Savin Goyal

Wall Street Journal article (https://lnkd.in/g_PksUsC) recently highlighted how AI is reshaping traditional org structures, forcing companies to rethink roles, responsibilities, and leadership models. One of the areas most impacted? Engineering management. That’s why we’re excited to back Actual AI, which just raised a $3.2M seed round led by AlleyCorp with participation from Irregular Expressions, G2C Venture Partners, and others. Actual AI (https://www.actual.ai/) is building the first agentic engineering management platform, giving managers “superpowers” to handle the explosion of AI-generated code. Their platform automates progress reporting, architectural reviews, and workflow oversight—freeing managers to focus on strategy, mentorship, and team culture. In a world where AI is disrupting not just how we code, but how we organize, Actual AI is pioneering the future of engineering leadership. Thrilled to partner with John Kennedy, Ethan Byrd and the Actual AI team as they redefine what it means to manage and build in the AI era. https://lnkd.in/gD-sVK3Y #AI #EngineeringManagement #VentureCapital #FutureOfWork

22

One idea from Lenny’s latest podcast with Anish Acharya, General Partner at Andreessen Horowitz (a16z), got me thinking about how quickly the gap between ideation and execution is narrowing. Acharya describes the AI-native company as a series of increasingly automated loops. 𝘈 𝘤𝘶𝘴𝘵𝘰𝘮𝘦𝘳 𝘳𝘦𝘱𝘰𝘳𝘵𝘴 𝘢 𝘣𝘶𝘨. 𝘈𝘯 𝘢𝘨𝘦𝘯𝘵 𝘳𝘦𝘱𝘳𝘰𝘥𝘶𝘤𝘦𝘴 𝘪𝘵, 𝘱𝘳𝘰𝘱𝘰𝘴𝘦𝘴 𝘢 𝘧𝘪𝘹, 𝘵𝘦𝘴𝘵𝘴 𝘪𝘵, 𝘴𝘩𝘪𝘱𝘴 𝘪𝘵, 𝘢𝘯𝘥 𝘯𝘰𝘵𝘪𝘧𝘪𝘦𝘴 𝘵𝘩𝘦 𝘤𝘶𝘴𝘵𝘰𝘮𝘦𝘳. If AI makes that execution loop dramatically faster, the bottleneck moves upstream. What should we build next? Which problems actually matter? When should we rethink the workflow entirely? Acharya describes agents as quickly climbing toward a local maximum, while human intuition helps identify the next hill. I think that has significant implications for product and technology organizations. As execution gets cheaper, judgment becomes more valuable. Product leaders spend more time defining the right problems and outcomes. Engineering leaders focus more on designing systems that can safely execute at speed. Planning itself changes when execution is no longer the scarce resource. We have spent years asking how AI can help us work faster. The more interesting question may be: What changes when execution is no longer the constraint? Highly recommend listening to the full conversation. Lots of interesting perspectives on how AI may reshape the way companies operate. (Source: https://lnkd.in/gdY7FYaK) #AgenticAI #AILeadership #ProductManagement #EnterpriseAI #FutureOfWork

1

1 Comment

Two patterns i've observed in generational founders: 1. Perpetual / persistent dissatisfaction with the status quo. 2. Ambition / TAM that expands as the company grows.

47

2 Comments

At Google, Prashant Jalan built profilers to understand where AI compute went. His startup Guickly applies that measurement problem to enterprise AI spending. Guickly has launched with $4.2M in seed funding led by Engineering Capital, with Converge VC, Neon Fund and angel investors participating. Jalan's launch announcement also names Christian Szegedy as a backer. The platform maps AI tools, agents, users and costs across an organization, including unapproved usage. It attributes spending by team and employee, sets budgets, and flags unused licenses and overpriced models. Its privacy approach could help with deployment: Guickly says sensitive prompts, source code and other confidential content stay on premises. Jalan spent eight years in applied AI at Google, including work on Google Maps and TPU performance. That experience brings a useful discipline to AI budgets: measure usage before deciding what to optimize. Quick facts👇 ● founders: Prashant Jalan ● HQ: San Francisco, California ● Investors: Engineering Capital; Converge VC; Neon Fund; Christian Szegedy Can a shared view of AI costs help finance and engineering agree on which deployments are earning their keep?

14

Hyperscalers work extremely well for Saas companies. Their demand is very spiky, relatively light on memory, and easy to multiplex across many users. More importantly, it hides away the burden of capacity planning for developers. Inference, on the other hand, is quite different. Inference workloads are persistent and memory-heavy. Model weights need to stay resident in memory, and latency matters far more than raw throughput. Once a GPU is serving inference, it cannot easily be shared or swapped out without performance penalties. As a result, the burden shifts back to the developer. Teams are forced to over-provision GPUs, manage cold starts, and pay for idle capacity just to be ready when traffic arrives. This is not because cloud infrastructure failed, but because the inference does not match the workload profile it was built for. This is why inference creates room for alternative execution models. We are convinced that decentralised compute is uniquely suited to service this growing inference need. And that's why we’ve created the Hellas Network so that developers can have their needs met with the smallest latency possible.

6

Running LLMs locally feels straightforward until you try to use them as system components. After processing thousands of ASR transcripts, the same issues kept showing up: • overconditioned prompts • oversized models doing small jobs • brittle, single-pass pipelines This article covers what worked instead. It looks at multi-pass pipelines with validation, how to right-size models without losing accuracy, and where local execution actually makes sense. Full article in the comments 👇

13

1 Comment

This is one of those things that just makes sense. The Linux Foundation (and by extension the PyTorch Foundation too) is a great home to nurture open-source projects.

4

20% confidence prediction I've had for a while: labs will stop offering API access to frontier models altogether to force people to use their proprietary products. Builders that aren't competing will be fine (OSS, previous-gen models, etc), but it'd very tough to compete directly in the niches they care about.

49

6 Comments