Co-founder & CEO of @brilliantorg. Working at the frontier of human + AI learning design.
San Francisco, CA
- Tweets 1,959
- Following 349
- Followers 15,314
- Likes 626
How learning products give feedback: “Incorrect. You pressed the jump button too early.”
Super Mario: Fall into the pit and die.
1/ Since Brilliant’s early days, we’ve been chasing that same visceral learning experience – the kind where understanding clicks through doing, not reading.
In more recent years, we’ve been focused on the question: Can AI accelerate how we make interactive games for learning?
⬇️
14
19
97
24,299
I believe most people do want to work on more pressing problems.
It will require a big shift in personal and collective values to escape the current attractor.
It’s a trite observation now, but we’ve raised our best to be averse to risk and real possibility of failure.
1
1
31
8,735
Another childhood SNAP recipient here. Food assistance makes a huge difference. My family went from poverty to both children becoming founders and job creators.
You’re much more likely to pull yourself up by your bootstraps if you have enough to eat.
1
1
30
2,006
This definitely works. She was 3 at the time, we did 10 mins/day of reading. Now 4 and reads chapter books.
She also likes watching TV, which I suspect is likely net-good in reasonable doses. In early childhood, reading ability is gated by oral/aural ability. High listening comprehension is a prerequisite to reading well.
This makes me skeptical about strict restrictions on (high-quality) TV. It's hard to get equivalent high-density language immersion in a complex story, with real-time visual cues to reinforce meaning.
1. This spring I've been working on a guide for teaching very early reading outside of school. It's simpler than some might think (fun progress video attached).
The first steps are *finally* published today.
3
1
30
3,671
A beautiful essay on the future of learning online, based on interactive play and the 4 freedoms of play.
We need to talk about how important interactive play is in learning. An essay.
pcho.medium.com/the-key-to-l…
5
22
Things mathematicians do, that LLMs still struggle to do:
1/ Mapping a problem to a simpler, isomorphic problem; having novel insights about what problems are isomorphic
2/ Reducing the complexity of the problem to be solved by solving a special case of it first
3/ Checking the result by a variety of mechanisms to ensure that the correct answer is overdetermined
4/ Successfully stepping back and asking: Does this make sense? Did I apply the right procedure the right way? Should the result look like this?
These critical thinking skills seem to transfer outside of math, too, especially for those who have learned math deeply enough that they’re reflexive.
2
6
27
2,197
I think about this a lot.
How much the best scientist on Earth 400 years ago would have given to know what a competent 15 year old knows today.
How this knowledge on its own doesn’t make us better scientists and truth seekers.
How today, LLMs can explain all human knowledge on demand, and this hasn’t yet replaced the need or satisfaction of learning by humans.
1
2
25
3,639
Thought of the perfect name for a venture fund founded by the super team of @SBF_FTX @ElizabethHolmes @MartinShkreli:
Conviction Capital
6
23
3,147
Incredible that, 17 years in, @salkhanacademy is still plugging away at making YouTube teaching videos.
He uploaded 28 in the past week. Most of them accumulate <200 views in the first few days, then slowly compound over time.
This is schlep blindness at its finest. I suspect this trait is non-trivial for making a successful learning product, if not a requirement.
2
3
22
3,493
"Will there be demand in the future for what we do? What must remain true about the world?"
I've been struck by how even execs at multi-billion dollar, profitable companies are privately, vigorously questioning the future relevance of their entire category.
It's important to keep asking the uncomfortable questions.
1
2
20
2,087
Paradoxes in learning found in research:
- Repetition is good for learning but bad for motivation
- People are naturally curious, but resist thinking hard about something new
- Methods that require struggle are more effective, but they feel less effective
What else?
2
2
18
1,740
4 yo: Can I do math workbooks until the bus gets here?
Me: It's almost here, let's do that after school.
4 yo: But I want to think about stuff.
Me: Great, you can think at school.
4 yo: No I want stuff TO think about. I wish I could just stay home and do workbooks all day.
Is this just my weird kid, or do brains this age crave "book learning" and new intellectual ideas? Should I be worried she isn't getting this at school?
4
1
18
2,542
Karpathy: A lot of what looks like learning is actually more maturation of the brain.
A zebra gets born, and a few minutes later it’s running around and following its mother. That’s an extremely complicated thing to do.
That’s not RL. That’s something that’s baked in. Evolution has some way of encoding the weights of our neural nets in ATCGs.
I think there’s actually very little reinforcement learning for animals.
A lot of the RL is more like motor tasks, not intellectual tasks. Humans don’t actually use RL, roughly speaking.
2
2
19
3,474
Wow. Khan Academy yearly active learners down from 153 million in 2023-24 to 105 million in 2024-25.
AI is eating education.
3
1
21
1,986
Mississippi's turnaround shows how much progress is possible.
Any state that's serious about equity should implement a phonics-based reading curriculum and teacher training.
wrote today about how even if you've heard of Mississippi's surge in reading scores, you're probably underestimating it:
1
2
19
2,789
It’s an existential question for consumer startups: Who exactly are you building for?
In education, for example, a product built for people who love language learning is a *different product* than one built for casual learners. Duolingo intentionally built for the latter.
Khan Academy grew because it was free, not because its pedagogy was novel. They only had to get one thing right, which was fundraising to keep it free.
Neither company cared about early adopters, they cared about ease of adoption for the 14-year-old in Wichita.
This might be obvious but the hardest thing about AI (or frankly anything) is not technology, it’s primarily humans..
It’s generally very hard to change human behavior to adopt something different.
This is why startups find it a lot easier to sell to startups and developers who are used to being early adopters.
As they move upmarket, they get frustrated on why customers don’t get it even though it’s a massive win-win.
3
16
3,473
In Orwell's 1984, people are controlled by inflicting pain.
In Huxley's Brave New World, they're controlled by inflicting pleasure – adoring the technologies that undo their capacity to think.
Orwell feared those who would ban books.
Huxley feared there’d be no reason to ban one — no one would care to read.
Orwell feared the truth would be concealed from us.
Huxley feared the truth would be drowned in a sea of irrelevance.
Orwell feared that what we hate will ruin us.
Huxley feared that what we love will ruin us.
Huxley thought those on alert to oppose tyranny "failed to take into account man's almost infinite appetite for distractions."
The excellent 40-year-old book “Amusing Ourselves To Death” is about the possibility that Huxley, not Orwell, was right.
3
6
18
2,503
Work is *love made visible*.
Amazing to get to work with such passionate people every day to invent the future of learning.
This programming series starts with the basics and will go all the way up to hard interview algorithms problems.
New content every week!
.@strickinato @forestfari go behind-the-scenes of our new interactive Python course, where breaking things is expected.
With smart autocomplete handling syntax, you focus on the real work: logic, iteration, and debugging.
By the end, you’ll have built a cybersecurity system, while becoming skilled at learning from real-world errors.
1
17
2,650
This @paulg essay is the rap song I listen to before gametime (public speaking or an important meeting).
16
1,154
@brilliantorg is built for learners ages 10-110! We’ve interviewed users as young as (very precocious) 7 year olds, all the way up to age 92!
5
15
625
This is an actually good take. What Google has done is very impressive, useful, even revolutionary. But it’s not what people usually think of when describing “generating a novel hypothesis.”
They didn’t ask the model for ideas and it responded. More like:
1/ They start with a big foundation model that can form embedding representations of cells
2/ They then fine tune it on datasets of drug interactions
3/ They use the embeddings from the fine tuned model to predict the effect of each molecule on antigen presentation in the presence/absence of cytokines
That makes a ranked list of candidate molecules.
I've always believed new knowledge can be (1) built on existing knowledge but connecting the dots in unique ways (2) creating pure de novo knowledge (through hypothesis, experiments, etc) that might go against current thinking (e.g. darwin/evolution or galileo/mass-gravity or copernicus/heliocentricity). LLMs are likely great at #1 and that's where perhaps a vast majority of the net new knowledge lies. Even if LLMs as they stand today never get to #2, their impact on research will be tremendous.
3
15
6,818
When ChatGPT came out, it changed user learning behavior almost overnight.
Since then, our team has spent much more time imagining possible futures and making predictions about what might happen next.
We started as builders solving a need, and it’s been uncomfortable to think of ourselves as forecasters. But it’s been necessary, and also fun.
Everyone got better at asking, “What would have to be true for X to be true?” and running mental simulations.
Being right about where things were headed in the next 3 years was important in 2022, and feels even more important today.
3
1
16
2,248
Some good stuff in this new research study on 5-7 yo children, showing that making your own predictions boosts conceptual learning more than observing others' predictions.
This minimal act of agency supports deeper learning by increasing internal commitment and cognitive engagement.
The technique they used is interesting – they created Bayesian learner-generated predictions, to isolate the effect of agency (doing it yourself) vs the information in the prediction.
They split children into 2 groups:
DIY (active) condition: children made their own predictions before seeing outcomes.
Yoked (control) condition: children observed predictions supposedly made by another child, with predictions matched to their prior beliefs.
Sample size is small (n=95) and effect size is modest, but it tracks with common sense that the generative act of making your own predictions is helpful for learning.
3
3
15
1,915
This totally worked! Pipeline is 🔥
“I saw the @soleio tweet” has been a common refrain.
Just incredible reach.
If you’re passionate about education and have a knack for designing elegant, new interfaces:
@brilliantorg is an awe-inspiring product and team to join
Few things inject optimism into my veins like the work @suekhim @pcho are leading rn
2
15
7,944
Restored an I.M. Pei townhouse as a little side quest… and then Dwell magazine showed up to photograph it yesterday. AMA!
5
14
2,782
3/ We kept coming back to Mario. Text explanations, no matter how perfect, just aren’t our goal. We want learners to develop intuition through interaction, to build understanding through experimentation, and to have fun.
We want them to feel the math.
1
12
738
Haunting new essay by Derek Thompson:
"Young people are already degrading their cognitive capabilities by outsourcing their minds to machines...
For most people, the tragedy won't even feel like a tragedy. We’ll have lost the wisdom to feel nostalgia for what was lost.
Humans take away their own capabilities in the presence of new machines. We are so fixated on how technology will out-skill us that we miss the many ways that technology can de-skill us."
1
3
15
2,188
Talked to a UChicago prof today who studies the human genome. He said some of his research would never have happened without ChatGPT.
He found a mathematical relationship in genomic data, asked ChatGPT about it, and learned the same equation had been discovered in econometrics decades ago. He would have never made that connection.
Just caught myself thinking “ugh, the AI couldn’t even compute this cohomology group correctly.” OK, admittedly things have come pretty far, pretty fast.
1
14
3,007
Simply Piano app is the 4 yo’s latest obsession.
I helped her get started, but after a week or so she became independent.
Not getting 3 stars on a song annoys her so much that I’ve seen her practice 40 minutes across 3 sessions in a day to get the last star.
Well-executed, definitely worth the subscription.
4
4
17
2,603
17/ Again, if inspiring people to think and solve problems is something you think about day and night and you'd like to learn more about joining a team that lives and breathes learning games – shoot me a DM!
3
11
982
Learning environment compounds talent in a way genes alone never could.
Families where multiple siblings reach world-class levels are counterintuitively a strong argument that nurture matters a lot.
The Polgar sisters all became elite chess players – a 1-in-a-million outcome repeated 3 times! Genes aren’t that precise.
What was consistent was the dad’s experiment in deliberate training, a chess-rich home environment, and siblings pushing each other. And the youngest reached the highest rank.
15
1,429
We suck at this test and yet we’re crushing the OECD on growth.
Teaching isn’t a prestige profession in the US, so far that’s worked out “great” for us, and that’s the problem.
The people who would have been absolutely amazing teachers instead being hard at work tagging training data for much higher pay – is that dystopic, rational, or both?
Probably both – perhaps it’s rational that topline growth for the US is more driven by how certain talented individuals are utilized, than by the average level of education. But it’s dystopic that it results in incentives and organization that lower the quality of general education and exacerbate inequality.
This is also why I find the case for UBI both rational and dystopic. If wealth and power is increasingly concentrated in the 5% who create ~all the GDP growth, average education won’t be a national priority. We can cheer on the LeBron Jameses of capitalism at FAANG, who will win our games. But for individual quality of life and healthy civic participation, it matters a lot.
As long as we reduce things to crude optimizations (like GDP) to decide whether better education is a national priority, we’ll continue to have massive disparities in opportunity, income, health, etc.
3
12
2,366
Wow. Incredible read on learning, children, and a beautiful, winding life.
Take curiosity seriously, and most of what we call “teaching” starts to look insane.
1
17
6,153
‘Does this mean companies can hire stupider people and pay them less, now that AI can solve even Terry Tao’s math problems?’
Not really, no.
Terence Tao + AI for solving hard math problems🤯
In this example the insight comes from Terence, and the muscle—via an hourlong conversation with GPT-5 and some python code it wrote—comes from AI.
"Here, the AI tool use was a significant time saver... Indeed I would have been very unlikely to even attempt this numerical search without AI assistance."
1
1
12
3,116
That question at the start of the paragraph? Classic LLM behavior — a subtle yet telling sign.
1
11
760
Very exciting.
That serious indications of past life on Mars published in Nature this week didn’t pierce my newsfeed until today tells me I need to curate better information sources.
Thanks for sharing.
1
1
12
2,319
On our Tutor team, 80-90% of the backend code is written by Claude.
For us, it doesn’t yet mean we write 10x the code we did before. Human thinking and review time is the limiting factor.
But it has sped things up a lot.
This is still wrong though.
The fact that AI produces 95% of the code, doesn't translate into humans being 20x more productive.
If the AI produced 100% of the code, would this mean that humans are infinitely more productive?
1
2
12
2,956
Experts have a richer inner monologue than novices.
They have more room in working memory to talk to themselves.
This allows them to abstract the situation and think about what's important.
ALT Excerpt from Why Don't Students Like School?
1
1
10
1,670
Thinking in a word-centered, reasoning culture:
"Most of the first fifteen presidents of the US would not have been recognized in the street.
To think about those men was to think about what they had written, to judge them by their public positions, their arguments, their knowledge as codified in the printed word.
This is the difference between thinking in a word-centered culture and thinking in an image-centered culture."
ALT Amusing Ourselves To Death
9
1,764
“One of the major limitations on human performance is not bias, it is just noise… Most of the errors that people make are better viewed as random noise.”
How would we learn and perform differently if we could care less about our errors, because the AI is always there to correct us? Like a global spellcheck.
@professor_ajay at NBER AI points out an amazing short comment by Danny Kahneman back at our 1st Econ of AI conference in 2017. Main point: AI has less noise than humans. Most discussion is about bias, but take that noise comment seriously...it's very deep.
1
10
2,094
Something odd: A year ago, Google and Amazon both promised to bring AI-powered assistants to every smart speaker.
They still haven’t rolled it out to more than a small percent of testers. Why? An LLM can help with frontier research, but can't reliably turn my lights on?
Maybe it's accuracy. "Good enough" is fine for search, but not for unlocking your door accidentally.
You could throw a reasoning model at it, but that’s expensive and too slow for good UX.
11
710
Listened to this today, and now I’m struggling to think of *any* examples where incumbents successfully absorbed a fundamental technology shift by making it a feature.
Are the companies bolting on "AI features" already cooked?
Evans: With any of these fundamental technology changes, the incumbents always try to make it a feature and they try to absorb it… Then over time, you get new stuff. You unbundle existing companies because of something that’s possible because of this new technology.
Here are 9 ideas that came up in my conversation with Benedict Evans.
Judging by the comments on YT, he has some very controversial AI takes.
1. Platform Shifts Are Normal: "[AI] is the biggest thing since the iPhone," Evans says. "But it's only the biggest thing since the iPhone." Every decade brings a new platform that seems civilization-altering. The internet was going to change everything. Mobile was going to change everything. Now AI will change everything. Until it becomes just software. The revolutionary always becomes routine.
2. The Data Moat Is a Mirage: Everyone thinks Google and Meta have an unbeatable advantage because of their data. But LLMs need such enormous amounts of text that everyone needs all the text there is. "All the text there is is kind of equally available to anyone," Evans observes.
3. Brand Beats Product (For Now): ChatGPT dominates while the models underneath are becoming indistinguishable. Do a blind test between Claude, Gemini, and ChatGPT: most people can't tell the difference. Yet ChatGPT captured the Google position in AI. In commodity markets, the first brand that sticks usually wins. Being best matters less than being first in the mind.
4. Incumbents Make Everything a Feature: Kodak went all-in on digital cameras and still died because the business model changed. Every incumbent tries to absorb new technology as a feature of what they already do. But platform shifts don't add to your product. They replace your product.
5. Pretty Close Isn't Close Enough: Evans tried AI for research in his domain. The error rate? "Dozens per page." He's blunt: "Today it has zero value for quantitative analysis." The gap between 99% accurate and 100% accurate is infinite when trust is binary. Until AI crosses that gap, it's a toy for brainstorming, not a tool for truth. Near-zero errors aren't zero.
6. AI Can't Judge Its Own Originality: AlphaGo could make original moves because it had a scoring system. LLMs don't. "Variance is bad. Originality is a lower score," Evans explains. Without external feedback on what's actually good, AI generates variations, not innovations. You can't create what you can't evaluate.
7. The Commodity End Game: Meta made Llama open source. Amazon wants models sold at cost. "They want to make it a commodity," Evans notes, "and they differentiate on top." When your competitors want your product to be free, you're in a tough spot.
8. Writing Is Thinking: "If you're thinking without writing, you only think you're thinking," Evans quotes. Students using AI for homework skip the mental work that builds reasoning. Delegating writing means delegating thought. The shortcut around thinking leads nowhere worth going. We build our minds by using them, not by having machines use them for us.
9. You Can't Pattern Match on Summaries: Reading other people's compressions of data creates an illusion of knowledge. Evans sees patterns because he reads source material, not summaries. "You're consuming a compression of the work but not the actual raw work." Second-hand insight isn't insight. Pattern recognition requires requires high quality data.
Search for "Benedict Evans, The Knowledge Project"
1
2
13
5,556
New paper observes GPT-5 doing “original mathematics”, but it required human steering.
AI is great at executing concrete steps. But essential insights, obvious to the mathematicians overseeing it, ultimately led to the interesting novel result.
1
1
10
1,518
Thought experiment: LLMs may be as transformative for thought as literacy was for language.
In this world, *being interested* is a precious instinct to preserve.
There’s a jetpack available for people who prolifically generate and ask questions about things they’re interested in.
The best education would prioritize closing a satisfying loop from the feeling of interest → discovery and learning → compounding.
8
1
10
1,672
This has been a long time in the making and I enjoyed going through it so much! A fun, approachable, interactive deep dive into quantum computing, made in partnership with Microsoft Quantum.
We've partnered with @BrilliantOrg to teach you about #quantumcomputing in a fun, interactive online course—which even includes a quantum Plinko board. Read more: msft.social/9zry67
1
3
10
My 3 yo learned so much math with this $8 toy from Amazon.
Recognizing numbers, counting forwards and backwards, skip counting, plus/minus, etc.
Tactile toys are great at this age, and she took to it right away.
She was counting well into the 30s or so when we got this, then easily got to 100 in a few days.
2
10
1,474
2/ When ChatGPT hit the scene, there was a massive wave of AI tutoring chatbots – but we decided to stay on the sidelines.
Our internal product testing convinced us that this wasn’t the way forward, even though we were impressed with the explanations they were able to give. Hallucinations were certainly a problem, but there was something more that didn’t feel right…
1
8
802
Nature or nurture? 4 yo is inventing "math games" for her classmates.
From a class parent: "She writes down a number and you have to say what it is. If you get it right, you get a point. At the end of the day, she'll make that number of drawings for you.
Her friends love it and they're getting better with numbers between 10 and 100."
10
588
I’ve been frustrated by interviews with AI skeptics because they somehow just ... sound so much dumber than their AGI counterparts. Why?
Some theories:
1/ It’s hard to make money if the skeptics are right, so the message itself is unwelcome.
2/ No bench of business leaders amplifying skeptic arguments. Being bearish signals you’re not riding the wave, even if privately many “AI-first” companies think it’s a bubble.
3/ Skeptics often pin their case to goalposts that AI quickly surpasses, then later dismiss those goalposts as irrelevant. Embarrassing, even if the goalposts are illusory.
4/ The ‘it’s a bubble’ crowd tends to narrowly focus on the high valuations of AI companies and the limits of LLMs, vs macroscopic observations of where the money is going and who bears the risk.
Anyway, Derek Thompson’s new piece “How the AI Bubble Will Pop” is a great read, featuring smart people inferring things from numbers. Link in thread.
1
11
1,743
Twitter = moralizing/outrage chamber.
I have a theory that’s why the great consumer product founders are usually quiet here.
Their superpower is listening, not speaking. Their instincts tell them this chamber doesn’t reflect reality, or what people really want.
A new meta analysis of millions of online posters finds that moral-emotional language is perhaps the number 1 rocket-booster for traffic.
Most ppl who spend time building an audience by posting online about politics (or anything) come to the same realization: high dudgeon outrage and out-group condemnation are easy tickets to going viral.
We're not ourselves here. The physical-world upholds a principle of kindly reciprocity (If you're nice to me, I'm nice to you) and negative punishment (Removing ourselves from ppl and places that are mean to us). Social media, for a variety of reasons, inverts this principle and encourages ppl to actively seek and fixate on out-groups who can serve as fuel for outrage, condemnation, and other moral-emotional communication. I think this is somewhat obvious, and I have no plan to fix it, but sometimes it's just worth pointing out the obvious.
2
20
3,843
14/ Our philosophy has approximated Bloom’s phases of learning
Stage 1: Develop interest and excitement
Stage 2: Deliberate practice
Stage 3: Solve problems you haven’t seen before / solve new problems
In STEM, nearly all learners are in Stage 1. Ever heard of someone who hates math? It's probably almost everyone you know.
The cultural change we’re excited to be a part of, that perhaps the coming wave of AI adoption will accelerate, is changing people’s relationship with STEM. We aspire for STEM learning for enjoyment to become the equivalent of running to stay fit.
1
10
442
So proud to have brought this to life over the past 2 years with @Greenstein_Joe and Carole Robin. Amazing to see the impact it's had on people's lives and leadership ✨
Introducing Leaders in Tech: leadersintech.org. Learn more about our beginnings at medium.com/@leadersintech/in…
1
9
7/ Unlike many AI applications in education, we’re not doing this in real-time. Every generated problem still goes through human review.
In learning, correctness isn’t just nice to have – it’s essential. A single wrong problem can shake a learner’s confidence or reinforce misconceptions.
1
7
561
8/ But the difference is that much of the manual, non-creative work of making a learning game is now done at the push of a button by AI. The AI handles the heavy lifting of implementation, but always under human creative direction and review.
Our course authors can focus on what they do best: prototyping ideas, designing pedagogical sequences, and hunting for those Mario moments where learning happens through play.
1
7
545
5/ That’s where AI comes in, but not in the way you might expect. Instead of using AI to explain concepts to learners, we’ve been using it to help us build better learning games, faster.
1
8
634
Sunday morning guilty pleasure: coffee + printed Substack essay (husbandt has taken the children outside).
"In his view, education tends to ruin talent, not enhance it... Johnstone began to think of children not as immature adults, but of adults as atrophied children."
3
1
9
2,094
6/ The capabilities have come a long way since our first attempts with GPT-2 in 2019, to ways that R1 is surprising us with what it can do just in the last 24 hours. Here’s a peek behind the curtain with some demos of the workflows in action.
blog.brilliant.org/hand-craf…
1
5
638
We are hiring designers for AI tutor and for learning interactives. We also have an internship open!
3
8
721
Loved this earthy, primal creation story.
Feels like a myth remembered, humming with something ancient and tender.
I’ve designed software, backed startups and helped founders. Now I’m writing fiction.
This piece is from a collection of fables I began this year. How we do one thing is how we do everything.
For people with 11 quiet minutes. Enjoy.
docs.soleio.com/keeper
1
1
11
4,522
18/ We have an absolutely cracked team of megabrains and diverse creative talents – multiple IMO medalists, ex-Editor-in-Chief at The Onion, Cannes Lion winner, knitwear designer for Marvel, an Amazon top 50 book of the year writer, and a large number of PhDs and dropouts from MIT/Caltech/Stanford/Harvard/etc, to name a few.
1
8
939
16/ A large percentage of people at Brilliant put Lockhart’s Lament in the list of reading that has most inspired them.
We’re passionate missionaries working hard to bring a different relationship to STEM to the world – to get people excited to figure things out on their own and learn to solve problems for the fun of it.
1
8
440
Told 4 yo I'd pick her up early if she could figure out her own things to do while I finish work.
She's doing her workbook and periodically asking ChatGPT for spelling help. Totally natural way to unblock herself.
Learning for her generation will be so much more autonomous.
8
3,727
Started using @WisprFlow voice keyboard today. Finally, an app that does the one thing I want and actually works (no affiliation).
It is such a non-intuitive UI insight that not showing the words *as you talk* is actually much better. You tap when you’re done with the thought.
I've become very attuned to what breaks flow and causes nervous system activation, and continuously having to watch for whether the voice dictation is doing what I want adds cognitive load.
Now that I've had enough reps with Wispr to know that it's going to nail the transcription pretty much perfectly, it’s great to not have to think about that and just talk like I would talk to a normal person, without monitoring the screen.
It feels much closer to natural turn-taking in a conversation, and results in an output that sounds much more natural, like how I actually talk.
1
7
545
4/ But the thing about building interactive learning is: it’s expensive. Designing the core gameplay loop – the part where learning feels like flow – that’s the fun part.
But then you need a thousand carefully calibrated problems across dozens of difficulty levels. Every level needs to teach something new while keeping learners engaged.
1
6
730
In rural Wisconsin, overhearing the 4 yo and 2 yo playing with cousins.
They’ve spent 2.5 *hours* drawing and narrating an epic tale:
They fly to the land of edible chairs and eat delicious chairs. But a spell is cast - they *become* edible chairs. Then they remember they can fly, and escape to another land…
[2 hrs later] Negotiating with the hairy cherry fairy, who wishes for Mary’s dairy.
😂 Please, let’s hold off on the singularity just a few more years.
7
743
12/ AI has greatly reduced the implementation constraint, and the new constraint is creativity.
Pedagogical design is hard. It’s hard to find amazing pedagogical designers. If that sounds like you, and you’d like to learn more about our team – shoot me a DM!
1
6
440
13/ Why and how people learn will change a lot with the rise of reasoning models. If we can offload much of our thinking, why might we learn? Our answer has been and continues to be: For enjoyment, first and foremost.
1
5
414
Those are now in the Archived Courses section at the bottom of the Courses page on web. Sorry for the inconvenience! We’re working on new and improved versions of our math and science offerings — some have already launched and there’s lots more on the way.
1
78
10/ Let's break down what it takes to create just one course – say, an introductory course on core pre-algebra concepts:
- 50+ core concepts to teach
- 20+ problems per concept
→ ~1,000+ individual problems
1
5
471
The US abandoned literacy as a primary goal of schools around the 1940s. Schools replaced proven methods like phonics with pseudoscientific 'whole word' reading. Literacy fell off a cliff, and never really recovered.
This disproportionately affected black students, because white students had more resources at home to compensate for dumbed-down schools.
Bring back phonics!
6
288
9/ This lets us iterate much, much faster. STEM is 10,000 topics. One of the biggest bottlenecks to providing high-quality learning games across 10,000 topics was sheer implementation cost.
1
4
501
15/ Cars can go 100 mph, and yet people run marathons. Humans don’t need to do distance running to hunt gazelles. There’s no purpose, other than that it's enjoyable and ennobling to achieve the potential of your body.
This idea has been around since Ancient Greece, and running ultra long races is more popular than ever, although people need it less than ever.
1
6
711
The hardest part of shipping this was the internal debate over whether the tweet would rile up the Phoenicians.
We remain committed to peace in Carthage.
1
6
168
Alabama is doing this for math, too, and it's working.
It's the only state where 4th-grade math scores are higher now than they were in 2019, before the pandemic.
Pedagogy matters!
wrote today about how even if you've heard of Mississippi's surge in reading scores, you're probably underestimating it:
2
6
1,944
11/ That last number used to keep us up at night.
With learning, you can’t just explain a concept once and move on. You need enough variations to let learners truly master each idea, enough edge cases to build real understanding, and enough difficulty levels to create the right learning curve.
2
5
463
Me: Eff. Have to bring a dish with “surprising textures”
Lisa: It's ok I cooked for Mario Batali, Jean Georges, & Saison. Can murder this
😳
1
5
Cultural programming happens so fast.
4 yo, 2 months ago, annoyed: Even the kids who lost got something.
4 yo, today, professorially: I came in 2nd, but everyone got a medal because that’s more fair.
Having many feelings about this…
5
458
Everyday indignities that children face.
Me: But *why* don’t you like the class?
4 yo: I have to ask permission to go to the bathroom.
Part of what school teaches is that children are not to be trusted, and order comes from the schedule and oversight of an authority figure.
6
802
Dwarkesh’s framework for picking podcast guests is also great for curating a media diet: Whose work do I want to spend 2 weeks learning about?
Most aren’t on X, or have ghostwriters.
Substack? The format’s better, but still low density...
My husband: “Are you just re-deriving the importance of books?”
6
1,272
Berkshire Hathaway shareholders meeting livestream is much better than watching tennis, but worse than watching football
4
