AI Computer Scientist • @harpercarrollai on IG • taught AI/ML @Stanford • built @Meta • AI is less scary under the hood ↓ 10-day AI basics + more (free)
Austin / SF / NYC
Joined September 2016
- Tweets3.5K
- Following1.2K
- Followers33.6K
- Likes10.5K
What are “embeddings” in AI and machine learning? How are they fundamental to neural networks and AI?
To me, these are one of the most fascinating and literally awesome parts of machine learning: the fact that regular items and concepts in our universe can be modeled accurately as numbers that machines can then interpret and make predictions on.
@harperscarroll @stanford @meta
Why is everyone suddenly buying Mac minis? Part 2
Part 1 covers the other BIG reason people are buying them.
Do you have a Mac mini, or similar? How are you using it?
Why was/is everyone suddenly buying Mac minis?
Part 2 covers the other BIG reason people are buying them.
I really believe we know who we are & what we’re destined for
i forgot about this until @ScottWu46 reminded me, but back in january 2017 when we were 19, scott and i had lunch in SF
at lunch, we talked about how our IMO/IOI/math competition gang could be the next paypal mafia. i knew the competition kids were geniuses by any standard, and felt confident in time we could take the tech world by storm
fast forward nearly 10 years later, and many of our friends from that time are some of the most exciting people in AI and technology (@ScottWu46, @akshat_b, @chameleon_jeff, Jerry Ma, @randomjohnnyh, @0xShitTrader, @stevenkplus1, @thejessezhang, @vickie_ye_, @alexwei_, @demi_guo_, @d_m_ziegler, @hughbzhang, @_mohansolo, @MarkSellke, and many others i'm sure i'm forgetting)
i'm now at that age where i reminisce about everyone i grew up with, and i've known most of these people since middle school (via MATHCOUNTS!). even though we were scattered around the country, we would gchat about math problems and hang out on AoPS, then we did math/algo camps together, and then we did internships together.
very crazy to think about the degree of sheer brilliance of this group when we were just teenagers, and so special to see so many go on to leave their mark on the world.
note to parents: if your kid is on gchat all day solving math problems with some anonymous group of cross-country asian kids, (s)he may just turn out alright.
3
19
2,743
Neural Network ExplAIned Series, Pt. 3: Activation Functions!
We didn’t cover this in the video, but they are called activation functions because they “activate” neurons to pass data forward. Perhaps in video 4 we can do an example of an actual activation function like ReLU, and see visually how it “activates” (and conversely de-activates) neurons as they pass data of a particular input through the neural network in the forward pass. Please let me know if that would be of interest to you!
As always, please let me know your questions & topic requests!
Growing up, I thought my dream car was a Porsche. Recently, I almost got one but then was a passenger in my friend’s self-driving Tesla.
Immediately knew I needed a Tesla.
Got my own white-interior self-driving Tesla and now have absolutely 0 desire for a Porsche. Would NEVER want any other non-self driving car.
Porsche owners are missing out
Intro to neural network output layers!
@harperscarroll @stanford
In this example, we are doing supervised learning with binary classification. Supervised learning - if you remember from my “3 ways AI learns” video - is when you learn from labels, like flash cards. There are two main categories for supervised neural network models: classification models and regression models. Classification models figure out percentage probabilities for all its possible “classes” - ie buckets for its data. With binary classification, there are only two classes, so you only need to predict one class probability because the probability of Class 2 is inferred via 1 - (probability of Class 1). Regression models output an actual number, like, for example, predicting house prices given zip code and square footage and age of development.
Can you guess what type of model - regression or classification - large language models are? For each, what is it predicting (ie if classification, what are the classes? If regression, what is the final number?)?
This touches on a concept that I would like to talk about more which is activation functions. The output layer here applies one activation function: it squashes the layer to a number between 0 and 1 - our probability output. Activation functions between the hidden layers (not covered here) do something more interesting and crucial - they allow for non-linearity. As you saw in the first video, the core math is weighted sums - multiply each input by a weight, add them up - and that’s “linear,” meaning a constant rate of change. Stack those forever and you still just get a straight line. The hidden activations are what break that. More on this soon!
In these 90 seconds you can learn a large portion of how neural networks actually work. It's a very simplified picture, but let me show you the actual math.
Every neural network has three kinds of layers: an input layer, one or more hidden layers, and an output layer. Each layer is made of neurons (the dots), connected by edges. Every edge carries its own weight, a number.
Your input word or text input goes through a transformation into number form to enter the AI model - check out my “AI is math” video. That transformation is called an embedding, and I have an explAIned video on those as well.
In real models an embedding can be hundreds or even thousands of numbers long; here I will use just two.
Say the input is 5 and 2.
Now give each edge a weight. To reach one neuron in the hidden layer, two edges feed into it: one with weight 10, one with weight 2. The neuron’s value is 5 x 10 + 2 x 2, which is 50 + 4, so 54 (remember PEMDAS, my friends!!).
The neuron beside it has its own edges, say weights 20 and 4. Its value is 5 x 20 + 2 times 4, which is 100 + 8, so 108.
Multiply each input by the weight on its edge, add the results, and pass the number forward. Do it again for the next layer, and the next. Basically every generative AI model you use runs on this calculation.
Btw - we don’t go over them in this first video, but we’re missing the *nonlinear* functions in this demonstration. I can make another video on those & why they are crucial to the neural networks. Also, the output layer works a little differently from the hidden layers, and I will cover that in Part 2. I can also break down how these weights are actually learned, through backpropagation. If that is of interest to you, or if you like this format, or if I can do better somehow, please let me know!
The last two weeks in Berlin and NYC!
I was in Berlin for NVIDIA GTC with Nebius and in NYC to speak at MetLife! Got to see family and one of my best college friends. He’s the CTO of a major gaming company, and we recorded a vlog about life while walking through Central Park. We may or may not release it 🙃
It was my first time in Germany and I learned that Germans like to communicate directly and…. honestly… I loved it. One of my other very best friends from college is German and it all made sense.
Hope you are enjoying the beginning of the holiday season 🤗🍁🍂🎃🎄
How does AI actually work?
Let’s talk about what’s happening under the hood of agentic chatbot models like @claudeai. I’ve heard from my students that understanding the two components of these types of models has been the biggest unlock for them in using these AI models confidently.
In the prior video, we discussed how LLMs (large language models) work, and how they are “probability machines.” Today, we’ll discuss the second component: agent harnesses.
This was created in partnership with Anthropic for other platforms, but I'm posting this to X on my own.
Harper Carroll retweeted
I got Qwen 3.8 running on a laptop and I'm blown away
It's Opus 4.6 performance, but unlimited, private, and local
It's on the RTX Spark and it's amazing
In this video I cover how to load local models on ANY device you have using Hermes and all the cool projects you can build
How does AI actually work?
Let’s talk about what’s happening under the hood of chatbot models like Claude. I’ve heard from my students that understanding the “probability machine” mental model was, by far, the biggest unlock for them in using these AI models confidently.
Thrilled about this new series with @claudeai. Stay tuned for episode 2.
This was created in partnership with Anthropic for other platforms, but I'm posting this to X on my own.
Going to be recording a bunch of long-form videos. What are your questions & topic requests? Please share!
10 Days of AI Basics, Day 2: Data & ML Models
@HarperSCarroll
Welcome back! It’s so exciting to connect with so many of you after introducing the series last week.
Today I want to talk about data. This is a vast topic and I’m realizing just how little I can fit into only one minute and 30 seconds. I will have the YouTube series out soon – still need to get some things set up – so stay tuned for that. I’m excited to go more in depth into this topic - there is SO much more to cover.
Please let me know in the comments if my explanation was clear, and if you have any questions regarding this topic - I can answer them here and/or take them into consideration for the longer YouTube video. You can also let me know in the comments if you have anything else you’d like to learn about AI. I read all of them and I’m excited to connect with you!
Also, I’m looking for tiny microphone recommendations! I got a little clip-on one, but it sounds terrible compared to these wired headphones; however, they still generate some scratchiness with movement.
@stanford @meta @facebook @HarperSCarroll
Introducing: 10 Days of AI Basics!
Day 1: What is AI, and how does it relate to ML (machine learning)?
This is, by far, the most requested series, so I’m really excited to begin!
@harpercarrollai
if you don’t know me, I’m Harper, and I received my Bachelor’s and Master’s degrees from Stanford University in computer science in artificial intelligence. I have always loved to teach, and at Stanford I TA’d multiple computer science courses, including a PhD-level core AI course, Decision Making Under Uncertainty. I was also at Facebook/Meta for 4 years doing AI/ML engineering for some really cool projects that I’ll talk about in the coming videos.
Now onto today’s topic: Artificial intelligence is the intelligence of computers, and machine learning is a *subset* of intelligence, which means that all machine learning is AI but not all AI is machine learning.
⭐️An example of AI that isn’t ML is programming (“hard-coding”) a machine to say “hi” when the user says “hi”, rather than the machine *learning* that “hi” is an appropriate response to “hi” based on a ton of data it’s received and processed (which would be ML).
ML involves complex math to allow the machine to actually learn from data has been given.
Please let me know in the comments if this video was clear and made sense, and also what other topics are interested in learning about or other questions you have!
@stanford @harpercarrollai
3
502
Zuckerberg’s Biohub Wants to Cure All Disease. Here’s What’s Missing
Can AI help cure all disease?
Alex Rives (@AlexRives), Head of Science at @Biohub, says the biggest holdup is likely not the AI but rather the data: we have barely measured how cells behave. In this conversation, he explains the $1.8 billion effort to generate the data needed to build a "virtual cell."
We talk about how a language model trained on proteins started revealing biology on its own, what a virtual cell actually is, and why so much of this money goes to measurement. We also get into zero-shot protein design, what could shorten the 10-year path to a new drug, what custom medicines for a single patient would require, why Alex says biorisk today comes down to the lab rather than the AI, and the idea of a universal immune system.
Biohub is a nonprofit founded by Mark Zuckerberg (@finkd) and Priscilla Chan, with a mission to cure or prevent all disease. It releases its main protein models openly, including the ESM family that Alex started at @Meta's fundamental AI research lab.
CHAPTERS
0:00 Preview
0:42 Introducing Alex Rives and Biohub
1:48 The $1.8 billion announcement
2:52 Alex on the effort to map the cell
3:37 What is a virtual cell?
4:58 The data gap: why proteins are ahead of cells
7:34 ESM: the first large language model for proteins
9:43 Proteins explained: a 20-letter alphabet
10:30 ESM vs. AlphaFold
13:14 How ESM learns biology by filling in the blanks
14:39 From proteins to cells
15:23 What data do we need to generate?
16:55 Transcriptomes and spatial transcriptomics explained
17:44 Why this needs wet labs
18:56 Will AI replace researchers?
21:01 Why drugs take 10 years, and how to shorten that
23:01 The future of custom medicines
23:40 Zero-shot protein design
25:20 A custom therapy for every patient
26:57 From discovery science to engineering science
29:05 What's missing for custom medicines?
30:22 N-of-one therapies and baby KJ
31:46 Do we need more wet labs?
33:42 Why wasn't anyone generating this data?
38:07 How serious is AI biorisk?
39:51 Gene synthesis screening explained
40:51 A universal immune system
43:37 A call to action
KEY TAKEAWAYS
- Alex calls the core problem a "fundamental data gap." AI for proteins was built on about half a century of data, and nothing comparable exists yet for cells.
- A virtual cell is a model that predicts what a cell will do if you change something. The test is whether its predictions match what happens in a real experiment.
- ESM learned by guessing hidden amino acids. Nobody told it about structure or function, but that information emerged in what it learned.
- Zero-shot design narrows hundreds of thousands of computer-generated candidates to about 100 to test in the lab. Alex describes it as 24 hours of computation and seven days of lab work.
- Custom medicines aren't here yet. Therapies still have to be safe and developable, and the infrastructure and regulatory rules weren't built for a drug only one person will take.
- On biorisk, Alex says "today it's not a risk" and that the key barrier is the experimental laboratory, not AI. He says we need gene synthesis screening, and that defensive applications need to stay ahead.
- Most of this isn't built yet. Alex expects a first dataset in about a year and useful models within five, and those are expectations rather than promises.
Harper Carroll retweeted
every social/media platform needs to trade off for what's good for
a) *you*
b) what's good for the rest of the users/platform
c) what's good for the business model of the platform even if user-hostile
my favorite example of (c) is amazon refuses to show me what I've ordered in my shipping emails, ostensibly to not let Google compile data on me.
if agents can circumvent this and construct your own experience, expect a bunch of companies to take action. mobile might be much harder as local models can't take over the browser.
it does seem like the arc of the internet with frontier tokens will bend towards what the user wants.
- Claude fighting for you. Short form video and rec sys attractor basins are the cigs of our generation. Your time can be spent in a so much more fulfilling way. It’s insanely user hostile that you can’t turn off YouTube shorts or Instagram reels permanently.
- Way better recs! Current algorithms are primitive.
- Look at what you can do in a weekend now.