Aviv Sheriff

thought experiment: what would you do if you had LLMs at work 10 years ago?

the stories are true

You can get a D on fractions, and public school will still move you up to algebra. In games, if you fail, you RETRY. So demotivating..? Except the hardcore versions of this (e.g., Elden Ring) are some of the most popular games in the world. People LOVE challenge. In learning science, that's called a mastery gate. You don't move forward until you master the level. It's such a natural concept. To learn higher-level concepts, you need to master the prerequisites first. Mastery gates reduce students' cognitive load as they tackle complex tasks. Mastery also builds confidence and increases motivation, while advancing students prematurely tends to demotivate them. At Alpha, kids have per-subject mastery gates and don't advance until they pass them. We're building games to teach kids business and entrepreneurship with the best principles from game design and learning science.

"The Book of Elon" author @EricJorgenson showed up at @alpha_founders to talk to the students. Some people claim Alpha students don't read books. In fact, Founders School students read 30 minutes a day after lunch, finish a book every ~2 weeks, and prepare for a group discussion. Then they get to meet the authors!

32% of Gen Z workers plan to take PTO to play GTA 6

Got to listen in on Chris Voss mentoring our students @alpha_founders today. Pure gold

Players churn if you force your goals on them. When I led the roadmap for Yahtzee with Buddies, every game launched a battle pass. So we built one too: "Prize Pass". It sucked. We used it to push the game modes we thought would increase metrics, but those weren't the modes players wanted to play. Retention dropped. Players have their own motivations, and you can't just force their hand. We revised the Prize Pass goals and rebuilt our entire strategy around motivational "clusters", building content for players who liked to collaborate, compete, collect, etc. This was a resounding success and led to record-breaking revenues, a positive reaction from our players, and industry-leading stickiness metrics. This influenced my philosophy on education and learning games. We have to start with motivation, as @PepsMccrea says. If a student loves to compete, give them a business-building battle royale. If they like to be immersed, give them a history role-playing game where they are there when it happened, make history-changing decisions, and form personal episodic memories. If the learning ties to motivational goals, they'll engage and retain.

Educators can learn from puzzle games. Level design in puzzle games is an art form. When I led the roadmap team on Yahtzee with Buddies, we built a complexity and difficulty framework: - complexity: the number of elements to juggle - difficulty: how hard it is to win The way it works: - introduce a new puzzle element, like a special tile - the player masters it - introduce another one, in isolation - the player masters it - put them both together (complexity increases) - the player masters that - now you can add more things on top of that This applies perfectly to education. It's called decomposition. Decompose the material into units, isolate them, teach. Then consolidate them into the shared context where the players put it all together. This is how we're building our financial literacy games at @alpha_founders.

When I was director of Apex Legends mobile my goal for the team was 95% tutorial completion. Yeah basically we needed to be fun, Interesting, AND also teach the mechanics well enough that basically everyone does it. Now let's do the same in education: - everyone wants to play - everyone learns how to play

Americans just scored 47% on a basic personal finance test, the lowest in the 10-year history of this test. Gen Z: 38%. Share of adults in "very low": 20% (2017) -> 25% (2026). Weakest area every year? Understanding risk. Great timing.. Don't worry, though; 30 states now require a finance class. Kids memorize IRA rules that will change next year and what's the minimum auto coverage in Iowa. Sweet. The education system missed the memo again. Nobody learns to write by memorizing essays. You need to learn HOW to write. At @alpha_founders we believe entrepreneurship is the lens through which you learn about business. We're building a finlit game that teaches: - Principles instead of trivia. - Learning placed in context kids will find engaging and immersive. E.g., "You're FDR in March 1933; banks are failing- what do you do?" - Starts with motivation: Some players like to compete, others like to express themselves, while others like to collect stuff. Our games will provide those paths. - Rigorous application of learning science. Grateful to collaborate with Alpha's insane learning science team, including Carl Hendrick, Becky Allen, and Zach Groshell. The world is at an inflection point. Major forces are coalescing into drastic changes in the world order (see @RayDalio). Kids need to understand money now more than ever.

Most games try to maximize ARPU: average revenue per user. My goal is to maximize ARPU: the amount of money the player makes after playing my game.

This just in: Playing video games increases kids' IQ by 2.5. Don't tell my mom. That's commercial games, mind you. Imagine what could be done if we built learning games with cutting-edge learning science. @alpha_founders

This is so helpful when learning how to prototype with Arduino. Imagine trying out your components in a virtual environment before ordering or connecting them, and debugging. You can be as ambitious as you want in your design, and the software will guide you as you build the hardware. The same applies to many learning problems. It's a great time to be building in education! @nateliason Wonder what hardware products we can start building @alpha_founders

Research: Fidgeting helps people with ADHD think. Yet we still make dozens of kids sit still in rows while one person talks at them for hours. The "teacher in front of a classroom" model was built for the industrial age. Watch the clip: the same kid sits still through Star Wars and squirms through a math lesson. Kids with ADHD move more when a task demands harder thinking. When researchers put motion sensors on kids and adults with ADHD during an attention task, they moved the most on the questions they got right. Fidgeting and moving created nervous system arousal that helped them process information better. A Stanford study also found people came up with about 50% more ideas while walking than sitting. I have ADHD. I take walking meetings, pace on calls, and still can't sit still in meetings. I learn best when I'm active or "playing." We urgently need to rethink the 100+ year old learning model, and that's what we're doing at @alpha_founders (part of Alpha School).

I picked up @bot a couple of weeks ago at @nateliason 's recommendation. Now I have ~25 agents organized into strike teams: - Researchers - Game prototypes - Recruiters - Admin - Other/catchall My favorite part is the cross-agent coordination and context sharing.

Must-read paper from Wharton (UPenn) on what ChatGPT does to learning. (bookmark it) The question: do AI tutoring help with skill acquisition? The setup: a randomized trial with nearly 1,000 high schoolers in Turkey. About 50 math classes in four 90-minute sessions: teacher review, practice, then a closed-book exam. Classes practiced with either: A) Textbook and notes B) "GPT Base": GPT-4 in a ChatGPT-style chat C) "GPT Tutor": the same GPT-4, told to give hints, never the full solution, and to ask for the student's work first What they found: With AI, practice scores rose 48% (Base) and 127% (Tutor). Then the AI was taken away. On the exam, GPT Base students scored 17% worse than students who never had it. GPT Tutor students scored about the same as the no-AI group. Why? The most common message to GPT Base was "What is the answer?"... Even worse.. The Base group didn't think they learned less. The Tutor group thought they did better on the exam. My thoughts: LLMs weren't built for learning. The products around them (i.e., "ChatGPT") also weren't. They are missing key ingredients of learning science: no scaffolding, no Socratic questioning, no mental modeling. The key risks are: - Cognitive offloading (the AI thinks, you don't) - Cognitive overload: walls of output nobody can process - Overblown confidence (as shown above) AI tutors have immense potential, but they require specialized products built with learning science in mind. doi.org/10.1073/pnas.2422633…

We’re hiring. DM if you want to learn more or join the team.