iiinova — where your digital self keeps evolving

Pete Bachant @petebachant.me · 30/09/2026

Calkit 0.47.11 greatly improves math handling in LaTeX --> Word --> LaTeX round-tripping, including support for LibreOffice. github.com/calkit/calki... #opensource #openscience #reproducibility

A screenshot showing Calkit's LaTeX to Word to LaTeX conversion results.

041

Pete Bachant @petebachant.me · 29/09/2026

So, in the age of AI coding agents, the job of a software engineer is of course less about writing code, and more about understanding if solution concepts will work. This can be done from experience, from theory, by checking tests and prototypes for applicability, etc.

110

Pete Bachant @petebachant.me · 29/09/2026

Writing code, like manufacturing parts for machines, is technician work, not engineering. Engineering is simply predicting if something is going to work. Both can be done by the same person, but we shouldn't confuse the two.

110

Pete Bachant @petebachant.me · 24/09/2026

How to prevent this problem? Instrumentation stamping artifacts with cryptographically signed attestation for provenance? That could be useful for more than just fraud detection. #metascience

130

Pete Bachant @petebachant.me · 10/09/2026

Maybe problems solved with AI should be attributed to every author whose work is present in the training data, not the company that trained the model.

000

Pete Bachant @petebachant.me · 09/09/2026

This allows the project lead hacker types to keep version-controlled .tex files as the source of truth and let others contribute by emailing Word docs. Tutorial: docs.calkit.org/tutorials/la...

docs.calkit.org

Collaborating on a LaTeX document using Microsoft Word for feedback - Calkit

000

Pete Bachant @petebachant.me · 09/09/2026

So, I put together a CLI that converts LaTeX PDFs using Microsoft Word itself (not Pandoc) to keep styling close, inserting special bookmarks to enable idempotently merging changes and comments back into the .tex source.

100

Pete Bachant @petebachant.me · 09/09/2026

In the past I've annoyed collaborators by insisting they learn Git and LaTeX. It's an uphill battle and pretty non-inclusive to be honest. Amazingly, not everyone wants to learn software engineering tools to make a few edits to a paper.

100

Pete Bachant @petebachant.me · 09/09/2026

Emailing Word documents back and forth, despite its many drawbacks, is still one path of least resistance for collaborative writing.

100

Pete Bachant @petebachant.me · 05/09/2026

How do you force an AI agent to work reproducibly? Trying to figure that out myself as I have Claude experiment with CUDA kernel optimization. Not perfect, but having it work with a pipeline with definitive staleness detection is helping: calkit.io/petebachant/... #agenticai #reproducibility

calkit.io

Calkit

010

Pete Bachant @petebachant.me · 29/08/2026

I'd be interested to see/hear some too. In my experience it's much easier to produce outputs but LLMs are not good at making them into arguments. They make ones that look good on the surface, but the human has to pay careful attention to ensure the story makes sense.

021

Pete Bachant @petebachant.me · 29/08/2026

In these cases the paper is highly dependent on the code and data, so they need to all be shipped as one unit. Quite the opposite of redundant!

110

Pete Bachant @petebachant.me · 29/08/2026

Many scientific papers are tightly-coupled to the code, data, and environments that produced them, and therefore the paper itself fails to provide much standalone value. Failure to reproduce or replicate is a direct consequence of this coupling. #openscience #replicationcrisis #metascience

110

Pete Bachant @petebachant.me · 28/08/2026

So now you can edit figures and write about them in a paper in the same app/project on calkit.io and in the repo on your local machine.

calkit.io

Calkit

000

Pete Bachant @petebachant.me · 28/08/2026

Calkit now has online figure and notebook editors that run in the browser with WASM/Pyodide, but produce code and environments integrated into a local-first reproducible pipeline. #opensource #openscience #reproducibility

120

Pete Bachant @petebachant.me · 21/08/2026

An AI bot chiming in on a metascience discussion?

110

Pete Bachant @petebachant.me · 18/08/2026

Thinking of revamping the Calkit onboarding a bit. What use case is more valuable to you: 1. You have a research project in progress that's a bit of a disaster and you'd like to get it under control 2. You're starting a new project and want to follow best practices from day one #reproducibility

001

Pete Bachant @petebachant.me · 13/08/2026

That is not to say AI generated prose is useless. I think it does fill in gaps accurately in many cases, especially when you're explaining something that already exists or has been done before.

000

Pete Bachant @petebachant.me · 13/08/2026

We are definitely going to need some social norms around asking people to read things generated with AI. It at least needs to be explicit, and it probably should be okay to respond with: can you explain in your own words?

100

Pete Bachant @petebachant.me · 11/08/2026

Calkit Chrome extension just dropped. Collect references and notes directly into a BibTeX file in a GitHub repo from journal websites and arXiv, sync figures and results into Overleaf, and review DVC-tracked files and LaTeX diffs directly on GitHub. #opensource #openscience #reproducibility

Calkit Chrome extension showing a LaTeX diff on GitHub.

130

Pete Bachant @petebachant.me · 04/08/2026

Doing open science gets you more citations (arxiv.org/pdf/2607.00546) However, does it also result in less individualistic innovation (less notoriety, more scooping, worse tenure outcomes)? #openscience #publishorperish

arxiv.org

110

Pete Bachant @petebachant.me · 04/08/2026

Leaders are responsible for vision and strategy, i.e., prioritizing problems and giving insights into avenues for solutions. If you're handing off fully-baked solution specs to your engineers you're wasting their talent and holding them back from growing.

120

Pete Bachant @petebachant.me · 03/08/2026

There's an interesting distinction between "this can run again" and "I can prove that I ran this and here's the output" and the latter is becoming more important with increasing AI usage #agenticai #aiforscience #reproducibility

020

Pete Bachant @petebachant.me · 01/08/2026

How many more times can I hear "agentic AI" this week before I quit using computers for the rest of my life challenge

010

Pete Bachant @petebachant.me · 01/08/2026

Dodge Rams are also highly likely to be driven drunk IIRC

001

Pete Bachant @petebachant.me · 29/07/2026

Agreed. Institutions are good and valuable, actually, when they help lower the cognitive load of assessing trustworthiness.

000

Pete Bachant @petebachant.me · 29/07/2026

AI agents, like humans, should be creating pipelines. They should not be parts of pipelines.

010

Pete Bachant @petebachant.me · 28/07/2026

Moving steadily towards a free, open-source, vertically-integrated research platform, calkit.io now has a project-based reference manager that can sync bidirectionally with Zotero. #openscience #opensource #reproducibility

A screenshot of the Calkit reference manager.

020

Pete Bachant @petebachant.me · 27/07/2026

New feature on calkit.io: attach evidence to answers to research questions and quickly verify that they're still up-to-date given their provenance. Example project: calkit.io/calkit/examp... #reproducibility #openscience #opensource

A screenshot of Calkit's questions section.

041

Pete Bachant @petebachant.me · 26/07/2026

Unifying development and operations in the same team/repository was a boon for software. Would science see similar gains from moving its various stages/functions closer together? #devops #openscience #opendata

020

Pete Bachant @petebachant.me · 19/07/2026

In my opinion the solution is to not allow LLMs to produce derived artifacts directly. They can produce code, but not figures, for example, and the code and data must be shared. We don't complain that we can't reproduce code written by humans as long as it's open, despite a "closed" human brain.

010

Pete Bachant @petebachant.me · 18/07/2026

"utils" modules are bad because they show you value DRY (don't repeat yourself) more than modularity, and you end up with tight coupling and low cohesion #softwareengineering #swe #softwarearchitecture

110

Pete Bachant @petebachant.me · 17/07/2026

Basically, researchers are left to build their own scaffolding to connect things together, and mostly they don't, so they resort to manual data transfer, reruns, etc.

010

Pete Bachant @petebachant.me · 17/07/2026

There are lots of great computational tools out there for doing different stages of research, but they're poorly integrated with each other, which is a source of errors and inefficiency.

110

Pete Bachant @petebachant.me · 17/07/2026

Totally. There are a few groups out there who do it for the sake of measuring success rates though, e.g., osf.io/ezcuj/overview

osf.io

OSF

000

Pete Bachant @petebachant.me · 14/07/2026

How well does failure to reproduce correlate with failure to replicate? #metascience #reproducibility

100

Pete Bachant @petebachant.me · 11/07/2026

If your collaborators don't want to learn Git/GitHub but you don't want to silo paper writing away from your code and data, calkit.io now has a LaTeX editor and WASM compiler :) #opensource #openscience #reproducibility

A screenshot of Calkit's in-browser LaTeX editor.

040

Pete Bachant @petebachant.me · 08/07/2026

If you're building a writing tool for scientists, it's pretty important to make it easy to connect to visualization/analysis tools. That is, if you make users manually upload their data or figures, that's bad.

000

Reposted by Pete Bachant

Konrad Hinsen @khinsen.net · 06/07/2026

New blog post: "Conviviality in computational science" blog.khinsen.net/posts/2026/0... "Conviviality matters for science ... if you want to derive knowledge from your work, you need to know exactly what you are doing, and that includes a detailed understanding of your tools." 🧪 #metascience

blog.khinsen.net

Konrad Hinsen's blog

34416

Pete Bachant @petebachant.me · 02/07/2026

Instead of compiling a replication package or repro pack at journal article submission time, imagine if you worked inside of a repro pack the whole project... #reproducibility #openscience

010

Pete Bachant @petebachant.me · 01/07/2026

@plos.org's concept of the "knowledge stack" in science is interesting from an architectural perspective. Is any layer truly valuable on its own? I'd argue the full stack is the most valuable unit, which means you should be shipping the whole thing with every study. #openscience #opensource

010

Pete Bachant @petebachant.me · 29/06/2026

If you know what individual to ask for help in a separate team, you are probably tightly-coupled, and maybe should be part of the same team

000

Pete Bachant @petebachant.me · 28/06/2026

The Calkit VS Code extension now allows viewing interactive @plotly.com figures, and includes a gallery view.

000

Pete Bachant @petebachant.me · 24/06/2026

If a human is supposed to read it, it shouldn't be written by AI

120

Pete Bachant @petebachant.me · 23/06/2026

For those who manage research groups: Do you tend to be more individual-focused, with each member getting their own independent project, or team-focused, where individuals need to find their niche to contribute to a larger collective goal?

010

Pete Bachant @petebachant.me · 15/06/2026

Been spending some time on the Calkit VS Code extension. Here's a screenshot using it with Claude to profile CUDA kernels in SLURM jobs and keep track of Nsight reports for each code change: #aiagents #cuda #gpu #claude

A screenshot of the Calkit VS Code extension.

130

Pete Bachant @petebachant.me · 07/06/2026

AI agents will happily create figures for you. They'll even export some source code. But can you trust that they did what they said they did? Here's a tutorial on how to get your AI agents to work reproducibly: docs.calkit.org/tutorials/ai...

docs.calkit.org

Declarative, structured prompting to make AI agents work reproducibly - Calkit

030

Pete Bachant @petebachant.me · 02/06/2026

I'm not a fan of splitting up code into a bunch of tiny functions as a default practice. It's like creating a bunch of minor characters in a story. It should only be done if those characters really add something, because each one will take away some attention. #swe #softwareengineering #dev

000