HCC 3030 · Week 2
AI Technologies 101:A Map of the Field
Rules, search, machine learning, neural networks, foundation models, robotics, and the economy underneath them.
AI
data
compute
learning
action
This lecture maps the field before zooming in. Students should leave able to distinguish rule-based AI, search and planning, machine learning, deep learning, reinforcement learning, foundation models, and embodied systems.
We will then connect those technical approaches to the products, markets, labor, infrastructure, and institutions around them. The details will change quickly, but the underlying questions are durable.
[Sources]
Opening question
Which of these systems count as AI?
Route planner Fraud score Recommender Industrial robot Chatbot
All five can count. They solve different problems with different combinations of rules, search, learning, and control.
Take a quick vote on each example. Students may reserve the word AI for chatbots or humanlike behavior. Use the disagreement to surface the fact that AI is an umbrella term for many kinds of systems.
A route planner may rely on graph search. A fraud detector may use supervised learning. A recommender may combine ranking models and business rules. A robot must perceive and control. A chatbot may use a foundation model plus retrieval and tools.
[Sources]
Start with the whole stack
Every AI output sits on layers of technical and economic choices.
01
Data What examples, labels, records, media, and feedback shaped the system?
02
Model What architecture and training objective turn inputs into outputs?
03
Compute Which chips, data centers, energy sources, and networks make it run?
04
Product What context, tools, defaults, permissions, and safeguards surround it?
05
Market Who owns the stack, sets the price, captures value, and bears the risk?
This is the organizing picture for the lecture. Most public discussion collapses these layers into “the AI.” That makes it hard to diagnose failure or understand why two products using similar models behave differently.
Technical architecture and market structure interact. A long context window changes inference cost. A closed API changes who can inspect or adapt the model. An agent with paid tools creates costs and dependencies that a text-only model does not.
[Sources]
The AI umbrella
AI is a field of approaches, not a synonym for language models.
Explicit structure
Rules & knowledge Represent facts, constraints, and logic so a system can infer conclusions.
State spaces
Search & planning Explore possible moves or action sequences to reach a goal.
Examples
Machine learning Fit patterns from data to classify, predict, rank, or detect.
Representations
Deep learning Use multilayer neural networks to learn useful features at scale.
Feedback
Reinforcement learning Learn a policy for sequential decisions from rewards and consequences.
World interaction
Robotics & embodied AI Connect perception, planning, and control to physical systems.
Walk across the map from systems with explicit human-authored structure toward systems that learn more of their internal representation from data and feedback. This is not a timeline in which each new family replaces the old one.
Hybrid systems are normal. A warehouse robot can use a neural vision model, a symbolic task planner, a path-search algorithm, learned control, and hard safety rules in one product.
[Sources]
Three ways to produce intelligent behavior
Different paradigms put the intelligence in different places.
01
Reason over structure Humans define states, rules, goals, constraints, or a knowledge base.
Expert systems · theorem proving · graph search · scheduling
02
Learn from data An algorithm fits parameters that make useful predictions on examples.
Regression · trees · neural networks · recommenders
03
Optimize through feedback An agent chooses actions, observes consequences, and improves a policy.
Game playing · robotics · resource control · personalization
Classical AI makes much of the structure explicit. Machine learning shifts more of the behavior into parameters estimated from data. Reinforcement learning focuses on sequential choice and delayed consequences.
The boundary is porous. Search can use a learned heuristic. A learned model can operate inside a rule-governed workflow. A reinforcement-learning agent can be trained in a simulator built from explicit physics.
[Sources]
Capabilities are not roles
Modern AI systems can sense, predict, generate, decide, and act.
Perceive Recognize patterns in text, images, audio, video, or sensor data.
Interpretation
Classify Assign a label, category, state, or risk group.
Categorization
Predict Estimate an outcome, value, demand, or future condition.
Forecast
Rank Order cases, options, documents, or people by a score.
Attention
Generate Create language, images, audio, video, code, or plans.
Proposal
Act Call tools, send messages, transact, allocate, or control.
Execution
Students often use “AI” to mean a chatbot. This capability map restores the wider field. Classification, ranking, forecasting, perception, and control remain central even when generative AI receives most of the attention.
A model that can produce a plan does not automatically have permission to execute it. Authority comes from tools, credentials, product design, and policy.
[Sources]
Machine learning in one sentence
The system learns a mapping from examples instead of receiving every rule by hand.
Input
What the system can observe
Pixels, tokens, sensor readings, records, clicks, or prior actions.
→
Output
What the system is trained to produce
A class, number, ranking, action, or probability distribution over possible next tokens.
Machine learning does not remove rules. It changes where the rules appear. Designers still choose the task, representation, objective, training data, evaluation metrics, thresholds, and deployment conditions.
A learned mapping can be extremely complex, but it is still optimized against an objective. If the objective is an imperfect proxy for the real goal, the system can improve its score while making the application worse.
[Sources]
Where the learning signal comes from
The training setup determines what patterns a model can learn.
Supervised
Inputs paired with labels or targets
Spam detection · price prediction · diagnosis support
Unsupervised
Structure inferred without target labels
Clustering · anomaly detection · dimensionality reduction
Self-supervised
Targets created from the data itself
Next-token prediction · masked inputs · representation learning
Reinforcement
Rewards from actions and consequences
Control · games · sequential optimization
Supervised learning uses desired answers. Unsupervised learning looks for structure without a supplied target. Self-supervised learning manufactures prediction tasks from raw data, which makes it possible to train on enormous unlabeled corpora. Reinforcement learning learns from consequences over time.
These categories describe training regimes, not product categories. One deployed system can contain models trained in several ways.
[Sources]
The machine-learning workflow
Most ML failures begin before the training loop.
01 Frame Define the decision, target, baseline, and cost of error.
02 Represent Choose examples, labels, features, sampling, and time window.
03 Split Separate training, validation, and test data before tuning.
04 Fit Train candidates and tune only with permitted information.
05 Evaluate Compare with baselines across groups and deployment conditions.
06 Monitor Watch drift, failures, feedback loops, cost, and human response.
Leakage: the model receives information during development that would not exist at the moment of real-world prediction. The score looks excellent because the test is no longer honest.
Use a concrete example such as predicting which students may need support by week four. A final course grade cannot be a feature because it occurs after the prediction point. Randomly splitting repeated records from the same student can also leak identity-specific patterns.
The validation set supports model and threshold choices. The test set estimates final performance after those choices. In forecasting, time-aware splits are usually more credible than random splits.
[Sources]
The model toolbox
Model choice should follow the data, task, and operating constraints.
There is no universal ranking of model families. A boosted-tree ensemble can outperform a neural network on many tabular tasks. A logistic regression can be the right production model when data is limited, decisions need explanation, and a simple model already clears the performance requirement.
Compare accuracy, robustness, latency, memory, calibration, interpretability, update frequency, and maintenance burden, not only a single benchmark score.
[Sources]
Classification is a decision pipeline
A threshold turns a model score into a real consequence.
Decision threshold = 0.70
Lowering the threshold catches more positives but usually creates more false alarms.
Precision Of the cases flagged, how many were actually positive?
Recall Of the real positives, how many did the system catch?
Specificity Of the real negatives, how many did it leave unflagged?
Calibration Does a score of 0.70 correspond to roughly a 70% event rate?
Classification models often produce scores or estimated probabilities. Product teams choose the threshold. That means precision and recall are partly policy choices, not fixed properties of the model.
Ask which error is worse for a cancer-screening tool, a fraud hold, or a content moderation queue. The answer changes across applications and groups. Calibration matters when people interpret scores as risks.
[Sources]
Not every ML problem is classification
Different task families demand different evidence.
↗
Regression & forecasting Predict a number or future value. Compare MAE or RMSE against a credible baseline and respect time order.
Can it beat “same as last week”?
◎
Clustering Group cases by similarity without target labels. A cluster is a mathematical pattern, not automatically a meaningful social category.
Does the grouping remain useful and stable?
◇
Dimensionality reduction Compress or visualize high-dimensional structure. The projection can distort distance and neighborhood relationships.
What information did the projection discard?
!
Anomaly detection Find unusual cases when labeled failures are rare. Unusual does not mean wrong, harmful, or fraudulent.
Who investigates the alert?
For regression, show why a lower error is only meaningful relative to a baseline, horizon, and unit. In time series, evaluate on the future rather than a random mixture of past and future records.
For unsupervised methods, emphasize that the algorithm will return structure even when the result has little domain meaning. Analysts supply names and interpretations, so validation must include domain evidence.
[Sources]
Sequential decisions change the problem
Reinforcement learning optimizes behavior through delayed feedback.
Environment
Agentpolicy π(a|s)
action at →
← state st+1 , reward rt
Credit assignment Which earlier action caused a later outcome?
Exploration When should the agent try an uncertain action?
Reward design Will optimizing the proxy produce the behavior we actually want?
Safety How can the system learn without causing unacceptable harm?
Reinforcement learning differs from ordinary supervised prediction because actions affect what happens next. The agent learns a policy, a rule for choosing actions from states, to maximize expected cumulative reward.
Delayed effects create credit-assignment problems. Exploration creates risk. A misspecified reward can be optimized in unintended ways. These issues matter in games, robotics, recommender systems, resource allocation, and post-training of foundation models.
[Sources]
Two phases, two different problems
Training creates the model. Inference uses it.
Training
How should the parameters change?
Input Large datasets and objectives
Work Forward pass, loss, gradients, updates
Cost Large, concentrated, paid up front
Result A model checkpoint
Inference
What output should this model produce now?
Input A prompt, image, record, or sensor stream
Work Forward passes and optional tool calls
Cost Small per request, repeated at scale
Result A prediction, response, or action
Keep training and inference separate. A deployed model usually does not update its core parameters during every conversation. It may store context or retrieve records, but that is not the same as retraining.
The economic distinction matters. Training may require a large capital commitment. Inference turns use into a recurring cost, so latency, caching, batching, model size, and output length become product decisions.
[Sources]
What “learning” looks like mechanically
Training repeats a simple loop at enormous scale.
01 · FORWARD
Make a prediction Pass examples through the current model.
ŷ = fθ (x)
02 · COMPARE
Measure error Use a loss function to score the mismatch.
L(ŷ, y)
03 · BACKWARD
Compute gradients Estimate how each parameter affected the loss.
∇θ L
04 · UPDATE
Adjust parameters Take a small step intended to reduce future loss.
θ ← θ − η∇L
Explain the symbols without turning this into a calculus lecture. The model is a function with parameters θ. The learning rate η controls the size of an update. Backpropagation efficiently assigns credit or blame to parameters throughout the network.
Optimization finds parameters that reduce the chosen loss on observed examples. It does not prove that the model learned the intended concept or will behave well in a new environment.
[Sources]
The real test is generalization
A model can fit its training world and fail in ours.
Performance depends on where the model is used.
Populations change. Sensors change. Language changes. Incentives change. People adapt to the model.
Did we test the conditions that matter here?
Generalization is the ability to perform on relevant cases beyond the exact training examples. Distribution shift occurs when the deployment environment differs from the data used to train or evaluate the model.
Shift can be technical, social, or created by the system itself. A ranking changes what receives attention, which changes what gets recorded. The model can therefore help create its own future data.
[Sources]
Neural networks learn representations
Layers transform raw input into features useful for the objective.
The network does not store a human-readable rulebook.
It distributes learned patterns across many parameters and intermediate representations.
Early layers often capture simpler regularities. Later layers combine them into task-relevant structure. Interpretability is possible, but rarely complete.
A neural network composes many parameterized transformations. During training, the network learns internal representations that help minimize the objective.
Avoid claiming that one node cleanly represents one concept. Features are often distributed, context-dependent, and entangled. Mechanistic interpretability can reveal meaningful structure, but the full behavior of frontier models remains difficult to explain.
[Sources]
Now zoom in on one branch
Foundation models are one powerful part of AI, not the whole field.
They combine deep learning, self-supervised pretraining, post-training, retrieval, tools, and product design. Language models are the most visible example.
Deep learning Scale Generative modeling Multimodality Tools
Pause and locate the next section on the map. Foundation models are deep neural networks pretrained on broad data and adapted to many downstream tasks. They are unusually general, but they do not replace search, optimization, classical ML, control, or domain-specific systems.
This framing matters because product teams often reach for a language model when a simpler classifier, rules engine, database query, or optimization method would be more reliable and economical.
[Sources]
Language models do not read words the way we do
Text becomes tokens, and tokens become vectors.
Human - centered AI changes the question .
Token ID A discrete entry in a vocabulary
Embedding A learned vector for computation
Position Information about order
Context Meaning changes with nearby tokens
Tokenization converts raw text into discrete units. Common words may be one token, while rare words, names, or code can split into several. That affects context usage, multilingual performance, latency, and price.
An embedding is a learned numeric representation. Similarities in the embedding space can capture useful relationships, but the dimensions usually do not have simple human labels.
[Sources]
The transformer’s central move
Attention lets each token weigh which other tokens matter right now.
The council denied the agent access because it lacked authority.
agent → authority permission relationship high
agent → council actor relationship med
agent → the grammatical context low
Attention is a learned routing mechanism.
Each token produces three vectors:
Q What am I looking for?
K What information do I contain?
V What information should I pass forward?
Self-attention compares a token’s query vector with the key vectors of other tokens, then uses the resulting weights to combine value vectors. Multi-head attention allows different relationships to be represented in parallel.
The original transformer was designed for sequence transduction. The architecture scaled well because attention operations can be parallelized during training. Modern systems add many engineering changes, but attention remains central.
[Sources]
Pretraining
Predicting the next token forces the model to learn broad statistical structure.
NEXT
Given the context so far, assign probabilities to possible continuations. Repeat across enormous datasets until the model becomes useful for many tasks.
Autoregressive pretraining uses the next token as the target. Because predicting text requires information about syntax, style, facts, code, and relationships, the model develops broadly useful representations.
The model does not retrieve a stored sentence for every answer. It computes a probability distribution over continuations. This explains both its flexibility and its capacity to produce plausible but unsupported text.
[Sources]
Post-training
A pretrained model becomes an assistant through additional training and system rules.
Instruction tuning
Learn patterns from demonstrations of desired responses.
Following tasks and formats
Preference learning
Use comparisons, ratings, rewards, or verifiers to favor some outputs.
Helpfulness, style, reasoning behavior
System scaffolding
Add prompts, safety policies, retrieval, tools, monitors, and routing.
Product behavior and authority
Pretraining gives the model broad capabilities. Post-training shapes how those capabilities are expressed. Supervised fine-tuning uses demonstrations. Preference optimization uses rankings, reward models, direct objectives, or automated verifiers.
InstructGPT showed that a much smaller post-trained model could be preferred to a larger base model. The lesson is that data quality, objective design, and post-training can matter as much as parameter count.
[Sources]
Reasoning models
More inference-time computation can improve difficult answers, but it is not free.
High
more latency and tokens
Maximum
best effort · highest cost
Spending more compute changes the odds. It does not turn the output into a proof.
Current model families expose reasoning effort as a selectable parameter. Internally, the system may use more reasoning tokens, search over candidate solutions, call tools, or apply verifiers before producing the visible answer.
More test-time compute can improve performance on hard tasks, but it increases latency and cost. It can also produce a longer, more convincing wrong answer. Use external checks when the task allows them.
[Sources]
Context is not memory
A million-token window can hold more information without making the model remember forever.
Context window Information included in the current request or conversation.
Long prompts raise cost and can bury relevant evidence.
External memory Saved records, vector stores, profiles, files, or databases retrieved later.
Storage creates privacy, correction, and retention questions.
Model weights Statistical patterns learned during training and post-training.
Updating weights requires training, not just another conversation turn.
Separate three ideas that products often blur. The context window is temporary input to inference. External memory is stored outside the model and retrieved later. Model weights change through training.
Current frontier systems advertise context windows around one million tokens. That expands possible applications, but retrieval quality, attention over long inputs, privacy, and pricing still matter.
[Sources]
Retrieval-augmented generation
RAG gives the model evidence at inference time instead of asking its weights to contain everything.
01
Question The user asks for information or a decision.
02
Search The system retrieves documents, passages, or records.
03
Select A ranking step chooses what enters the context.
04
Generate The model synthesizes an answer from prompt and evidence.
05
Check Citations, rules, or people verify the result.
RAG combines parametric memory in the model with non-parametric memory in an external collection. It can improve freshness, domain specificity, traceability, and access control.
There are at least three failure points: retrieval can miss the right evidence, ranking can select misleading evidence, and generation can misread or ignore the retrieved material. Citations help only if they actually support the claim.
[Sources]
Tools change what a language model can do
A model proposes text. A tool can search, calculate, send, buy, or control.
Generate
Draft an answer or plan.
Retrieve
Read current or private information.
Compute
Run code, calculations, or simulations.
Execute
Change another system or the physical world.
Tool use is not simply a smarter answer. It changes the action surface. Search adds current information. Code execution adds exact computation. Credentials may allow a system to send messages, alter records, spend money, or operate equipment.
The model usually generates a structured tool request. The surrounding software validates it, executes the tool, and returns the result. Permissions, schemas, logs, confirmation rules, and failure handling live outside the model.
[Sources]
An agent is a loop around a model
Agents observe, plan, act, and check until they stop or are stopped.
MODEL
Observe Plan Act Check The loop creates both usefulness and risk.
Each cycle can gather evidence and correct course. It can also compound a mistaken assumption, repeat a costly action, or continue beyond the user’s intent.
Every agent needs a budget, permission boundary, audit trail, and stop condition.
“Agent” is best understood as a system pattern, not a distinct model species. The model chooses or proposes actions inside a loop. Memory, tools, environment feedback, policies, and termination logic complete the system.
Current agents are improving rapidly, but independent benchmarks still show substantial failure rates. Long tasks also introduce cost accumulation, state drift, and hard-to-review chains of action.
[Sources]
Multimodal AI
Different media can be translated into representations a shared model can use.
SHARED
Text Tokens, documents, code, structured data
Images + video Pixels, patches, frames, temporal structure
Audio Speech, sound events, music, prosody
Action Tool calls, robot commands, interface events
Multimodal models learn relationships across text, images, audio, video, and sometimes action. Some systems use separate encoders connected to a language model; others are trained more jointly.
Current research is moving from perception toward embodied action. The difficult part is not only recognizing a scene, but tracking progress, understanding physical constraints, and knowing when a task is complete.
[Sources]
Efficiency is a technical capability
The best model is often the smallest one that reliably fits the task.
Small and specialized Less model, tighter task
Fine-tuning and domain data can outperform a larger general model on narrow work.
Mixture of experts Route tokens through part of the network
Large total capacity without activating every parameter for every token.
Compression Distill, prune, and quantize
Trade some precision or breadth for lower latency, memory, energy, and cost.
Frontier performance receives attention, but most production systems care about the quality-cost-latency frontier. Smaller models can run on local hardware, respond faster, and reduce dependency on a remote provider.
Mixture-of-experts models route inputs through subsets of parameters. Distillation trains a smaller model to reproduce aspects of a larger model. Quantization uses lower-precision numbers to reduce memory and compute.
[Sources]
Uncertainty is not one number
A confident output can still come from the wrong model, data, or context.
Output
How ambiguous is this prediction or continuation?
Compare alternatives or calibrate scores.
Model
Does evaluation support use for this task and population?
Respect the competence boundary.
Data
What is missing, stale, selective, noisy, or mismeasured?
Inspect provenance and coverage.
World
Could the environment, behavior, or goal have changed?
Monitor, detect shift, and pause.
For classifiers, probability scores may be calibrated or miscalibrated. For generative systems, token probabilities do not directly measure whether a whole answer is true. A fluent output can be high probability and still be unsupported.
The design question is what different uncertainty states trigger. The answer might be retrieval, a second model, human review, a safer fallback, or refusal to act.
[Sources]
Evaluation is a narrowing funnel
A benchmark score is the beginning of evidence, not the end.
General benchmark
Task-specific test
Workflow simulation
Observed outcomes
Evidence becomes more relevant as it becomes more specific.
But specific evaluation is slower, more expensive, and harder to compare across products.
Strong evaluation uses several layers instead of one score.
Benchmarks are valuable for controlled comparison, but they can become saturated, contaminated, optimized against, or disconnected from the real task. Human preference rankings also mix capability, style, and user expectations.
A strong evaluation plan moves from general capability to the actual population, workflow, failure modes, and outcomes. It also includes adversarial and responsible-AI testing, which is still reported less consistently than capability.
[Sources]
The frontier is jagged
Impressive performance on one task does not tell you what happens on the next.
≈100%
Coding benchmark SWE-bench Verified rose from 60% to near 100% in one year.
50.1%
Analog clocks The top model was only slightly better than chance at reading them.
≈66%
Computer tasks Agents improved sharply on OSWorld and still failed about one in three attempts.
<20%
Paper replication Frontier models remained weak on full scientific replication tasks.
This slide should feel surprising. Frontier systems can excel at competition mathematics and coding while failing mundane perception or sustained scientific work. That unevenness is the jagged frontier.
Do not infer that any one benchmark describes a complete system. Even strong agent scores imply meaningful failure rates when actions are repeated at scale.
[Sources]
State of the art · August 2026
The frontier is moving from chat toward longer, multimodal, tool-using work.
Reasoning
Adjustable effort
Applications trade more inference-time compute for harder tasks.
Context
Million-token windows
Large codebases and document collections can enter one request.
Agents
Longer tool chains
Models increasingly operate browsers, code, and enterprise systems.
Multimodality
Real-time media
Text, vision, audio, video, and action are becoming more integrated.
Efficiency
Smaller and routed models
Cost-performance and specialization compete with raw scale.
This is the durable version of a state-of-the-art slide. Model names will age quickly, but these architectural directions are likely to remain useful: more controllable reasoning, longer context, richer modalities, more tool use, and greater attention to efficiency.
The AI Index also reports that capability benchmarks are advancing faster than responsible-AI measurement and that model transparency has declined.
[Sources]
A few signals from the current news
Recent releases are competing on agents, speed, multimodality, and cost.
Current
OpenAI GPT-5.6 family
A family spans frontier, balanced, and high-volume tiers, with selectable reasoning effort and a 1.05M-token context window at the top tier.
Jun 30
Claude Sonnet 5
Anthropic emphasizes cost-performance and agentic work, with permanent pricing of $2 input and $10 output per million tokens.
Jul 24
Claude Opus 5
The frontier tier targets long-running agents and professional work at $5 input and $25 output per million tokens.
Aug 13
Gemini 3.7 Flash
Google frames its newest workhorse model around coding and agents, while robotics releases push video understanding into physical action.
Use this slide as news literacy, not a product ranking. Each company frames its release around selected benchmarks and product priorities. Students should distinguish official specifications from independent evidence and actual system performance.
The notable economic shift is model families. Providers sell several intelligence, speed, and cost tiers rather than a single universal model. Reasoning effort adds another meter inside each tier.
[Sources]
There is no single “best model”
Model selection is a trade among quality, speed, cost, and control.
FIT FOR
Capability How well does it perform on the work and failure modes that matter?
Operations Latency, throughput, context, reliability, and integration constraints.
Economics + control Price, hosting, privacy, adaptability, provider dependency, and governance.
As leading models converge on some benchmarks, cost, reliability, domain performance, and deployment control become stronger differentiators. A small local model may be the right choice for a narrow private task. A frontier model may be justified for infrequent, hard, high-value analysis.
The answer can also be a routed system: use a cheap model for common cases, escalate difficult cases, and reserve people for exceptions or consequential decisions.
[Sources]
The economy underneath the interface
AI is a supply chain before it is a product.
Energy + land
Electricity, water, grid connections, buildings
local infrastructure
Chips
Fabrication, packaging, accelerators, memory, networking
specialized manufacturing
Cloud
Data centers, clusters, storage, orchestration
capital and scale
Foundation models
Pretraining, post-training, safety, evaluation
data, talent, compute
Applications
Retrieval, tools, workflow, interface, support
distribution and domain access
Work + institutions
Adoption, training, oversight, policy, consequences
legitimacy and trust
The consumer sees an interface, but the production chain begins with physical infrastructure. Each layer has different capital requirements, margins, bottlenecks, and regulatory exposure.
Applications can create value by owning a workflow or customer relationship even when they do not own the model. Model providers depend on cloud and chip supply. Cloud firms may also build models, which creates vertical integration and dependency.
[Sources]
Capital is flooding into the stack
Global corporate AI investment more than doubled in 2025.
$581.7B
global corporate AI investment
Up 130% from the prior year, according to the 2026 AI Index.
Global corporate investment $581.7B
Global private investment $344.7B
U.S. private investment $285.9B
China private investment $12.4B
The scale of investment explains both the speed of technical progress and the pressure to find revenue. Private investment grew 127.5%, and generative AI captured nearly half of private AI funding.
Do not interpret U.S.-China private investment as complete national spending. The AI Index notes that Chinese government guidance funds make private comparisons incomplete.
[Sources]
Scale creates concentration
The frontier depends on a small number of firms and physical bottlenecks.
>90%
Industry-built frontier models Industry produced over 90% of notable frontier models in 2025.
5,427
U.S. data centers The United States hosts more than ten times as many as any other country.
≈1
Leading chip foundry TSMC fabricates almost every leading AI chip, according to the AI Index.
Frontier AI requires scale in compute, data, talent, and capital. That moves notable model development toward industry and makes the supply chain dependent on a small set of cloud providers, chip designers, and manufacturers.
Concentration affects human-centered design. It changes bargaining power, transparency, audit access, provider lock-in, geographic distribution of environmental burdens, and national interest in AI sovereignty.
[Sources]
Inference has a price curve
Within one model family, the frontier tier can cost 25 times more per token.
Luna
Cost-sensitive, high-volume work
$0.20
$1.20
Terra
Balance capability and cost
$2.00
$12.00
Sol
Complex professional reasoning
$5.00
$30.00
Short-context standard pricing shown as of August 17, 2026. Long context, tools, priority service, and vendor changes can alter the bill.
This is not a recommendation for one provider. It is a clear example of model-tier economics. The same family offers different price-performance points, and output is priced much higher than input.
Applications can route easy work to a cheaper tier and escalate hard cases. They can also reduce cost through caching, shorter context, constrained outputs, batching, and fewer retries. Each optimization can affect quality or flexibility.
[Sources]
A simple unit-economics example
The model choice changes this hypothetical workload from $200 to $5,000 per day.
High-volume tier
$200
10,000 users × 20 turns
400M input × $0.20/M + 100M output × $1.20/M
25×
Frontier tier
$5,000
Exactly the same workload and token volume
400M input × $5/M + 100M output × $30/M
Walk through the arithmetic. Twenty turns for ten thousand users produces two hundred thousand requests. At two thousand input tokens each, that is four hundred million input tokens. At five hundred output tokens, it is one hundred million output tokens.
The cheapest model is not automatically the best, and the frontier model is not automatically worth the extra cost. The application needs evidence that the added capability changes outcomes enough to justify the price.
[Sources]
Compute becomes an energy-system question
Global data-center electricity use is projected to roughly double by 2030.
≈415 TWh · 2024
≈945 TWh · 2030
Demand grows about 15% per year in the IEA base case.
Growth is global, but grid congestion, water use, construction, prices, and emissions are experienced locally.
AI can also help optimize energy systems, so the net effect depends on deployment and rebound.
The IEA projects global data-center electricity consumption reaching about 945 TWh by 2030, just under 3% of global electricity use. The 2024 baseline is roughly 415 TWh.
Efficiency does not guarantee lower total consumption. Cheaper inference can unlock more applications and more requests. This is a rebound effect: lower cost per unit can increase total demand.
[Sources]
AI changes tasks before it cleanly replaces jobs
Productivity gains are strongest where work is structured and easy to check.
Measured productivity 14–26%
The 2026 AI Index summarizes gains of 14–15% in customer support and 26% in software development studies.
The gains vary with worker experience, task structure, and the ability to verify outputs.
Uneven labor effects 1 in 3
One-third of surveyed organizations expected AI-related workforce reductions in the coming year.
Aggregate job losses remain limited so far, but exposed entry-level hiring pipelines are changing.
A job is a bundle of tasks. AI can automate some, speed up others, create new checking work, and change the skills needed for the remaining tasks. That is why task-level analysis is more useful than a binary list of jobs that will disappear.
Productivity studies often occur in settings with measurable outputs and clear feedback. Gains are smaller in deeper reasoning tasks, and heavy reliance may reduce learning or skill development over time.
[Sources]
Value and burden are distributed differently
A useful AI product can create value for everyone and still divide that value unevenly.
Users Gain speed, access, convenience, and new capability.
May trade privacy, autonomy, skill, or dependence for that benefit.
Workers Gain leverage, support, and relief from routine tasks.
May face monitoring, deskilling, work intensification, or reduced entry paths.
Organizations Gain throughput, consistency, data, and lower marginal cost.
Take on integration, liability, reliability, and governance costs.
Infrastructure owners Capture rents from scarce chips, cloud capacity, models, and distribution.
Carry capital risk while external costs may fall elsewhere.
The AI Index estimates U.S. consumer surplus from generative AI at $172 billion annually by early 2026. Many tools are free or inexpensive to users, while providers spend heavily on infrastructure and use free access to build distribution, data, and future demand.
Human-centered evaluation asks who captures the benefit, who pays directly, who pays indirectly, and who has bargaining power when the system becomes essential.
[Sources]
A six-question diagnostic
task data learning evidence action economics
When someone says “AI,” ask what kind of system they actually mean. Task: What output is the system trying to produce?
Data: What examples and feedback shaped it?
Learning: What architecture, objective, and post-training created the behavior?
Evidence: Where does evaluation support this use?
Action: What can the output influence or execute?
Economics: Who pays, who owns the stack, and who captures value?
Return to the opening question. Before an AI answered, a chain of choices produced the data, objective, model, infrastructure, product, permissions, and price. The output makes sense only when we can trace that chain.
These six questions are the takeaway. They are technical enough to expose how the system works and broad enough to reveal power, incentives, and consequences. A model card answers only part of them.
[Sources]