The Evolution of Software Engineering in the Age of AI

QR code for the talk track
Young James Gress at a Radio Shack computer Young James Gress at a Radio Shack computer

The Evolution of Generative AI

Seven years of OpenAI models

GPT-2 to GPT-6:
What Actually Changed

How a text predictor became a model that reasons, calls tools and takes new instructions mid-task. Figures checked against OpenAI docs, November 2026.
2019 · Zero-shot

GPT-2: predict the next word

Text input
GPT-2
Continued text
Released
February 2019
Full 1.5B model published in November 2019
Scale
1.5B parameters
Trained on WebText, about 40GB of web pages
Training
Only pre-training
No instruction tuning: it continues whatever text you give it
2020 · Few-shot

GPT-3: learn from examples in the prompt

dog → chien
cat → chat
bird → ?
GPT-3
oiseau
Released
May 2020
Paper: "Language Models are Few-Shot Learners"
Scale
175B parameters
About 100× GPT-2
Key idea
In-context learning
Shown a few examples, it does the task without any weight updates
2022 · Chat

GPT-3.5 and ChatGPT: follow instructions, hold a conversation

User message
Conversation history
GPT-3.5
Chat reply
Released
November 30, 2022
ChatGPT launched as a free research preview built on GPT-3.5
Training
Instruction-tuned with RLHF
Same approach as InstructGPT: human feedback teaches it to follow requests
Context
~4K tokens
Why it matters
AI goes mainstream
2023 · Multimodal

GPT-4: instructions, alignment and images

Text
Image
System instructions
GPT-4
Answer
Released
March 14, 2023
Training
RLHF fine-tuned
Aligned with reinforcement learning from human feedback
Context
8K–32K tokens
The 128K window came later, with GPT-4 Turbo (Nov 2023)
Multimodal
Text + image input
Announced at launch, but image input wasn't public until later
2025 · Reasoning

GPT-5: think first, then answer

Prompt
Effort (reasoning)
Tools (external)
GPT-5
Reasoning tokens
Answer
Released
August 7, 2025
Default model in ChatGPT at launch
Key idea
Router in ChatGPT
Automatically picks the fast or the reasoning variant per request
Context
400K tokens (API)
272K input + 128K output. ChatGPT limits were lower and varied by plan.
2026 · Async tools

GPT-6: keeps working while you steer

Prompt GPT-6 New instruction (mid-turn steering) Keeps reasoning Continues work Async tool calls Answer tool results user Dashed lines: inputs that arrive while GPT-6 is still working
Released
~September 2026, as GPT-6 Astra
Family also includes GPT-6.1 Sol and GPT-6.1 Terra
Async tool calling
Reasons while tools run
Replaces "searchable notes," which OpenAI's docs don't mention
Mid-turn steering
Change course mid-task
Responses API over WebSocket; completed work is kept
Context
1.05M tokens
128K max output tokens

Release date from a third-party tracker; not confirmed by OpenAI

At a glance

Six milestones, six shifts

2019
GPT-2
zero-shot
Key ideaZero-shot task transfer
Parameters1.5B
Context1,024 tokens
InputText
OutputContinued text
2020
GPT-3
few-shot
Key ideaIn-context learning
Parameters175B
Context2,048 tokens
InputText
OutputAnswer
2022
GPT-3.5
chat
Key ideaChat (ChatGPT launches)
ParametersNot disclosed
Context~4K tokens
InputText
OutputChat reply
2023
GPT-4
multimodal
Key ideaSystem instructions
ParametersNot disclosed
Context8K–32K tokens
InputText (image later*)
OutputAnswer
2025
GPT-5
reasoning
Key ideaRouter in ChatGPT
ParametersNot disclosed
Context400K (API)
InputText + image
OutputReasoning tokens → answer
2026
GPT-6
async tools
Key ideaAsync tools + mid-turn steering
ParametersNot disclosed
Context1.05M tokens
InputText + image
OutputReasoning tokens → answer

* GPT-4 launched text-only; image input was demoed but shipped later (ChatGPT Sept 2023, API Nov 2023). OpenAI hasn’t published parameter counts since GPT-3.

How many days since ChatGPT Launch?

Copilot prompt: how many days since ChatGPT launched

What the Leaders Building AI Are Saying

Jensen Huang - CEO NVIDIA

"Software is eating the world, but AI is going to eat software."

— MIT Technology Review interview, May 2017

"AI is not going to take your job. The person who uses AI is going to take your job."

— Columbia Business School, October 2023

"You probably recall over the course of the last 10, 15 years almost everybody who sits on a stage like this would tell you that it is vital that your children learn computer science. Everybody should learn how to program. And in fact, it’s almost exactly the opposite."

— World Governments Summit, Dubai, February 2024

"Let’s say you have a software engineer or AI researcher, and you pay them $500,000 a year. We do that all the time. At the end of the year, I’m going to ask him how much did you spend in tokens… If that person said $5,000, I will go ape… If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed."

— All-In Podcast, March 2026

Eric Schmidt - Former Google CEO

"We believe as an industry that in the next one year the vast majority of programmers will be replaced by AI programmers."

— August 2024

Actual
  • Total jobs went up. US software developer jobs rose to 1.69M in May 2025 from 1.65M in May 2024, about 2% (BLS OEWS). BLS still projects 10% growth from 2025 to 2035.
  • Entry-level is where the pain is. Stanford's “Canaries in the Coal Mine” study (Brynjolfsson et al., November 2025) found employment of developers aged 22–25 is down nearly 20% from its late-2022 peak.

Dario Amodei - CEO of Anthropic

“I think we'll be there in three to six months, where AI is writing 90% of the code.
And then, in 12 months, we may be in a world where AI is writing essentially all of the code.”

— March 2025

Actual 90%+

“90 plus percent of our code is actually written by Claude Code.”

— Krishna Rao, CFO of Anthropic, Invest Like the Best podcast, May 2026

Sundar Pichai - CEO of Google

75% of all new code at Google is now AI-generated and approved by engineers.

— Google Cloud Next, April 2026

Sam Altman – CEO of OpenAI

Questions was asked "What does my future look like?"

"There are sort of two views you can take.
You can say, ‘Oh man, it’s doing everything I can do,’
or you can say, ‘Look at all the new things I can now do.’"

— April 2025

"The thing about coding is that people went through this: there's no more jobs as coders… Coders are never going to be busier than they've been and coders are never going to have more of an impact and make more money. Now, the texture of the job changed a lot."

— Matthew Berman podcast, October 2026

Andrej Karpathy's post coining the term vibe coding

Vibe Coding

February 2, 2025

FROM VIBES TO VERIFICATION

From Vibe Coding to Harness Engineering

“How do we trust code we didn’t write?”

Then: vibe coding

  • Code accepted on vibes, not understanding
  • Insecure code and exposed sensitive data
  • Architecture and design principles ignored
  • Testing treated as an afterthought
  • Teams lose track of what their code does

Now: harnesses that encode the organization

  • Coding standards, security and compliance policies
  • Architecture decisions, patterns and API contracts
  • Requirements, user stories and acceptance criteria
  • Tests, linters and security scans as automated gates
  • Human review at the key decision points
Key message

The model writes the code. The harness defines what correct looks like: your standards, your artifacts and the checks every change must pass before it ships.

THE ENTERPRISE ADOPTION PATH

AI Software Engineering in the Enterprise

Established Expanding Emerging Created Consumed
PHASE
01PlanningExpanding
02DesignExpanding
03CodingEstablished
04TestingExpanding
05InfrastructureEmerging
06ReleaseEmerging
07OperationsEmerging
Goals & user stories
Created Consumed Consumed
Architecture & API contracts
Created Consumed Consumed Consumed
Source code & documentation
Created Consumed Consumed Consumed
Test suites & quality reports
Consumed Created Consumed Consumed
Infrastructure as code & config
Created Consumed Consumed
Build & release package
Created Consumed
Telemetry, alerts & runbooks
Consumed Consumed Created

WHERE THE TIME ACTUALLY GOES

The Work Is Fast. The Waiting Is Slow.

Creating the artifact Human review Queue or approval Prioritization Intake rejected for quality, rework Illustrative feature · 63 calendar days
Today
63 days · 8 days of work (13%)
AI on the work
57 days · 2 days of work (4%)
4× faster workonly 10% faster delivery

AI shrinks the blue bars, not the queues, intake checks, approvals or backlog. Fix the wait, not just the work.

AI Writing Code ≠ Ignoring DevOps

"When we say AI will write the code, we don’t mean we’re skipping the discipline."

AI is transforming how code is created
but code quality, security, and delivery pipelines still matter.

  • DevOps isn't going away
  • Guardrails are still essential
  • Testing, observability, and governance remain critical

We’re not removing the need for rigor —
we're elevating it to meet the speed of AI.

WHEN THE ASSUMPTIONS CHANGE

The Architectural Reversal

Earlier

Centralized

CVS · Subversion

Optimized for a shared source of truth

Git era

Distributed

Git · GitHub · pull requests

Optimized for developer autonomy and local work

AI era

Server-oriented

Anthropic Scion

Optimized for agent-scale operations

Case study · Anthropic, 2026
2,500 → 30,000+changes merged per day, May to September
~25MCI checkouts per day on one monorepo
Git → Scionrepo operations moved server-side

Architectures aren’t permanent best practices. They encode assumptions about workload, scale and economics, and AI is changing those assumptions.

DEVOPS AT MACHINE SPEED

We Built DevOps for Agile. Now It Must Keep Up with AI.

Built for human paceBreaks at AI speed
Human-paced commitsHigh-volume agent changes
Human review of every changeReview capacity becomes the bottleneck
Fixed CI/CD capacityBursty, parallel workloads
Manual security approvalsAutomated policy enforcement
Developer-owned changesAgent identity and accountability

DevOps for agents

  • Agent-aware source controlBuilt for thousands of parallel changes
  • Elastic CI/CDScales with bursts of agent work
  • Risk-based automated testingHumans review where risk is highest
  • Continuous security controlsPolicy as code, not approval queues
  • Agent-to-outcome observabilityTrace every agent action to production

Don’t Just Automate the Past—Reimagine the Future

“It’s like putting a jet engine on a horse-drawn carriage. Sure, it moves faster, but you’re missing the point—the whole vehicle can be reinvented. Generative AI gives us a chance to reimagine how we deliver, not just speed up what we already do.”

THE HUMAN SIDE OF AI TRANSFORMATION

AI Fatigue

AI is changing what we do, how fast we do it, and how we see our future. How do we make that change sustainable?

1

Hype Fatigue

“Is everything really changing?”

Too much noise
2

Change Fatigue

“How do I keep up?”

Too much change
3

Cognitive Fatigue

“Why am I so mentally exhausted?”

Too much mental processing
4

Identity Fatigue

“Where do I fit?”

Too much uncertainty

SUMMARY

Four Dimensions of AI Fatigue

DimensionCentral questionWhat helps
Hype FatigueIs everything really changing?Focus on outcomes, not announcements
Change FatigueHow do I keep up?Involve employees and provide time to adapt
Cognitive FatigueWhy am I so mentally exhausted?Reduce cognitive overload and protect focus time
Identity FatigueWhere do I fit?Provide credible career and responsibility roadmaps

Sustainable AI transformation requires more than adopting technology. It requires helping people understand, adapt to, and find their place in the future of work.

The four fatigue categories are a presentation framework, not clinical diagnoses. Jevons Paradox is an economic analogy, not a prediction of future employment.

Preparing for the Future

  • Learn agentic systems, system thinking, and architecture
  • Embrace change: experimentation and transparency
  • Grow internal champions for AI adoption

HOW THE JOB IS CHANGING

The Transformation of Work and New Roles

FromTo
Writing code line by lineSpecifying intent and directing agents
Closing ticketsOwning outcomes end to end
Manual testing and reviewDesigning evals, guardrails and pipelines
Individual contributorLeading a team of humans and agents
Knowing the syntaxKnowing the business and the system

Roles taking shape

  • Agent orchestratorPlans, delegates and integrates agent work
  • Context engineerCurates the knowledge, specs and memory agents use
  • AI evaluation engineerDefines “good” and measures quality and safety
  • AI platform engineerBuilds the paved road: tools, agents and guardrails
  • AI governance leadOwns risk, compliance and responsible use

EMERGING CAREER PATHS

Trending Roles in AI Software Engineering

01 / PRODUCT-CENTERED

AI-native
software engineer

Changes how software gets built.

IntentAI agentsProduction
  • Starts with intent; agents handle most coding and testing
  • Owns architecture, review, and judgment; directs agents
  • Coordinates agent work and integrates it into production
CORE QUESTION“Can you build it?”
Typical workLonger engagementsSuccessFeatures and products shipped
02 / CUSTOMER-EMBEDDED

Forward Deployed
Engineer

Changes what gets built and how the business runs.

ContextPrototypeDeploy
  • Applies AI-native skills to ambiguous client problems
  • Chooses what to reinvent; designs agentic systems and workflows
  • Pairs business judgment and platform depth with client/C-suite influence
CORE QUESTION“Can you figure out what needs solving, get the client there, build it, and produce the outcome?”
Typical workShorter engagementsSuccessMeasurable business value
Both direct AI agents; FDEs extend the work upstream into problem framing and business change.

INTERESTING OBSERVATION #1

Where Should We Save Tokens?

“Token costs are going up. Should we cap how much our developers use?”

Reality
  • Developers aren't the real cost. A few hundred dollars a month per developer is small next to the productivity it buys.
  • The real cost is agentic processes. A workflow that runs AI over a million records every night uses far more tokens than any developer.
  • Optimize there. Pick the right model, trim the context, cache and batch, and use plain code wherever AI isn't needed.

INTERESTING OBSERVATION #2

“How Accurate Is It?”

“If it isn't 100% accurate, we can't use it.”

Reality
  • Humans aren't 100% either. Most businesses don't track how accurate the people doing the work today are.
  • Without a baseline, there's no real threshold. 100% becomes the bar by default, and nothing can meet it.
  • Measure the current process first. Then ask whether AI with human review of the exceptions beats it on accuracy, cost and speed.

The End of "Tech for Tech’s Sake"

  • Old mentality: "I just code, not business"
  • AI needs business context to generate useful output
  • Engineers must understand goals, not just syntax

“In the future, the best engineers won’t just know how to code — they’ll know why the code matters.”

Closing: A Call to Lead

"AI is the runtime that is going to shape all of what we do…" — Satya Nadella

  • Will you be a passenger — or help engineer the future?
  • Start learning. Start building. Start leading.

“The next chapter of software engineering is being written now — and we’re all holding the pen.”

QR code linking to James Gress

Questions?

Appendix

AI Fatigue: a deeper look

Satya Nadella - CEO Microsoft

"AI is the runtime that will shape all of what we do"

— December 2024

"I think what'll happen is these crud I mean SAS applications are a crud database with a lot of business logic so the crud database will then get orchestrated outside of the business logic tier of just the SAS application"

— January 2025

AI FATIGUE · 1 OF 4

Hype Fatigue: Too Much Noise

“Is everything really changing?”

What is happening

Constant announcements, sweeping predictions and pressure to follow every new model leave people skeptical or disengaged.

For years we’ve heard that technology would make developers obsolete. Yet better technology can also create new opportunities and increase demand for software.

What helps

  • Focus on measurable outcomes, not announcements
  • Curate trusted information sources
  • Celebrate real business and engineering wins
  • Make AI adoption selective and purposeful
  • Don’t make every breakthrough an immediate priority
Key message
Jevons Paradox

When software becomes cheaper and faster to produce, organizations may want more of it. Efficiency doesn’t necessarily eliminate demand. Sometimes, efficiency fuels demand.

AI FATIGUE · 2 OF 4

Change Fatigue: Too Much Change

“How do I keep up?”

What is happening

AI is changing not just the tools people use, but the work itself.

Developers who spent most of their time designing and writing code are shifting toward reviewing, validating, integrating and maintaining AI-generated code.

Responsibilities change again before people have built confidence in them.

What helps

  • Involve employees in redesigning workflows
  • Provide realistic transition periods
  • Update job expectations to match new responsibilities
  • Build skills through practical experience
  • Adapting takes time, not just training
Key message

Sometimes the hard part isn’t learning the new tool. It’s losing the part of the job you loved. Change affects people as well as processes.

AI FATIGUE · 3 OF 4

Cognitive Fatigue: Too Much Mental Processing

“Why am I so mentally exhausted?”

What is happening

AI can generate work faster than people can reliably evaluate it.

  • Reviewing large volumes of AI-generated code
  • Validating AI-generated documentation
  • Evaluating multiple AI-generated solutions
  • Constantly switching between agents and workflows
  • Owning decisions made with AI assistance

Routine tasks used to give the mind a break. Automate them, and more of the day becomes high-concentration work.

What helps

  • Clear review and validation checkpoints
  • Automated testing and security checks
  • Limit concurrent AI workflows
  • Protect uninterrupted focus time
  • Design out unnecessary context switching
  • Measure cognitive workload alongside productivity
Key message
Productivity paradox

AI can reduce the effort of producing work while increasing the effort of understanding and validating it. Is that productivity sustainable?

AI FATIGUE · 4 OF 4

Identity Fatigue: Too Much Uncertainty About the Future

“Where do I fit?”

What is happening

Someone spent twenty years becoming a great software engineer: designing systems, writing code, solving hard problems.

Now AI is changing how much of that work gets done. Their experience still matters, but how they apply it is evolving.

“How do I keep up?” “Where do I fit?”

What helps: a credible roadmap

  • Future responsibilitiesWhat will my role look like?
  • Skills developmentWhat still matters, what’s new?
  • Career progressionWhere can roles lead?
  • MentorshipWho helps me navigate?
  • Practical experienceHow do I build confidence?
  • TransparencyWhat’s known and what isn’t?

Not a guarantee. Direction, support and a credible path forward.

Key message

Organizations can’t simply ask employees to be flexible. They need to show what people are evolving toward.

A QUESTION FOR LEADERS

Your organization has an AI roadmap.

But does every employee have a people roadmap?

AI transformation isn’t only a technology journey. It’s a people journey.

PANEL DISCUSSION

Top Concerns: AI & Data Governance

1

Shadow AI & Data Leakage

“What are people pasting into AI tools?”

Unsanctioned tools, personal accounts, source code and customer data leaving the boundary.

2

Oversharing & Access Control

“Can AI find what users shouldn’t see?”

AI exposes permission sprawl. RAG and copilots surface whatever a user can technically reach.

3

Prompt Injection & Agent Permissions

“What can an agent do on our behalf?”

Untrusted content steering agents with real tools, credentials and write access.

4

Vendor Use, Retention & Residency

“Where does our data go, and who trains on it?”

Model providers, MCP servers and plugins. Retention, training opt-outs and data location.

5

Lineage, Provenance & Audit

“Can we prove what data drove a decision?”

Logging prompts and outputs, tracing sources, and classifying AI-generated content.

6

Regulation & Accountability

“Who owns it when AI gets it wrong?”

EU AI Act, GDPR and sector rules. Privacy, deletion rights and clear human ownership.

AI doesn’t create new data governance problems. It finds the ones you already have, faster.