iiinova — where your digital self keeps evolving

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

At the same time, many AI education approaches exacerbate the tyranny of metrics: doubling down on trying to quantify everything, working only with discrete decomposable units, or discounting the value of human relationships. 10/

010

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

Too many people romanticize the past and the experience of analog school. Getting kids off screens isn’t going to improve their lives– particularly if they only read dull passages in basel textbooks or are endlessly drilled on detached tasks. 9/

110

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

“Students across all grade levels are really hungry for meaningful adult relationships. A lot of them love their teachers but can’t get the level of attention they want from them. This is not the teacher’s fault; this is the system that we created.” 8/ hollykorbey.substack.com/p/will-tutor...

hollykorbey.substack.com

Will tutoring change school forever?

"There are just too many kids who aren't getting where they need to go," says researcher Liz Cohen

110

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

When the measure becomes the target, it ceases to be a good measure. This is Goodhart’s Law. 6/ sketchplanations.com/goodharts-law

130

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

An Australian (now former) teacher: "I’m rarely required to ‘teach’ anymore. Apparently I’m more valuable as an assessor, an examiner, a data collector. I have had to dull my once-engaging lesson sequences... It is mechanical and rigid and driven.” 5/ griffithreview.com/articles/tea...

griffithreview.com

Teaching Australia – Gabbie Stroud

I AM THIRTY-EIGHT and tired. I’m only a third of the way through my class roll, a list that hurts my heart if I study it for too long. But I know what to do with these students. I’m an excellent teach...

110

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

I grew up in the birthplace of high-stakes testing: 1990s Texas under Gov George W. Bush. I saw the precursor to No Child Left Behind: art, music, & gym classes cancelled so students could spend more time drilling tedious multiple choice worksheets. 4/

110

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

The number of 9 year olds who read for fun almost daily declined from 53% (in 2012) to 39% in 2022. Atomized reading “skills” are taught without sparking a love of reading. The part is less than the whole. 3/ @karenvaites.bsky.social

120

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

For the most popular ELA curriculums in the USA, 3rd-6th graders go the entire year without reading a whole novel. Reading has been reduced to a discrete set of tasks: decoding words, summarizing, making inferences, identifying main ideas. 2/ curriculuminsightproject.substack.com/p/why-have-b...

curriculuminsightproject.substack.com

Why have books disappeared from many ELA curricula?

Some curricula have no books – and strangely, key curriculum influencers don’t seem to care.

110

Rachel Thomas @math-rachel.bsky.social · 16/02/2026

Often debates about education are framed as non-tech versus AI approaches, but too often, AI ed tech just magnifies the same failures of traditional school. 1/ My latest post: www.fast.ai/posts/2026-0...

fast.ai

fast.ai - I Don’t Want a Learning Dashboard for My Child

What analog and AI education both get wrong

171

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

I revisited the work of Csikszentmihalyi on flow and of researchers studying gambling addiction to better understand how ~vibe coding~ impacts us. Read more in my latest post: www.fast.ai/posts/2026-0... 11/

fast.ai

Breaking the Spell of Vibe Coding – fast.ai

Sinister variations on the positive state of flow

091

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

AI coding agents produce syntactically correct code. However, they don’t produce useful layers of abstraction nor meaningful modularization. They don’t value conciseness or improving organization in a large code base. We have automated coding, but not software engineering. 10/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

Part of the appeal of vibe coding is extrapolation about how effective it will be in the future (including its ability to manage ever increasing complexity). But the tech industry has a long history of overpromising & overhyping products. 9/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

Dario Amodei said 90% of all code would be AI-written by Sept 2025. Sundar Pichai & Jeff Dean said everyone would use neural architecture search by 2023. Geoffrey Hinton said AI would replace radiologists by 2021. None of these came true. Would you bet your career on tech ceo predictions? 8/

130

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

I work at an AI company, and we use AI every day. AI is useful! However, we approach vibe coding with caution and have seen that much can go wrong. Don't completely abandon the development of your current skillset. 7/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

With dark flow we lose our ability to accurately assess our productivity levels & work quality. When developers used AI tools, they estimated that they were working 20% faster, yet in reality they worked 19% slower. Study from @metr.org 6/

130

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

Not all highly absorbed focus is flow. Researchers on gambling addiction have coined the term “dark flow”. In a 2014 interview, Csikszentmihalyi defined a concept of "junk flow". 5/

Dark Flow, Depression and Multiline Slot Machine Play
Mike J Dixon 1,2, Madison Stange 1,2,✉, Chanel J Larche 1,2, Candice Graydon 1,2, Jonathan A Fugelsang 1,2, Kevin A Harrigan 2
Author information
Article notes
Copyright and License information
PMCID: PMC5846824 PMID: 28589480
Abstract
Multiline slot machines allow for a unique outcome type referred to as a loss disguised as a win (LDW). An LDW occurs when a player gains credits on a spin, but fewer credits than their original wager (e.g. 15-cent gain on a 20-cent wager). These outcomes alter the gambler’s play experience by providing frequent, albeit smaller, credit gains throughout a playing session that are in fact net losses. Despite this negative overall value, research has shown that players physiologically respond to LDWs as if they are wins, not losses. These outcomes also create a “smoother” experience for the player that seems to promote a highly absorbing, flow-like state that we have called “dark flow”.

110

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

Vibe coding & gambling violate several characteristics needed for true flow: - lack clear clues on how well you are performing (both provide misleading losses disguised as wins) - match between challenge level and skill level is murky - false sense of control 4/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

When coding, many of us experience a state of flow: full absorption & energized focus. One's skills are adequate to cope with challenges, in a rule-bound system that provides clear clues on performance On the surface, vibe coding seems to induce a similar flow. However... 3/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

“When [I first got] hooked on Claude, I did not sleep. I spent two months excessively prompting the thing & wasting tokens. I ended up building & building & creating a ton of tools I did not end up using much...” Armin Ronacher described his experience of "agent psychosis" 2/

120

Rachel Thomas @math-rachel.bsky.social · 28/01/2026

Vibe coding is the creation of large quantities of complex AI-generated code. Executives push lay-offs claiming AI can handle the work. Managers pressure employees to meet quotas of how much code must be AI-generated... yet results are far from what was promised 1/ www.fast.ai/posts/2026-0...

fast.ai

Breaking the Spell of Vibe Coding – fast.ai

Sinister variations on the positive state of flow

1238

Reposted by Rachel Thomas

Scott H. Hawley @drscotthawley.bsky.social · 22/01/2026

I had great experiences with close reading as part of this course last fall and using the solveit platform (though you can do it with other platforms, just maybe not as easily. ;-))

011

Rachel Thomas @math-rachel.bsky.social · 21/01/2026

It is early days of building the tools for close reading with an LLM. Hopefully, the above ideas & videos provide inspiration of how LLMs can help you go even deeper in your reading! 5/

020

Rachel Thomas @math-rachel.bsky.social · 21/01/2026

To retain new information as you read, spaced repetition learning is a useful technique. Create Anki flashcards within an LLM reading dialog. These sync to the same Anki deck you use on your phone & desktop. 4/ answerdotai.github.io/fastanki/

answerdotai.github.io

fastanki

Python tools for Anki

110

Rachel Thomas @math-rachel.bsky.social · 21/01/2026

Jonno reads an academic ML paper with an LLM. He sets up context and asks for explanations. Tools allow him to explore the source code repo from within the LLM environment. He then builds a simple demo to develop intuition. 3/ www.youtube.com/watch?v=U5ak...

youtube.com

How to Actually Understand Dense Machine Learning Papers - Solveit free lesson

YouTube video by Jeremy Howard

100

Rachel Thomas @math-rachel.bsky.social · 21/01/2026

Using an LLM, @howard.fm reads Eric Ries's new book. He sets up necessary context & creates handoff notes to maintain continuity across chapters. Jeremy follows rabbit holes, seeks counterexamples, & asks how the book's principles apply to his own startup. 2/ www.youtube.com/watch?v=zIqL...

youtube.com

The Best Way to Read a Book (That Nobody's Doing)

YouTube video by Jeremy Howard

110

Rachel Thomas @math-rachel.bsky.social · 21/01/2026

Close reading is a technique for careful analysis of a piece of writing, practiced by many ancient cultures, major religions, & academic scholars. The latest fastai course experimented with using AI to go deeper when reading. 1/ www.fast.ai/posts/2026-0...

fast.ai

How To Use AI for the Ancient Art of Close Reading – fast.ai

Experiments in reading with LLMs

1194

Reposted by Rachel Thomas

Alex Crits-Christoph @acritschristoph.bsky.social · 04/06/2025

A great blog post by @math-rachel.bsky.social on this preprint describing failures of an ML approach for gene annotation: rachel.fast.ai/posts/2025-0... www.biorxiv.org/content/10.1... "how challenging (or even impossible) it can be to evaluate AI claims in work outside our own area of expertise"

rachel.fast.ai

Rachel Thomas, PhD - Deep learning gets the glory, deep fact checking gets ignored

an AI researcher going back to school for immunology

32010

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Claims that AI will cure cancer are often used as superficial marketing, ignoring how AI could further disempower patients (whose expertise is already disregarded by medical system) Hiding health-related data from the public does not improve the lives of patients. 8/ rachel.fast.ai/posts/2024-0...

rachel.fast.ai

Rachel Thomas, PhD - “AI will cure cancer” misunderstands both AI and medicine

an AI researcher going back to school for immunology

060

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Promises about what AI can achieve with electronic health records must be tempered with the awareness that the data within is too often biased, incorrect, or missing. (studies on some of the diagnosis delays, pain mismanagement, and biases that are recorded as fact in medical data) 7/

Gender inequalities in the promptness of diagnosis of bladder and
renal cancer after symptomatic presentation: evidence from
secondary analysis of an English primary care audit survey a
Georgios Lyratzopoulos1, Gary A Abel1, Sean McPha Time to Take Stock: A Meta-Analysis and
Systematic Review of Analgesic Treatment
Racial Disparities in Pain
Disparities for Pain in the United States FREE
Management of Children With
Salimah H. Meghani, PhD, MBE M, Eeeseung Byun, PhD(c), Rollin M. Gallagher, MD, MPH
Appendicitis in Emergency
Pain Medicine, Volume 13, Issue 2, February 2012, Pages 150-174,
Departments
"Brave Men" and "Emotional Women": A Theory-Guided Literature Review
Monika K. Goyal, MD, MSCE1,2,3; Nathan Kupperm on Gender Bias in Health Care and Gendered Norms towards Patients
with Chronic Pain
Anke Samulowitz,“ 1 Ida Gremyr, 2 Erik Eriksson, 2 and Gunnel Hensing 1
Age and Gender Variations in Cancer Diagnostic Intervals in
15 Cancers: Analysis of Data from the UK Clinical Practice
Research Datalink
Nafees U. Din , Obioha C. Ukoumunne, Greg Rubin, William Hamilton, Ben Carter, Sal Stapley, Richard D. Neal

270

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

One thing holding back the application of AI to medicine is lack of the *right* data. It is not just that data is scattered & hard to access; many interesting variables aren’t being measured or collected at all. 6/

130

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Too many AI projects begin at the wrong starting point. They start with an existing dataset, and ask “what can we do with this data?” The harder question is, “what are the biggest questions in your area, and what data would be useful to answering them?” 5/

2123

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

“What Alphafold2 pulled off — applying a clever model to a large body of pre-existing data to revolutionize a field — is something that will be extremely hard to replicate. Why? Because we’re almost out of that pre-existing data.” -- @owlposting1.bsky.social www.owlposting.com/p/wet-lab-in... 4/

owlposting.com

Wet-lab innovations will lead the AI revolution in biology

1.9k words, 9 minutes reading time

130

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Some mistakenly believe AI can easily create magic solutions, without understanding the need for high-quality data. The success of AlphaFold was made possible by 50 years of prior work gathering protein structures into a rich database (Protein Data Bank launched in 1971) 3/

NATURE NEW BIOLOGY VOL. 233 OCTOBER 20 1971
CRYSTALLOGRAPHY
Protein Data Bank
A repository system for protein
crystallographic data will be oper-
ated jointly by the Crystallographic
Data Centre, Cambridge, and the
Brookhaven National Laboratory.
The system will be responsible for
storing atomic coordinates, structure
factors and electron density maps
and will make these data available
on request. Distribution will be
on magnetic tape in machine-read-
able form whenever possible. There
will be no charge for the service
other than handling costs. Files

2258

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Choices about which data to collect & which to neglect have long been shaped by power disparities and social & financial influences. Missing Data Sets are “blank spots that exist in spaces that are otherwise data-saturated” (Mimi Onuoha, 2016) github.com/MimiOnuoha/m... 2/

github.com

GitHub - MimiOnuoha/missing-datasets: An overview and exploration of the concept of missing datasets.

An overview and exploration of the concept of missing datasets. - GitHub - MimiOnuoha/missing-datasets: An overview and exploration of the concept of missing datasets.

110

Rachel Thomas @math-rachel.bsky.social · 24/01/2025

Two (conflicting) executive orders were made this week: - freeze on comms from scientific & health agencies - investment in medical AI w/ bold promises These highlight common misunderstandings & obfuscations about the relationship between data, AI, & power My post rachel.fast.ai/posts/2025-0... 1/

rachel.fast.ai

Rachel Thomas, PhD - The Missing Medical Data Holding Back AI

an AI researcher going back to school for immunology

1223

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

In addition to the challenge of gathering more data, another challenge of pathology models is needing to capture both local patterns (that show up in a small tile within a slide) and global patterns across the whole slide. 6/ (Image from Chen, et al, 2020, Hierarchical Image Pyramid Transformer)

Image showing a microscopic slide 150,000 x 150,000 pixels, subdivided into smaller and smaller squares. The 4096 x 4096 squares are labeled "tissue phenotypes", 256x256 squares are labeled "cellular organization", and 16x16 labeled "cellular features"

030

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

The Cancer Genome Atlas (TCGA) was an ambitious project begun in 2006 by the National Cancer Institute. Samples were collected from > 11,000 patients w/ 33 cancer types. All 3 of the above papers (UNI, Prov-GigaPath, & kaiko ai) concluded TCGA is not large enough for effective foundation models 5/

110

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

Another interesting paper evaluated the impact of scaling model size and training dataset size. They found limited need to scale *model size* beyond a certain point, but that *larger datasets* continued to lead to increased performance. 4/ arxiv.org/abs/2404.15217

arxiv.org

Towards Large-Scale Training of Pathology Foundation Models

Driven by the recent advances in deep learning methods and, in particular, by the development of modern self-supervised learning algorithms, increased interest and efforts have been devoted to build f...

110

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

Two big pathology foundation models were published last year: UNI and Prov-GigaPath. They achieved state-of-the-art results on dozens of tasks (although were not directly compared). 3/ 🟣 www.nature.com/articles/s41... 🟣 www.nature.com/articles/s41...

nature.com

A whole-slide foundation model for digital pathology from real-world data - Nature

Prov-GigaPath, a whole-slide pathology foundation model pretrained on a large dataset containing around 1.3 billion pathology images, attains state-of-the-art performance in cancer classification and ...

110

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

The powerful idea behind *foundation models* is to train a on many datasets (e.g. tissue images from many organs) and on multiple tasks (e.g. recognizing cancer, segmenting cells, predicting treatment outcomes) Patterns learned from one dataset or one task are likely to generalize to others. 2/

Image from Wikimedia Commons showing a human body, with 4 different types of tissue highlighted: nervous tissue, muscle tissue, connective tissue, and epithelial tissue.

110

Rachel Thomas @math-rachel.bsky.social · 16/01/2025

What AI can tell us about microscope slides: classifying cancer cells, predicting prognosis, and identifying genetic mutations that may drive treatment choices. My latest post is a friendly introduction to Foundation Models for Computational Pathology: rachel.fast.ai/posts/2025-0... 1/

rachel.fast.ai

Rachel Thomas, PhD - What AI can tell us about microscope slides

an AI researcher going back to school for immunology

2195

Rachel Thomas @math-rachel.bsky.social · 07/12/2024

"With LLMs the main uses are to produce outputs that are extremely similar to the training set. But the point of computational drug design is to deal with cases that are new... Copying from what we know already will only get you so far." -- @dereklowe.bsky.social www.science.org/content/blog...

science.org

Computational Care

0101

Rachel Thomas @math-rachel.bsky.social · 07/12/2024

Nature Method of the year for 2020 was Spatial Transcriptomics and for 2024 it is Spatial Proteomics. Sounds like this is the decade for spatial -omics!

1131

Rachel Thomas @math-rachel.bsky.social · 03/12/2024

An unmet need in lung cancer research: how to integrate -omics to understand extracellular matrix (ECM) remodeling (This is the first talk I've seen incorporating the ECM with omics-- it's an interesting perspective!) Amelia Parker 6/

020

Rachel Thomas @math-rachel.bsky.social · 03/12/2024

The extracellular matrix is a collection of proteins changing over time and space. It has different profiles for different cancer subtypes & profiles. -- Amelia Parker #multiomics2024 /5

141

Rachel Thomas @math-rachel.bsky.social · 03/12/2024

I can't share videos with 🦋, but there were some neat videos of 3D spatial information from various cancers & the additional info 3D imaging can provide. Zoe West 4/

100