Voxel51

• Explore, tinker with, and push the boundaries of the latest AI and machine learning technologies to showcase what's possible with open-source tools and frameworks

• Build inspiring prototypes, proof-of-concepts, and example applications that demonstrate creative and impactful uses of AI to the developer community

• Contribute meaningfully to the open source ecosystem by creating and maintaining libraries, tools, datasets, tutorials and guides that enable developers to more easily leverage AI capabilities

• Curate, clean, annotate and publish unique multimodal datasets spanning text, images, audio, video, 3D and other emerging data types to facilitate novel machine learning research and applications

• Create engaging technical blog posts, social media content, videos, and presentations that educate and inspire developers and data scientists to build with AI

•  Collaborate openly with our internal teams and the broader AI/ML community to cross-pollinate ideas, share knowledge, and collectively advance the state-of-the-art

• Evangelize our technology and vision by representing the company at hackathons, conferences, meetups, universities, and other developer-focused events

•  Influence the product roadmap by providing input on the developer experience, uncovering opportunities for new features and improvements based on community feedback and your work

• Continuously learn, experiment with, and provide feedback on beta versions of our latest models and capabilities to help shape the future of our platform

Hacker-in-Residence

LinkedIn

1) Introductory computer vision course for LinkedIn Learning

2) Prompt Engineering with LangChain (releasing March 2024)

3) RetrievalAugemnted Generation with LlamaIndex (releasing Q2 2024)

4) Fine-tuning LLMs with Hugging Face TRL (est release Q3 2024)

5) Text Generation with Hugging Face Transformers (est release Q3 2024)

LinkedIn Learning Instructor

The Artists of Data Science

The ONLY personal growth and development podcast for Data Scientists

Host

Deci AI

Developer Relations Manager

Pachyderm Inc.

Senior Developer Advocate

Comet

ML Developer Advocate

Data Science Dream Job

Principal Data Science Educator

Price Industries

Lead Data Scientist

Bold Commerce

Senior Data Scientist

Emergent BioSolutions

Biostatistician

I create content about deep learning | The DevRel and Deep Learning Dude

Hacking on LLMs, curating multimodal datasets, and writing a book about RAG.

Harpreet Sahota 🥑

sam3 is seriously a step change improvement over sam2

the model is seriously good, glad they brought back text prompts. the automatic segmentation isn't as good as i hoped it to be...but it still works quite well. 

i also got it to output image embeddings and that was cool to visualize in the app

learn more here: https://lnkd.in/gzqJzj6G

🤖 Hacker-in-Residence @ Voxel51| 👨🏽‍💻 AI/ML Engineer | 👷🏽‍♀️ Technical Developer Advocate | Learn. Do. Write. Teach. Repeat.

qwen3vl is literally the best model i've used for video usecases. blows my mind

hosting a virtual event tomorrow about document ai

you may or may not find it interesting, but i guarantee you'll learn something

join me here: https://lnkd.in/dFXPimVc

sony ai released a pretty cool dataset called the fairness human centric image benchmark, super high quality labels

instructions for downloading and parsing the dataset into fiftyone format can be found here: https://lnkd.in/g_VwxJx4

physical AI is top of mind as the next frontier for deep learning. but it'll only succeed if you have good datasets

language models scaled with web-scale text. 

vision models got ImageNet. 

now Voxel51 and NVIDIA are trying to do the same thing for physical AI 

autonomous vehicles, humanoid robots, anything that moves around atoms.

building these systems means millions of hours of testing in high-fidelity sims before you can actually deploy anything. 

petabytes of multimodal sensor data from various modalities

 LiDAR, cameras, radar, IMUs. 

but over 50% of simulations end up completely unusable because the input data is bad.

even small calibration or timing errors between sensors cascade through your neural reconstructions. 

you waste weeks and burn through millions in compute costs before you realize what went wrong.

the new FiftyOne Physical AI Workbench. it integrates with NVIDIA Omniverse and NVIDIA Cosmos and basically sits at the start of your sim pipeline to validate, clean up, and prep your multimodal data before it hits neural reconstruction and synthetic data generation.

what it does:

- automatically audits physical AI inputs across 75+ critical checkpoints

- transforms raw sensor streams into structured, searchable datasets

- triggers accurate neural reconstructions and generates synthetic scene variations

this means fewer wasted compute hours. 

iteration loops go from weeks to hours. 

you build safer, more reliable systems on data you can actually trust.

read more about what we've been building: https://lnkd.in/gC6_zn-X

i recently integrated 4 ocr models into fiftyone as remote zoo models these handle text extraction and document parsing. 

all available as remote zoo sources, you can get started with a few lines of code

different approaches for different needs:

1. mineru-2.5

1.2b params, two-stage strategy: global layout on downsampled image, then fine-grained recognition on native-resolution crops. 

handles headers, footers, lists, code blocks. strong on complex math formulas (mixed chinese-english) and tables (rotated, borderless, partial-border).

good for: documents with complex layouts and mathematical content

https://lnkd.in/dSwimdvp

deepseek-ocr

dual-encoder (sam + clip) for "contextual optical compression." 

outputs structured markdown with bounding boxes. has five resolution modes (tiny/small/base/large/gundam). gundam mode is the default - uses multi-view processing (1024×1024 global + 640×640 patches for details). 

supports custom prompts for specific extraction tasks.

good for: complex pdfs and multi-column layouts where you need structured output

https://lnkd.in/dhJn2h-v

olmocr-2

built on qwen2.5-vl, 7b params. outputs markdown with yaml front matter containing metadata (language, rotation, table/diagram detection). 

converts equations to latex, tables to html. labels figures with markdown syntax. reads documents like a human would.

good for: academic papers and technical documents with equations and structured data

https://lnkd.in/dNdTQYKg

kosmos-2.5

microsoft's 1.37b param multimodal model. two modes: ocr (text with bounding boxes) or markdown generation. automatically optimizes hardware usage (bfloat16 for ampere+, float16 for older gpus, float32 for cpu). handles diverse document types including handwritten text.

good for: general-purpose ocr when you need either coordinates or clean markdown

https://lnkd.in/dCTfGfV8


two modes typical across these models: detection (bounding boxes) and extraction (text output)

i also built/revamped the caption viewer plugin for better text visualization in the app: 

https://lnkd.in/dBebWdga

i've also got two events poppin off for document visual ai:

- nov 6 (tomorrow) with a stellar line up of speakers: https://lnkd.in/gkBEY4_r

- a deep dive into document visual ai with just me: 

https://lnkd.in/dFXPimVc

Open article: Document Visual AI with FiftyOne - November 14, 2025 by voxel51.com

Document Visual AI with FiftyOne - November 14, 2025

hype for this new integration, and proud of the team for shipping it at lightning speed. big ups to Daniel and AdonAI for goin beast mode on this one

🚀 Introducing Physical AI Workbench — integrated with NVIDIA Omniverse NuRec and NVIDIA Cosmos to solve the biggest bottleneck in physical AI.  

More than half of all physical AI simulations fail — not from bad models, but from bad input data. A timestamp off by milliseconds or a calibration drift of a few degrees can cascade through your neural reconstructions, leaving you with unusable results after weeks and millions in wasted compute. 

FiftyOne Physical AI Workbench sits at the start of your simulation pipeline and provides turnkey access to NVIDIA technology to ensure every simulation run begins with trustworthy data: 

✅ Audits 75+ critical checkpoints to ensure every reconstruction starts with validated inputs 
🎯 Enriches raw sensor streams into structured, searchable datasets 
🔗 Integrates directly with NVIDIA Omniverse and NVIDIA Cosmos for high-fidelity reconstruction and synthetic data generation. 

The result: an end-to-end workflow that prevents downstream failures — cutting iteration cycles from weeks to hours and eliminating GPU waste on failed reconstructions. We’re excited to see this workflow power autonomous vehicles NVIDIA DRIVE and robotics NVIDIA Robotics development teams move faster with higher accuracy. 

🔗 Read the full technical breakdown: https://lnkd.in/gTy5-TNF

QIMA

As the global Head of Data for QIMA, I oversee the company's artificial intelligence (AI), business intelligence (BI) and data engineering teams with a focus on generating actionable, data-driven insights for the business.  

QIMA is one of the world's leading supply chain analytics companies with labs and offices in 100+ countries around the world.

Head of AI and Data

Kolea Consulting LLC

📊 Designed, built, and implemented client solutions related to data strategy & data project prioritization, predictive analytics, and data visualization for clients in North America.

🤝 Worked closely with the Vice President of Analytics at CGI to create the data strategy for one of Canada’s largest provincial governments. The project had a 5-year budget of $40 million to develop a Data Management Office for the province. Data Quality, Data Literacy, Cybersecurity, and Organizational Change Management were my primary focus areas. 

🖥️ Co-led an Agile analytics team with CGI to design, build, and implement an analytics system using R Shiny web applications for one of British Columbia’s public health departments.

Director Data Science

upGrad

🎓Designed and delivered live teaching sessions for students enrolled in upGrad's professional data science and business analytics certificate program in partnership with the University of Maryland Robert H. Smith School of Business.

📊 Topics taught included Python, Deep Learning, Convolutional Neural Networks, NLP, Text Analytics, Exploratory Data Analysis, and Statistics.

🗣️ Feedback from students:
     "Courtlin was excellent in his presentation and shared his experience well."
     "Loved the session. One of the best sessions so far."

UpGrad Expert - Advanced Machine Learning, Data Visualization and Statistics Instructor

Philip Morris International

📈 Led the data-driven decision-making transformation as head of the Enterprise Analytics department (reporting directly to the Finance Director for Indochina) while promoting the use of self-service analytics.

🧑‍🤝‍🧑 Recruited and supervised a team of junior, senior, and manager-level analysts while guiding their professional development.

📋 Developed comprehensive standard operating procedures for the analytics division to improve the consistency of operations and assist with new hire onboarding.

🛠️ Managed the project team that built the company’s business intelligence system (covering its sales, distribution, and operational functions) using SQL, Python, R, and Microsoft Power BI. Created the data model design and overall data architecture for the BI system.

🤝 Operated as the primary analytics business partner to support internal stakeholders with comprehensive analyses of issues, including the performance of new product launches, geospatial analysis for distribution routes, customer segmentation, and customer profitability (ROI).

🎯 Developed an improved sales forecasting methodology using Python that was more accurate than the company’s preexisting, manually derived sales forecasts. (Forecast error was reduced by approximately 67%)

⏱️ Reduced the time spent by analysts in the Sales Development, Distribution, and Marketing departments gathering data and preparing routine management reports by 10-20 hours per month per department by measuring report usage, eliminating unnecessary reports, and leveraging Power BI automation. The time saved was used to conduct deeper, more insightful analysis of the underlying data.

👥 Led the development of a B2B customer segmentation model using unsupervised clustering algorithms that challenged several long-held company beliefs regarding which customer segments were growing and where they were located. The results were used to generate various B2B customer personas to improve trade promotion program design.

Director of Enterprise Analytics Vietnam

Data Scientist

En Pointe Management Corporation

💼 Provided internal business plan consulting, financial modeling, business process optimization, competitive marketing research & business intelligence to multiple subsidiaries within our group of companies in Vietnam.

🏆Accomplishments 

🔍 Designing, overseeing, and analyzing the competitor, market, and consumer intelligence data collected by the department’s research analysts to generate actionable insights and strategic recommendations for the Board of Directors.
💹 Creating financial forecast models for potential acquisitions.
📊 Developing KPIs, dashboards, and business scorecards for subsidiaries to align their strategic and operational priorities.
📑 Conducting business feasibility studies, developing business plans, and designing sales forecasts for potential new ventures.
🤝 Structuring and negotiating partnership agreements with counterparts from the US, EU, and South East Asia region.
📅 Assisting subsidiary companies with the creation of annual strategic plans and budgets.

* Skills Required
Business intelligence  data analysis, financial analysis, financial modeling, strategic planning

En Pointe provides core business functions including Human Resources, Digital & Information Technology, Finance & Accounting and Business Development to enable each company in our group to excel at what it does best. Our Group companies include:

1) Purpose Media 
2) Novel Production
3) Amberstone Media
4) Billboard Vietnam
5) INQ
6) Renaissance Collection
7) Embassy Education
8) AMPA Education
9) Soul Music & Performing Arts Academy - SMPAA

Director Of Business Intelligence

Kolea Consulting Group manages a portfolio of machine learning and AI-powered businesses serving customers in North America and Southeast Asia.

Director Consulting Services

Diamond Head Financial Advisors

Managed the firm's value-oriented investment strategy by conducting primary research on investment ideas, selecting securities, executing trades and managing portfolio risk.

Created the firm's summer intern equity investment analyst training program. Many of graduates from the training program went on to secure investment banking job offers.

Developed strategic partnerships with 3rd party financial services providers.     

Software: 
Bloomberg Professional, Factset, TD Ameritrade Institutional ThinkPipes, TD Ameritrade ThinkOrSwim, AssetBook, Morningstar

Portfolio Manager & Vice President of Business Development

Analyzed and pitched long/short equity investment ideas to the firm's Chief Investment Officer. Built financial models and valued companies on the basis of DCF, relative value and LBO methodologies. Covered companies in the technology, healthcare and industrials sectors. Executed trades and rebalanced client portfolios on behalf of the CIO.

Registered Investment Advisor (RIA)

Investment Research Analyst

Gigaisland LLC

Created an in-house firewall Global Management System; reduced service delivery costs by 87%.
Developed and implemented the firm’s cybersecurity assessment product offering.
Refined and documented the firm’s core operational procedures into a series of SOP manuals.
Remotely managed several hundred client PC and Mac workstations and several dozen Windows servers.
Conducted cybersecurity assessments for corporate clients, identified vulnerabilities and recommended improvements.
Designed hardware and software solutions to protect clients from malware, ransomware and spyware.

IT Systems & Cyber Security Analyst

State of Hawaii Employees' Retirement System

Served as the State of Hawaii public pension fund's Investment Specialist throughout the Great Recession of 2008-2011. Assisted the Chief Investment Officer in the management of the State's $10.5 billion employee pension plan. Monitored the monthly performance and operations of the Plan's 30+ outside investment managers. Conducted reviews of several hundred prospective third-party investment managers. Created a comprehensive cash flow model for the pension fund to model various investment return scenarios. Rebalanced the Plan's asset allocation on a quarterly basis. Analyzed RFP responses for the Plan's private equity, venture capital and large-cap equity mandates.

Investment Specialist

As the Head of Data for QIMA, I lead the design, development, and implementation of data-driven solutions for the company and its clients. I have over 10 years of experience in data analytics, business intelligence, and machine learning, as well as various Microsoft certifications in data analysis and data science. 

My mission is to help organizations leverage data to optimize their operations, enhance their performance, and achieve their goals. I have successfully delivered data strategy and BI projects for one of Canada's largest provincial governments, one of British Columbia's public health departments, and Philip Morris International, where I headed the Enterprise Analytics department for two years. I also have a solid background in finance, risk management, and consulting from my previous roles. I enjoy teaching and sharing my knowledge on advanced machine learning, data visualization, and statistics as an instructor for upGrad.

Courtlin Holt-Nguyen

“Nobody in their right mind would use LLMs of the type that we have today" within 3-5 years, he said at MIT last month. (Yann LeCun)

His alternative: world models. Systems that learn from visual data and build causal understanding, like a child learning physics by dropping things thousands of times.

Global Head of AI and Data @ QIMA | AI Strategy & Innovation Leader | Driving Business Value through AI/ML & Data | ex-Fortune 500 Analytics Head, AI Engineer, Hands-on developer

Founder & CEO, Curiouser.AI | Berkeley Instructor | Building Values-Based, Human-Centered AI | LinkedIn Top Voice in AI

Yann LeCun just left Meta.
Let that sink in.

One of three people who invented modern deep learning. Turing Award winner. 

Chief AI Scientist at a company spending billions on AI.

Gone.

Not to a competitor. Not to retirement. 
To a startup building technology 
Meta apparently doesn't prioritize.

This matters more than most people realize

LeCun has spent 40 years being early and right. 

In the 1980s, he couldn't find a PhD adviser because machine learning was "a dead end." 

Three decades later, that dead end became the foundation of modern AI.

Now he's saying large language models, the technology behind ChatGPT, Gemini, and Meta's Llama, are another dead end.

"Nobody in their right mind would use LLMs of the type that we have today" within 3-5 years, he said at MIT last month.

His alternative: world models. Systems that learn from visual data and build causal understanding, like a child learning physics by dropping things thousands of times.

Current LLMs? 
They predict the next word. 
They don't understand gravity. 
They don't know that objects exist when you're not looking at them.

LeCun compares their intelligence to a cat's. He thinks the cat is smarter.

Here's what troubles me:

Meta is spending countless billions on technology that its own Chief AI Scientist publicly called a dead end. 

Either LeCun is wrong (possible, but his track record suggests otherwise), or Meta is chasing market narratives over technical reality.

His departure suggests he's betting on the latter.

The implications for current AI investments are significant. If LeCun's 3-5 year timeline holds, companies built entirely on LLM scaling may face architectural obsolescence.

That's not a guarantee. Even geniuses miss sometimes.

But LeCun has never been wrong about something this big before
********************************************************************************
The trick with technology is to avoid spreading darkness at the speed of light.
Stephen Klein
Founder & CEO, Curiouser.AI
Instructor, UC Berkeley Haas School of Business
WeFunder (link in comments)

Sources:

Wall Street Journal - LeCun departure and startup plans (November 2024)
MIT Symposium - LeCun remarks on LLM limitations (October 2024)
Meta FAIR Research Publications - World model architectures
Turing Award Archives - LeCun, Bengio, Hinton recognition (2018)

Impressive how much can be done now with minimal pre-trained CV tools.

🤖 Generative AI Lead @ AWS ☁️ (150k+) | Startup Advisor | Public Speaker | AI Outsider | Founder Thinkfluencer AI

AI Engineer builds a real-time potato counter using just ONE training image.

Most people assume you need huge datasets and large models to build anything useful with Al. Well, this project proves otherwise.

A senior Al engineer created a real-time potato counter for a conveyor belt using a tiny setup.
No custom dataset. No heavy model. Just one annotated frame.

He used Ultralytics ObjectCounter with a YOLO11 nano model, labeled a single image with SAM 2, and trained it.

That one frame was enough for the system to track every potato in the entire video.

This is a solid reminder of what real industrial Al looks like. Simple. Efficient. Built for one clear job.

These focused systems save time, cut errors, and work without big infrastructure. In other words, this is the future!

#ai #physicalai

There is a strange comfort in the chaos of AI projects.

You walk in with a plan, a roadmap, a stack of best practices. You think you know how the story will go. But then the data is messier than you expected. The business priorities shift. The model that looked so promising in the lab suddenly falls apart in production, like a sandcastle at high tide.

And you realize, again, that the real work is not in the code or the math. It is in the conversations. The late-night Slack threads with product managers who are quietly panicking. The whiteboard sessions where you try to explain why the model needs more data, not just more compute. The moments when you have to say, "I don't know yet, but we'll figure it out."

This is the part nobody puts in the case studies. The uncertainty. The improvisation. The way a team learns to trust each other, not because everything goes right, but because they survive the things that go wrong.

Truth is, the best AI teams are not the ones with the fanciest algorithms. They're the ones who can adapt, who can laugh at the absurdity, who can keep moving when the ground shifts under their feet.

If you want to build real value with AI, start by building that kind of team. The rest is just details.

There is a myth that AI teams are made up of lone geniuses.

The reality is much messier. The best teams I have worked with are a patchwork of personalities, backgrounds, and skill sets. Some are obsessed with the math. Others are relentless about the business case. A few just want to build cool stuff and see what happens.

The magic happens in the friction. The arguments over feature selection. The debates about what "success" really means. The moments when someone asks a question that no one else thought to ask.

You cannot fake this kind of collaboration. You have to create space for it. Psychological safety is not a buzzword, it is the foundation.

If people are afraid to be wrong, you will never get the right answers.

So yes, hire for skill. But more importantly, build for trust. That is how you get teams that do more than just ship models. They move the business forward.

And honestly, that is the only metric that matters.

This is a key piece of the explanation why so many studies (MIT, McKinsey) are reporting little financial impact from genAI. 

When the complexity of a coding project exceeds the POC stage, continuing to rely on a flaky AI coding assistant that refuses to follow instructions becomes a serious problem. The 5-sentence prompt that vibe coded the POC is not going to cut it. 

Really interesting Alex Turnbull , thanks for sharing.

Bootstrapped Groove from $0–$5M ARR solo. Now rolling it into a holding co. for CX SaaS. Launching Helply, InstantDocs & ZeroTo10M to scale $0–$10M ARR w/ 50%+ margins. Sharing it all at ZeroTo10M.com.

AI coding tools traffic collapsed 76% in 12 weeks.
10,000 startups got prototypes working, then 57% hit the complexity wall. That's a $400M-$4B rebuild market nobody's talking about.

Base44 spiked 950% in May, then crashed to 95% by October.
Lovable went from +207% to -37%.
Cursor dropped from +62% to -19%.

The pattern was identical across every AI coding platform.

Massive adoption surge, then 60-80% user decay within three months.

These tools made it easy to build prototypes fast.
But integration, governance, security, and reliability stayed hard.

The average user tried for 12 weeks, hit the wall, and churned.

Most founders who went all-in on "vibe coding" is now paying the rebuild tax.

They got an MVP up.
Found traction, raised a pre-seed, or onboarded early customers.
Then everything broke when real users showed up.

Now they're spending $200-300K hiring senior engineers 
to re-architect everything.

Multiply that across thousands of teams.
Hundreds of millions in wasted spend.

Everyone could suddenly build something.
Now everyone's learning what it costs to make it actually work.

This is a market correction. 

You can't outsource architecture to autocomplete.

System design
Error handling
Deployment, compliance.

All the unsexy stuff that keeps products alive 
multiplies the need for engineers. 

The gold rush now is rescuing companies that got stuck 
with working demos they can't scale.

If 10,000 startups experimented
say 8,000 will need help getting unstuck.

With budgets between $50K and $500K each.

That's the real market AI coding created
The money's in the valley of this chart.

Your BI program is burning millions of dollars.

But not on the technology.

On the invisible costs nobody talks about.

I've watched Fortune 500 companies spend 60-80% of their analytics budget on something that doesn't appear in any dashboard or ROI calculation:

Data reconciliation.

Here's what's actually happening:

Your sales team trusts the CRM numbers. Finance believes the ERP data. Marketing swears by their attribution platform.

They're all looking at the same customers. But seeing different revenue figures.

The real cost?

Your analysts spend their days playing data detective instead of delivering insights. Reports arrive weeks late because someone needs to "validate the numbers" first.

Executive meetings turn into debates about whose data is correct rather than what decisions to make. Or worse, every meeting generates follow-up meetings to validate the data.  

The brutal truth about BI failure:

It's not a technology problem. It's an alignment problem.

You bought the best analytics platform money can buy. But your data still lives in silos. Your metrics still lack clear definitions. Your insights still don't connect to strategic goals.

I've seen this time and again with the companies I've worked for; world-class tools, talented teams, substantial budgets but the same result. 

Our BI program struggled until we stopped asking "what data do we have?" and started asking "what decisions do we need to make?"

The shift?

Define 3-5 strategic outcomes first. Then build the data infrastructure to support those specific decisions.

Not the other way around.

When you have 20 minutes, watch this on 1.5x speed. It’s a fascinating video summary of the book “If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All”. 

This isn’t doom and gloom AI where Terminator-style robots come for humanity (see Musk’s plan to have “helper” police bots following each of us around all day to stop crimes before they happen for that scenario.) 

Instead, it’s a logical, step-by-step analysis of how super intelligence could escape from the AI research lab and then continue to obtain the resources it needs to improve itself beyond human control. The use of social media influencers (some human, some AI-generated) to manipulate public opinion, encoding messages to its future self inside its own made-up mathematical language, creating bioweapons to weaken humanity’s response while simultaneously positioning itself as our savior with DNA-based vaccines, the intentional sabotaging of rival AIs (and even hobbling smarter future versions of itself so they don’t deceive it the same way it would have deceived us) will give you plenty to ponder this weekend. 

Link to the YouTube and original book are in the comments.

Flashy investments may grab the headlines, but the boring, mundane ones often have the best economics.

CEO at Novaro Capital • $9bn+ of Transaction Experience • Investing in Niche Real Estate Opportunities

Everyone's fighting over gigawatt data centers. The real AI infrastructure play is sitting empty in your city.

That old warehouse with 5 MW of power? It's worth 10x what it was three years ago. And nobody's paying attention yet.

-The Hidden AI Gold Rush-

While Meta builds $29B facilities for training models, the actual AI revolution needs something different: distributed compute for real-world applications.

Self-driving cars can't wait for data to travel to Virginia and back. Drones need sub-10ms response times. Manufacturing robots require local processing to avoid catastrophic delays.

These applications don't need 500 MW campuses. They need 1-10 MW facilities within 50 miles of where they operate.

-Why These Sites Are Suddenly Valuable-

Power availability in urban areas is the constraint. New utility connections take 2-4 years. But existing sites? Already connected.

An old warehouse with 5 MW can become an edge data center in 9 months. Try getting that power allocation new - you'll wait until 2029.

The economics:
- Acquisition: $5-15M
- Conversion: $10-20M  
- Value post-conversion: $50-80M
- Timeline: 9-12 months vs 3-4 years

-Who's Already Moving-

Vapor IO is dropping micro data centers at cell towers - supporting autonomous vehicle corridors. Partnering with Hangar for drone operations.

EdgeConneX raised $1.9B to convert suburban facilities. They're buying existing industrial sites with power in place.

DataBank is turning dead office buildings into edge facilities. 2-10 MW conversions where nobody's looking.

Locus Robotics needs edge compute for warehouse robots. Every fulfillment center running AMRs needs local processing.

-The Use Cases Driving Demand-

Autonomous Vehicles: Waymo and Cruise need compute every 10-15 miles. Can't process in the cloud at 45mph.

Drone Delivery: Amazon Prime Air, Zipline require local processing for collision avoidance.

Smart Manufacturing: Boston Dynamics robots need sub-5ms latency. Compute must be within 10 miles.

Healthcare: Surgical robots can't risk network delays. Hospitals are building their own edge facilities.

-What Makes a Site Valuable-

Power: 1-10 MW available capacity
Location: Within 50 miles of population centers
Fiber: Proximity to major routes
Zoning: Industrial/flex allowing data center use
Loading: Existing docks for equipment delivery

-The Window Is Closing-

We're seeing 3-5x appreciation on these sites. A client bought a 3 MW Dallas building for $8M in 2022. Current offers exceed $25M.

This arbitrage won't last. Once institutional capital realizes edge is where AI deployment happens - not training - these sites disappear.

Smart money is quietly accumulating these properties. Not for industrial use. For the infrastructure that makes autonomous systems work.

The hyperscalers are building AI brains. These edge sites? They're the nervous system making AI useful.

Who else is tracking this shift from centralized to distributed AI infrastructure?

Before joining a new company, you should take a look at its corporate values to make sure they are aligned with your values, right?
Maybe, but not for the reasons you might think. According to an intriguing study from MIT of 1.2 million employees at 500 large companies:

"The analysis reveals that there is no correlation between the cultural values a company emphasizes in its published statements and how well the company lives up to those values in the eyes of employees. All of the correlations between official and actual values were very weak, and four of the nine — collaboration, customer orientation, execution, and diversity — were negatively correlated." 

The negative correlations are fascinating. If a company claims to be customer-focused, diverse, collaborative or results-oriented, it is LESS LIKELY to actually exhibit those values according to its own staff. 

This argument makes sense if we consider the perspective that corporate values are aspirational rather than descriptive of the current environment. In other words, companies seek what they don't currently have (or do).

Link to the full article is in the comments.

New HBR piece from Vipin Gopal, Tom Davenport, and Randy Bean makes an argument I've been watching play out across enterprise AI implementations:
The proliferation of C-level tech roles (CDO, CDAO, CAI Officer) is creating fragmentation exactly when organizations need unified execution.

Their proposed solution: The CDAIO (Chief Data, Analytics, and AI Officer).

What caught my attention:
-- 53% of companies have either appointed a Chief AI Officer, believe one is needed or are expanding existing data roles to include AI. That's a lot of organizational restructuring happening simultaneously.

-- Only a small fraction report positive P&L impact from gen AI so far. The authors argue this underscores the need for business-first AI leadership, not technology-first.

-- 42% of leading organizations now have AI/data leadership reporting to business or transformation leaders rather than IT. JPMorgan made their CDAIO part of their 14-member operating committee.

The uncomfortable truth buried in here: Many early CDO roles failed because expectations were unclear and ROI was hard to demonstrate when the mission focused solely on foundational data investments. 

The authors state "The failure of data initiatives to create commensurate business value has likely contributed to the short tenures of data leaders (to say nothing of doubts about the future of the role entirely). CDAIOs need to focus on business value from day one."

AI provides the channel to finally show the value of data investments. But it requires someone who can be both evangelist AND realist, inspiring the organization while quickly terminating projects that don't deliver returns.

The authors propose a 7-part mandate for CDAIOs:
1) Owning the AI strategy 
2) Preparing for a new class of risks 
3) Developing the AI technology stack for the company 
4) Ensuring the company’s data is ready for AI 
5) Creating an AI-ready culture 
6) Developing internal talent and external partner ecosystems 
7) Generating significant ROI for the company

The seven-part mandate they outline is ambitious. But after watching coordination problems kill promising AI initiatives, I think the consolidation argument has merit.

No offense to my Chief AI Officer friends out there, but my coauthors Vipin Gopal and Randy Bean and I think the job is best combined with a Chief Data and Analytics Officer role. Read this new article for our rationale and what we think the job should involve.
https://lnkd.in/esT4snwz

Open article: Why Your Company Needs a Chief Data, Analytics, and AI Officer by hbr.org

Why Your Company Needs a Chief Data, Analytics, and AI Officer

Sometimes I wonder if we overcomplicate AI strategy on purpose.

We draw elaborate frameworks, invent new acronyms, debate the merits of this platform versus that one. We talk about MLOps pipelines and model governance and responsible AI, and somewhere along the way, we forget what we were trying to solve in the first place.

I've seen teams spend months perfecting their infrastructure, only to realize the business problem was never clearly defined. Or worse, that nobody actually cared about the answer the model was supposed to provide.

Here's a radical thought: what if we started with the problem, not the solution? What if we spent more time with the people who actually use the output, and less time arguing about which cloud provider has the best GPU pricing?

The best AI projects I've seen are the ones where the tech fades into the background. Where the focus is on impact, not architecture diagrams.

Maybe it's time to get back to basics. Less noise, more signal.

What do you think?

This was true when I posted it 2 years ago and it still rings true today ;)

Chief Data, Analytics, AI Officers—it's not your fault; you were set up to fail

⚠️That’s the conclusion of a recent HBR article titled, Why Chief Data and AI Officers Are Set Up to Fail (June 2023), which everyone in a data leadership position should take a few minutes to read. 

💰Despite massive investments over the past decade in people + technology, and assurances to shareholders that “digital transformation” and “data-driven decision-making” initiatives were underway, senior corporate executives admit that only: 

1) 23.9% have a data-driven organization 
2) 20.6% have an established data culture 

🚫The other 80% of companies are beginning to blame their CDAIO for the lack of “insightful insights” and analytics-driven transformation. This is one reason for the high turnover in CDO / CAO / CAI roles I wrote about previously https://lnkd.in/gKYhQNws 

As the HBR article mentions:
“The CDAIO has been left on the hook for major projects that require huge investment and reach into all corners of the company, but often fail to deliver measurable benefit.

Even when they’ve done exactly what was asked of them, it can be hard to argue they’ve succeeded.

We believe that two factors largely led to this situation: the wrong focus and a lack of trust.”

📝 The continuously expanding scope of the CDO role has also contributed to the problem.
CDO Version 1 = primary focus was on data risk governance and regulatory compliance 
CDO Version 2 = CDO V1 responsibilities + data monetization expectations  
CDO Version 3 = CDO V2 + generative AI adoption 

❓What can CDAIOs do to fix the situation? 
The article suggests the following actions:

1) Make data everyone’s business - a sole CDAIO cannot be responsible for an entire organization’s data practices (governance + literacy).

2) Make business leaders champions for data projects - CDAIOs need to find a handful of leaders within the company who are ready and willing to invest in data & analysis with realistic expectations of what is possible. 

3) Review all data and AI investments to make sure funds are well spent - CDAIOs should concentrate their attention & budgets on must-have data projects with near-term business value instead of nice-to-have POC projects. 

4) Shift to an ecosystem mindset - CDAIOs need to cultivate relationships with vendors, universities and outside partners to build a successful analytics program and talent pipeline. 

5) Proceed with caution - be careful with new technologies like generative AI (ChatGPT) that bring new data-related risks to an organization. 

📍Here’s the link to the full HBR article: 
https://lnkd.in/gcU5wekH

Open article: Why Chief Data and AI Officers Are Set Up to Fail by social.koleaconsulting.com

Why Chief Data and AI Officers Are Set Up to Fail

I’ve lost count of how many times I’ve seen AI projects stall at the same place. 

The proof of concept works. The demo is slick. Everyone’s excited. 

Then reality hits. Integrating with legacy systems. Data that’s messier than anyone admitted. Stakeholders who suddenly have a hundred new requirements. 

It’s like running a marathon and tripping on the last mile. 

Here’s the thing nobody tells you: operationalizing AI is 80% negotiation, 20% code. 

You need translators, not just engineers. People who can bridge the gap between what the model can do and what the business actually needs. 

I’ve started to think of these folks as the unsung heroes of AI. The ones who keep the project alive when everyone else is ready to give up. 

If you’re building an AI team, don’t just hire for technical skills. Look for people who can navigate ambiguity, build consensus, and keep moving when the path isn’t clear. 

That’s how you get models out of the lab and into the real world.

No alternative text description for this image

A 55% smaller model. Same performance. Zero censorship. Here's how quantum physics made it possible and why it matters for your AI infrastructure costs.

Multiverse Computing just demonstrated something enterprise AI teams should pay attention to:

They took DeepSeek R1, one of the most capable open-source reasoning models, and compressed it by 55% using tensor networks borrowed from quantum physics.

The result performs nearly as well as the original while requiring significantly less compute.

Here's why this matters:
⚡ THE EFFICIENCY ANGLE
Most LLMs are bloated. Multiverse's approach identifies and removes redundancy with mathematical precision. That translates directly to lower GPU costs and energy savings.

🎯 THE CUSTOMIZATION ANGLE
The same technique that removed Chinese censorship can selectively add or remove other behaviors. Domain expertise. Safety guardrails. Brand-specific alignment.

🧭 THE STRATEGIC ANGLE
If you can compress and modify open-source models at the weight level, you're no longer dependent on API providers or original model creators for customization.

This is where AI infrastructure is heading: not just fine-tuning, but surgical model editing.

For enterprise teams running open-source models in production, the question becomes: are you optimizing what you deploy, or just accepting default weights?

Open article: Quantum physicists have shrunk and “de-censored” DeepSeek R1 by technologyreview.com

Quantum physicists have shrunk and “de-censored” DeepSeek R1

LLMs out of the box continue to be “often wrong but never in doubt.”

WRONG 60% OF THE TIME:
ChatGPT. Gemini. Grok. Perplexity. Copilot. DeepSeek 

Premium models, $20 to $40 per month, performed WORSE than free versions.⁵ Higher cost. Higher confidence. Higher error rates.

Columbia University's Tow Center tested eight AI search engines on the simplest possible task: Given a direct excerpt from a news article, identify the headline, publisher, date, and URL.¹

Collectively, they gave incorrect answers more than 60% of the time.²

Not on complex reasoning.
Not on nuanced analysis.
Not on predicting the future.

On citing their sources.

A basic Google search gets this right instantly.

The individual results:

Grok 3: 94% wrong
Gemini: 1 correct answer out of 200
ChatGPT: 67% wrong
Perplexity: 37% wrong (best performer)

But the error rate isn't the real problem.

The confidence is

ChatGPT gave incorrect information 134 times. It expressed uncertainty just 15 times.³ It never once said "I don't know."

Never. Once.

Grok 3 didn't just get sources wrong, it invented them. 
Out of 200 queries, 154 citations led to fabricated URLs.⁴ 
Pages that don't exist. 
Evidence that was never there.
They're not making mistakes.
They're manufacturing proof.
And here's where it gets absurd:

Now think about where these tools are being deployed.

Writing code. Summarizing contracts. Informing investment decisions. Drafting legal briefs. Generating medical information.

If they can't accurately cite a news article, the most basic intellectual honesty task, what else are they confidently fabricating?

The researchers didn't mince words: "Most of the tools we tested presented inaccurate answers with alarming confidence."⁶

This isn't a bug to be patched.

This is the architecture working as designed.

Large language models don't know things. They predict what sounds plausible. 

They've learned that confidence gets rewarded, accuracy be damned.

We're building critical infrastructure on foundations that fail 60-94% of the time
********************************************************************************
The trick with technology is to avoid spreading darkness at the speed of light.
Stephen Klein
Founder & CEO, Curiouser.AI
Instructor, UC Berkeley Haas School of Business
WeFunder (link in comments)

Sources:
¹ Columbia Journalism Review, Tow Center for Digital Journalism, "AI Search Has A Citation Problem," March 6, 2025 - Study analyzed 1,600 queries across 8 AI search tools
² Ibid. - "Collectively, they provided incorrect answers to more than 60 percent of queries"
³ Ibid. - ChatGPT "signaled a lack of confidence just fifteen times out of its two hundred responses, and never declined to provide an answer"
⁴ Ibid. - "Out of the 200 prompts we tested for Grok 3, 154 citations led to error pages"
⁵ Ibid. - Premium models "demonstrated higher error rates... tendency to provide definitive, but wrong, answers"
⁶ Ibid. - Direct quote from study authors

Your million-dollar AI project just failed.

The Product Manager said "we just need better predictions."

Let me tell you where this goes wrong.

I've led AI initiatives at various global companies and the the pattern is always the same:

Non-technical product managers create systematic failures through misaligned expectations, unrealistic timelines and poor data strategy.

Not because they're incompetent or unskilled.

Because they lack the technical vocabulary to bridge the divide.

Here's what actually happens:

Your business stakeholder says: "We need better customer recommendations."

Your PM nods and adds it to the roadmap.

Your data scientist asks: "Better how? What's the success metric?"

Silence.

So the AI engineer builds something. Ships it. 

And six months later, everyone's confused why it didn't deliver business value.

The real problem?

Your PM couldn't translate "better" into measurable technical requirements.

They approved the project without understanding:
- Data availability constraints
- Model performance trade-offs  
- Computational costs
- Ongoing maintenance needs

I've watched teams waste entire quarters this way.

Here's what changes the game:

Product managers who understand the full AI/ML lifecycle don't need to be data scientists.

But they do need technical literacy.

They need to know when a request is feasible versus impossible.

They need to set success metrics that align model optimization with business outcomes.

They need to communicate in both languages.

Because that translation layer?

That's where AI projects either create value or burn cash.

From my experience building and mentoring cross-functional AI teams: the organizations that succeed invest in technical foundations for their product managers.

The ones that fail keep treating AI like traditional software.

It's not.

And your product managers need to understand why.

Watson AI was a nice demo but it was very difficult to work with. This is an interesting post-mortem about the platform. 

The part about Watson Health’s failure is a good cautionary tale for organizations:

“Watson didn't fail alone.
It forced organisations to confront their own complexity too.
Hospitals discovered through Watson that:
- clinical notes are inconsistent
- workflows are fragmented
- ontology alignment is non-trivial
- "ground truth" in medicine is often contested

Watson revealed that organisations were nowhere near ready for the level of Al they claimed to want.
The data wasn't ready.
The workflows weren't ready.
The governance wasn't ready.”
The expectations certainly weren't ready.”

25+ years in AI | 39k+ Followers | MBA (AI) - BSc (Stats) - BCom (Finance) | Follow for real-world insights on AI, data & automation

IBM’s Watson Health wasn’t just one of the biggest AI failures - it was a $4 billion reminder that ambition without alignment is indistinguishable from delusion.

People keep flexing “I worked on Watson” like it’s a badge of honour.
It shouldn’t be.
Watson wasn’t early. Watson was wrong.

It was a great Jeopardy demo that collapsed the moment it touched the real world.

And it didn’t collapse quietly. We just stopped mentioning it loudly.

The trail it left behind was brutal:
- MD Anderson spent $62m and got nothing deployable
- Australian hospitals abandoned it after unsafe recommendations
- IBM sold off Watson Health in 2022 after years of losses

That isn’t innovation.
That’s a pattern.

People still parade Watson as proof of their AI credibility.
It should be the opposite.

Watson didn’t move the field forward.
The field simply moved past it.

While modern AI was built on deep learning, attention, transformers, self-supervision, scaling laws, RLHF & foundation models, Watson stayed anchored to symbolic scoring pipelines, brittle ontologies & offline indexing.

None of the breakthroughs that define today’s stack came from Watson.
None reference it.
None stand on its shoulders.

Watson didn’t evolve into the modern AI ecosystem.
It became a dead branch on the tree.
A shiny lab demo that never adapted to real-world conditions.

The story is not just a technical failure.
It’s also a story of what the field learned by watching the failure unfold.

And Watson certainly taught us something important.
It taught us what not to do.

It clarified the ingredients modern AI actually needs:
- end-to-end learning
- self-supervision
- scale
- probabilistic reasoning
- flexible representations
- uncertainty estimation
- RLHF-style alignment with human judgment

Watson failed at these, but in failing, it made them visible.
Its collapse did more to define the boundaries of the field than its success ever could.

Watson didn’t fail alone.
It forced organisations to confront their own complexity too.

Hospitals discovered through Watson that:
- clinical notes are inconsistent
- workflows are fragmented
- ontology alignment is non-trivial
- “ground truth” in medicine is often contested

Watson revealed that organisations were nowhere near ready for the level of AI they claimed to want.
The data wasn’t ready.
The workflows weren’t ready.
The governance wasn’t ready.
The expectations certainly weren’t ready.

So yes, Watson misled the market.
It overpromised.
It oversold.
It pushed the wrong paradigm.

But it also revealed the hidden complexity that everyone had been pretending was simple.

Watson is a reminder of what happens when:
- vendors oversell
- executives under-examine
- models overreach
- complexity is ignored
- narrative replaces reality

So the next time someone cites Watson as proof of their AI credentials, ask a simple question:
“What did it achieve in production?”

The silence is the real legacy of Watson.

Not the demo.
Not the hype.
The silence.

There’s a weird paradox in AI strategy right now. 

Everyone wants to talk about the next big thing. GenAI, autonomous agents and self-improving models. 

But when I look at the projects that actually deliver business value, they’re almost always... boring. 

Fraud detection. Churn prediction. Inventory optimization. 

The stuff that doesn’t make headlines but quietly saves millions. 

I get it, the hype is fun. But I’ve seen too many teams chase the shiny object, only to realize six months later that the ROI just isn’t there. 

The best AI leaders I know? They’re obsessed with the unsexy problems. They ask, "Where are we bleeding money? Where are we wasting time?" Then they build solutions that actually move the needle. 

It’s not glamorous. But it works. 

So here’s my question: Are we too distracted by the promise of AI to focus on the problems that matter?

Managing AI teams is not easier than managing software teams.

It's fundamentally different. I've learned that the advantage isn't reduced complexity; it's a reallocation of effort.

You spend less time micromanaging coding tasks. You spend more time validating outputs, maintaining model performance, and ensuring ethical deployment.

Once you introduce AI agents into the mix, the human-AI collaboration dynamics look nothing like traditional software development.

Here's what actually works:

1) Define specific responsibilities for each team member (human or AI) to eliminate confusion. Document decision-making authority for each role.

2) Create cross-functional collaboration structures. Weekly syncs between data scientists, engineers, and product teams ensure alignment.

3) Implement monitoring systems that track both technical metrics and business outcomes. Neither set alone tells the complete story.

The teams that succeed treat AI management as a distinct discipline requiring its own frameworks, metrics and leadership approaches.

Those that fail assume traditional software management practices will transfer directly. 

They won't.

The claim that managing AI teams is easier than managing traditional software teams misses the point entirely. AI tools reduce certain management over

Open article: Managing AI Development Teams: What Actually Changes by Courtlin Holt-Nguyen • 6 min read

Managing AI Development Teams: What Actually Changes

If your AI project sailed smoothly into production, you missed something.

Seriously.

Smooth projects are a myth.

Real innovation in AI isn’t born from pristine checklists or perfect plans.
It’s born from chaos, mistakes, and the stubborn refusal to give up when nothing works the way it should.

Let’s be brutally honest:
If you checked every box and nothing broke, you probably just automated yesterday’s problems.
You didn’t push boundaries.
You didn’t ask the uncomfortable questions.
You didn’t find out what your data was hiding.

I’ve led global teams through enough AI projects to know that you're going to be surprised sooner or later. 

Here’s what actually happens on the front lines:
→ Half your data is missing or wrong.

→ The business changes direction midflight.

→ Someone “forgets” to tell you about a critical process (until after go live).

→ Your model outperforms in test but falls flat in the real world.

Sound familiar?
Good.
You’re where the good stuff happens.

Here’s what to do when chaos takes over:
Embrace failure (quickly.)

↳ If your first model works perfectly, dig deeper.
Assume you missed an edge case or a data quirk.
Celebrate when things break early.
Build feedback loops, everywhere.

↳ The real world isn’t your Jupyter Notebook.
Get business users involved from day one.
Force your model into the wild early and often.
Let it get bruised.
Prioritize ugly prototypes over polished decks.

↳ That slick slide isn’t saving you when real data starts flowing.
The ugly, hacked together prototype that survives chaos will likely deliver ROI.
Document what doesn’t work.

↳ Every failed model, every rejected feature is gold.
This is your anti-checklist.
Share these stories with your team.
Celebrate the mess (publicly).

↳ If your team isn’t talking about the biggest disaster of the week, you’re not innovating.
Make it safe to share failures and weird discoveries.

Here’s my confession:
The most impactful AI solution my team delivered last year started as a Frankenstein prototype.
It broke three times in staging.
It made a business analyst question their career choices.
But it forced us to see the real problem.
It exposed the ugly truth in our data.
It worked because it failed first.

So, stop romanticizing perfect launches.
Stop chasing the fairy tale of “seamless production.”

You want ROI? Hunt for the chaos.

Remember:
If your AI project feels a little too easy, ask yourself, what did we overlook?

Tag someone who loves a messy project ;)

taplio

The Battle

Harpreet Sahota 🥑

Courtlin Holt-Nguyen

Followers

Posts

Engagements

Likes

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance

Performance ScoreCourtlin Holt-Nguyen's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Harpreet Sahota 🥑

Courtlin Holt-Nguyen

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Harpreet Sahota 🥑's LinkedIn Performance

Performance ScoreCourtlin Holt-Nguyen's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreHarpreet Sahota 🥑's LinkedIn Performance