AI singularity:
how engineers and students can adapt.

AI singularity and AI supremacy: what do they mean?

IBM describes the technological singularity as a hypothetical period of runaway technological change, often associated with AI improving itself. It concerns what happens to our ability to anticipate change. Artificial superintelligence, or ASI, concerns capabilities beyond human intelligence. The ideas are related, but one does not establish the other’s timetable.

“AI supremacy” can refer to several different concerns. Keep these three questions separate:

A high test score answers only part of the first question. It grants no moral authority. The February 2026 International AI Safety Report treats loss of control as an uncertain risk involving capabilities, behavior, and access to real systems. That assessment predates the September model launches discussed in Part 1.

Before connecting a model to a codebase, decide what it can do. Reading files, deploying changes, managing payments, and deciding who receives a service carry different responsibilities. Make those permissions visible and name the person responsible for reviewing them.

What optimism about AI means for people

I’m drawn to Sam Altman’s vision of AI expanding human agency. I want it to help a student understand a difficult subject or a small team build something useful. That’s a direction I’d like to work toward. It doesn’t tell us how close a singularity might be.

In The Gentle Singularity (June 2025), Altman imagines more abundant intelligence and energy while people continue choosing goals, creating, and finding meaning with others. His speculative vision of the 2030s depends on alignment and widely distributed access. It is not an established timetable. His earlier Three Observations emphasizes agency and adaptability, while arguing that broad benefits require deliberate choices about access and power.

Dario Amodei’s Machines of Loving Grace describes another possibility: powerful AI could accelerate health research, while relationships and personal achievement retain meaning. He also describes physical and institutional bottlenecks. His scenario assumes powerful AI already exists. It does not establish when those benefits arrive.

For engineers, Greg Brockman’s 2022 essay It’s time to become an ML engineer connects progress to software infrastructure, machine learning knowledge, and willingness to revise familiar assumptions. It describes engineering and research as complementary work. Read it as a perspective on skills, not a guarantee of employment or a requirement that everyone change roles.

For a student or engineer, there’s a useful question here: after using the tool, can you explain the decision, challenge the result, and tackle a new problem? Finishing sooner is only part of the answer.

Try a problem with and without AI

Use the campus-assistant fixture below for this proposed learning exercise. Its documents are synthetic. This is a way to examine your process, not a completed study or a model benchmark.

1. 01 / WORK SOLO

   Explain the deadline

   Read the current and archived handbooks without AI. Explain why 30 September applies instead of 31 October. Cite the current document. Answer the fee question using only the supplied evidence. Say clearly when the information is missing.

2. 02 / USE AI

   Check the model’s answer

   Ask a model to answer the same questions from those documents. Keep its prompt, response, model name, and date. Check every citation and record where you accepted or rejected its reasoning. Revise your explanation in your own words.

3. 03 / TRANSFER

   Answer a different question

   Ask a study partner to create another fictional handbook excerpt with a changed rule or missing detail. Close the chat and answer it. Explain which evidence controls the decision and when you should ask for clarification. Compare how you reached each answer.

Keep all three responses in your portfolio. Add notes on what you knew, what the model helped you change, and what you could explain independently afterward.

Human–AI evolution through skills, tools, and institutions

Here, human “evolution” means changes in skills, tools, and institutions. It does not assume that people must merge biologically with machines to keep up. A society can become more capable through better education, more accessible software, and better ways of coordinating work.

One example comes from a 2025 Nature Neuroscience study that used brain recordings and deep learning to stream synthesized speech for a clinical-trial participant who could not speak because of severe paralysis. This was assistive communication in a trial participant. It did not demonstrate general cognitive enhancement.

Restoring communication gives this work a concrete purpose, regardless of whether broader forms of enhancement become possible.

People, AI, and institutions

Human–AI collaboration has its own limits. A 2024 meta-analysis of 106 experiments found that human–AI combinations improved on humans alone on average, yet did not outperform the better of the human or AI working alone. The studies covered 2020–2023. They do not measure today’s frontier models.

For a product team, assign responsibilities and test the handoffs. Give reviewers enough time and information to question a result. An “approve” button does little if the person clicking it cannot assess the work.

Job growth and AI exposure measure different things

Two questions often get mixed together: how the labor market might change, and which tasks AI could affect. These charts measure different things. Neither is a countdown to the disappearance of software engineers.

The WEF survey drew on more than 1,000 employers. Its outlook lists both AI and machine learning specialists and software and application developers among fast-growing roles. That supports preparing for several engineering paths. It does not establish that one job title will replace another. Read the jobs outlook.

ILO’s analysis emphasizes transformation because occupations contain tasks that still need human input. Exposure also varies by the kinds of work an economy contains. The same technology can therefore create different training needs in different places. Read the exposure study.

At the team level, measure completed work. Drafting speed can hide time spent checking, fixing, integrating, and maintaining a change. METR’s February 2026 update explains how selection effects and changing work patterns made its newer developer-productivity estimates difficult to trust.

Software engineer, AI engineer, or researcher?

These roles overlap. Titles differ between organizations, and small teams often combine responsibilities. You can grow within software engineering, work on AI products, or pursue research. The right direction depends on which problems you want to spend time solving.

If you already write production software, start with one recurring failure in an AI feature: weak retrieval, unsupported answers, unreliable tool calls, or excessive cost. Learn enough about it to design an evaluation and improve the system. Document the failure, your change, and how you checked it.

Research takes practice too. Read a paper carefully, implement a small baseline, and ask why a result changes when an assumption changes. A focused study group or a supervised project can help you discover whether you enjoy that process before committing to a research degree.

If routine starter tasks become automated, junior engineers still need opportunities to learn through investigation, review, and gradual responsibility. Teams should assign this work deliberately and give experienced engineers time to supervise it.

What should universities teach in the AI era?

A degree should teach students to understand a problem, investigate competing explanations, and defend their decisions. Curricula need room for AI, alongside algorithms, systems, mathematics, writing, and working with other people.

UNESCO IESALC’s higher-education working paper argues for coherent technical, critical, and ethical competencies. Its review and study of 16 institutions highlight the limits of fragmented institutional responses. Adding an isolated AI workshop leaves many decisions about teaching and assessment unresolved.

UNESCO’s separate 2024 student AI competency framework organizes learning around understanding, applying, and creating, with attention to human agency and ethics. It targets school curricula. The university learning ladder below is my adaptation of that progression, rather than an official degree standard.

Funs Jacobs’s April 2025 commentary on Altman’s observations emphasizes adaptability and questions how long specific technical curricula remain useful. My view is that students need durable foundations and room to revise their methods. Understanding types, data flow, and debugging helps you judge an unfamiliar framework or an AI-generated solution.

Assessment can reflect that reality. Pair an AI-assisted project with an oral defense, a code walkthrough, and a small task completed independently. Ask students to explain an error they found and a design they rejected. That gives instructors evidence of learning across both assisted and independent work.

Access deserves attention as well. A required assignment should not quietly depend on a premium subscription some students cannot afford. Provide shared access or viable alternatives, clear data rules, and accessible tools. Students should know which uses are permitted before they begin.

For students choosing courses, look for sustained projects, feedback, sound fundamentals, and opportunities to work with researchers or real users. Learn enough probability and statistics to question a chart. Learn enough systems thinking to understand what happens beyond the model call. Keep practicing the ability to explain a difficult idea clearly.

What Cyberpunk, Raised by Wolves, and Ex Machina ask about AI

These stories put characters in situations involving power, education, and trust. The readings below are my interpretations of fictional worlds. They do not predict what AI will do.

Build a campus document assistant

The brief below includes a small task, test cases, and failure conditions to check during review. Use it for a course project or study group. It is a teaching exercise, not a finished university service.

Adapt the emphasis to your direction. A software engineer can focus on a reliable interface and integration. An AI engineer can compare retrieval and evaluation choices. A research engineer can make the experiment repeatable. A researcher can formulate a narrower question about a failure pattern and test alternative explanations.