Rachel Thomas's Recent LinkedIn Posts

Many people incorrectly assume that AI is only for an elite few– a handful of Silicon Valley computer science prodigies with monthly budgets larger than most people’s lifetime earnings, turning out abstruse academic papers. This couldn’t be more wrong. Deep learning (a powerful type of AI) can, and is, being used by people with varied backgrounds all over the world. A small taste of that variety can be found in the stories shared here: a Canadian dairy farmer trying to identify udder infections in his goats, a Kenyan microbiologist seeking more efficiency in the lab, a former accountant expanding use of solar power in Australia, a 73-year old embarking on a second career, a son of refugees who works in cybersecurity, and a researcher using genomics to improve cancer treatment. Hopefully this may inspire you to apply deep learning to a problem of your own!

This year, I have explored many of the ways that AI is being applied to immunology. These applications include predicting T cell binding, mapping immune communication networks, and discovering new antibiotics. In my latest post, I round up my writing on AI & Immunology from 2024: https://lnkd.in/gEeJiyJU

I wrote an explainer on some of the most common types of AI. Two major categories of AI tasks are: - Classification: e.g. is this a picture of a chihuahua or a blueberry muffin? - Generation: e.g. create a picture of Kanye out of tiny Captain Picard faces (this was work from a fast ai intern in 2017) Some problems can be framed as either classification OR generation, such as discovering new antibiotics. You can use AI to classify millions of compounds to predict which have desired qualities, or to generate new molecules... Read more details here, including the most useful neural net architectures:

My latest essay: In topics ranging from HIV research to efforts to downplay covid-19 to the long history of wrongly assuming women’s illnesses are psychosomatic, we have seen again and again that medicine, like all science, is political. This shows up in myriad ways, such as: who provides funding, who receives that funding, which questions get asked, how questions are framed, what data is recorded, what data is left out, what categories included, and whose suffering is counted. https://lnkd.in/g4WGJDQG

I'm leading a free workshop this week: debunking myths about innate math ability, and exploring misconceptions, cultural influences, & research studies on the factors that get in the way of enjoying & succeeding with math. You can sign up here:

My new post is on surprises of the microbiome. There are trillions of microbes in the air, in the soil, traveling along plant roots, and inside of us. In humans, the gut microbiome can help protect us from infection. But friendly gut microbes can also turn our immune systems against us if they escape into the bloodstream where they don't belong. One component of the plant microbiome are fungi that surround plant roots and the bacteria that travel along these “fungal highways.” The fungi share sugars with the bacteria, and the bacteria mineralize phosphate for the plants. Read more here:

Personalized cancer vaccines are like a WANTED poster to let the immune system know which outlaw cancer proteins to look out for. Neural networks can be used to identify the most promising candidates. Several different types of deep learning architectures are being applied to this problem, including ones: based on AlphaFold or not based on AlphaFold, for variable-length/sequences or for fixed-length inputs. It is not yet clear what the best approach is! Data is central to any machine learning project but gathering & curating appropriate datasets can be expensive & time-consuming. Data work is particularly challenging in the area of T cells, due to their incredible diversity. Read more in my latest post:

Chronic boredom is harmful to adults, causing stress, disengagement, and poor well-being. Academic researchers have shown that boredom in the workplace can be just as damaging as burnout. But search for information about childhood boredom and you’ll find the opposite message: articles describing boredom for kids as “fantastic”, “important”, and full of benefits. My latest post:

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. For the most popular language arts 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. This approach is not sparking a love for reading, with the number of 9 year olds who read almost daily for fun having declined from 53% to 39% over the last decade. One (now former) teacher wrote, "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.” When the measure becomes the target, it ceases to be a good measure. This is Goodhart’s Law. It can occur in non-tech systems, such as the obsession with standardized test scores driving curriculum and teaching constraints. AI often further exacerbates Goodhart's Law, since AI can be *too effective* at optimizing metrics. Too many people romanticize the past and the experience of analog school. Getting kids off screens isn’t necessarily going to improve their lives– particularly if they are only reading dull passages in basel textbook readers, if they are endlessly drilled on detached tasks, or if their teachers are constrained against including creative, open-ended activities. At the same time, many AI education approaches are exacerbating the problems created by the tyranny of metrics... Read more in my latest essay on education:

We are providing diversity scholarships for our updated part-time, in-person Deep Learning for Coders part 2 course presented in partnership with the University of San Francisco Data Institute, to be offered one evening per week for 7 weeks, starting March 18, 2019, in downtown San Francisco. Women, people of Color, LGBTQ people, people with disabilities, and/or veterans are eligible to apply. We are still looking for additional financial sponsors, so please contact datainstitute@usfca.edu if your company is interested in donating. Please share with anyone who may be interested (in applying or, in sponsoring!) Details here:

I was on NPR this week, together with Danielle Citron and Jack Clark, to discuss deepfakes: compelling fake videos which are being used to harass women (by creating fake pornographic videos of them) and to impersonate political leaders (spreading disinformation). You can listen to the episode here:

AI can give us lot of information from slides of human tissue. It can classify cancer cells (including subtypes and stages), predict prognosis, and identify genetic mutations that may impact treatment decisions. I wrote a friendly introduction to Foundation Models for Computational Pathology, including two big models that came out in the last year. This includes background on the field, some of the key challenges, and the impact of dataset size. My latest post:

I appreciate the mention of my AI & Biology blog in this list, thanks Muazma Zahid!

I was included in this article on "Machine Learning & AI Main Developments in 2018 and Key Trends for 2019" My take is that two main AI developments in 2018 were: 1. the successful application of transfer learning to NLP 2. growing attention to dystopian misuses of AI (including surveillance and manipulation by hate groups & authoritarians)

I am hosting a journal club (on zoom) tomorrow on the use of deep learning for enzyme function prediction and limitations of such approaches. Details on how to join in the post below. I am looking forward to a lively discussion!

My keynote at SciPy conference covered some recent developments in machine learning and the dangers of algorithms that create realistic prose and videos: https://lnkd.in/gqMS7jN

The human immune system is impressive, but so are the mechanisms pathogens use to evade it. In my latest post, I cover 5 surprising and ingenious ways that viruses and bacteria can sabotage our defenses. This includes surviving in bags of acid within our cells, seeking out our most powerful immune cells to subvert them, commandeering our transport machinery, and more. https://lnkd.in/g4vXFBeE

I wrote about Medicine's Looming Machine Learning Problem for the Boston Review:

A thoughtful write-up of our panel last week on Operationalizing AI Ethics at TWiML Conference: https://lnkd.in/gVYfeJN cc: Craig Newmark

My latest post: Decoding T cells with AI. T cells can rearrange their genes to recognize billions of different types of infected or rogue cells. Some cancer therapies are based on teaching T cells to better recognize tumors. Predicting what a T cell will bind to would be a vital medical breakthrough. Many are using AI to tackle this problem...

I wrote about some surprising and fun microbiome facts: 🔹 Wildfire smoke seems inhospitable, yet it is a good habitat for some microbes (which can be bad for humans). 🔹 There are microbes that thrive on microplastics in our oceans, and other microbes that can help remediate pollution. 🔹 Different groups of hyenas have distinctive microbiomes, which create group-specific scents. 🔹 A major pathogen that destroys agricultural crops also serves a helpful & necessary role in weather precipitation cycles. 🔹 Playing contact sports, like roller derby, changes your skin microbiome. Read more here:

There is much confusion about the "hygiene hypothesis". It is good to play in the dirt and to be exposed to microbes, but you should also wash your hands after using the toilet. When is hygiene good and when is it bad? A clearer refinement is the framing of "old friends" vs. "crowd infections." The evidence for pathogens that can be beneficial to the immune system (reducing risk of autoimmune diseases and allergies) is almost entirely for parasitic worms and friendly bacteria. In contrast, many viruses can even trigger the onset of autoimmune diseases or allergies. The "Old Friends" hypothesis distinguishes between microbes that we co-evolved with for 300,000 years and "crowd infections" that arose recently with the shift to crowded cities, poorly ventilated indoors spaces, and massive global travel. Read more in my latest post:

Registration is open for Tech Policy Workshop, Nov 16-17 at University of San Francisco and I'm so excited about our speaker line-up! Anyone interested in the impact of data misuses on society & the intersection with policy is welcome to join us. Details here: https://lnkd.in/g5wvmFY The Data Institute, University of San Francisco Craig Newmark Tawana "Honeycomb" Petty Catherine Bracy Rumman Chowdhury, PhD Y-Vonne Hutchinson Kristian Lum Guillaume Chaslot Irina Raicu Chris Riley

My latest blog post: Tech’s long hours are discriminatory & counter-productive

Advances in artificial intelligence are transforming many fields, including medicine and immunology. Some of these breakthroughs are beneficial, such as identifying promising new drug candidates, predicting T cell binding, and detecting diseases in retinal images. But AI can also amplify risks and cause harm, particularly when its implementations centralize power and ignore crucial perspectives from those closest to the task—mirroring rather than correcting existing power imbalances in the medical system. I recently gave a talk weaving together many of the topics I care about most, including: the potential of AI in immunology, racial & gender bias in medicine, the systemic dismissal of patient experience by doctors, and the promise of participatory machine learning approaches. You can watch my 29-minute talk in full here: https://lnkd.in/g5D3Ynfc Or read an edited transcript of my talk here (with links to many of the references): https://lnkd.in/gpsf-b6i

Deep learning is a powerful family of AI algorithms which is revolutionizing image recognition, text generation, and protein folding. Immunology is a complex and high-impact field, becoming ever more relevant as climate change and global mega-travel are leading to increasing pandemics, fast viral mutations, and antibiotic resistance. Here, I introduce basic concepts of what deep learning is and how it is being applied to 3 crucial problems in immunology. We also cover the gaps, risks, and opportunities of these approaches. This talk is geared towards a general audience. https://lnkd.in/gSjuE86g

my latest post: Was this Google Executive deeply misinformed or lying in the New York Times? YouTube has played a significant role in radicalizing people into conspiracy theories that promote white supremacy, anti-vaxxing, denial of mass shootings, climate change denial, and distrust of mainstream media, by aggressively recommending (and autoplaying) videos on these topics to people who weren’t even looking for them. YouTube recommendations account for 70% of time spent on the platform, and these recommendations disproportionately include harmful conspiracy theories. Given all this, you might expect that Google/YouTube takes these issues seriously and is working to address them. However, when the New York Times interviewed YouTube’s most senior product executive, Neal Mohan, he made a series of statements that, in my opinion, were highly misleading, perpetuated misconceptions, denied responsibility, and minimized an issue that has destroyed lives. Mohan has been a senior executive at Google for over 10 years and has 20 years of experience in the internet ad industry,... so we can’t just dismiss this interview.

I'm excited to announce our next CADE data ethics seminar, Monday Sept 23 at 6pm in downtown SF: Street-Level Algorithms, and Seeing the Forest for the Trees in AI, by Ali Alkhatib Algorithmic systems, especially those enabled and empowered by artificial intelligence, are rapidly encroaching on our lives on various fronts. They're deciding where we work, whom we date, what we read and see, and ultimately how we live. Perhaps unsurprisingly, this has led to widespread frustration and anger as AIs mistreat and malign us. I hope to dispel the common myth that AIs are inscrutable, opaque technical problems, and reframe them as tractable, albeit complex, social problems. We'll conclude with some reflections on agendas of research into AI for social good, and historical and social factors that tend to problematize that work. I hope you can join us! More details here: https://lnkd.in/gwa2pQW University of San Francisco Craig Newmark

"Tech companies are fighting over the same few people when they could be looking at a much broader group. The AI talent shortage is partially a perception problem. "Companies think, ‘Oh, I have to hire the Stanford PhD,’ and that’s not actually what they need,” Thomas said. “In-house talent is being undervalued right now." I'm quoted in this Fast Company article:

Close reading is a technique for careful analysis of a piece of writing, practiced by many ancient cultures, major religions, and academic scholars. It might come as a surprise that a technique associated with such a long history could see a revival with the use of Large Language Models (LLMs). Here are two examples of close reading with an LLM: Jeremy Howard reads a chapter of Eric Ries's upcoming new book, and Johno Whitaker reads a cutting-edge machine learning paper from Yann LeCunn (LeJEPA). They share their processes and reflect on the experience.

I'm excited to share the latest fast.ai course: A Code-First Introduction to Natural Language Processing. The course teaches a blend of traditional NLP topics (including regex, SVD, naive bayes, tokenization) and recent neural network approaches (including RNNs, seq2seq, attention, and the transformer architecture), as well as addressing urgent ethical issues, such as bias and disinformation. Topics can be watched in any order. All videos and code are available for free. https://lnkd.in/g-X4tPd

Twenty to thirty years ago, politicians, scientific leaders, journalists, and even Nobel laureates predicted that sequencing the human genome would revolutionize how we treat disease. And while the advances in DNA sequencing that have occurred since then have improved recognition and treatment for some cancers and rare diseases, on the whole the field has not lived up to earlier hype. Most common diseases are not linked to a single gene, but rather to small effects from hundreds of genes, environmental exposures, and microbiome influences. Two other key limitations of earlier forms of DNA sequencing are that it ignored protein expression and destroyed spatial information. In my latest post, I explain why these matter and how newer technologies are addressing them. The convergence of multi-omics, spatial information, and user-friendly programming libraries means that it is an exciting time to be working at the intersection of data science and microbiology. Read more in my latest post: https://lnkd.in/gYm2nZdb

As a former mathematician, I was used to nobody reading what I wrote. So when I first began blogging in 2015, I never expected that several of my blog posts would go viral or to have multiple journalists contact me (including from NPR, Wired, and Fortune), make the front page of Hacker News (over 10 times), receive conference keynote invitations, and be interviewed on podcasts. I do not consider myself a “natural” writer. In college, I tried to avoid classes that required essays, because writing was a struggle for me. It wasn’t until I was 30 that I set out to practice writing more. After 10 years of blogging, I look back on some of my most impactful posts challenging conventional wisdom, debunking AI hype, and democratizing machine learning.

Bacterial resistance to existing antibiotics is an urgent threat. Developing new drugs is an expensive, time-consuming, and failure-prone process, with only a small percentage of potential drug candidates making it from discovery through clinical testing to regulatory approval. Antibiotic research has been particularly neglected (compared to other types of drugs) due to unfavorable financial incentives. Some AI-assisted approaches to antibiotic discovery have built on protein folding prediction. Other work has created deep learning models directly for the underlying problem, to predict which drugs will have antibacterial activity. This approach has been used to discover new antibiotics which are structurally distinct from existing classes. Antibiotics discovered through deep learning-assisted processes is still in preliminary stages. In other fields, machine learning models have sometimes performed well in testing but failed to work as expected when deployed in the real world... My new post: https://lnkd.in/g_sFJXXw

As data scientists, Jeremy Howard and I have done our best to look at the data around covid-19, and what it means to you and your community. Our view: it is appropriate to be very concerned, and significantly change your lifestyle, right now. Epidemiologists are experts in studying the spread and impact of disease. Their modeling shows that covid-19 is *not* at all like the flu. If you're under 50 and in good health, that's not a reason for complacency. Responding to covid-19 is an ethical obligation for us all. Read more in our post:

I wrote a post with my advice for creating an ERGONOMIC computer setup ON A BUDGET. It doesn't have to be as expensive as you think. Please don't permanently destroy your back, neck, & wrists while working-from-home during the pandemic! https://lnkd.in/gHjZ8wE

Deep learning is glamorous and highly rewarded. If you train and evaluate a Transformer (a state-of-the-art language model) on a dataset of 22 million enzymes and then use it to predict the function of 450 unknown enzymes, you can publish your results in Nature (a very well-regarded publication). Your paper will be viewed 22,000 times and will be in the top 5% of all research outputs scored by Altmetric (a rating of how much attention online articles receive). However, if you do the painstaking work of combing through someone else’s published work, and discovering that it is riddled with serious errors, including hundreds of incorrect predictions, you can post a pre-print to bioRxiv that will not receive even a fraction of the citations or views of the original. In fact, this is exactly what happened with a recent pair of papers... In my latest post, I walk through two papers on enzyme function prediction that make for a fascinating case study on the limits of AI in biology and the harms of current publishing incentives. This contrast is a stark reminder of how hard it can be to evaluate the legitimacy of AI results without deep domain expertise. Read more here (with links to the 2 papers): https://lnkd.in/eUcPwWUw

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 of their code must be AI-generated... yet results are far from what was promised. To understand what is going wrong, it is helpful to look at research on gambling addiction and to return to Mihaly Csikszentmihalyi's original work on flow. While vibe coding and gambling can be highly absorbing, they violate several characteristics needed for the positive flow state we associate with creative work: - 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 With "dark flow", we lose our ability to accurately assess our productivity levels and work quality. When developers used AI tools, they estimated that they were working 20% faster, yet in reality they worked 19% slower. Read more in my latest post:

Last year, I became captivated by a new topic in a way that I hadn’t felt since I first discovered machine learning. I have been studying immunology daily for the last 6 months, and I was delighted to be accepted to a MSc in Immunology grad program. I have officially gone back to school. My ultimate goal is to apply my machine learning & data ethics skills to immunology, but I want to make sure I fully understand the underlying domain & relevant context first. With ML, it’s important to not just be a hammer searching for a nail. Immunology is complex, vast, and full of open questions and not-yet-understood phenomena. It was only in 2021 that researchers proved Epstein-Barr virus causes multiple sclerosis. Researchers are making new discoveries about links between viral infections and neurodegenerative diseases, such as Alzheimer’s. A study in late 2022 found a possible mechanism to explain the fact that the chicken pox/shingles virus significantly increases risk of stroke. In addition to the still-ongoing, devastating covid pandemic, we can expect more pandemics in coming years, with climate change rapidly raising the risk of viral spillovers. Read more of my thoughts & journey here: https://lnkd.in/gNxDVseg

Hype around AI in medicine often ignores two key risks: - patterns of automation contribute to centralization of power - medical knowledge is limited by the systemic refusal to trust patient expertise my new post: https://lnkd.in/gZD7v2Zg

"We’re drowning in an ocean of data, but we are starved of knowledge." - Nobel laureate Sydney Brenner "It is through deep familiarity with the biology — not simply a drive to collect more and more data — that important questions will be asked." - Nobel laureate Paul Nurse I wrote about some of the gaps and risks of how AI is applied to the life sciences, including: 🔵 Often the data we would be most interested in does not exist or is too difficult to gather. As a result, data scientists may instead use imperfect proxies or alter their research questions to make use of the data they have. 🔵 ML systems are trained in clearly defined environments, while the physical world often has complex & volatile underlying phenomena, which are not fully accounted for during training. This can create a brittleness to ML-generated solutions. 🔵 Too often the arduous task of curating quality datasets has been neglected, undervalued, and not given the same currency for career advancement as algorithmic work. 🔵 High profile competitions have led to huge breakthroughs, but over-focusing on competitions can lead to some research areas being neglected. In other cases, models created for benchmark tasks may be inappropriately applied to tasks for which they are less suited. Read my full post here:

Friends with no previous interest in AI ethics have begun asking me questions in the wake of the release of ChatGPT4, Google Bard, and Bing Chat. To consider the risks posed by new AI applications, it is useful to first understand several underlying concepts about how AI contributes to the centralization of power... People don’t just need an *explanation* of automated decisions (“why was my loan denied?”), they need *actionable recourse* (“what can I change to get my loan approved?”). [Berk Ustun] My new post:

Last week, I gave a headliner talk at the Stanford AI in Medicine Symposium. I spoke about several topics close to my heart: - racism & sexism in medicine - ways that machine learning can unintentionally contribute to the centralization of power - what is missing from many conversations about bias & fairness - a bit of my personal story of being dismissed & disbelieved by doctors (I was invited to speak at this event as an AI researcher, but I believe my experience as a patient is just as valuable) My talk is the first 17 minutes of this video, and is accessible to a general audience. This was my first time working personal experience into this type of talk, and I would love for you to watch it! https://lnkd.in/dVGQUzk

Qualitative humanities research is crucial to the field of AI. In examples ranging from measuring racial bias to understanding recommendation systems (like YouTube's), any quantitative issue quickly leads to a host of qualitative questions. Unfortunately, there is often a large divide between computer scientists and social scientists, with over-simplified assumptions and fundamental misunderstandings of one another. My latest essay, together with Louisa Bartolo: https://lnkd.in/dcvecmBe

Two executive orders were made this week which contradict each other: - a freeze on communication from scientific and health agencies, including infectious disease data on growing outbreaks - a large investment in AI, including promises to use AI to analyze health records and treat cancer This pair of announcements highlights common misunderstandings & obfuscations about the relationship between data, AI, and power. Choices about what data is collected and who it is made available were shaped by social influences, financial factors, and power disparities long before this past week. A decade ago, Mimi Onuoha coined the concept of Missing Data Sets to refer to “blank spots that exist in spaces that are otherwise data-saturated.” Many AI projects begin at the wrong starting point, ignoring the data we do not have and thus overlooking many important questions. Read more in my new post:

Most people catch many viruses in their lives– for example, over 90% of adults have Epstein-Barr virus, and adults catch the flu about once every 5 years. For a long time, catching frequent viruses was considered both inevitable and harmless. But it turns out that common, seemingly-mild viruses have disturbing long-term health impacts. Common respiratory viruses increase the risk of heart attacks and strokes. Viruses are linked to dementia and Alzheimer’s Disease. They can re-awaken cancer cells in patients whose cancer was previously in remission. Persistent infections accelerate aging and undermine longevity. Viruses can be the trigger that kicks off life-long autoimmune diseases. New studies come out each week confirming that viruses can harm the health of your heart, blood vessels, brain, nervous system, and gut. Please pause and let this sink in. If we were to truly internalize this information, there would be massive shifts in the practice of medicine, scientific research, and public policy... https://lnkd.in/g5jKrT4u

There are several ways in which AI hype is outrunning reality in the life sciences. Here are a few that worry me: - Many people assume that we already have all the data we need, and we just need to throw it into a model for amazing outputs. We are underinvesting in thinking about new types of data and exploration of causal mechanisms. AlphaFold didn't happen overnight. It was made possible by 50 years of data collection (the Protein Data Bank was begun in the 1970s) and 25 years of a well-structured & well-run competition. - The field does not reward error-checking or domain expertise. For example, an enzyme classification paper using AI was published in Nature Communications and contained hundreds of errors. The microbiologist who discovered these errors had great difficulty getting her rebuttal published. - Scale can be misleading, especially when data has systematic biases rather than random noise, when particular types of data are missing, or when the underlying paradigm is incorrect. Bigger doesn't mean better in these cases. - Our current AI systems are doing a fuzzy interpolation between existing data points. This is valuable, but won’t give us something truly outside the scope of the training data. We still need research where new paradigms or different causal mechanisms are required. For more details and concrete examples, check out my discussion with Kamayani G.! Here is a link to the YouTube video and a condensed transcript: https://lnkd.in/gEvNXwxG

I am looking forward to this discussion with KAMI Think Tank!

Thanks for creating these visual notes from my TEDx talk Danesh Mohiuddin!

16 Things You Can Do to Make Tech More Ethical (a 3-part series) https://lnkd.in/g3MpEYx https://lnkd.in/gYMzZBN https://lnkd.in/g4HuFJD