Ronan Weinberg Waks

Do we get some a free DGX Spark depending on the number of companies we have worked in our career that showed up today during Jensen Huang's NVIDIA GTC keynote? 🤣

Embracing the AI Hype: Why I’m Not Being Replaced (Yet) I spent February leaning into the AI hype to see if it actually lived up to the noise. In our industry, it’s easy to get caught up in the binary debate of "AI is everything" versus "AI is just an autocomplete." This month, I decided to treat it as a literal member of my team. By building custom Copilot CLI skills and Python agents to automate "high-toil" tasks like OKR tracking, service reporting, and telemetry validation, I managed to reclaim about 8 hours of my week. It’s one thing to use LLMs for boilerplate or quick shell commands, but it’s another to watch a tool I built autonomously validate OTLP (OpenTelemetry Protocol) emissions across our cloud infrastructure. This automation ensured 100% build stability even as we navigated a massive, ongoing transition to ARM-based hardware—a move that requires precise observability to prevent performance regressions. However, this month was also a reality check on why AI isn't a replacement for an engineer. While my agents were great at the "how"—generating scripts and flagging potential issues—they couldn't replace the need for critical human oversight. For instance, when it came to correcting misleading architectural guidance on Kubernetes arch taints for ARM clusters, a script alone wouldn't have caught the nuance required to prevent scheduling errors. Nor could an AI step into a technical sync to negotiate the "Sidecar" bridge strategy that was stalling our modernization. AI can scale my execution, but it still lacks the nuance to navigate cross-team friction or turn a messy technical debt discussion into a clear path forward. The biggest takeaway? Embracing AI hasn't made my job "simpler"—it’s just allowed me to focus on the high-leverage work that actually matters. By offloading the administrative and repetitive technical overhead, I finally have the bandwidth to tackle the architectural gaps and infrastructure alignment that a script just can't solve. AI is the engine, but you still need an engineer to know where to drive it. If you're still on the fence about AI, my advice is simple: stop worrying about it replacing you and start using it to unblock you. Besides, until an LLM can figure out how to navigate a multi-team consensus meeting without hallucinating a reason to leave early, our jobs are probably safe. I’ll take the extra bandwidth for the ARM transition—I've got enough real-world bugs to deal with without the "AI-generated" ones. https://lnkd.in/gir7M_jp #AI #careergrowth #productivity #modernization #ARM

2

Our team at Meta just launched SAM 3 and SAM 3D - the next generation of Segment Anything Models! Check out these models in our demo playground: https://lnkd.in/eymcBgBa SAM 3 - Builds on SAM 2 with two major upgrades: text prompting and visual example prompting! SAM 3D - Reconstruction of objects and people into the 3rd dimension from a single 2D image! Learn more: SAM 3: https://go.meta.me/d74ec5  SAM 3D: https://go.meta.me/f019f3

54

4 Comments

The drive to improve AI models, while balancing expectations, is definitely top of mind right now. Meta is hosting LlamaCon, and it seems like the pressure's on to win back developers after some mixed reactions to Llama 4. But here's the thing: Yelp is already deploying AI-powered "voice agents" using OpenAI's Realtime API to handle calls for restaurants and service providers. These agents can answer questions, add customers to waitlists, and even forward reservation details. It seems that while the cutting edge may be debated in benchmarks, the practical applications are already here, streamlining everyday tasks. How are you seeing the gap between AI hype and real-world utility play out in your own projects or industry? #AI #LLM #Innovation #Meta

1

New Tellusim Core SDK Tutorial Explore how to implement zero-copy resource interoperability between CUDA, ROCm, and Vulkan/Direct3D12 - all in just a few lines of code. https://lnkd.in/gH3f_KSE

19

I just finished A New Set of Eyes by Joshua A. Rodriguez, and it’s one of the few AI books that doesn’t feel like hype or doom. What I took from it as a software developer: my workflow is not your workflow. And it shouldn’t be. The way you think, pause, get stuck, and recover is personal. Copying someone else’s “AI system” is a fast way to build a process you hate. And it brings back the rubber duck principle. Coders have always talked through problems out loud. AI is the duck that talks back. Useful when you treat it like a mirror and a sparring partner, not an authority. My favorite frame is the flashlight vs paint roller idea. Flashlight: aim it at the exact spot you’re stuck. Paint roller: coat everything and pretend it’s all the same. “The future isn’t less human. It’s more intentional.” That’s the real takeaway. AI isn’t here to erase craft. It’s a companion to it. Where do you draw the line today between using AI for clarity vs outsourcing the thinking? a.co/d/hRmQebi

2

Today we're excited to announce Profound's $20M Series A led by Kleiner Perkins with Khosla Ventures, Nvidia Ventures, Saga Ventures, South Park Commons, and SV Angel. We're witnessing a monumental shift in consumer behavior. Chat is becoming the search interface of the future. Traffic from ChatGPT and other LLMs is increasing every month as more users skip the blue links and expect direct answers instead. At Profound, we believe that brands will always want to show up where their customers are searching. But search is no longer just Google, it's everywhere users are asking questions. Profound has helped brands like Ramp, Indeed, and Clay ensure that they're always showing up in the answers. This new funding helps accelerate our ability to help brands not just surface their visibility in AI results, but take action on it. When anyone can write a blog post with AI, only the most unique, interesting, and useful content will rise to the top. Original research based on your company's first party data will be one of the only marketing moats available to brands in 2025 and beyond. This massive opportunity is exactly why I joined Profound. At Profound I've had the opportunity to lead the development of groundbreaking tools such as our Conversation Explorer to help marketers, brand strategists, and researchers understand and improve their visibility in AI responses. This is only the beginning. If you've seen brands like Vercel talk about getting 10% of their signups from LLMs and you're wondering "how are they doing that?", Profound can help.

130

27 Comments

Gemini Embedding 2 is now available! The team did a great job using the multimodal power of Gemini to represent all modalities in one single embedding space. A big leap forward in performance and capabilities! More details in the GDM blog post: https://lnkd.in/g_p8wBXp

17

2 Comments

Thursday, I spent most of my day at Google attending their "Build with AI" workshop, organized as part of the NY Tech Week. The workshop was intense and fast-paced, in about 5 hours we - built multiple agents using Google’s ADK dev kit - set up an MCP server and defined agent cards - deployed agents to Cloud Run using Gemini and Vertex AI - orchestrated workflows with LangGraph and A2A protocol - integrated tools, memory, and agent tracing for observability But funny enough, my biggest takeaway from the workshop wasn’t about how to build agents, but finally understanding that Google is not trying to compete with MCP through their A2A protocol. Back in April when A2A was announced, I remember thinking, “ugh, do we really need yet another AI agent protocol? Can’t we all just rally behind MCP?” But the workshop made it very clear that Google sees A2A as a complement, not a competitor. That they see MCP as more of an agent-to-server (or tools, if you want) protocol, and A2A more focused on (client)agent to (client)agent communication. MCP is here to stay. For folks who aren't as deep in the tech side of things, or just not following this space as obsessively as I am: MCP (Model Context Protocol) is basically a standard that lets AI agents understand how to talk to tools, APIs, or services, even if they were built by different companies or use different models. Just like how your web browser can access any website using shared rules like HTTP, MCP provides a common set of instructions that lets AI agents interact with different tools and services. When Anthropiс gave us MCP back in November, I was hesitant, because I felt in order for a protocol like this to become an industry standard, other AI heavy-hitters (OpenAI, Google..) would have to embrace it. But thankfully they do, and that became clear also on Wednesday at a great MCP event hosted by Stainless, with panelists from both OpenAI and Anthropic — and others — showing strong support and discussing real world use-cases and experiences they've had so far. Long live MCP!

80

7 Comments

For the last 5 years at Meta, nobody has assigned me a project, and yet I’ve been able to create huge impact in products like Rayban Meta Glasses, Generative AI, Video Calling, and countless other domains. Recently, we had 4 to 5 new E6s join my org. It took around 3 weeks for people to figure out where they wanted to plug in. And that was completely fine. Nobody rushed them. Nobody dumped random work on them just to keep them busy. Nobody said, “Here is your project, now go execute.” Because that is not how things work here. If I believe a project matters, I do not wait for a manager to hand it to me. I pitch it. I explain the problem. I explain why it matters. I explain the impact. Then, the other engineers decide if they want to join. Yes, that means a senior engineer can ask: “Do you want to work on this with me?” And the answer can be no. Agency here is not a nice-to-have. It is the job. You are expected to look at the product, spot what matters, and move toward it. You are expected to create momentum, not wait for instructions. You are expected to choose your direction, then back it up with impact. That also means this culture is not for everyone. If you want a manager to define your roadmap every quarter, you will struggle here. If you want to be spoon-fed work, you will struggle here. If you need constant clarity before moving, you will struggle here. But if you like ownership, if you like thinking for yourself, if you like building your own path, this kind of environment is hard to beat. That is the trade. More freedom. More responsibility. Much less hiding. And honestly, that is what personally has helped me do so much in the last 5 years here. 

237

20 Comments