaced during an apprenticeship relationship is promoted to one or more documents. Documentation over time is the history behind why code works a certain way: special cases, environment-specific behaviors, why some IPC was chosen over another. While this information is often available in commit history or cryptic comments, it's less often made first class. Artifacts like ADRs capture the evolution of code. This historical documentation provides counterbalancing evidence against what is intentionally absent from the current codebase version. Automated feedback mechanisms include style, code organization, linters, compiler warnings as errors, semgrep policies, OWASP Web Testing Guidance, fitness functions for correctness, unit tests, fuzzers, and similar types of functional and non-functional feedback mechanisms. These are guardrails that operate at agent speed and provide actionable corrective signals to ensure alignment These are also hallmarks of pre-AI high-performing engineering teams. They highlight the importance of shared understanding, history, and actionable feedback. In my experience these practices help most teams to be more effective, independent of how they are constituted. 

17

5 Comments

rqlite -- and the way I develop it -- is fundamentally changing. LLMs are in the loop -- from ideation to Pull Request. https://lnkd.in/ejwXGddR

27

1 Comment

If you're keen on real AI use cases, this is actually quite powerful and I encourage you to read further (and help us build: https://lnkd.in/gsYnAfKK) At Hex, we've found our Threads product (https://lnkd.in/gJKCurJB) is the best way to summarize customer support load, because we're not limited to what's in the customer support portal - we can ask in terms of active users, customer size, feature adoption, and still get great answers. This is based on a strategy of keeping your data warehouse as a source of truth, and all related data work in one place - Hex - so that AI understands concepts in your business and makes smart connections between things. Arguably, ALL data analysis for a particular vertical is eventually best achieved in a tool like Hex, connected to curated data in your data warehouse. Wait - can't I do this directly in the customer support tool? Can't I connect an MCP server and get a functionality like this in one of my other AI enabled products? Maybe, but you have 100+ SaaS apps, so play this scenario out: - How many places are you going to sync your core data sets? - How many cross product integrations are going to exist, and how many of them are you going to enable? (100 choose 2 is ~5000) - In how many of these products will all users have seats? - In how many of these products will these seats have permissions to use AI features? (Our support tool just told me I don't have permission to ask the agent questions, probably because I don't have the right seat type, which means more $$) My bet is a on a world with ONE place for all analysis.

33

2 Comments

Gemini 3 is here. It’s our most intelligent model – and it’s available now in Google Search, starting with AI Mode. Gemini 3 brings incredible reasoning power to Search, making it easier to ask anything on your mind. It also unlocks incredibly rich new generative UI capabilities, so AI Mode can now dynamically create interactive tools and simulations on the fly to help you understand complex topics. Take for example asking AI Mode to help solve a Rubik's Cube. With Gemini 3's new gen UI capabilities, it created an interactive visualization offering strategies to master the puzzle. Proud of the phenomenal collaboration across Google Research, Google DeepMind and Google Search on this launch Dive deeper into the core principles that enabled our implementation of generative UI in our recently published Google Research paper: “Generative UI: LLMs are Effective UI Generators.” https://lnkd.in/gJmJR-EH  It’s also an exciting week because today we launched Nano Banana Pro in Search via AI Mode. Built on Gemini 3, Nano Banana Pro can generate precise, detailed and contextually rich visuals with factual accuracy. Access Gemini 3 and Nano Banana Pro in Search by selecting the “Thinking” model in the AI Mode drop-down menu. It’s available now to all Google AI Pro and Ultra subscribers in the U.S. and is coming to everyone in the U.S. soon.

357

15 Comments