Hey, it’s Andreas.
Welcome back to Human in the Loop.

The year is winding down and the news flow is thin. People are starting to reflect on the year and prepare for what comes next - so let’s do the same here, briefly.

First of all, this feels like the right moment to say THANK YOU for reading, commenting, and sharing your perspectives throughout the year. This newsletter grew far beyond what I expected and is now close to 30,000 readers. I didn’t think that would happen when I started earlier this year in May.

Secondly, it’s also the perfect time to look ahead to 2026. Making predictions about AI is difficult - especially in a field that moves this fast. Still, having a point of view matters. Below are eight predictions that frame how I see AI evolving next year. A year from now, we’ll come back to them and see what held up.

This is also my last newsletter of the year. I’ll be back after January 1st.

I wish you and your family a peaceful holiday season. We’ll see each other again in 2026 - refreshed and ready for what’s next.

Let’s dive in.

🔮 AI: 2026 Edition
My AI Predictions for 2026

2025 was a turning point for AI. Not because all the hype was fulfilled, but because AI was tested in real environments, at scale, and under pressure. That exposed both progress and limits.

Looking ahead, I believe that 2026 will be defined by consolidation, capability, and consequence. Fewer experiments. More systems that are expected to work reliably, repeatedly, and in production.

Here is my 2026 thesis

1. Agentic AI Gets Real

2025 was widely labeled “the year of AI agents.” In reality, it was the year we learned what agents can and cannot do. We discovered that giving a model a tool is not the same as giving it a job. Agent hype ran into familiar limits: error propagation, context drift, and fragile execution.

Most agent systems work well in constrained environments - single objectives, limited context, short horizons. They break down when tasks become multi-variable, long-running, or require real trade-offs. Context degrades, decisions drift, and recovery logic is weak. That does not make agents a failure. It defines a new baseline. In 2026, the shift will be from autonomy to reliability. We will see the emergence of much more “System 2” agents - systems that use extended inference-time reasoning to evaluate, correct, and stabilize decisions before acting.

2. Scientific Acceleration Becomes Tangible

2026 will be a year of scientific acceleration through AI. Systems are beginning to handle parts of scientific reasoning that previously required deep human expertise - evaluating competing hypotheses, proposing experiments, interpreting results, and iterating with human oversight.

Early benchmarks are already emerging to measure this shift, testing expert-level reasoning across domains like physics and chemistry. The impact is not marginal productivity gains, but shorter research cycles, higher experiment throughput, and more scalable R&D operations.

3. Inference-Time Compute Becomes a Strategic Lever

Inference-time compute will no longer just be about latency and cost. It is becoming a lever for intelligence.

Until recently, models ran inference with fixed compute budgets, regardless of task complexity. The shift now is toward variable inference - allocating more compute when reasoning is needed and less when it is not. This decouples intelligence gains from constant retraining. Better outcomes increasingly come from how a model reasons at inference time, not just from model size or data volume.

Smarter systems will not come only from bigger models. They will come from better use of compute at the moment it matters.

4. (Very) Large Language Models Quietly Win Adoption

Most frontier AI models now operate at the trillion-parameter scale, unlocking capabilities that were not feasible just a few years ago. The next generation will push this frontier further, with models scaling significantly beyond today’s architectures.

These very large models will not dominate every workload. But where deep (scientific) reasoning, multimodal understanding, and creative synthesis matter, they will set the ceiling of what is possible.

Their impact will be less visible than headline demos suggest, but foundational. They will serve as the capability backbone for advanced systems, while smaller models handle execution closer to the edge.

5. (Very) Small Language Models Gain Importance

While ultra-large models push the frontier, smaller, domain-specific models have quietly gained momentum. These models, often with only a few billion parameters, are efficient enough to run on laptops or even smartphones, enabling targeted tasks without heavy compute requirements.

The advantage is practical, not theoretical: lower cost, faster deployment, and tighter integration into real workflows. In 2026, these models will continue to win where focus, efficiency, and control matter more than raw scale.

6. More Advanced Use Cases Move Up the Value Chain

In 2025, AI proved its value in areas like customer experience, simple automation, and virtual assistants. In 2026, the focus shifts toward more proactive and complex applications closer to core operations.

A clear example is AIOps. With enterprises spending 70–80% of their IT budgets on operations amid rising complexity and technical debt, the pressure to stabilize, automate, and scale is significant. When applied with discipline, AI can reduce noise, improve resilience, and lower operational load. This is one of the big use cases where adoption will accelerate in 2026.

7. Near Infinite Memory Becomes a Core Design Question

AI systems are moving toward persistent memory - the ability to retain and recall long-term context across interactions. This enables more personalized, context-aware, and continuous experiences over time.

But memory is not just a feature. In 2026, it becomes a design decision with real consequences: what is remembered, for how long, and under whose control. Done well, it can redefine customer engagement and continuity. Done poorly, it introduces new risks around trust, privacy, and failure modes.

8. Human-in-the-Loop Automation Becomes the Default

The most durable AI systems will not remove humans from the loop. They will redesign the loop.

In 2026, human-in-the-loop approaches will mature beyond prompt engineering and manual oversight. Systems will increasingly amplify expertise without requiring users to think like engineers. The focus shifts to better handoffs, clearer accountability, and tighter collaboration between human judgment and machine execution. This is where trust, adoption, and real impact converge.

In a year from now, we’ll look back at these theses and see what held up - and what didn’t.