The DHI Blog.
Perspectives on edge AI, video privacy, and deploying safety analytics on the cameras you already have. No hype, just how this actually works.
Ask the model less
A new tracking paper reformulates a hard vision task as a yes or no question instead of free text generation. DHI's own alert verifier made the same bet, and the paper shows us the next step we have not taken.
Paying full price for every frame is a choice
Two independent papers published three days apart both argue for spending compute unevenly across a video stream instead of uniformly. DHI arrived at a coarser version of the same idea from an 8 GB memory budget.
When your verifier stops looking
A new paper finds vision-language models often answer without using the image evidence in front of them. That is a direct risk for any system, DHI's included, that uses a VLM as a verification gate.
The guardrails are arriving through bargaining, not legislation
More than 175 union contracts now contain AI guardrails, most aimed at keystroke and screenshot monitoring. The distinction between that and camera-based monitoring will not hold, and it should not.
Sixteen AI bills, and one of them names the bathroom
California's legislature passed 16 AI bills and 8 privacy bills in 2026, none of them law yet. One would ban workplace surveillance in bathrooms, and two would turn AI verification into a licensed function.
The models are good enough. The plumbing is the product.
Ambarella and Capgemini just built an integration practice, not a research lab. That is a tell about where physical AI deployment actually breaks, and it matches what we see installing DHI one camera at a time.
The next people to ask about your cameras will be underwriters
Verisk just built a database to help insurers price data center risk. The same logic that produces that database eventually asks what condition a facility is in, continuously, not just at the last inspection.
The government wrote down the risk itself
A federal privacy assessment for a system that fuses video, facial biometrics and plate data contains a sentence most vendors would never publish. It also describes exactly where the risk in computer vision has moved.
A fab makes silicon, not a deployable edge system
Tesla and SpaceX are putting an initial 16.8 billion dollars into a Texas fab, partly for edge inference chips. One number in that announcement does not survive scrutiny, and the harder problem is not wafer supply anyway.
Fully engulfed on arrival
A seafood warehouse in Galveston was already fully involved when the first truck pulled up. The cause is still unknown, but the timeline itself is worth sitting with, because it is the same timeline on most industrial sites.
The problem was never the camera
Cities around Houston are exiting their license plate reader contracts. We build license plate recognition, so this one deserves a straight answer about architecture rather than a defence of anyone's product.
OSHA's warehouse program just got longer and narrower at the same time
OSHA renewed its warehousing emphasis program for five years instead of three. Most coverage called it an expansion. The directive itself says something more specific, and the difference matters if you are the one being inspected.
Our thermal model's evaluation was wrong twice before it told us the truth
A thermal perception backbone looked broken, then looked useless, then turned out to be neither. The bug was never in the model. It was in the test we used to judge it, and in a data pipeline that let RGB photos into a corpus we called thermal.
DHI joins Rice University's 2026 Summer Venture Studio
DHI is one of nine ventures selected for the 2026 Summer Venture Studio at Rice University's Liu Idea Lab. Here is what we are building with it, and why edge AI on the cameras you already own is the thing we came to prove.
The camera backlash is a trust problem, not a camera problem
On the Fourth of July, residents in Lubbock protested to have their city's license-plate cameras removed. The lesson for everyone building safety tech is not about the hardware. It is about what you collect, where it goes, and who controls it.
The real divide in camera AI isn't capability. It's control.
A US city wants facial recognition on its buses, and the fight that broke out misses the point. The line between safety and surveillance was never the technology. It is who controls it, where it runs, and what it is pointed at.
Every warehouse has cameras. Almost none have early warning.
A Los Angeles cold-storage warehouse burned for eight days. The hard lesson for safety teams isn't about fire codes: it's the gap between a camera that records an incident and one that catches it early enough to matter.
Who can turn your AI off?
In June 2026, the most powerful AI models on the market were gated, pulled, and partially reinstated by parties their customers don't control. For most software that's a policy story. For safety-critical systems, it's an ownership question.
Nobody answers the alarm anymore
When most camera alerts are shadows, rain, and headlights, your team learns to ignore all of them, including the one that mattered. Why alarm fatigue is a detection problem, not a discipline problem.
You don't need new cameras
Rip-and-replace is what kills safety projects. How to add AI to the CCTV and VMS you already run, starting with a single camera.
Your "edge AI" might just be the cloud with extra steps
A cloud provider just discontinued its edge-vision product. Here's what it reveals about how most camera AI is really built, and the one question that separates real edge from rented edge.
Where does your video actually go?
The question that quietly stalls AI camera projects, and why keeping footage on-site turns privacy from a liability into the default.
Use the guide, then validate it on your cameras.
Don't let the guide be the end of it. Take the checklist, a clear deployment path, and a direct line to ask implementation questions.
The checklist is built for operators evaluating a live pilot in the next 30 days.
Request a demo
See the flow on a real operating scenario and scope a pilot around one facility or corridor.
See deployment architecture
Review camera ingest, edge inference, alert routing, and what stays on-premises.
Get the implementation checklist
Download the deployment checklist buyers use before green-lighting an industrial AI pilot.
Talk to an engineer
Bring camera count, VMS constraints, latency expectations, and privacy requirements to a technical review.