back

local AI beyond basic applications

A side effect of keeping up with AI since GPT 3.5 Turbo is that it becomes easy to forget how far things have moved.

There is still a long way to go, but the progress is already strange. Models that would have felt close to frontier level not long ago can now run on consumer laptops, desktops, and phones. Some local models today are competitive with frontier models from recent history, and in specific domains, small specialized models can outperform much larger general models.

That changes what local AI means.

The usual arguments for local inference are privacy, latency, cost, and self hosting. Those are all valid. It is useful to have a model that does not send your data to a cloud API. It is useful to avoid paying per token. It is useful to run inference without depending on someone else's server.

But the more interesting argument is that local AI gives machines intelligence even when the internet is unavailable, unreliable, expensive, or too slow.

That matters because many future AI applications will not look like chatbots.

A chatbot can fail gracefully. If Claude, ChatGPT, Gemini, or a hosted model API goes down, the user is annoyed, but the failure is usually not catastrophic. The user waits, refreshes, or tries again later.

But if the AI is inside a robot, a vehicle, a medical device, a home assistant, an industrial machine, or a tool that people rely on in the physical world, the failure mode is very different. A caretaker robot cannot stop working because of a cloud outage. A warehouse robot cannot pause every time WiFi drops. A surgical assistance system cannot wait for an API response before deciding how to stabilize a camera or warn a surgeon.

Local AI matters because physical systems need autonomy at the edge.

the first wave of local AI

The first wave of local AI is mostly about replacing cloud chat.

Cloud chatbot:
   user prompt
        ↓
   API call
        ↓
   hosted model
        ↓
   text response


Local chatbot:
   user prompt
        ↓
   model on device
        ↓
   text response

This is useful, but it is still a basic application. The interface is the same. The output is the same. The only difference is where the model runs.

That is a good starting point, but it is not the most important version of local AI.

The more interesting version is when local models become part of the operating layer of a machine.

Local AI as a chatbot:
   text in
   text out


Local AI as infrastructure:
   text, audio, vision, sensor data in
   decisions, actions, controls, code, geometry, or commands out

Once the output is not just text, local AI becomes much more powerful. It becomes a way for machines to perceive, reason, and act without constantly depending on the cloud.

why local inference matters

Local inference has four obvious advantages.

1. Privacy
   Data can stay on the device.

2. Latency
   The model can respond without a network round trip.

3. Reliability
   The system can keep working without internet.

4. Control
   The user or developer owns the runtime.

Privacy is the most commonly discussed. If a model runs locally, private documents, medical data, CAD files, source code, camera feeds, and microphone input do not need to leave the machine.

Latency is also important. Even a fast API still has network overhead. For interactive tools, robots, and real time systems, the difference between local inference and cloud inference can change the product experience.

Reliability may be the most underrated. A system that depends on a hosted model inherits the uptime of that model provider, the network, the user's WiFi, and every layer between the device and the API.

Cloud dependent system:

   device
     ↓
   local network
     ↓
   internet provider
     ↓
   cloud region
     ↓
   model API
     ↓
   response


Local system:

   device
     ↓
   model
     ↓
   response

That shorter dependency chain matters. The more physical the application becomes, the more important it is.

A writing assistant can rely on the cloud. A robot should not need to.

local AI and robotics

Robotics is one of the clearest reasons local AI matters.

A robot is not just a computer with a chat window. It has sensors, motors, actuators, cameras, microphones, constraints, and safety boundaries. It exists in the world. It needs to understand its environment and act within it.

Language models are useful for robots because they provide a communication interface. A person can tell a robot what they want in natural language instead of using buttons, menus, or rigid commands.

Human:
   "Pick up the red cup and put it on the counter."

Robot:
   parse instruction
   identify objects
   plan motion
   execute task

But language alone is not enough. A robot needs vision, audio, spatial understanding, and action generation. It needs to map human intent into physical behavior.

The compelling version of local AI for robotics is multimodal and action oriented.

Inputs:
   language
   camera frames
   audio
   depth
   force sensors
   joint positions
   environment state

Model output:
   plan
   motion command
   tool command
   warning
   correction
   control policy

This is where local inference becomes more than a privacy feature. The robot needs low latency understanding. It needs to keep operating if the network drops. It needs to make decisions near the hardware.

A caretaker robot is a simple example. If it helps an elderly person move around a home, reminds them to take medication, detects falls, or calls for help, it cannot be out of commission because of a model outage. It should be able to do the core tasks locally, then use the cloud for optional upgrades, search, remote monitoring, or heavier reasoning.

The same principle applies to many robotic domains:

Hospital robots:
   transport supplies
   assist nurses
   monitor patients
   clean rooms

Home robots:
   help with chores
   support elderly care
   interact naturally with people

Industrial robots:
   inspect equipment
   perform repetitive manipulation
   react to changing environments

Surgical robots:
   assist with camera control
   stabilize motion
   identify anatomy
   support procedure specific workflows

In all of these cases, cloud intelligence may be useful. But local intelligence is the foundation.

the output should not only be text

Most AI products still treat the model as something that outputs text, images, or maybe code. That is a limited view.

For local AI to become important in physical systems, models need to output actions.

In software, an action might be:

write file
run command
edit spreadsheet
open browser
query database
generate code

In robotics, an action might be:

move arm to position
adjust gripper force
rotate camera
follow path
stop motion
avoid region
hand control back to human

This changes the model from an assistant into a control layer.

The simplest architecture is to split the system into two parts:

┌────────────────────────────┐
│ Understanding model        │
│                            │
│ Inputs:                    │
│   language                 │
│   vision                   │
│   sensor data              │
│                            │
│ Output:                    │
│   plan                     │
└──────────────┬─────────────┘
               │
               ▼
┌────────────────────────────┐
│ Action model or controller │
│                            │
│ Inputs:                    │
│   plan                     │
│   robot state              │
│                            │
│ Output:                    │
│   motor commands           │
│   tool commands            │
└────────────────────────────┘

This separation is easier to reason about. One model understands the task. Another system executes it safely.

But for latency, some systems may eventually combine more of this stack.

┌────────────────────────────┐
│ Multimodal action model    │
│                            │
│ Inputs:                    │
│   language                 │
│   vision                   │
│   robot state              │
│   sensor data              │
│                            │
│ Outputs:                   │
│   motion instructions      │
│   control commands         │
│   safety decisions         │
└────────────────────────────┘

The right architecture depends on the domain. In low stakes environments, direct action models may be acceptable sooner. In high stakes environments, the model will likely sit inside a larger safety system with validators, constraints, simulators, and human override.

surgical robotics as a long term example

Surgical robotics is a useful example because it shows both the promise and the difficulty of local AI.

A far future surgical robot might take in vision, patient data, instrument state, force feedback, imaging, and procedure context, then output precise control instructions for robotic arms and end effectors.

Surgical robot loop:

   camera feed
   instrument positions
   force feedback
   patient data
   procedure state
          ↓
   surgical model
          ↓
   motion plan
   warnings
   camera movement
   instrument commands
          ↓
   robot arms

That is not a near term product. Surgery is too complex, too regulated, and too high stakes for full autonomy to appear all at once.

But the direction is important. A surgical system does not need to become fully autonomous immediately for local AI to matter.

Near term uses are more realistic:

camera stabilization
anatomy identification
tissue tracking
no go zone warnings
suture guidance
instrument motion smoothing
procedure checklists
skill assessment
automatic video annotation
training feedback

Many of these features benefit from local inference. They need to run close to the operating room hardware. They need low latency. They may involve sensitive data. They should not depend entirely on a remote model call.

Over time, these assistance features can become a data collection and training pipeline. Every procedure can generate structured data:

video
instrument trajectories
surgeon actions
procedure phase labels
errors and corrections
patient context
outcomes

That data can train better surgery specific models. Those models can assist more effectively. Better assistance can make the robot more useful. More usage creates more data. The loop compounds.

robotic assistance
        ↓
procedure data
        ↓
surgery specific models
        ↓
better assistance
        ↓
more adoption
        ↓
more data

This is why local AI matters beyond chat. It can become the intelligence layer for machines that collect data, act in the world, and improve over time.

local AI for CAD

A nearer term example is local CAD generation.

CAD is a good test case because it is not as high stakes as surgery, but it still has a real world output. The model cannot only produce a nice sentence. It has to produce geometry.

The workflow looks like this:

User:
   "Make a wall mounted phone holder with two screw holes."

Model:
   generates CAD code

Renderer:
   executes the code

Viewer:
   displays the 3D object

This is the idea behind C3D, a project I am working on.

C3D is a local text to CAD editor. It uses a fine tuned model to generate CadQuery code, runs the code locally, renders the object, and opens it in a browser based 3D viewer.

C3D pipeline:

   natural language prompt
          ↓
   local CAD generation model
          ↓
   CadQuery code
          ↓
   local renderer
          ↓
   3D object
          ↓
   viewer and iteration

This is a much smaller problem than autonomous robotics, but it points in the same direction. Local AI is not just answering questions. It is generating artifacts that can become physical.

A CAD model can be printed, machined, simulated, edited, or used inside a larger engineering workflow. The model output becomes part of a physical design process.

That is a different category from a chatbot.

the pattern

The general pattern is:

Basic local AI:
   local model replaces cloud chatbot


More interesting local AI:
   local model becomes part of a tool


Most interesting local AI:
   local model becomes part of a machine

The first category is already useful. The second category is where developer tools, creative tools, CAD tools, personal operating systems, and offline agents become interesting. The third category is where robotics, medical devices, industrial machines, vehicles, and other physical systems start to change.

The shift is from conversation to control.

Conversation:
   user asks
   model answers


Control:
   system observes
   model reasons
   system acts

That shift requires more than a model. It requires sensors, validators, safety systems, user interfaces, data pipelines, and domain specific training. But local inference is one of the enabling pieces.

cloud and local will coexist

Local AI does not mean cloud AI disappears.

Cloud models will remain useful for heavy reasoning, large context, search, training, and orchestration. Frontier models will probably stay ahead of local models in raw general capability for a long time. When a task needs the most capable model available, and latency and privacy are not the binding constraints, the cloud is the right place to run it.

Local models are valuable for different reasons. They are fast, private, always available, and close to sensors and hardware.

Local model:
   fast
   private
   always available
   close to sensors and hardware


Cloud model:
   larger
   more general
   easier to update
   better for heavy reasoning

These are not competing answers to the same question. They are suited to different work. A research task that needs a large context window and the strongest possible reasoning belongs in the cloud. A perception loop that has to run every frame, next to a camera, with no guaranteed network, belongs on the device. Neither one makes the other unnecessary.

Cloud AI is not going away, and local AI is not a cheaper imitation of it. Both will matter, because the range of useful AI applications is wider than either one can cover alone.

The point is not that everything should be local. The point is that the core loop of important tools should not always require the cloud.

conclusion

Local AI is often discussed as a cheaper or more private version of cloud AI. That is true, but it is not the full story.

The more important idea is that local AI lets intelligence live inside tools and machines. It allows software to keep working without WiFi. It allows robots to respond with low latency. It allows private data to stay near the user. It allows models to generate not just text, but actions, code, geometry, commands, and control signals.

The first local AI applications are chatbots. The next ones will be tools. After that, machines.

Robotics is the clearest long term example. A robot needs language, vision, audio, spatial understanding, and action. It cannot depend entirely on a remote API. Surgical robotics is an extreme version of that idea, where local intelligence could assist with perception, training, camera control, and eventually narrow forms of autonomy.

CAD is a nearer term version. A local model that turns language into CAD code is already useful because it produces something that can become physical. That is why projects like C3D are interesting. They test whether local models can move beyond conversation and into creation.

The future of local AI is not just an offline chatbot.

It is intelligence embedded into the tools and machines around us.