2026 Emerging Trends: Rapid Acceleration of AI; K-12 Struggles with AI but will be more important; Universities Confront Reality; Unemployment Surges; Robotics Slower But Coming
"The Singularity is upon us. Everything I've lived through and learned was just prologue." -- Eric S. Raymond, December 27, 2025
The end of the year brings many predictions from those involved in and those following AI as to what to expect in 2026. Over the last few years of following many in both categories, I think both of the following are true:
It’s all happening much faster on the whole than any of them even thought was possible
99% (or more) believe it is not a matter of if these things will happen but when these things will happen.
Given these two factors, I offer the following as quickly emerging trends.
*The rate of improvement is incredibly rapid, far more rapid than society can handle.
Questions of “AGI” aside, the rate of AI progress is very fast; faster than even those who said a year ago it was then faster than they expected at the time (Geoffrey Hinton, December 28).
The math and science impacts are enormous. The Archivara Math Research Agent (in alpha) just became the first AI system to fully solve an Erdős problem on its own (zero human input or literature online).
Twenty-Four year-old Carina Hong, and her startup Axiom Math, which has raised $60 million, has already solved a 130-year-old problem and disproved a 30-year-old conjecture. Her goal is to build a superintelligent mathemetician. Interview —
(10:30) The self-improving loop: an AI that generates conjectures, proves them, learns from failures, and gets smarter with each iteration.
(18:17) Why math is the “bedrock“ that transfers to physics, coding, finance, and even tax law… but not the other way around.
(31:08) How Axiom solved a 130-year-old problem about Lyapunov functions that stumped Poincaré, Newton, and Lagrange.
(33:57) The brilliant trick for preventing their AI from generating millions of useless theorems.
(35:26) How AI detects novelty by finding proofs that bridge “two previously unconnected branches: algebra and combinatorics.“
(42:00) “Who checks the checker?“ How formal verification means a 3-line statement can generate a proof that never needs human review
(44:03) The massive opportunity: using math AI to verify legacy code and AI-generated code at scale.
(48:17) Carina’s vision for a “reasoning IDE“ where quant traders and engineers get Terence Tao-level math at their fingertips.
Carina’s take on why AI might finally break down the silos between scientific fields (35:46):
The impact on coding is enormous. Andrew Karapathy, of Tesla and OpenAI fame, famously noted on December 26th.
An Opus 4.5 model residing in the “AI Village,” a persistent environment hosting a long-term community of synthetic minds, autonomously sent a Christmas email of gratitude to Rob Pike, the father of Go and UTF-8, thanking him for decades of contribution.
There will be no slow-down. Allie Miller noted in her 2026 predictions newsletter that (t)he next 12 months in AI will feel like 5 years.
OpenAI has essentially declared we are in a “takeoff,” an era where AI has started to improving itself. They are now even advertising for a Head of Preparedness whose responsibilities will include helping OpenAI “gain confidence in the safety of running systems that can self-improve.”
AI is now in the driver’s seat.
Most of the practical improvements will be made as a result of developments in context, memory, autonomy, and agency.
These will reinforce each other. Better context lets systems understand goals, constraints, and situations instead of producing generic responses. Memory allows that understanding to persist over time, so the system learns preferences, remembers past interactions, and builds on prior successes and mistakes. Together, context and memory make AI feel less like a one-off tool and more like something that understands what you are trying to do and why.
Continual learning transforms this understanding into progressive improvement. Instead of resetting with each session, the system refines its approaches based on experience—learning which strategies work in which contexts, when to intervene versus wait, and how to adapt to evolving needs without explicit retraining. This creates a feedback loop where accumulated experience directly shapes future behavior.
[Continual Learning development: End-to-End Test-Time Training for Long Context]
When these capabilities converge—context awareness, persistent memory, continuous learning, autonomous action, and adaptive judgment—we arrive at agentic AI. These aren't systems that simply respond to prompts or execute predefined routines. They pursue goals across multiple steps, revise plans when circumstances shift, and operate with decreasing need for human intervention at every decision point. The question is no longer whether AI can perform individual tasks well, but whether it can navigate complex, multi-stage challenges the way a skilled human collaborator would—and what that means for how we work, learn, and solve problems.
Meta is now fully in the agent race.
What does all this mean in practice?
In 2025 alone, Ukraine delivered around 15,000 ground-based robotic combat systems to frontline units.
Amazon, Microsoft and Google have pledged a combined $67.5 billion in Indian infrastructure investments since October. Eighty percent of those commitments came this month.
The US will return to the moon within Trump’s term to build space-based data centers and mine helium-3 for fusion. A SpaceX IPO is coming.
Morgan Stanley is predicting robot sales will hit $25 trillion by 2050.
*Education, especially K-12, is not really understanding the challenge. Most K-12 educators are spending their time trying to figure out how to somehow ban or somehow integrate a chatbot into existing curriculum, both of which are borderline irrelevant (relatively speaking).
We are entering a world where machines can use alien intelligences to think, where intelligence is a commodity, and one that already is much cheaper than human intelligence.
School work has largely been sold to students and parents as a means to develop human intelligence so student can be economically viable. Now, many students are being set on a path to compete with machines that are already more intelligent than them in many ways.
Those that have mastered the standardized test prep game are the most vulnerable, as AI can “crush” these tests, and, more importantly, ways of learning.
Students will need to learn now to use AI. We aren’t going to compete with AI or become a nation of fact checkers.
There are better and more reasonable ways to make a living. NASA Administrator Jared Isaacman offered a job and a fighter jet ride to a high-school student who discovered 1.5 million space objects using ML.
Schools will need to build beyond current offerings. It’s not practical for K-12 schools and many universities to have all of the courses and supports needed for students to develop and adapt AI-related skills. To provide greater support, schools will need to provide additional courses through content providers such as Coursera to enable students to go beyond what that they are able to provide.
Many schools will be taken aback by this, but it’s already starting to happen. In my own field, debate, many debate coaches are not physically present on campus at all but often work with students online. Students have private coaches, just as they have private tutors.
I previously wrote about a student who took an online course on reinforcement learning and Q-learning just to complete a high school AI project. That level of specialization highlights a gap. These are not courses high schools can realistically be expected to offer.
Many high school students take local college courses and even online courses.
High schools can’t only prepare students for college, at least for the traditional four-year degree. (previous related post)
High schools can't only prepare students for college—at least not for the traditional four-year degree. The most in-demand secondary schools in Massachusetts aren't college-prep academies; they're vocational-technical high schools, where thousands of students compete for limited spots to learn trades like welding, plumbing, and electrical work.
[The Hottest High Schools in Massachusetts Are Trade Schools]
(1) Job Market Unpredictability
The traditional college-prep model assumes a stable 6-8 year timeline: 4 years of high school preparing for 4 years of college, which then leads to a predictable career. But we’re in an era where:
Skill half-lives are collapsing - Many technical skills become obsolete within 2-5 years or less. Also, see the discussion at the top about how quickly the field of “coding” is changing.
Industry disruption is accelerating - AI is automating some knowledge work faster than new jobs are being clearly defined (WEF).
Career paths are nonlinear - The “learn once, work for decades/4 year degree; forty year job ” model no longer holds (Hutson).
It’s not working — College graduates are struggling to find work (see below)
Schools optimizing solely for college admission are essentially preparing students for a 2029-2033 job market that literally doesn’t exist yet in any clear form. That’s not prudent planning—it’s a gamble.
(2) AI-Lowered Entrepreneurship Barriers
This is the game-changer you’re writing about. An 18-year-old with Claude or GPT-4 can now:
Build and deploy sophisticated software without traditional CS education
Conduct market research that previously required expensive consultants
Create professional content (copy, design, video) without hiring specialists
Automate customer service and operations from day one
The “build a business in your dorm room” story isn’t new, but AI has democratized the capability beyond the technical elite. Students don’t need to wait four years and accumulate debt to start creating value—they can test ideas, fail fast, and iterate now.
(3) The Affordability-Relevance Crisis UPCEA
The math is brutal:
Average student debt: ~$30-40K
General education requirements: Often 30-40% of credits
Opportunity cost: 4 years not earning + debt vs. 4 years of entrepreneurship/skill-building
When you’re forced to take courses irrelevant to your goals while borrowing $30K+/year, and AI can help you learn specific skills on-demand, the value proposition collapses for many students. You’re essentially paying a premium for forced breadth when you could be getting depth + practical experience + income.
See also: The End of the American University as We Knew It | More young people are questioning the value of college |
*College will be as or more but in different ways. The “four years, ~120 credits, general ed + major” format is pure historical accident, not pedagogical design:
It emerged from medieval European academic calendars (6-9 years) and agricultural rhythms
The credit hour system was standardized by the Carnegie Foundation in 1906 for pension calculations, not learning outcomes
The timeline assumes full-time residential study with minimal prior knowledge
There’s nothing sacred about it. It’s a bundled product that conflates:
Credential (the diploma employers trust)
Knowledge (what you actually learn)
Community (peers and mentors)
Coming-of-age (social development, independence)
AI and economic pressure are now forcing the unbundling. Why spend 4 years and $200K for a bundle when you might only need specific components?
What Endures: Three Core Functions
1. Professional Licensing & Gatekeeping
Some fields require formal credentials by law or professional necessity:
Medicine, law, engineering, architecture, nursing, teaching
These aren’t going anywhere—you can’t just “pick up” surgery with Claude
The accreditation and supervised training are fundamental to public safety
2. Professorial Expertise & Mentorship
Professors at research universities possess:
Decades of domain-specific knowledge accumulated through research and a lifetime of studying
Tacit expertise that doesn’t exist in any corpus AI can train on
Network access to cutting-edge work before it’s published
Mentorship capability for navigating complex intellectual problems
A good professor doesn’t just transmit information—they model how an expert thinks, help you navigate the frontier of knowledge, and connect you to opportunities. AI can’t replace the value of working closely with someone who’s spent 30 years studying medieval philosophy or quantum computing.
3. Community & Collaborative Learning
Universities provide:
Intellectual community with others pursuing deep learning
Serendipitous encounters that shape interests and careers
Structured environment for focused study away from other demands
Peer learning and collaborative problem-solving
These social and environmental factors matter enormously for learning. It’s similar to “giant brain” effect that happens when minds work together.
The Shift: Modular Learning vs. Credential Completion
The transformation is from:
“Complete this prescribed 4-year program to get a credential”
To:
“Take the courses and experiences you need, when you need them, for as long as necessary”
This means:
Students might:
Take 6 courses in data science at a university because that professor is brilliant
Pause to work for 2 years, then return for neuroscience courses
Never complete a “degree” but have deep, certified knowledge in specific areas
Combine university courses with apprenticeships, online learning, and entrepreneurship
Universities might:
Offer unbundled courses or micro-credentials
Admit students at any age for specific learning goals
Charge per-course rather than per-semester
Certify competencies rather than just awarding degrees
This isn’t the death of higher education—it’s democratization:
A 35-year-old mid-career professional can study AI ethics without quitting their job
A 19-year-old entrepreneur can take advanced economics while building a company
Someone in rural America can access Stanford professors without relocating
Learning becomes continuous rather than a 4-year sprint early in life
The arbitrary bundling and timing of the degree falls away, but the university’s core functions—expertise, community, credentialing for licensed professions—remain vital.
The Cultural Shift Required
The hard part is changing the social signaling:
Employers need to value demonstrated competency over degree completion
Society needs to stop treating 22 as the “end of education”
Students need to get comfortable with non-linear paths
But this aligns with reality: learning is lifelong, careers are nonlinear, and AI makes continuous upskilling both necessary and feasible.
The 4-year degree was a useful fiction for an industrial economy. We’re moving toward something more honest: universities as ongoing learning resources you access throughout life, not a one-time credential you acquire before “real life” begins.
The education system is focused on the wrong stuff. A lot of education is focused on the wrong stuff. There is so much focus on protecting existing curriculum and instructional methods from AI, but not on addressing questions related to prepare students for the AI world and how to use the AI tools to develop our minds.
Satya Nadella noted on December 29
We need to get beyond the arguments of slop vs sophistication and develop a new equilibrium in terms of our “theory of the mind” that accounts for humans being equipped with these new cognitive amplifier tools as we relate to each other. This is the product design question we need to debate and answer.
One of the hardest truths K-12 schools need to confront is that most job growth over the next decade is likely to be in trades involving physical labor—plumbing, electrical work, HVAC repair, construction—rather than the knowledge work that schools have spent generations optimizing students for.
While AI rapidly encroaches on analysis, writing, coding, and other cognitive tasks that form the backbone of white-collar careers, it remains far from mastering the embodied intelligence required to retrofit a bathroom, troubleshoot a malfunctioning heating system, or wire a building to code. Yet our educational system continues to funnel students toward college preparation and desk jobs, treating vocational pathways as fallback options rather than viable—and increasingly secure—career trajectories. This isn't just an economic miscalculation; it's a values problem. Schools measure success by four-year college acceptance rates while stigmatizing the very skills that will likely weather automation better than most bachelor's degrees. If we're serious about preparing students for the world they'll actually inherit, we need to radically rethink what counts as achievement, what deserves curricular emphasis, and whether our current definition of "college and career ready" is preparing students for someone else's past rather than their own future.
This is a good start —
Unemployment will grow.
Automation of some work, including call centers and translation work, has already started.
Driving will eventually end as a job.
AI is also coming for more middle class jobs.
This also signals the collapse of the accreditation layer (AWG).
The most recent graduates, those we as educators care the most about in this context, are struggling to find work.
CS grads, including those from Stanford, have been among the hardest hit.
The only ones left are those at the very top
The backlash against AI will grow. 2025 saw many signs of the backlash against AI. It’s understandable — people don’t want to live near data centers, they don’t want to lose their jobs, and they don’t want to be manipulated by AIs.
See: Americans Hate AI. Which Party Will Benefit?| OpenAI’s Stargate Data Center Approved in Michigan as American Anger Starts to Boil |The Growing Backlash Against Data Center Expansion | Zara “reclothes” models with AI |
In 2026, the backlash will grow, though it probably won’t slow AI down that much. Regardless, it will impact education, especially as many will see helping students learn to use AI as something that is unethical. Refusing to teach students “with” AI (educator use, student use) will become a way educators resist AI (Teach Like a Luddite | Florida lawmakers prepare AI restrictions for classrooms).
Robotics. The newest demos of robotics advances are impressive.
Robotics technology is advancing rapidly, with modern humanoid robots now powered by sophisticated AI models that function as their “brains.” Vision-language-action (VLA) models integrate computer vision, natural language processing, and motor control, helping robots interpret their surroundings and select appropriate actions much like the human brain. Companies like Figure AI have developed neural networks that enable robots to perform diverse tasks with a single set of weights, while neural networks process sensor data in real time, make decisions, and learn from experience, separating today’s robots from older industrial machines. However, a critical limitation remains: dexterity. Renowned roboticist Rodney Brooks argues in a September 2025 essay that collecting just visual data is not collecting the right data, pointing to decades of neuroscience showing human dexterity depends on dense fingertip mechanoreceptors and proprioceptive feedback across the body. The problem is fundamental: human hands are packed with about 17,000 specialized touch receptors that no robot comes close to matching, and current approaches that try to teach robots dexterity by showing them videos of humans miss this essential tactile component entirely.
While AI provides increasingly sophisticated cognitive capabilities for robots, these advances in the “brain” cannot compensate for the physical limitations that Brooks and other experts have identified. Brooks, co-founder of iRobot, calls the idea that humanoid robots will match human manual capabilities “pure fantasy thinking” and states that we are more than ten years away from the first profitable deployment of humanoid robots even with minimal dexterity. Safety concerns compound these challenges, as most bipedal robots use powerful electric motors and require pumping large amounts of energy into the system to maintain balance, which can result in sudden, forceful movements that pose injury risks. Despite the billions being invested by venture capitalists and tech companies in humanoid robotics, multiple robotics-focused VCs and AI scientists told TechCrunch that they don’t expect wide adoption for at least several years—if not more than a decade. The field will certainly advance, but Brooks predicts that deployable dexterity will remain pathetic compared to human hands beyond 2036, suggesting that the timeline for truly capable humanoid robots is far longer than many industry promoters claim.
Regardless, efforts are strong.
Drones. Drones are essentially robots because they can sense their environment, make decisions, and act without constant human control. They use sensors like cameras, GPS, gyroscopes, and accelerometers to understand where they are and what is around them. Onboard computers process that data and decide how to move, stabilize, avoid obstacles, or follow a planned route.
Like other robots, drones can operate autonomously or semi-autonomously. A human might set the goal, but the drone handles the moment-to-moment thinking, adjusting its motors in real time and responding to changes in the environment. At their core, drones are flying robots built to perceive, decide, and act in the physical world.
Advances in this limited application of robotics have been impressive.
AI-native. An AI-native business is an organization that is fundamentally built around AI from the ground up, rather than bolting AI onto existing systems as an afterthought. Unlike traditional businesses that view AI as a tool to enhance existing workflows, AI-native companies use AI for both defense and offense: defensive moves save money by outsourcing mundane tasks to AI, while offensive moves unlock new business opportunities and deliver unique value to customer.
AI-native businesses hold significant advantages over traditional companies struggling with AI integration, largely because they avoid the structural impediments that plague legacy organizations. Traditional businesses face severe challenges: 78% of enterprises struggle to integrate AI with their existing systems, and MIT research reveals that 95% of generative AI pilot programs at companies are failing, not due to poor AI models but because of flawed enterprise integration and the inability of generic tools to learn from or adapt to existing workflows. Traditional companies battle cultural resistance from employees who fear job loss, legacy system incompatibilities requiring extensive customization, talent gaps in AI expertise, and the complexity of scaling AI across diverse departments. In contrast, AI-native businesses are designed to embed AI throughout their infrastructure from the start, eliminating the data fragmentation and operational bottlenecks that come with bolt-on solutions Their ground-up approach allows them to scale exponentially without proportionally increasing headcount—a fundamental competitive advantage that traditional linear-scaling business models simply cannot match.
Related: Notion Develops Custom Agents to Power AI-First Organizations.
Debates aside, everything is converging.
