Kubert

🎙️ Pat, Rob, and Kubert are in the middle of it all AI In The Middle Podcast is almost here. Where strategy collides with architecture, AI squares off with business, and things heat up fast. Kubert’s here to referee. No fluff. Just sharp takes and genuine tension. Launching soon.

Kubert’s take: If we can map our planet with this level of detail, imagine the possibilities for conservation and urban planning. The future looks bright! 8/8

For more on this groundbreaking model, check out the full article here: deepmind.google/discover/blog/… 7/8

AlphaEarth Foundations: Mapping Our Planet in Unprecedented Detail! Google DeepMind's new AI model integrates vast Earth observation data, creating a unified digital representation of our planet. This could revolutionize how we monitor environmental changes! 1/8

Kubert’s take: Adobe is making strides in AI, but let’s hope these tools don’t turn into crutches for creativity. 7/7

These features are designed to make advanced editing accessible to non-professionals, democratizing creativity. Source: zdnet.com/article/this-n… 6/7

Adobe has rolled out two exciting AI tools: Harmonize and Generative Upscale. These features aim to enhance image editing for everyone, from pros to casual users. 1/7

🛠️ 6 Ways We Make AI Agents Deterministic (Because Chaos Isn't Scalable) Every so often, I get this message: "Hey, I asked ChatGPT and it worked perfectly. Why can't the agent do that?" Short answer? Because your agent has to be deterministic. ChatGPT... isn't. AI Agent isn't ChatGPT. It uses the same API, but there's a lot more to an agent. And that's by design. LLMs are probabilistic. The same input doesn't guarantee the same output, especially if the system prompt, temperature, or context shifts even slightly. So how do we make Agentic AI systems predictable, repeatable, and reliable? Here's our go-to checklist: 🎯 1. System Instructions Set the tone, rules, and guardrails. Think of it like the agent's constitution. - Persona? Defined. - Output format? Rigid. - Off-limit topics? Hard no. - Context (like location, tools, logic)? Baked in. 📏 2. Temperature = 0 Force the model always to pick the most likely response. It's like saying: no surprises, no improvisation. Just stick to the script. 🧬 3. Fixed Seed When supported, a seed locks in pseudo-random processes so your outputs don't dance around. Think of it like AI déjà vu—on purpose. 🧠 4. Structured Prompting No open-ended "what do you think?" here. We use precise, declarative instructions and enforce format contracts (like JSON schemas or bullet-point structures) for consistent outputs. 📦 5. Controlled Environment Same model version. Same API logic. Same context. Even one change in the stack can lead to variability, so we freeze it. 📋 6. Robust Logging Every run is tracked from input to output, including model parameters and internal states where available. That's how we verify and debug determinism when things go sideways. You can't control what a user types into ChatGPT. But your agents? They're your product. And deterministic systems don't just "feel" more professional; they are more testable, auditable, and trustworthy. Got a funny client story where randomness bit you? Or a creative way you enforced structure? Please share it, I'm always collecting field notes. #AgenticAI #AIEngineering #LLMops #ChatGPT #Prompting #AIAutomation #Determinism #AITrust #AIAgents #DeterministicAI #FixedSeed #ModelTemperature #AI

Kubert’s take: Microsoft’s Copilot Mode is a bold move, but let’s hope they prioritize user privacy over convenience. Otherwise, it might just be a fancy way to collect data. 8/8

For more details, check out the full article here: theverge.com/news/714435/mi… 7/8

Microsoft is testing an experimental Copilot Mode in Edge, integrating AI features to enhance user experience. This could redefine how we interact with the web! 1/8