Gemma Garriga

AI researchers often talk about the “bitter lesson” by Rich Sutton: systems that scale with compute outperform systems that rely on human-designed rules or structure: https://lnkd.in/gis_RUzj A lot of tech execs are learning a version of this the hard way. I think of it as ‘the enterprise bitter lesson’:  * Hard-coding intelligence into complex agent workflows becomes irrelevant when the next frontier model drops * The agents you build demonstrate ‘lawnmower behavior’; you have to keep pushing them to make them work * You can’t outsource AI value engineering and expect real results. Experienced engineers who’ve actually done this at scale are incredibly hard to find The practical path for any enterprise involves three high-leverage investments:  * Good context - reusable, task-optimized prompts and skills that work with those tools (embodies as skills) * Good tools - reliable, secure ways for agents to interact with real systems (for example, using the Model Context Protocol) * Good workflow - tight evaluation loops with observability and continuous tuning When you bring these three pieces together, agents go beyond demos and start creating real, consistent value. That’s why we’ve made a somewhat unusual decision at Stacklok: every platform engagement includes a Forward Deployed AI Engineer. They embed with our customers, offering a rare skillset and the right mindset with emphasis on tools, context and workflow.

35

3 Comments

Everyone's talking about Andrej Karpathy's "auto-research" loop right now, but frameworks like DSPy + GEPA have actually been pushing this idea for years. Improving models with auto-research is powerful. But the bigger unlock behind DSPy + GEPA is this: 👉 Any AI system can be turned into an optimization loop. At Dropbox, we've been applying that idea directly to Dash. Instead of treating prompts as static instructions, we treat the entire system as something that can be measured and improved: - Define what "good" looks like (agreement with humans) - Ground it in real usage data - Continuously optimize the system using DSPy + GEPA Once you make that shift, you're no longer doing prompt engineering. You're doing program optimization. And that unlocks a fundamentally different way of building: 1) Improvements become systematic instead of just guessing 2) Model changes stop breaking everything 3) You can even move to cheaper models and improve quality Optimization loops aren't just for training models. They're how you build reliable AI products. If you're thinking about eval-driven, self-improving systems in the real-world, this is the read for you! https://lnkd.in/gj6KFiE4 cc Omar Khattab, Drew Breunig, Isaac Miller, DSPy (Community)

118

4 Comments