About
Over the past few decades Jud Valeski has enjoyed a successful career in technology. As a…
Activity
2K followers
-
Jud Valeski reposted thisExcited to start sharing a little more information about what we've been building . 🚀Jud Valeski reposted thisGuild.ai The control plane for AI agents. Discover, share, and run agents as shared production infrastructure. Now in beta. Sign up below.
-
Jud Valeski reposted thisJud Valeski reposted thisAI is making something very clear to me: AI + domain expertise = real productivity. AI without expertise = convincingly wrong output. The dangerous part is that it often looks right. I’m seeing more and more work get called ‘good’ because it reads well or was produced quickly, not because it would actually hold up under scrutiny. It just looks good on the surface. I notice this most outside of areas where I have deep expertise. It’s much easier to be impressed by output that sounds right when you don’t have the context to really evaluate it. I even catch myself doing this. This feels especially dangerous in startups. By definition, you don’t have deep expertise in every area yet, which makes it tempting to just ask Claude or ChatGPT to handle something in the interest of time and convenience. So I’ve been trying to hold a higher bar: Principle: No expert to sign off? It’s not done. AI-assisted work still needs to be owned and validated by someone with domain expertise. If no one in the room can confidently say “this is correct and here’s why,” it’s not done. And for coding: If you can’t explain it, you can’t ship it.
-
Jud Valeski reposted thisJud Valeski reposted thisA founder emailed me: she committed to Cursor, went heads down for 4 months, and shipped a healthcare platform with 400 users and 50 paying customers. 100% built with AI. Seasoned engineers keep telling her it's not possible. Investors are reluctant to fund it. I've watched this exact movie four times - the Internet was a toy, web software didn't work, the cloud wasn't safe, and mobile would never replace a computer. Steve Ballmer literally laughed at the iPhone on camera. Every time, both camps were wrong. The new thing doesn't replace the old thing. The old thing doesn't survive unchanged. New categories emerge that neither camp could describe. https://lnkd.in/gN_MD-rX
-
Jud Valeski reposted thisJud Valeski reposted thisMost people try to earn the next job by getting really good at the one they have. But careers aren’t really about the rung you’re standing on. The fastest ones happen when you step onto the next rung before anyone asks you to. Most management careers only have two structural shifts: IC → manager, and manager → manager of managers. After that, the pattern mostly repeats. The scope grows, and the work shifts. What used to be operational becomes tactical, and eventually strategic. At some point in every leadership career, your brain makes a small shift. HOW → WHO I think of it like a ROR instruction in the CPU. The bits don’t disappear. They just move to a different position. Early in your career your value comes from figuring out how to do the work. As you move into leadership, your leverage comes from figuring out who should own the work, who should decide, and who needs context. The job gradually shifts from solving problems yourself to building the system that solves them. I actually discovered this by accident. I was annoyed that we weren’t doing the things that mattered. So I raised the pirate flag and asked myself: “What should my boss be doing here that isn’t happening?” Then I started doing those things where it helped the team most. I didn’t ask for permission. Honestly, I thought I might get fired. It was uncomfortable. Instead, I got promoted. Two levels at once. Along the way I learned a few lessons. Here are five of them. 1. Start delegating the work. Your job is to give away every piece of your job. It’s no longer about being the fastest problem solver. It’s about creating more problem solvers. 2. Own outcomes, not tasks. Don’t just finish your piece. Make sure the outcome actually happens, and give people room to figure out how. 3. Start making decisions. If something is stuck because a decision isn’t being made, step in and make it even if you’re not sure it’s yours. Just communicate so you don’t create chaos. 4. Fix systems instead of incidents. If the same issue keeps showing up, the problem usually isn’t the people. It’s the system. For example, if every release breaks the same way, fix the release process. 5. Think one level higher. Eventually you stop solving problems yourself and start designing the system that solves them. ICs solve problems. Managers orchestrate execution. Leaders design systems. Often that means defining principles, rules that make good decisions repeatable. For example: “The team closest to the problem makes the decision.” As Israelmore Ayivor wrote, “The higher you are, the further you see.” The people whose careers accelerate spend time on problems one level above their role, even before it’s formally their job. Leadership rarely begins the day the title changes. It begins when you step onto the next rung.
-
Jud Valeski reposted thisI'm excited to see this release of a real-time translation tablet that runs with no internet connection, powered by Moonshine AI. If you need fast, private, accurate speech translation, check it out!Jud Valeski reposted thisIn the field, in the clinic, or in the classroom—clear communication can be the difference between success and setback. Yet language often stands in the way. At LILT, we believe no mission should fail because people can’t understand one another. Introducing LILT Converse: an AI-powered device that gives you real-time speech-to-speech and speech-to-text translation, anywhere, anytime. Imagine a first responder coordinating with local teams during a crisis, or a doctor explaining critical information to a family—all effortlessly, across languages. Unlike traditional translation tools that rely on the cloud or bulky hardware, LILT Converse runs 100% offline on secure, air-gapped Android devices. No data leaves the device, no connection is required, and no conversation is ever exposed. With mission-tuned AI models, Converse is built for professionals who need accuracy, speed, and privacy when it matters most. From defense and law enforcement to healthcare and education, LILT Converse lets teams speak the same language (literally). Because when the stakes are high, every word should connect. Read more: https://lnkd.in/gBhxgQC3 #PublicSector #TranslationInnovation #AITranslation
-
Jud Valeski reposted thisJud Valeski reposted thisToday I'm proud to launch Moonshine Voice, a new family of on-device speech to text models designed for live voice applications, and an open source library to run them. They support streaming, doing a lot of the compute while the user is still talking, so your app can respond to user speech an order of magnitude faster than alternatives, all while continuously supplying partial text updates. Our largest model has only 245 million parameters, but achieves a 6.65% word error rate on HuggingFace’s OpenASR Leaderboard compared to Whisper Large v3 which has 1.5 billion parameters and a 7.44% word error rate. We are optimized for easy integration with applications, with prebuilt packages and examples for iOS, Android, Python, MacOS, Windows, Linux, and Raspberry Pis. Everything runs on the CPU with no NPU or GPU dependencies. and the code and streaming models are released under an MIT License. We’ve designed the framework to be “batteries included”, with microphone capture, voice activity detection, speaker identification (though our diarization has room for improvement), speech to text, and even intent recognition built-in, and available through a common API on all platforms. As you might be able to tell, I’m pretty excited to share it with you all! We’ve been working on this for the last 18 months, and have been dogfooding it in our own products, so I can’t wait to see what you build with it. It's been quite a ride training all this with a seven-person team and a monthly GPU budget in the low tens of thousands, so please consider giving the repository a star on GitHub, that helps us a lot. https://lnkd.in/diz_SNSwGitHub - moonshine-ai/moonshine: Fast and accurate automatic speech recognition (ASR) for edge devicesGitHub - moonshine-ai/moonshine: Fast and accurate automatic speech recognition (ASR) for edge devices
-
Jud Valeski reposted thisJud Valeski reposted thisEvery time a product adds ads, people lose it. Instagram “sold out.” YouTube “unwatchable.” Spotify “unusable.” Netflix “betrayed us.” Prime Video “wait… I pay for this.” Facebook “ruined my feed.” Snapchat “killed the vibe.” Twitter/X “just a billboard now.” Reddit “they’re everywhere.” Cue the rage posts. The “I’m quitting” comments. The end-of-days hot takes. If you’ve ever launched something and panicked at the comments, this is for you. Because ads are a particularly disliked change. But really, people dislike most change. People hate surprises. They hate losing something they had. And they really hate having a habit disrupted. Initial backlash tells you about expectations. Retention tells you about value. Watch what people do once the dust settles. Outrage is loud. Habits win.
-
Jud Valeski reposted thisJud Valeski reposted thisI've been wrestling with embeddings for speaker identification recently, and it's not an area I know well. To improve my own understanding, and help anyone else who needs to get up to speed on the practical details, I've put together a Python notebook on Colab. It walks you through how the embeddings work, with inline examples that build into a simple "is this audio from the same person or not?" system. I hope it's helpful to a few of you out there! https://lnkd.in/gE8qWATB
-
Jud Valeski shared thisYou’re getting to the nut of it; trust. Because of its non-deterministic nature, the trust-gap in using language models (large or small) is real (I like your callout to context windows being opaque state). If trust is going to be an issue, then narrowing and breaking down model usage via agent orchestration becomes a big part of the game. And large foundation models are important (they’re foundational after-all), but, they serve only one, increasingly small, role. An all-singing-all-dancing model/agent in a demo sounds neat, but, its impractical.Jud Valeski shared thisOne Agent Is a Mainframe Problem I’ve been thinking a lot about agents lately. And the more I look at them, the more they remind me of the early days of computing. Back then, we bet on mainframes: one powerful system that did everything. Then PCs showed up. Then clusters and racks of servers won. Not because any single machine was smarter, but because coordination scaled better than capability. AI feels like it’s in that same mainframe moment. Everyone’s pitching “one agent to rule them all”: a universal assistant that holds all context, understands every domain, and acts autonomously on our behalf. It’s a compelling demo. It’s also a fragile system. Demos love monoliths. Production doesn’t. Here’s the thing: the minute AI touches real workflows—real users, real data, real consequences—trust, cost, and control start pushing you toward specialization. And that’s when the cracks show up: – context windows turn into hidden state – prompts turn into brittle contracts – failures get hard to isolate – and pretty soon, humans don’t know what to trust If you can’t isolate failure, you can’t scale trust. What works in practice looks different. A concrete example: An agent ships a pricing change in a fintech system. It reads the ticket, updates the code, flips a config, and opens a PR. Looks magical. Until it misses one requirement: pricing changes need a customer notice window and an approval trail. Now the failure isn’t a bug. It’s a compliance incident. And the worst part: you can’t even tell which step in the agent’s reasoning went wrong. The blast radius is the entire workflow. So teams do what they always do in production: they split the work. One agent lives in the codebase. Another reasons about requirements. Another tries to break the change. Another checks risk or compliance before anything ships. Each one is narrower. Each one is more predictable. And together, they’re more powerful than a single “do-everything” agent. This doesn’t mean foundation models go away. Mainframes never disappeared. Large models will still be the source of raw intelligence. But intelligence alone isn’t a system. Autonomy is the demo. Operability is the product. The real innovation surface is orchestration: how work gets decomposed, how agents coordinate, and how humans stay in the loop. We’ve seen this cycle before. Centralization feels inevitable early because it’s easier to imagine. Decentralization wins later because it’s easier to operate. The future of AI probably doesn’t look like a single mind. It looks more like a rack of agents.
-
Jud Valeski liked thisExcited to start sharing a little more information about what we've been building . 🚀Jud Valeski liked thisGuild.ai The control plane for AI agents. Discover, share, and run agents as shared production infrastructure. Now in beta. Sign up below.
-
Jud Valeski liked thisJud Valeski liked thisGuild.ai The control plane for AI agents. Discover, share, and run agents as shared production infrastructure. Now in beta. Sign up below.
-
Jud Valeski liked thisJud Valeski liked thisA founder emailed me: she committed to Cursor, went heads down for 4 months, and shipped a healthcare platform with 400 users and 50 paying customers. 100% built with AI. Seasoned engineers keep telling her it's not possible. Investors are reluctant to fund it. I've watched this exact movie four times - the Internet was a toy, web software didn't work, the cloud wasn't safe, and mobile would never replace a computer. Steve Ballmer literally laughed at the iPhone on camera. Every time, both camps were wrong. The new thing doesn't replace the old thing. The old thing doesn't survive unchanged. New categories emerge that neither camp could describe. https://lnkd.in/gN_MD-rX
-
Jud Valeski liked thisJud Valeski liked thisIt took me longer than expected to write this, but I finally did. This is why my first startup failed. https://lnkd.in/gN6m7Bk7
-
Jud Valeski liked thisJud Valeski liked thisYou know it's a good dinner when you leave with a head full of ideas, a list of to-dos, and realize a 3-hour dinner flew by and you didn't take a single photo! 😅 Thanks to my co-hosts Malte Witt, Conor Swanson, John Kozura for humoring me with a picture before we left the restaurant! To the amazing group of Boulder CTOs who joined us for a delicious farm-to-table dinner at Bramble & Hare tonight - Appreciate you taking time out of your busy day! It's a great reminder of how important it is to carve out time and mental space from back-to-back meetings for strategic thinking and connecting with peers on similar journeys. Whether we're leading teams of 7 or 200+, the challenges are real, and the stakes are high! The conversations covered a lot of ground: - Big picture questions about where all this is heading - What problems feel solved - and which ones remain unsolved - How people are adapting to new tools and expectations - Fundamental changes to orgs, processes, and units of work From advancements in science to existential crisis - and everything in between! We all agreed: the pace of change is unprecedented, and part of our role as leaders is providing clarity and support as teams navigate AI transformation. It's 2am, my brain is running 100+ mph, and I'm far too wired to sleep... instead, I'm jotting down ideas while my AI agents work quietly in the background, lol - the future is here! 😂 Next up: Austin, Texas. We're taking CTO dinners on the road, and will be co-hosting with the team at Howdy.com on April 14th. Jacqueline Samira found us the perfect venue (sneak peak below), and I'm excited to hear Frank Licea talk about his experience driving AI transformation at scale. Austin area CTOs interested in attending, please DM me for details and an invitation - space is limited! #CTO #AITransformation #StartupLife
Experience & Education
-
Sovrn
View Jud’s full experience
See their title, tenure and more.
or
Publications
-
Bad Data - Chapter 18 "Social Media: Erasable Ink?"
O'Reilly
See publicationWhat is bad data? Some people consider it a technical phenomenon, like missing values or malformed records, but bad data includes a lot more.
Patents
-
Web Page Performance Scoring
Issued US 7,475,067
See patentThis patent was a result of brainstorming around how to better measure end-user performance of web browsing back at Aol / Netscape.
Languages
-
French
-
Recommendations received
-
LinkedIn User
7 people have recommended Jud
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content