“The only control or ‘guardrails’ that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT”
“WHOEVER WINS AI, WINS!”
EPISODE 001 / BEN LYDDANE
SPONSORED BY![BuildVision]()
01 / UNDERSTAND AIAI in the
HVAC world
for reps and
manufacturers
01 / UNDERSTAND AIAI Is a Five-Layer Cake
Jensen Huang’s framework · read from the foundation up
Infrastructure
Data centers, networking and cooling bring the chips together.
01 / UNDERSTAND AIToday’s conversation
From AI models to the equipment and work behind them.
01
Understand AI
Models, capabilities and the regulation debate
02
Follow the heat
Chips, data centers, power and cooling
03
Connect the tools
Agents, context and systems that work together
04
See it in practice
BuildVision and construction documents
01 / UNDERSTAND AIHow people feel about AI
Different surveys answer different questions.
01 / UNDERSTAND AIContext windows have grown
Context-window growth · 2018–2025
01 / UNDERSTAND AIThe Transformer
Tokens, attention and next-token prediction
Text pieces → relationships → next-token probabilitiesOpen live explainer ↗
01 / UNDERSTAND AIThe model landscape
Models and the companies behind them. Updated September 14, 2026.
OpenAI
GPT-6 Astra
$852BFunding valuationMar. 31, 2026
Anthropic
Claude Fable 5.1
$965BFunding valuationMay 28, 2026
IPO planned
Gemini 3.8 Flash
$4.21TAlphabet market capSept. 14, 2026
xAI / SpaceXAI
Grok 4.7
$250BxAI acquisition valueFeb. 2026
Release pending
01 / UNDERSTAND AIAI regulation: who sets the pace?
Calls to pace the frontier meet presidential pushback.
Frontier modelOutside reviewShared rules
“We Must Pace
the Frontier” ↗
Embedded outside evaluators.
Coordinated limits on unchecked progress.
A frontier AI standards body
Independent testing before release.
Government oversight.
1,386 signatories
Calling for international tools to pace automated AI development.
01 / UNDERSTAND AIFour lenses on an AI system
Each lens answers a different question about the same system.
Architecture
How the model computes
Thefanmovesair
AttentionFeed-forward
Next token through
Transformer
GPT-2 connects tokens through attention.
Modality
What the model can take in
“What equipment is in this image?”
IMAGE + TEXTOne model reads both inputs
Image + text
Gemini can interpret a photo and a question.
Function
What job the system performs
EQUIPMENT SCHEDULE
AHU-1 12,000 CFM
EF-2 1,500 CFM
| Tag | Airflow |
|---|---|
| AHU-1 | 12,000 CFM |
| EF-2 | 1,500 CFM |
Document extraction
BuildVision turns schedules into equipment data.
Deployment
Where the model runs
OLLAMAModel runs here
Local inference
Ollama runs supported models on your Mac.
01 / UNDERSTAND AICapability and compute
Measure the result, and the resources needed to deliver it.
Start with the task
A benchmark gives the system a defined problem and checks its answer.
01 / UNDERSTAND AIHLE: tools change the result
Humanity’s Last Exam · September 2026
01 / UNDERSTAND AIARC-AGI-1: accuracy and cost
Verified leaderboard, September 14, 2026
Static patterns
Infer a rule from example grids. Apply it to a new grid.
Accuracy and cost per task
01 / UNDERSTAND AIARC-AGI-2: a harder test
Verified leaderboard, September 14, 2026
Harder puzzles
A new task set tests gaps that earlier puzzles can miss.
Compare systems within the same benchmark.
01 / UNDERSTAND AIARC-AGI-3: interactive reasoning
Verified leaderboard, September 14, 2026
Interactive tasks
Explore the environment and learn how to win.
Standard harness
| GPT-6 Astra Max | 62.7% |
|---|---|
| Claude Opus 5 High | 30.2% |
| GPT-5.6 Sol Max | 7.8% |
Provider Adapter
99.9%GPT-6 Astra (High)
Different harnesses, different results.
01 / UNDERSTAND AITrends in AI
Data centers in space and Musk’s Terafab vision
Data centers in space
Solar-powered satellites could run AI workloads in orbit. Google’s Project Suncatcher is exploring the idea.
Musk’s Terafab vision
Bring logic, memory and advanced packaging together to make chips for Earth and orbit.
1 TW / yearStated target for annual chip output
01 / UNDERSTAND AIInside an ASML machine
Laser pulses on tin droplets create the light that prints advanced chips.
From tin to tiny circuits
- Laser pulses strike tiny tin droplets.
- The plasma emits 13.5 nm extreme ultraviolet light.
- Mirrors guide the light to pattern a silicon wafer.
01 / UNDERSTAND AIChatJimmy.ai: why it is so fast
Taalas builds one model into silicon, so each token needs less data movement.
NVIDIA GPU
Taalas HC1
Taalas: weights and compute together
Move the weights to the math
Each new token needs model weights from HBM memory. Moving them can limit speed.
Build the model into the chip
Fixed weights and compute sit together. The chip avoids repeated weight transfers from HBM.
HC1 / LLAMA 3.1 8B
17,000
tokens / second / user
Reported by Taalas
Less data movement.
Less waiting per token.
The tradeoff: one fixed base model. Fewer bits per weight save space but can reduce answer quality.
2.5 kW server6 nm53B transistors
02 / FOLLOW THE HEATOne Megawatt Is a Mechanical Problem
An illustrative steady-state IT load
1.0 MW
IT electrical load
≈
3.41 million Btu/h
284 refrigeration tons
284 refrigeration tons
02 / FOLLOW THE HEATThe $7 trillion buildout
Investment and electricity demand describe different parts of the same buildout.
02 / FOLLOW THE HEATWhere projects cluster
U.S. data-center construction · March 2026
02 / FOLLOW THE HEATWhat buyers need from suppliers
Reliability, specification compliance and delivery
02 / FOLLOW THE HEATThe handoffs that bring a data center online
Example: cooling stays available while one unit is offline.
01 / OWNER / OPERATOR
Define the need
Keep racks cool during maintenance.
02 / ENGINEER
Design the system
Specify backup capacity and controls.
03 / MANUFACTURER / SUPPLIER
Deliver equipment
Match the spec and delivery date.
04 / OWNER / OPERATOR
Data center online
Power and cooling keep the racks running.
A missed requirement can delay the day the GPUs go online.
02 / FOLLOW THE HEATThe shift toward liquid cooling
Cooling technology · 2020–2026 estimate
02 / FOLLOW THE HEATFollow the Heat, And Both Water Loops
Capturing heat at the chip is different from rejecting it outside
Cold plates absorb chip heat into the IT coolant loop.
02 / FOLLOW THE HEATData center cooling without evaporation
Keep the fluid in a closed loop and reject the heat to outdoor air.
YORK YDAM chiller ↗
Air cooled. Closed water loop.
No evaporative water use.
Closed circuit tower
Closed process loop.
No evaporation in dry mode.
Dry mode matters. A closed circuit tower still evaporates spray water in wet mode.
Year-round dry operation requires enough capacity for the climate and load.
02 / FOLLOW THE HEATGPU coolant temperatures
From chilled-water plants to warm-water GPUs. The system design sets the temperature.
02 / FOLLOW THE HEATSpecify the Temperature Interface
Facility-water classes are not universal GPU inlet temperatures
SELECTED ASHRAE FACILITY-WATER CLASSES
W27
80.6°F
W32
89.6°F
W40
104°F
W45
113°F
60°F120°F
Maximum facility supply temperature for each selected class.
A NAMED IT INLET EXAMPLE
113°F
45°C · NVIDIA Vera Rubin NVL72
Confirm coolant, flow and the supplier’s complete operating envelope.
02 / FOLLOW THE HEATWhen Does the Compressor Turn On?
Example: dry cooler + chiller plant, with no evaporative assist
DRY COOLING
IT INLET113°F
−CDU APPROACH5.4°F
−COOLER APPROACH9°F
Maximum outdoor dry bulb98.6°F
50°F 120°F
3.6°F below the dry-cooling limit
DRY: fans + pumps
ASSIST: compressor cooling
ASSIST: compressor cooling
Illustrative dry cooler + chiller plant. No evaporative assist or selection margin.
02 / FOLLOW THE HEATWhat the Compressor Actually Does
The compressor raises refrigerant pressure and temperature so heat can move outdoors
ILLUSTRATIVE ENERGY BALANCE
Heat absorbed100 kW
Compressor input+ kW
Heat rejected kW
The evaporator absorbs heat from the water loop.
03 / AGENTS & TOOLSAgents: from answers to actions
Construction documents become work your team can use.
“Get this bid ready for our team.”
A project folder, ready to work
- Read the RFQFind the bid date, scope and plan links.
- Open the connected servicesDownload the drawings, schedules and specs.
- Organize the projectSave the files and flag anything missing.
THE WORK YOU GET BACKOrganized project files + an intake record
BUILDVISION
“Pull the equipment requirements.”
An equipment schedule with evidence
- Find the right pagesLocate equipment schedules and governing specs.
- Extract and normalizeCapture airflow, pressure and cooling capacity.
- Keep the source attachedLink each field to the original row or passage.
THE WORK YOU GET BACKStructured equipment data + source links
BUILDVISION
“Prepare a selection from this request.”
Selection inputs become equipment
- Read the requestCapture duty, application and project location.
- Use selection softwareEnter the requirements into the connected tool.
- Return a reviewable resultBring the selection back to the rep.
THE WORK YOU GET BACKAn equipment selection ready for review
BUILDVISION
“Build the submittal package.”
The package comes together
- Start with the selectionUse the selected equipment and project duty.
- Gather the supporting pagesCollect product data, curves and dimensions.
- Assemble the deliverablePrepare the package for technical review.
THE WORK YOU GET BACKA submittal package ready for review
BUILDVISION
03 / AGENTS & TOOLSThe agent is more than the model
A model chooses the next step. Connected software carries it out.
Project contextDocuments + prior work
ModelChooses the next step
Connected toolsRead, write and run
Result + feedbackCheck, continue or stop
One assignment.
Several real actions.
- Context tells it what mattersThe current drawings, specifications and your instructions.
- The model chooses a toolFind a page, extract a field or run a selection.
- The application runs the actionAPIs, connectors or browser controls do the work.
- The result drives the next stepInspect the output. Retry, ask or finish.
You define the outcome and access.The team reviews technical judgments and commitments.
03 / AGENTS & TOOLSOpenClaw and personal agents
The growth of personal agents
03 / AGENTS & TOOLSModel Context Protocol
A shared connection between AI applications, tools and data
03 / AGENTS & TOOLSConnect the assistant to real systems
An AI application can use MCP connectors to work with tools and data in other systems.
AI APPLICATION
Your AI assistant
A model, your conversation
and enabled connectors.
Tool callsMCPResults
Model Context
Protocol
Meeting recorders & notes
Search transcripts and decisions Direct MCP servers
Calendars
Find meetings and create events Via Zapier MCP
Customer records
and opportunities
Look up records and take action Salesforce or Zapier MCP
One request
“Read my last customer call, check the next meeting,
and prepare a Salesforce update.”
Available connectors depend on the AI app or API, your plan and account permissions.
03 / AGENTS & TOOLSChoose one workflow. Make it useful.
A trigger starts the workflow. Claude uses MCP tools to read and act.
Parts inbox
Trigger: a new email
Outlook
Read the parts email
A new customer request
hits the inbox.
Outlook MCP→via Zapier
Claude skill
Identify the request type
Use your categories
and example requests.
Outlook MCP→via Zapier
Outlook
Apply a category label
Quote, order status,
return / warranty, technical help.
Weekly sales plan
Trigger: every Sunday night
Salesforce CRM
Review account status
Read each rep’s accounts
and active projects.
CRM MCP→via Zapier
Claude skill
Choose weekly targets
Target projects and accounts
for each salesperson.
Outlook MCP→via Zapier
Outlook
Email every salesperson
Send their priorities
for the coming week.
Same pattern with Gmail MCP: read an email and apply a Gmail label.
04 / BUILDVISION EXAMPLESAn example from my work
BuildVision bid desk automation
The video could not load. Please reload the page and try again.
04 / BUILDVISION EXAMPLESThe equipment-sales workflow
The work around a piece of equipment
04 / BUILDVISION EXAMPLESBuildVision example: equipment extraction
LaGuardia Community College Library Renovation Phase II
04 / BUILDVISION EXAMPLESBuildVision example: drawing takeoff
Symbols, schedules and source drawings
04 / BUILDVISION EXAMPLESBuildVision example: manufacturer × category
Analytics across selected projects