About
Hands-on AI / Data / Engineering leader with 20+ years at startups to Microsoft &…
Articles by Vlad (Volodymyr)
Activity
22K followers
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisNEW AI report from Google. Every prior intelligence explosion in human history was social, not individual. These authors make the case that the AI "singularity" framed as a single superintelligent mind bootstrapping to godlike intelligence is fundamentally wrong. This is directly relevant to anyone designing multi-agent systems. They observe that frontier reasoning models like DeepSeek-R1 spontaneously develop internal "societies of thought," multi-agent debates among cognitive perspectives, through RL alone. The path forward is human-AI configurations and agent institutions, not bigger monolithic oracles. This reframes AI scaling strategy from "build bigger models" to "compose richer social systems." It argues governance of AI agents should follow institutional design principles, checks and balances, role protocols, rather than individual alignment.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisGoogle's latest paper on Compression is the future. Here's why. They compressed LLM memory 6x with zero accuracy loss. When ChatGPT writes a reply, it remembers every word you've said. That memory is stored in a growing notebook (KV cache). A 100,000-word conversation can eat 16 GB of GPU memory. That's half of what most high-end GPUs even have. This is the #1 cost of running AI. Not the thinking. The remembering. TurboQuant shrinks each number in that notebook from 32 bits to just 3. That's like replacing a full paragraph with three words and losing nothing. No retraining. Works on any model instantly. Compressing numbers usually destroys their meaning. Here's how they solved it: 1. Rotate the numbers randomly so they all land on a predictable curve (PolarQuant) 2. Use one extra bit to fix the tiny errors left behind (QJL) Once numbers are predictable, you need far fewer bits to store them. The results: > 8x faster on Nvidia H100 GPUs > 16 GB notebook shrinks to under 3 GB > Search indexing drops from 500 seconds to 0.001 > Accuracy identical to the uncompressed model There's a proven math limit on how good compression can get. TurboQuant is only 2.7x above that floor. We're near the ceiling. Every company running LLMs spends most of its budget on memory. This cuts that cost by over 80%. The race is no longer about bigger models. It's about cheaper inference. Models that needed a $200K server cluster start fitting on a single $2K GPU. AI agents run 24/7 without burning budgets. The companies that win won't just have the best models. They'll have the best compression. Papers are open-access on arXiv, presented at ICLR.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisI stayed up till 2 am, finished TurboQuant, and just pushed it to my GitHub. 4 Million tokens KV cache 1 Million tokens context 716.7 tokens/sec Qwen3.5-35B AWQ All of it on one tiny ZGX GB10 I implemented real TurboQuant-packed KV cache support in vLLM’s Triton attention path, meaning the KV cache is actually compressed in memory, not just academically optimized in algo. Real packed KV. Real memory reduction. Real speed. You can clone it from my GitHub and ruin your weekend hacking. Happy Friday! Make sure you own your AI. AI in the cloud is not aligned with you; it’s aligned with the company that owns it.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted this"When LLMs take on the bulk of code writing, engineers finally have the space to do all of the things that make software good, not just functional." This is the point I try to get across whenever I meet an engineer who fears that AI makes us obsolete. LLMs aren't replacing us; they're abstracting away some of the tedious parts so we can focus on the fun bit. Software engineering has been continuously evolving since its inception; none of us are out here making punch cards any more. This is a natural progression in the leverage that engineering tools hold and the way in which we use those tools. If you're a language purist I do have some bad news for you, because those details are about to become less critical to most folks. But if that's what you geek out over, maybe it's time to design a language optimized for LLMs? AI isn't perfect. Is any programming language, database implementation, or cloud platform perfect? Not even remotely. But each iteration problem solves for some of the constraints that make it difficult to make progress. We're builders; progress is what we're here for 🚀
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 shared this"...Also, I spend a lot more time thinking a lot farther out into the future of the effort than the AIs typically do..." - but can we make AI better in it Sam Schlesinger?Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 shared thisAnything you use AI assistance for you should also be practicing by hand. If you’re not gaining human knowledge in the process of your work, you’re wasting human potential and limiting yourself in the long run. In that vein, I’ve been working on a zero-dependency, handwritten formalization of Turing machines in Lean here: https://lnkd.in/eVew-yHP. I love writing Lean without assistance so far, though I’ve made much slower progress I am having a lot more fun than using AI assistance. Coming from Haskell and Rocq, it is my favorite functional programming language I’ve used by a long shot. Also, I spend a lot more time thinking a lot farther out into the future of the effort than the AIs typically do, setting myself up for a much easier time for theorems I know are coming down the line. To me, this means that with the current technology, even if you’re using AI assistance, investing heavily in human-driven design processes to ensure high quality foundations is going to be very important for the time being. The AI is much more willing to tolerate a slog than I am, which leads to a long-term compounding in the complexity of proof terms. This also leads to a reduction of the utility of AI generated proofs to the advancement of human knowledge. As we proceed with these new, automated scientific discovery tools we should keep an eye towards maintaining and expanding the knowledge of humanity. Humanity and our collective knowledge should be prioritized over flashy displays of AI prowess.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisAs someone who leads both the AI group and the Clinical Data Science organization at Gilead Sciences — spanning biostatistics, clinical pharmacology, real-world evidence, development systems and statistical programming — I sit at an interesting intersection: responsible for both the new capability and the established disciplines it needs to work alongside. Leaders in this position must carefully navigate fostering innovation through AI while simultaneously delivering on an exciting clinical pipeline. This article is intended for fellow leaders across biopharma who are navigating the early days of realizing value through AI. What follows are four hard-learned, actionable lessons — not theory — based on a cross-functional team’s journey of moving from AI promise to AI impact in a Clinical Development organization. #ai #pharm #datascience #genai #innovation #value #lifesciences #aiinhealthcareAI in Clinical Drug Development: Accelerating the Transformation Already UnderwayAI in Clinical Drug Development: Accelerating the Transformation Already Underway
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisRussia Bombs Beloved Ancient Church in City Center, Lviv We were there - meters from ancient church, when Shahed hit it destroying the annex. Below is a fast eyewitness report I created. All footage from my iPhone.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisAnthropic just shipped 73 product releases in 52 days. But almost everyone is taking the wrong lesson from this. The timeline is filled with people saying, "Anthropic just has the best engineers" or "Claude is just a smarter model." That completely misses the point. 73 releases in 52 days isn't just a flex about talent. It’s a warning shot about the new physics of software engineering. Anthropic is internally eating their own dog food. They are using unlimited tokens and agentic workflows to build Claude with Claude. And they just proved a terrifying reality for traditional tech companies: 1. The bottleneck is no longer writing code. The bottleneck is product velocity—deciding what to build and reviewing the PRs. 2. Two-week sprints are obsolete. If your engineering team is still doing quarterly roadmapping, you are competing against companies operating on daily, compounding autonomous cycles. 3. Scale isn't the moat; architecture is. They aren't shipping this fast because they have brute-force compute. They are shipping this fast because they've built a tight, deterministic feedback loop where AI agents handle the boilerplate, and humans handle the architecture. Siri shipped less in 5 years than Anthropic did in 52 days. If you are a founder or an engineering leader, you need to stop asking "How do we integrate AI into our product?" The only question that matters now is: "How do we integrate AI agents into our engineering org chart?" If you don't figure it out, someone else is going to ship your 2027 roadmap next Tuesday.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 reposted thisThis recent paper presents Mamba-3, the most advanced iteration of Mamba, a non-transformer architecture that provides the SOTA text modeling quality and speed. Read with an AI tutor on ChapterPal: https://lnkd.in/esJ9r-q2 PDF: https://lnkd.in/eBwUXfNp
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 replied to a comment
2mo
If complexity is the real moat, the outlook for human engineers is grim. Human capacity scales slowly, almost linearly. AI’s capacity appears to scale exponentially. At some point, AI will operate in abstraction spaces we cannot even understand - forget about competing with them.
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 commented on a post
3mo
"The rise of the #RenaissanceDeveloper ... Less syntax ... More intent ... Stronger systems thinking ... Real ownership ... You build it. You run it." - insights form the event by Aswini Atibudhi: https://www.linkedin.com/posts/aswiniatibudhi_renaissancedeveloper-ai-sdlc-activity-7409631616265908224-l1A0
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 commented on a post
3mo
Continuing the conversation from our recent panel: Arun Gupta - one of our awesome speakers - captured several standout points (and more) in this post: https://www.linkedin.com/posts/arunpgupta_sdlc-agents-renaissance-activity-7407129846314610688-6qwf/
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisNEW AI report from Google. Every prior intelligence explosion in human history was social, not individual. These authors make the case that the AI "singularity" framed as a single superintelligent mind bootstrapping to godlike intelligence is fundamentally wrong. This is directly relevant to anyone designing multi-agent systems. They observe that frontier reasoning models like DeepSeek-R1 spontaneously develop internal "societies of thought," multi-agent debates among cognitive perspectives, through RL alone. The path forward is human-AI configurations and agent institutions, not bigger monolithic oracles. This reframes AI scaling strategy from "build bigger models" to "compose richer social systems." It argues governance of AI agents should follow institutional design principles, checks and balances, role protocols, rather than individual alignment.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisGoogle's latest paper on Compression is the future. Here's why. They compressed LLM memory 6x with zero accuracy loss. When ChatGPT writes a reply, it remembers every word you've said. That memory is stored in a growing notebook (KV cache). A 100,000-word conversation can eat 16 GB of GPU memory. That's half of what most high-end GPUs even have. This is the #1 cost of running AI. Not the thinking. The remembering. TurboQuant shrinks each number in that notebook from 32 bits to just 3. That's like replacing a full paragraph with three words and losing nothing. No retraining. Works on any model instantly. Compressing numbers usually destroys their meaning. Here's how they solved it: 1. Rotate the numbers randomly so they all land on a predictable curve (PolarQuant) 2. Use one extra bit to fix the tiny errors left behind (QJL) Once numbers are predictable, you need far fewer bits to store them. The results: > 8x faster on Nvidia H100 GPUs > 16 GB notebook shrinks to under 3 GB > Search indexing drops from 500 seconds to 0.001 > Accuracy identical to the uncompressed model There's a proven math limit on how good compression can get. TurboQuant is only 2.7x above that floor. We're near the ceiling. Every company running LLMs spends most of its budget on memory. This cuts that cost by over 80%. The race is no longer about bigger models. It's about cheaper inference. Models that needed a $200K server cluster start fitting on a single $2K GPU. AI agents run 24/7 without burning budgets. The companies that win won't just have the best models. They'll have the best compression. Papers are open-access on arXiv, presented at ICLR.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisI stayed up till 2 am, finished TurboQuant, and just pushed it to my GitHub. 4 Million tokens KV cache 1 Million tokens context 716.7 tokens/sec Qwen3.5-35B AWQ All of it on one tiny ZGX GB10 I implemented real TurboQuant-packed KV cache support in vLLM’s Triton attention path, meaning the KV cache is actually compressed in memory, not just academically optimized in algo. Real packed KV. Real memory reduction. Real speed. You can clone it from my GitHub and ruin your weekend hacking. Happy Friday! Make sure you own your AI. AI in the cloud is not aligned with you; it’s aligned with the company that owns it.
-
Vlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked thisVlad (Volodymyr) Pavlov 🇺🇸 🇺🇦 liked this"When LLMs take on the bulk of code writing, engineers finally have the space to do all of the things that make software good, not just functional." This is the point I try to get across whenever I meet an engineer who fears that AI makes us obsolete. LLMs aren't replacing us; they're abstracting away some of the tedious parts so we can focus on the fun bit. Software engineering has been continuously evolving since its inception; none of us are out here making punch cards any more. This is a natural progression in the leverage that engineering tools hold and the way in which we use those tools. If you're a language purist I do have some bad news for you, because those details are about to become less critical to most folks. But if that's what you geek out over, maybe it's time to design a language optimized for LLMs? AI isn't perfect. Is any programming language, database implementation, or cloud platform perfect? Not even remotely. But each iteration problem solves for some of the constraints that make it difficult to make progress. We're builders; progress is what we're here for 🚀
Experience & Education
-
PMG
View Vlad (Volodymyr)’s full experience
See their title, tenure and more.
or
Licenses & Certifications
Publications
-
Early Bug Detection for Software Analysis and Design
Software Test & Performance Conference, San Mateo, California, USA
-
An Experience of Integrating INTSPEI P-Modeling Framework with Microsoft Solutions Framework for Agile Software Development
IASTED Software Engineering 2008 Conference, Innsbruck, Austria
-
Applying Pantomime and Reverse Engineering Techniques in Software Engineering Education
The 37th annual IEEE/ASEE Frontiers in Education Conference, Milwaukee, Wisconsin, USA
-
Global Demand For Human Productivity Impacts Software Companies
Software Business 2007 Conference, Santa Clara, California, USA
-
How to Optimize Your Software Development Process
"Open Systems" Magazine, Moscow, Russia
-
Best Practices for Launching and Managing Outsourcing Subsidiaries
First CEE Outsourcing Summit, Kiev, Ukraine
-
Is There Still a Room For Programmers' Productivity Improvement?
The 5th annual IEEE East-West Design and Test Symposium, Yerevan, Armenia
-
"Virtuoso" - Educational Project for Russian Universities
The First International Conference "Information Technologies and IT-Education", Moscow, Russia
-
Using Pantomime in Teaching OOA&OOD with UML
The 18th Conference on Software Engineering Education & Training (CSEET'05), Ottawa, Canada
-
"The Babel Experiment": An Advanced Pantomime-based Training in OOA&OOD with UML
The 36th "ACM Technical Symposium on Computer Science Education" (SIG CSE 2005), St. Louis, Missouri, USA
Patents
-
Enabling Seamless Remote Access to Online Applications Through a Web Server
Issued US 9,369,504
Other inventors -
-
Shared Simultaneous Access to A Single Instance of a Remotely Executing Application
Filed US 62/218,769
Other inventors -
-
Enabling Seamless Remote Access to Online Applications from a Device
Filed US 13/651,217
Other inventors -
Honors & Awards
-
Silver Medal at Kaggle 2019 Data Science Bowl
-
Silver Medal (top 5% solution among 3,497 competition participants) at Kaggle 2019 Data Science Bowl
-
Bronze Medal at Kaggle ASHRAE Great Energy Predictor Data Science Competition
-
Bronze Medal (top-10% solution among 3,614 participants) at ASHRAE Great Energy Predictor III Data Science Competition
-
Chair-Emeritus at CEE-SECR
2010
-
Marquis' Who's Who in the World
2009-2012
-
Top 10 the most popular papers from ACM’s refereed journals and conference proceedings
June - August 2006
in 2006 “Communications of the ACM” magazine three times included Vlad’s paper into the monthly list of top 10 the most popular papers from ACM’s refereed journals and conference proceedings
Languages
-
English
Native or bilingual proficiency
-
Russian
Native or bilingual proficiency
-
Ukrainian
Native or bilingual proficiency
Recommendations received
-
LinkedIn User
27 people have recommended Vlad (Volodymyr)
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