Jaiden Reddy

building a great teacher

[This Week in Startups @twistartups]

Frontier AI models are built to do your homework, not teach you how to do it. Shan Reddy of Aristotle joins Jason to discuss a voice-first tutor built to run a session like a human tutor would, then Ask Jason covers how to break into finance in the AI era, what to do with $500K and no obligations, and why AI doomerism might be coming from the labs themselves.

@jason @lons @rshanreddy @hey_aristotle, @JustinBrady @DaveJayRose

0:00 Live Ask Jason kicks off
0:14 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin S. You can check it out at https://t.co/AhhYi7Bzen and use code TWIST for 10% off!
1:40 Fordham-bound caller asks how to break into finance in the AI era
5:59 Jason's gift: read one biography a month
10:06 Vanta: Get $1000 off your SOC 2 at https://t.co/RpdW9AtnAj
12:06 Returning caller David: what would you do with $500K?
19:53 DigitalOcean: Want to see what building on a true AI-native platform looks like? Head to https://t.co/YCLFZ4OMDZ to start building on DigitalOcean's AI-Native Cloud today - and cut your AI workload costs by up to 50%.
24:40 Bringing an AI protein folding tool into drug discovery
27:19 How to do customer discovery with big companies
29:44 The "dark arts" of market research and corporate intelligence
30:13 Sentry: Your team should be focused on shipping features - not chasing down bugs. New users can get $240 in free credits when they go to https://t.co/Sr84r6e96I and use the code TWIST26
33:43 Is AI fear a dark PR campaign?
35:01 Anthropic meets with religious scholars
40:45 Are AI labs living through an Oppenheimer moment?
42:50 Why we need to talk about AI's upside in education
44:31 Sean Ready, co-founder and CEO of Aristotle, joins
45:42 Why AI models are getting worse at teaching
47:53 Aristotle demo: a voice AI tutor for AP Calculus
51:13 Predicting the next best teaching action
54:16 Bloom's two sigma problem and one-on-one tutoring
55:22 Aristotle's business model and pricing
57:53 Building a tutoring harness vs. relying on frontier labs
1:01:05 Why dyslexic students love Aristotle
1:02:40 Hand-grading thousands of hours of tutoring sessions
1:05:06 Socialization, classrooms and the future of public education
1:09:16 Where to try Aristotle and open roles

🎥 Watch the full episode here 👇

[Shan Reddy 🍎 @rshanreddy]

America is banning AI in schools.

China is using AI to create geniuses.

Introducing Aristotle: The AI tutor that solves America’s broken education system. https://t.co/hqPJN8cke0

[Jaiden Reddy @ReddyJaiden]

We are incredibly excited to share this with students worldwide. #LEARNOUTLOUD!

[Shan Reddy 🍎 @rshanreddy]

America is banning AI in schools.

China is using AI to create geniuses.

Introducing Aristotle: The AI tutor that solves America’s broken education system. https://t.co/hqPJN8cke0

[Jaiden Reddy @ReddyJaiden]

@rshanreddy LEARN OUT LOUD!! Incredibly excited to share this with the world.

[Jaiden Reddy @ReddyJaiden]

man, my older brother (& cofounder) asked me to show Aristotle how to teach about the history of Middle-earth.

guess I just have to reread Silmarillion immediately. for work.

[Jaiden Reddy @ReddyJaiden]

if u work in edAI, your job… the one you get paid for… is to read transcripts where people encounter Tolkien, or geometric congruence, for the first time. and then ponder: “How could magic language machines make this better?”

remind me why people work on anything else

[Jaiden Reddy @ReddyJaiden]

my takeaway from the OpenAI-HF “hack” is that RL harnesses should have model welfare requirements

prob shouldn’t put thousands of hyper intelligent beings in sealed white boxes with sticky notes in the exhaust vent

[Jaiden Reddy @ReddyJaiden]

@MParakhin As an engineer I think you’re right — context switching between each agent task gets super expensive

but I think for other roles where you need answers to questions, can run auto research loops, have longer running tasks, or where you can accept mediocre process ADHD wins

[Aristotle 🍎 @hey_aristotle]

Why isn't traditional AI improving education? CEO Shan Reddy talks about the gaps. Follow @hey_aristotle for more educational content. #edtech

[Jaiden Reddy @ReddyJaiden]

@bigsnugga I know they weren’t acting in the actual epic but lowkey felt like that when I was reading it as well

telemachus is not him

[Jaiden Reddy @ReddyJaiden]

This is one of the magical model capabilities that makes speech the best modality for students learning from AIs. You can ramble about photosynthesis for a minute and the LLM knows exactly what you mean… and exactly what’s next to learn :)

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Jaiden Reddy @ReddyJaiden]

@azlenelza would love to playtest!!

[Jaiden Reddy @ReddyJaiden]

codex is great for codewriting, but the model leaves so much to be desired. human conversations happen within a shared discourse, responses share some constructive logical chain

codex writes you an unreadably dense essay that at best loosely relates to something you referenced

[Jaiden Reddy @ReddyJaiden]

I wonder how resistant models are to jailbreaks where the real world (or real warfare) is cleverly presented as a videogame / simulation.

Kind of an “Ender’s Game” attack…

[Jaiden Reddy @ReddyJaiden]

@MasterTimBlais wow, the handoff is gorgeous. We’re the neighbor’s cat, right? Opening a new conversation, pretending it’s the same presence we spoke with the night before, but we know it isnt. and although we go to the LLM for help or even comfort, we don’t belong to it… no matter what we say

[Jaiden Reddy @ReddyJaiden]

@thsottiaux codex in terminal editing prev messages / navigating to prev messages is quite laggy. if I accidentally focus edit into an old message in history, I can’t move back forward

[Jaiden Reddy @ReddyJaiden]

LLMs are very much like psychics.

They’re both entities that say uncannily apt things with high confidence and make you wonder whether they’re deeply insightful, pattern-matching monsters, or just really talented frauds.

[Jaiden Reddy @ReddyJaiden]

@0xDevShah The problem is just how many knobs are available for tuning performance, some of which are difficult to build + require business specific logic

More importantly, evals - altho you CAN A/B test prompt changes on a single turn, complex, consumer facing agent loops are much harder

[Jaiden Reddy @ReddyJaiden]

knowing when we’re in the special case where the person communicating is likely to skew, or simply lack, the context needed to properly assist in a task feels critical.

e.g. when establishing a student’s knowledge before teaching a new concept

[Jaiden Reddy @ReddyJaiden]

“It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t so.”
- Mark Twain

Agents will need the same strange combo of confidence & humility that makes humans ask again, differently, rather than extrapolating our way to an answer

[Jaiden Reddy @ReddyJaiden]

I think this is especially important for consumer use cases, like therapy. For example, when advising a friend who’s having an argument, not wholly taking their perspective / trusting their information is a critical part of your role