From Pair Programming to Scientific Discovery:
AI as Research Partner

ORF-RE11 Project Team Meeting Keynote

Graham W. Taylor

[email protected]

University of Guelph / Vector Institute

March 6, 2026

Talk Overview

Part 1: AI for Biodiversity Research

• Our team’s work
• Connection to bioinformatics and biotechnology

Part 2: The Changing Landscape

• AI agents in software development
• Preparing for the workforce
• Practical integration strategies

Since I’ve been working in the AI for biodiversity space, I want to quickly share some of my team’s most important work and discuss how it relates to the symposium.

But then I’ll spend most of my time focused on how AI is changing the broader landscape and what that could mean for your careers.

AI for Biodiversity ↔︎ Biodiversity for AI

AI → Biodiversity

• Earth faces unprecedented biodiversity crisis
• <10% of ~20M species scientifically named
• Explosion of multimodal data:
  • 🧬 DNA barcoding
  • 📸 Imaging
  • 🎤 Bioacoustics
• Shift from lab to “in-the-wild”
• Foundation models across modalities

Biodiversity → AI

• Complex, hierarchical data structures
• Long-tailed species distribution
• Fine-grained recognition challenges
• Out-of-distribution performance
• Continual learning requirements

Let me start with why I think AI and biodiversity is such an important and interesting place to study for someone with my background. I come from deep learning and neural networks - I’ve done extensive work on vision and time series models, and more recently I’ve been getting into language models, particularly DNA sequence models.

Earth’s biodiversity faces an unprecedented crisis, with scientists warning of a sixth mass extinction. This decline threatens ecosystem health and human well-being, yet our ability to intervene is limited by vast knowledge gaps. Less than 10% of an estimated 20 million multicellular species have been scientifically named. What’s particularly exciting is that there’s an explosion of multimodal data being generated - DNA barcoding, imaging, and bioacoustics - shifting from lab-based to “in-the-wild” observations. We’re seeing the development of foundation models across the vision modality, the DNA modality, and the language modality, and it’s at this intersection where my team has been working.

[TRIGGER FRAGMENT] But here’s what makes this research area especially compelling: while AI promises to mitigate the biodiversity crisis, biodiversity science reciprocally offers a sandbox for tackling fundamental AI and machine learning problems at scale. The complex, hierarchical nature of biodiversity data presents challenges for core ML research. The long-tailed distribution of species abundance provides a real-world context for learning from imbalanced datasets. Taxonomic classification challenges our ability to distinguish subtle differences in images or acoustic data. Ecological systems’ dynamic nature demands innovations in out-of-distribution performance and continual learning. This bidirectional relationship means we’re not just applying AI to solve biodiversity problems - we’re advancing the fundamental science of AI itself.

Key Statistics

• 5 million specimens
• Images + DNA barcodes + taxonomic labels
• Challenge: Incomplete taxonomic annotations
  • Many labels only to family/genus level
  • Expert annotation is difficult

Now let me move on to actually putting the data together for machine learning and building models for biodiversity monitoring and species discovery. In collaboration with Guelph’s Centre for Biodiversity Genomics and researchers at SFU and U Waterloo we developed and released the BIOSCAN-5M dataset.

[TRIGGER FRAGMENT] It has impressive scale - we’re talking about 5 million specimens with paired images, DNA barcodes, and taxonomic labels.

One important challenge I want to highlight is that taxonomic labels don’t always go to the final level of taxonomic precision, which are the species. Sometimes, because it’s such a difficult problem to annotate this data, even experts can only classify to family or genus level for much of the dataset. Annotator uncertainty is a fundamental challenge in our field.

Three-Way Contrastive Learning: CLIBD

CLIBD: learning and inference

Key Innovation

• Three-way contrastive learning
• Vision ↔︎ Text encodings of taxonomic labels ↔︎ DNA
• Cross-modal retrieval

Advantage

Train with DNA, deploy with images only

Image courtesy of Gong et al. 2025

Next, I want to discuss how we go from dataset to building a model with our CLIBD project, which uses a type of contrastive learning popularized by CLIP. But while CLIP only worked with general natural images and their captions, we do three-way contrastive learning where we learn embeddings from vision, language, and DNA simultaneously.

We take examples from the BIOSCAN dataset and learn to pull together the embeddings that correspond to the same specimen while pushing apart the embeddings that correspond to different specimens. This allows us to learn representations such that we can do retrieval from any modality to any other modality.

DNA-Supervised Vision: Cost-Effective Deployment

Retrieval examples showing cost-effective deployment

Cost Comparison

• DNA sequencing: $$$
• Image capture: $

Key Insights

• DNA gives fine-grained supervision to vision
• Train expensive, deploy cheap

Image courtesy of Gong et al. 2025

Here’s an example of the trained model in-use. The images in the column on the left represent novel specimens, our “queries”. We can match that specimen’s DNA barcode to DNA barcodes in a reference database and use that to make a taxonomic classification. But we can also use it’s image to match to images that we have in the reference database. Or, because we use a shared embedding space, we can match image embeddings to DNA or vice-versa.

[TRIGGER FRAGMENT] What’s really exciting about this work is that we can supervise a vision encoder using DNA information. Even though DNA is often expensive to obtain, we can build image-based retrieval systems that only require images at inference time while utilizing DNA at training time. This makes deployment much more cost-effective.

Taxonomic RAG: Beyond Direct Classification

Example output from Module I (light blue) and Module II (light green) for an image of a yellow garden spider (Photo attribution: David Illig).

Image courtesy of Lesperance et al. 2025

The next project I’ll highlight is our work on taxonomic retrieval and caption generation. Rather than going directly from image to taxonomic label, we take a more sophisticated approach that represents a radically different way of doing taxonomic classification.

The key insight is that there’s a vast amount of textual knowledge locked up in databases and historical records. If we can get the “pixels” into the language space, we can unlock that knowledge. We generate a descriptive caption from the image, then take that text description and put it into the context of a large language model along with retrieved biodiversity text data from external sources. The language model then reasons over this evidence to provide a taxonomic classification to a level of specificity that it decides.

Handling Rare Species Through Retrieval

Taxonomic RAG Architecture: RAG-based approach for handling rare species

Note

Traditional (Vision-Only) Approach

• ✓ Great for common species
• ✗ Fails on rare taxa
• ✗ No uncertainty awareness

RAG Approach

• ✓ Handles rare species via text matching
• ✓ Leverages historical text-based records
• ✓ Knows its confidence level

Image courtesy of Lesperance et al. 2025

Here is a deeper view of the architecture.

[TRIGGER FRAGMENT] This approach is particularly powerful for rare and unknown species - while a naive vision model excels at classifying common species, our retrieval-augmented approach enables classification of rarer taxa by matching explicit textual descriptions of taxonomic features to contextual biodiversity knowledge. There’s a crucial aspect here - the model becomes aware of what it knows and what level of taxonomic reasoning it’s competent at for that particular organism.

From Research to Career Questions

“How should I be safely integrating AI into my workflows?”

While I’m genuinely excited by the progress we’re making in AI for biodiversity research, I think at a symposium like this, at this moment of radical technological change, it’s more important to address the career questions - even existential questions - that are on many of your minds.

I’ve had numerous students approach me recently asking: “How should I be safely integrating AI into my workflows?” This connects to the broader question of how AI is going to reshape the job market as you graduate and move into your careers. People are genuinely concerned about their future, and that’s what I want to spend the rest of our time discussing.

Why You Shouldn’t Dismiss AI

This is not about tomorrow.

This is about right now.

The pace of change is unprecedented.

You’ve probably been exposed to a lot of hype around AI by now, but let me motivate why you shouldn’t dismiss or ignore the advent of these systems, particularly as it pertains to your career.

The METR Study: A New Moore’s Law for AI

METR Study showing doubling time of AI task completion capabilities

Key Finding

“The length of tasks AI can do is doubling every 7 months” - METR Study

(Time Horizon 1.1 update suggests the rate may be accelerating)

Timeline Evolution

• GPT-4 era: Seconds to minutes of autonomy
• Current models (GPT-5, Opus 4.6): Hours of autonomy
• Trend: Exponential capability growth

Image courtesy of METR 2025, updated 2026

I want to show you a report by an organization called METR. I think this is one of the most sobering studies we’ve seen recently. While many studies describe the evolution of AI systems in terms of performance benchmarks, they often don’t translate to improvements in real-world work settings.

The METR study takes a different approach — it looks at the rate of doubling of the time it takes to complete industry-relevant tasks. Here’s what’s concerning: while models around the GPT-4 era could complete tasks that take humans anywhere from three seconds to a minute, the latest models like GPT-5 and Claude Opus 4.6 can handle tasks that take humans hours.

[TRIGGER FRAGMENT] The original study found capabilities were doubling every seven months. Their February 2026 update, with a larger task suite, suggests the rate may be accelerating. For those of you familiar with Moore’s Law, this is like a new Moore’s Law for AI capabilities.

It’s Not Just Software: GDPval

A fair criticism of the METR study is that it measures software tasks almost exclusively.

GDPval (OpenAI, Sep 2025) evaluates AI on real professional deliverables across 44 occupations and 9 industries — including manufacturing, healthcare, and finance.

Human experts blindly compare AI output against professional work.

Key Finding

AI output quality vs. human experts: 12% → 83% in under two years

Win Rate vs. Human Experts

GPT-4o (2024)

12%

GPT-5 (2025)

39%

Opus 4.1 (2025)

48%

GPT-5.2 (2025)

71%

GPT-5.4 (2026)

83%

OpenAI GDPval (2025) | GDPval-AA (Artificial Analysis)

Now, a fair criticism of the METR study I just showed you is that it’s overwhelmingly focused on software tasks. The arxiv paper title was even changed to include the word “software.” So you might be wondering — does this apply beyond coding?

GDPval, released by OpenAI in September 2025, directly addresses this question. It evaluates AI on real professional deliverables — documents, spreadsheets, slides, diagrams — across 44 occupations and 9 industries that collectively represent most of the U.S. economy. This includes manufacturing, healthcare, finance, and professional services — not just tech.

The evaluation method is rigorous: experienced professionals with an average of 14 years in their field blindly compare AI-generated deliverables against expert-created ones.

[TRIGGER FRAGMENT] What we see is a steep trajectory. GPT-4o in spring 2024 won or tied against human experts only 12% of the time. GPT-5 reached 39%. Claude Opus 4.1 hit 48%.

[TRIGGER FRAGMENT] GPT-5.2 with reasoning jumped to 71%.

[TRIGGER FRAGMENT] And the most recent model, GPT-5.4, reaches 83% — meaning human experts preferred the AI output or rated it equal more than four times out of five.

That’s a nearly 7x improvement in under two years, across occupations that include mechanical engineers, financial analysts, and compliance officers — not just software developers.

I should note that there’s also an independent version of this benchmark called GDPval-AA from Artificial Analysis, which runs models with agentic tools and scores them differently. The rankings are broadly consistent but the methodologies aren’t directly comparable.

Gradual Disempowerment: A Different Kind of Risk

“We might all find ourselves struggling to hold on to money, influence, even relevance. This new world could be more friendly and humane in many ways, while it lasts… But humans would be a drag on growth.”

— David Duvenaud, The Guardian, “Better at everything: how AI could make human beings irrelevant”

I want to bring your attention to a recent Guardian article by David Duvenaud from U of T on the question of gradual disempowerment. He points out that while a lot of attention has been placed on apocalyptic AI scenarios — humans suddenly losing power to AI systems — there could be much less sinister but equally disruptive scenarios.

The graphic shows three dimensions. Economically, as the METR study shows, AI is moving into territory that was the exclusive domain of human professionals — it will be hard to justify spending twice as much for human work that’s only half as good. Socially, AI companions and services are already being rated higher than their human equivalents — from AI doctors with better bedside manner to personalised counsellors and companions. Politically, Duvenaud draws on the idea of a “resource curse” — when governments can derive revenue and capability from AI rather than citizens, the incentive to invest in education, healthcare, and democratic institutions erodes.

The most unsettling part of his argument is that none of this requires a conspiracy — it happens by default if AI keeps improving.

The Competitive Reality

Organizations with Humans

• Speed: Slower
• Cost: 2x more expensive
• Reliability: Variable
• Outcome: Outcompeted

Organizations with AI

• Speed: Faster
• Cost: 50% cheaper
• Reliability: Consistent
• Outcome: Market dominance

Organizations that don’t adapt to AI will be displaced by those that do

What’s particularly concerning is the authors’ point that organizations choosing slow, expensive humans will be outcompeted by those who choose fast, cheap, reliable AI. And this isn’t some distant future scenario - it’s the logical endpoint of the exponential trend that METR documents.

The Solution: Strategic Integration

Avoiding AI accelerates displacement

Three pillars for thriving in the AI era:

• 🤝 Learn to work WITH AI
• 🧠 Maintain human capabilities
• 🚀 Stay competitive & relevant

Here’s the crucial insight from the gradual disempowerment research: trying to ignore AI or stay away from these systems won’t protect you - it will actually accelerate your displacement. The solution isn’t avoidance, but learning how to work with AI strategically while maintaining the human capabilities that remain valuable. Let me talk about how you, as young professionals, can safely and smartly integrate AI into your workflow while staying competitive and relevant.

Skills Being Automated

🔍 Web Search & Literature Review

• LLM-powered search replacing Google
• Synthesis in seconds vs hours
• Multiple sources browsed automatically

💻 Software Development

• GitHub Copilot, Cursor, Claude Code
• Coding like a junior PhD student
• From autocomplete to autonomous development

💡 Data Analysis & Ideation

• Creative tasks being automated
• Research ideation affected
• No longer “human-only” territory

When you think about our jobs as researchers, several skills that are part of our general skill set are quickly being automated, if they haven’t been automated already. Let me call out the main areas:

LLM-powered search tools are rapidly replacing traditional Google searches, delivering superior results by browsing multiple sources and synthesizing information in seconds rather than hours.

[TRIGGER FRAGMENT] Many of you know about the rapid advancement of AI coding tools like GitHub Copilot, Cursor, and agentic platforms like Claude Code.

[TRIGGER FRAGMENT] Even tasks that were firmly in the realm of creative thinking, including data analysis and proposing novel research ideas, are being affected. We’re not being spared from automation in any area.

Self-Driving Labs: AI Meets the Physical World

• AI + robotics in closed-loop experiments
• Autonomous materials discovery
• From weeks of lab work to hours

Image: Aspuru-Guzik Group, UofT | Video: Global News / Sinton Lab

Everything I’ve shown so far is about digital skills — search, coding, data analysis. But AI automation isn’t limited to the digital world.

Self-driving labs combine AI with robotic platforms to autonomously conduct physical experiments — synthesis, characterisation, and optimisation — in a closed loop. The diagram on the left shows the workflow from Alan Aspuru-Guzik’s group at the University of Toronto, where ChemOS orchestrates the entire experimental cycle: design a molecule, synthesise it, characterise it, and feed results back to the AI for the next iteration.

Aspuru-Guzik leads the Acceleration Consortium, a major initiative at UofT focused on accelerating materials discovery. I saw him present in February at Vector’s Remarkable conference and it was eye-opening. Their lab is also investing heavily into agentic systems — they’ve developed an agentic ecosystem called El Agente that orchestrates AI agents to plan and run experiments autonomously. The video here is from Global News, showing the Sinton Lab — also part of the Acceleration Consortium — doing autonomous chemistry experiments.

This is directly relevant to what many of you do — discovering and optimising new composite materials. The same AI that’s automating software development is now automating physical experiments in chemistry and materials science.

Web Search Revolution

“I can feel my usage of Google search taking a nosedive already. I expect a bumpy ride as a new economic model for the Web lurches into view.” - Simon Willison

Two Modes of AI-Powered Search

• Quick Search: 15-30 seconds, 20-40 sources
• Deep Research: 15-30 minutes, comprehensive report

GPT-5.4 Thinking search response

Deep research question interface

Deep research comprehensive report

Let me start with web search, since it’s probably something we all do dozens of times every day without thinking about it.

There are two recent articles that have caught my attention on this topic. Simon Willison, who’s a veteran web developer and co-creator of Django, has been writing extensively about AI’s impact on web technologies. Cal Newport, a computer science professor at Georgetown who’s written influential books like “Deep Work” and “Digital Minimalism,” has been analyzing technology’s broader effects on knowledge work. Both are arguing that web search is flying under the radar but will cause massive disruption because so many people’s jobs are based on web-based research.

I want to share what Willison discussed — specifically, the difference between deep research and standard LLM web search tools. I can tell you that I myself have completely migrated off of Google search and now use OpenAI’s GPT-5.4 with web search tools when I’m looking something up.

[TRIGGER FRAGMENT] Here’s an example of a quick search — I asked about LoRA hyperparameters and got a detailed, sourced answer in about 30 seconds. This is the kind of lookup that used to mean opening a dozen browser tabs.

[TRIGGER FRAGMENT] Deep research is a different beast. You give it a complex question and it asks clarifying questions before diving in — like a research assistant scoping a task.

[TRIGGER FRAGMENT] Then it goes away for 15–30 minutes, browses dozens of sources, and comes back with a comprehensive report with citations. This is the kind of literature synthesis that used to take a graduate student an afternoon.

Specialized Research Tools

Elicit interface for systematic reviews

Elicit - Systematic Reviews

• Extract specific attributes from papers
• High-quality sources only
• Cost: $$$ but streamlined

Other Tools

• Your Agent + Semantic Scholar API
• ResearchRabbit
• Connected Papers

Key Benefit

Quality UI/UX, easy export to other formats

Image courtesy of Elicit

Let me quickly mention specialized products for scientific research. For example, there’s a platform called Elicit, designed for systematic reviews. It has a fairly high cost, but if you need to extract specific and structured data from scientific papers and want to limit your research to high-quality content rather than generic web content, this could be really valuable for you.

Agentic Coding: The Paradigm Shift

Evolution of AI Coding Tools

1. Autocomplete (GitHub Copilot) → seconds of autonomy
2. Pair Programming (Cursor) → minutes of autonomy
3. Agentic Development (Claude Code) → hours of autonomy

And the result?

“Catnip for programmers”
- Armin Ronacher

Pairing with Zed AI agent in synchronous mode

Now let me move on to software development. As researchers or practitioners in biotechnology, I suspect almost all of you are writing software. Many of you are probably getting exposure to agentic software development.

What we’re seeing is a fundamental shift in how we write code. Armin Ronacher, the creator of the Python Flask framework, describes this as moving away from AI simply auto-completing your thoughts to something much more powerful - real-time collaboration between human and AI. You’re working together with an agent that can break down complex tasks, execute them step by step, and work autonomously for extended periods while you oversee and guide the process.

A friend called this “catnip for programmers,” and Ronacher says it really feels like that - incredibly energizing and addictive in a way that traditional coding tools aren’t.

Two Approaches: Synchronous vs. Asynchronous

Synchronous (Cursor)

• Work in IDE
• Seconds to minutes
• Pair programming
• Direct oversight

Asynchronous (Claude Code)

• Work via issues/PRs
• Minutes to hours
• Project management
• Multiple agents

Videos courtesy of Simon Willison and Kushagrasikka

This brings us to an important distinction in how these tools work. If you’re familiar with Cursor or GitHub Copilot, you’re experiencing one approach - pair programming where you’re working inside an IDE with code completion and chat interfaces. You’re overseeing every step, maybe letting the AI work for seconds to minutes at a time.

Developers are excited about Claude Code because it represents a different paradigm - it de-emphasizes the editor interface entirely. You can let agents work for minutes to hours on substantial tasks, potentially running multiple agents simultaneously. It’s less like pair programming and more like project management.

However, Claude Code can also be used in the synchronous/pair-programming mode. And I’ve included videos here to show the tool being used in these two very different ways.

The Future: Nerd Managers

Workflow Evolution

1. Open GitHub issue
2. Agent works autonomously
3. Returns with PR
4. Review and merge

Future Vision

“Developers will be empowered to keep work queues full in large fleets of coding agents” - Steve Yegge

Video courtesy of All Hands AI

There’s an interesting recent interview between Jack Clark from Anthropic and Tyler Cowen that I recommend, where Jack Clark describes a future of “nerd managers” - people who are effectively managers of AI systems, babysitting agents that are spinning away and responding to them when they come to you, almost like managing a team of human interns. Ethan Mollick has remarked that organizations will need to update their org-charts to reflect human and machine collaboration.

In this video from All Hands AI, Graham Neubig, who is also a professor at CMU demos his own development workflow, which is largely GitHub-based rather than in an IDE. He’s effectively opening issues, describing bugs or features he needs help with in plain language. The agent takes that over and works away while he goes and opens up another issue for another agent. He can interact with an agent if it gets stuck, but the idea is that the agent returns with a pull request in GitHub with its completed work, which he can then review and merge. He’s typically overseeing 3-5 agents at a time.

What skills are needed for agentic development?

Graham Neubig from All Hands AI suggests these key skills:

1. 🏗️🥧 Strong architecture skills and taste
2. 💬 Communication
3. 📖 Code reading
4. 🔄 Multitasking

I like that Neubig is documenting his workflow. But I like even more his comments about what skills are really needed for agentic development. He talks about strong architecture skills and taste, communication, code reading, and multitasking. Of these skills, taste is really special, but I think it’s the most nebulous.

Taste is essentially the ability to discriminate between good and bad outputs, to recognize quality and appropriateness in AI-generated content. It’s a fundamentally human skill that becomes crucial when you’re managing AI systems rather than doing all the work yourself.

Taste and RLHF

Figure 1

Imitation (SFT)	RLHF	Why the difference matters
Model copies full human output distribution. That includes occasional mistakes and mediocre phrasing. Ceiling ≈ human average.	Model samples its own answers and a human simply picks the better one. Humans don’t have to create perfection, only recognise it.	Humans are far stronger critics than creators. Preference-grading lets them steer the model away from the left tail (bad answers) and pull the whole distribution rightward.
Training signal = “produce exactly what a human would have written.”	Training signal = “move towards whichever candidate the human preferred.”	Over many iterations the reward model keeps nudging the policy towards the best-judged answers, eventually surpassing the median human.

Image courtesy of Ayers et al. 2023

And there’s a really cool connection between taste and reinforcement learning with human feedback (RLHF) training, the “secret sauce” of LLMs.

Thomas Scialom, who led Llama2 and Llama3 post-training at Meta, provides compelling insight into why this discrimination ability is so powerful.

How can it be that chatbots don’t just match the physician distribution on quality or empathy, they actually go beyond it?

This is not due to supervised instruction tuning, which trains the model to match the humans’ distribution. It’s due to RLHF.

RLHF works because people are better at saying which answer is better than at writing the perfect answer themselves. By constantly rewarding the model’s best candidate and punishing the worst, you squeeze out the bad tail and push the whole distribution past typical human performance.

Developing Taste in Graduate School

Practice Deliberate Comparison

• Generate multiple AI outputs for the same task and practice choosing which is better, articulating your reasoning
• Develop familiarity with different frontier models and learn what each excels at
  • Claude
  • ChatGPT
  • Gemini
• Platform selection itself is an exercise in taste

Engage in Structured Peer Collaboration

• Seek out reviewing opportunities for conferences and journals to practice evaluating others’ work
• Participate in code review with lab mates or open source projects to develop technical judgment
• Don’t work in isolation - co-author papers with both senior and junior students, getting comfortable with giving and receiving feedback

Seek Active Mentorship

• Ask experienced researchers to walk through their decision-making when they assess quality
• Submit your own quality judgments to mentors for validation and refinement

No “Taste 101” course - learn by doing

Taste isn’t traditionally taught directly - it’s gained implicitly. There’s no course in developing your taste. I notice this with students coming into master’s or PhD programs. Often when they’re collecting references and producing citations in early paper drafts or doing literature reviews, they’re not really discriminating among the quality of publication venues - where those works are published, or the authors or groups behind those papers.

So how are you supposed to develop this skill? Here are some concrete approaches:

Practice deliberate comparison and platform selection. When working with AI tools, generate multiple outputs for the same task and practice choosing which is better, articulating your reasoning. Develop familiarity with different AI platforms - Claude, ChatGPT, Gemini - and learn what each excels at. This platform selection itself is an exercise in developing taste.

Engage in structured peer collaboration. Seek out reviewing opportunities for conferences and journals. If that’s not possible you can paper review sessions where everyone reads the same paper, writes individual reviews, then discusses differences in judgment. Participate in code review with lab mates or in open source projects. Don’t work in isolation - co-author papers with both senior and junior students, getting comfortable with giving and receiving feedback.

And if you have the opportunity to work with experienced researchers directly, ask them to walk through their quality assessments with you. Have them give you feedback on your own quality judgments.

Are Junior Developers in Trouble?

MYTH: Junior developers are doomed

REALITY: You’re best positioned to succeed

• ✓ Quick to adopt AI-driven workflows
• ✓ Treat AI tools as on-the-job training
• ✓ Unburdened by legacy tooling

“Junior devs are vibing. They get it. The world is changing, and you have to adapt. So they adapt!”

“It’s not AI’s job to prove it’s better than you. It’s your job to get better using AI.” - Steve Yegge, “Revenge of the Junior Developer”

There’s been a lot of negative talk about career prospects for young professionals entering fields adjacent to bioinformatics and biotechnology. Are you all doomed? Is AI going to eliminate entry-level positions before you even get started?

I think this narrative is not only wrong, but backwards. Let me explain why I believe you’re actually in the best position to thrive in this new landscape.

Here I lean on an essay by Steve Yegge from Sourcegraph called “revenge of the junior developer”. His key argument is that it’s actually the junior developers who are positioned well for success in the AI era because they’re adaptable, dynamic, and embracing these new agentic platforms.

It’s actually the legacy developers who tend to be comfortable in their careers and tooling that face the most disruption by AI systems.

The AI Execution Gap

Study details:

• Prior studies found settings where LLM-generated research ideas were judged as more novel than human-expert ideas
• This study: 43 researchers, 100+ hours each
• Randomly assigned to execute AI-generated or human-expert ideas

Image courtesy of Si et al. 2025

And there’s even more good news. Recent research from Stanford reveals something called the AI execution gap, which shows where humans maintain their edge. While many studies have found that LLM-generated research ideas seem more novel than human ideas at first glance, this study went further - they actually had experts implement the ideas.

Here’s what they did: 43 NLP researchers each spent over 100 hours implementing either AI-generated or human-expert ideas, then wrote papers documenting their results. When these papers were blindly reviewed, something interesting happened.

Although LLM-generated ideas may seem more novel at the ideation stage, they suffer a larger drop in novelty, excitement, effectiveness, and overall score after execution, effectively closing—and in some metrics reversing—the gap between AI and human ideas.

Two Simple Rules

Rule 1

“AI can be used to avoid learning, and AI can be used to assist learning”

Rule 2

“It’s ok to ask AI to do things you already know how to do,
but don’t ask AI to do things that you don’t know how to do”

Summing this all up, my colleague Andrew Hamilton-Wright said to me recently, very nicely: “AI can be used to avoid learning, and AI can be used to assist learning.”

This distinction is so important that OpenAI just released “study mode” in ChatGPT, specifically designed to help students work through problems step by step instead of just getting answers. Rather than providing solutions outright, it uses Socratic questioning and guided prompts to help you build understanding. As one student put it during testing, it’s like having “a live, 24/7, all-knowing office hours.”

How do you use AI to assist with learning rather than avoid it? Another eloquent way of capturing this was articulated by one of my postdocs, Scott Lowe: “Just don’t ask AI to do things that you don’t understand.”

Because if you get AI to do something and you don’t have any clue how it was done, you’ve completely avoided a learning opportunity.

Threading the Needle

Finding the right balance between learning with AI and automating with AI

There’s a fine needle to thread here. You want to have AI help you accomplish things you don’t presently know. This hints back to the two paradigm way of using coding tools: using them as a pair programmer, reviewing their work, inspecting it, and if you don’t understand it, going back and having it explain it to you or consulting other sources - versus completely autonomous agents working for minutes or hours and checking things in, whether text or code, without thoroughly reviewing and understanding it.

This is not to say that you should never use full-blown agentic mode - just that you should be using it to automate skills that you already know how to do.

Key Takeaways

1. Unprecedented change - AI capabilities doubling every 7 months

2. Avoiding AI accelerates displacement - Engage strategically instead

3. Develop your taste - The ability to discriminate quality remains crucial

4. Junior professionals have advantages - Adaptability + experience

5. Use AI to assist learning, not avoid it - Build skills while automating

I’m wrapping up with the key takeaways of the presentation. I won’t read through the list, but if you keep this with you, I think you’ll be in an excellent position to thrive in this new AI future. The key is embracing these tools while maintaining your critical thinking and taste, using AI to assist your learning rather than replace it.

Thank You & Resources

Key Papers & Reports

• METR Report (Kwa et al.) on AI Task Completion 📄
• Kulveit and Douglas et al. on “Gradual Disempowerment” 📄
• Si et al. on “The Ideation-Execution Gap” 📄

Practical Resources

• Willison: “AI assisted search-based research actually works now” 📝
• Ronacher: “Agentic Coding” 🎥
• All Hands AI: “What will software dev look like in 2026?” 🎥
• Yegge: “Revenge of the junior developer” 📝

Contact

www.gwtaylor.ca

With that, I’ll leave you with a list of references and look forward to your questions.