I want to start by wishing you a Happy New Year! It’s been quite a year for the Rising Researcher Academy newsletter.

When I started this about a year ago, I wasn’t sure what would resonate.

I don’t have a machine learning PhD.

I wasn’t one of the authors of “Attention Is All You Need.”

I’ve never worked at a frontier AI lab like OpenAI or Anthropic.

I’m just a clinical researcher who decided to write about what I was learning, building, and thinking about. And hope it would be useful.

Turns out, a lot of you did find it useful.

The newsletter has grown to 16,000+ subscribers, which still feels surreal to type.

Thank you for being here and for all your support.

My goal for the newsletter is to help thousands of academics (people like me) keep up with the technology and use it to improve their careers, their research, and the way they work.

Because the pace of AI progress has been… wild.

Much faster than most of us expected.

And here’s the part I don’t think enough people say out loud:

Even if AI progress halted completely today (and there’s no real evidence that it will), we already have enough capability to change how research and work gets done.

I think the next decade will feel like a before-and-after moment for knowledge work.

Here are my top 5 (no-hype) predictions for 2026:

1. Academic adoption of Nano Banana Pro

Google Gemini Nano banana was one of the most incredible, paradigm-shifting models of 2025. It is a “Thinking” image model for advanced outputs and precise control. It includes advanced text rendering, precise editing controls (lighting/camera angle/aspect ratio), and 2K resolution for professional use (plus better “world knowledge” for things like infographics and diagrams) (Google Gemini product overview, 2025).

Until now, image generation was unreliable – something that instagrammers could use. But nowhere close to what you would want in academia.

Pretty pictures.

Unreliable text.

Hard to control.

But the underlying tech has matured fast.

Modern image models are built on diffusion foundations that made high-quality generation practical at scale (e.g., latent diffusion models) (Rombach et al., Latent Diffusion Models, 2022). Then the field started solving the control problem. How do you constrain generation so outputs are actually usable and reliable?

Two examples:

• Instruction-following image editing: InstructPix2Pix showed that you can take a real image + a natural language instruction and generate a coherent edit quickly—without per-image fine-tuning (Brooks et al., InstructPix2Pix, 2023).
• Spatial control: ControlNet showed that you can condition a diffusion model on structure (edges, pose, segmentation, depth) to keep layout and meaning stable while editing style/content (Zhang, Rao & Agrawala, ControlNet, 2023).

Nano Banana Pro feels like that trajectory made product-grade: “dial in every detail,” edit locally (including by drawing), render text more reliably, and output higher-resolution assets (Google Gemini product overview, 2025).

And suddenly, it doesn’t feel like a toy anymore.

For academics, it can help as:

• a figure and schematic ideation engine
• a visual abstraction tool
• a way to turn a messy concept into something that communicates

I expect we’ll see downstream effects in how papers are written and how ideas are presented:

• cleaner visual storytelling in manuscripts
• faster iteration on figures and study schematics
• fewer bottlenecks where “we know what we mean” but can’t show it clearly

The catch (and it’s a big one): publication ethics.

Right now, major publishers still restrict generative AI images in most circumstances.

• Springer Nature explicitly says it does not allow the inclusion of generative AI images in its publications (Springer Nature editorial policies, accessed 2026), and Nature Portfolio cites unresolved legal and research integrity issues as a reason they’re “unable to permit” AI-generated images/videos for publication (with limited exceptions) (Nature Portfolio AI editorial policy, accessed 2026).
• Science journals state that AI-generated images and other multimedia are not permitted without explicit editorial permission (Science Journals editorial policies, accessed 2026).

So: I’m not saying “everyone will publish AI images in 2026.”

I am saying academics will adopt Nano Banana Pro-like tools heavily for thinking, drafting, iteration, and communication, and journal policy will be forced to evolve (slowly) as both capabilities and disclosure norms mature.

ICMJE’s 2024 updates explicitly added guidance on how work done with AI assistance should be acknowledged (ICMJE updated recommendations, 2024). It’s doesn’t say anything about images specifically. But we will hopefully see this in the 2026 update.

2. On-the-job training and domain specificity will become the bottleneck (and the focus)

We’re going to see more attempts at on-the-job training for AI systems.

Because right now, this is the real bottleneck for adoption.

(Yes, ChatGPT and Claude have the memory features – very helpful but still in a primitive state and not very reliable right now.)

A simple example (something that one of my colleagues inquired about):

Someone wants an AI agent to help clean a dataset.

Sounds straightforward.

Until you realize:

No two datasets are the same.

Yes, there are general rules for cleaning.

But in practice, every dataset has its own “local logic”:

• how values were entered
• what missingness actually means
• how a clinic labels things
• what a given variable represents in that specific workflow

Data cleaning is not just code.

It’s judgment.

Humans can generalize across datasets because we carry context across experiences.

We adapt.

We learn the quirks.

We build intuition.

Most AI systems today still struggle with that kind of generalization.

This is the “domain shift” problem and it’s not theoretical.

Identifying a macrophage in a slide sounds like a classic machine learning problem.

But real lab workflows vary a lot:

• sample prep differences
• staining differences
• slide preparation differences
• preservation differences
• institutional quirks that change what “normal” looks like

A human can walk into a new lab and adapt over weeks.

Current AI struggles there.

Same story in radiology.

The decade-old promise of AI solving radiology hasn’t quite borne out.

External validation of deep learning diagnostic models has been a challenge. The vast majority of external validation studies have shown diminished performance on external datasets, sometimes with substantial decreases (Yu et al., external validation systematic review, 2022).

That means models that look impressive in one environment often degrade when they hit another environment.

Different scanners.

Different protocols.

Different patient populations.

Different labeling practices.

This is also why the famous 2016 prediction (“stop training radiologists”) aged poorly.

Not because AI didn’t improve.

Because real-world medicine isn’t a Kaggle competition.

It’s messy.

It’s contextual.

And “generalizing into the wild” is the hard part.

So in 2026, the focus won’t just be “build an agent.”

It’ll be:

How do we train it inside the messy reality of a domain without it breaking the moment it sees a different workflow?

3. AI adoption will accelerate by reducing entropy

We’ll see more real-world adoption not because workflows magically get cleaner…

…but because organizations will start making workflows cleaner on purpose.

To delegate reliably to AI, your workflow needs to be:

• simplified
• less variable
• more deterministic

When we were building Research Boost (an AI writing assistant for clinical researchers), this showed up immediately.

Writing is high-entropy.

It’s hard to break into a fixed format.

The path to “AI that writes well” isn’t just better models.

It’s better structure.

Clear constraints.

Explicit rules.

Defined inputs.

Defined outputs.

If you work in clinical research, you’ve already lived this logic.

We reduce entropy all the time.

That’s why reporting guidelines exist.

And the AI-specific versions are an even stronger signal.

SPIRIT-AI (Nature Medicine, 2020) was created because AI interventions need protocol-level transparency: what the model is, how it’s used, what data goes in/out, how humans interact with it, and how error cases are handled (SPIRIT-AI extension, 2020).

That’s entropy reduction.

Not for aesthetics.

For reliability, interpretability, and bias control.

We have evidence that “vague” tasks get brittle quickly.

METR’s work on measuring AI capability on realistic, longer-horizon tasks shows that success rates drop as task length/complexity increases. Models can look strong on short tasks while remaining unreliable on longer, messier ones (METR long-task evaluation, 2025).

So in 2026, reliability will be the competitive edge.

And the way to get reliability, in the real world, is to reduce entropy.

There’s of course a balance here (you don’t want to strangle creativity).

But in research workflows (manuscripts, grants, protocols) constraints are often what make progress possible.

4. This will be the year of vertical agents

If I had to make one singular prediction:

Agents will get 100x more useful.

We’re entering the era of vertical agents.

General-purpose chatbots are great.

But the real economic value comes when the system understands a specific role deeply (vertical integration):

• Clinical research.
• Trial ops.
• Radiology workflows.
• Grant writing.
• Data QC.
• Manuscript workflows.

In practice, these agents may orchestrate multiple LLM calls behind the scenes → chained together into a workflow (a DAG) to balance quality, speed, and cost.

They will provide an application-specific user interface for the human in the loop.

That’s how Research Boost was built: a focused system built around the real workflows of clinical researchers so you need less and less prompting over time.

Not because it’s “smarter.”

Because it’s more aligned.

More aware of the constraints and steps that matter.

And I don’t think agents need full autonomy to create value.

The biggest wins will come from semi-autonomous agents that handle tightly scoped tasks reliably with humans in the loop.

Because again: reliability is the bottleneck.

METR’s long-task evaluations are a reality check here: getting to “50% success on multi-hour tasks” is still non-trivial, and brittleness remains a core limitation (METR long-task evaluation, 2025).

So the 2026 shift will not be “agents take over.”

It will be:

• agents become normal in narrow domains
• organizations redesign workflows so agents can actually execute
• evaluation/benchmarking becomes less demo-driven and more outcome-driven.

5. SEO will keep dying, and AEO will become the new battleground

First some quick definitions:

• SEO (Search Engine Optimization) = how you optimize your website/content to show up in Google search results.
• AEO (Answer Engine Optimization) = how you optimize your content so it shows up inside AI answer engines (ChatGPT, Claude, Perplexity, Google’s AI Overviews) when people ask questions.

If SEO was about ranking “10 blue links,” AEO is about being included in the answer.

Search behavior has changed.

And it’s not subtle.

People don’t say:

“Let me Google that.”

They say:

“Let me ChatGPT that.”

This has shown up in my real life in a way I didn’t expect.

More than once, a patient has told me:

“Do you know how I found you? I asked ChatGPT for a psoriatic arthritis expert, and it recommended you.”

That is the new reality.

And it creates what I think of as a binary outcome problem:

Either you show up in the AI’s answer…

Or you might as well not exist.

There’s no “page 2.”

No “rank 11 but still gets some traffic.”

Here’s the evidence that something structural is happening:

• Pew Research Center analyzed Google usage and found that when an AI summary appears, users click traditional links less often (8% of visits with AI summaries vs 15% without) (Pew Research Center analysis of AI summaries, 2025).
• Similarweb reports that in 2025, nearly 80% of searches that trigger Google’s AI Overviews end without a click (Similarweb on zero-click searches & AI Overviews, 2025).
• Industry commentary and research summaries are increasingly describing organic traffic drops (often in the 20–40% range) in sectors where answer engines absorb informational queries (Digiday analysis on AI referral traffic vs zero-click search, 2025).

So yes: SEO is not exactly “dead” (yet).

But the old mental model is.

The default behavior used to be:

Search → click → read → decide.

Now it’s:

Ask → get answer → maybe click if you need depth.

That’s where AEO comes in.

Answer Engine Optimization is the practice of shaping your content so it shows up inside AI-powered answer engines like ChatGPT, Claude, and Perplexity.

For academics, the takeaway is simple (and uncomfortable):

Maintaining an online persona is becoming more important, not less.

Not for vanity.

For discoverability.

For credibility.

And for being legible to the systems people increasingly consult first.

(How to make your online persona and your research AEO friendly likely deserves its own post.

I plan to write more about it in the coming weeks.)

Special mention: robotics is going to creep into labs faster than we think

We’re going to see more robotics.

Not just humanoids.

Lab robots.

Automation for repetitive steps.

Workflow assistance.

Sample handling.

Nature’s “tech to watch” coverage of self-driving laboratories frames this as a real trend: automation plus AI isn’t just speeding up chemistry and materials science, it’s changing how experiments are planned and executed (Nature: self-driving laboratories to watch, 2025).

2025 felt like the year of training and experimentation.

I think 2026 is where we’ll see more implementation – still rudimentary, still uneven, but real.

I didn’t include this in the main 5 because I don’t think production-ready robotics for clinical research workflows is “right now.”

But it’s coming.

And it will matter.

💬 If an AI agent could do one thing for you perfectly (and you could only choose one), what would you delegate: screening abstracts, cleaning variables, drafting Methods, or polishing figures?

PROMPT OF THE WEEK

A mentee asked me last week:

“What skill would you recommend I pick up in 2026.”

I thought hard about it.

My answer was:

Learn how to learn new skills fast.

And in 2026, that starts with learning how to use AI well.

I created an updated “getting started with AI” guide for 2026.

Get it here (free):

How to start using AI for research today Guide

If you do one thing at the start of this year, make it this:

Set up your AI properly (follow the steps in the guide).

Sharpen your saw.

Then build.