Jon Vaughan

CEO at Isara (

[Aakash Gupta @aakashgupta]

A third grader just scored a 5 on AP Calculus BC. The system that trained him contains no LLM. The core is a knowledge graph one man spent 250 hours encoding by hand, two minutes per edge.

The platform is Math Academy. Its "AI" is an expert system that routes each student through nearly 3,000 math topics, from 4th grade arithmetic to the math behind machine learning. Every node, every prerequisite link, every weight was placed manually by a team of mathematicians.

The weights alone took Justin Skycak a full month: 1,500 topics at the time, roughly 5 prerequisite links each, 2 minutes to estimate each one. 8 hours a day of pure encoding, done before ChatGPT existed to ease the load.

Why go through that? Because the graph unlocks mastery learning, the closest thing education research has to a cheat code. In 1984, Benjamin Bloom showed that students with one-on-one tutoring perform two standard deviations above a regular classroom. The average tutored kid beats 98% of the lecture hall.

Nobody could afford a tutor per child, so the finding sat in journals for 40 years. A prerequisite graph with a mastery gate is the workaround. The system always knows the exact next topic a specific kid is ready for, drills it until proven, then moves on. Zero time spent waiting for 29 classmates.

That waiting is most of school. A year of classroom math is roughly 150 hours of instruction, and the majority goes to pacing, review, and re-teaching. Strip it out and a motivated kid covers six grade levels in one calendar year.

The origin makes it better: this grew out of a math program at Pasadena High School where 8th graders were passing AP Calculus BC, back when the founders were still hand-grading the whole thing.

The most effective education AI running today is a graph a few humans built by hand, one edge at a time.

[Brian Armstrong @brian_armstrong]

How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching.

Better Defaults (not Usage Caps) – Engineers can choose any model they want, but defaults matter. We’re experimenting with defaulting to open weight models like GLM 5.2 and Kimi 2.7 through our LLM gateway, while still encouraging engineers to choose the right model for the task. 91% of our employees were never hitting their usage caps, so instead of lowering caps and driving up alerts, we're moving to cheaper defaults. Note that code reviews use a diversity of models, so they can check each other's work.

Better Routing – In our custom harnesses, we preprocess prompts and route to the best model for the job, considering cache hits and model pricing. For instance, you may want a frontier model for planning, but not for execution where they can be overkill. Ultimately, humans shouldn't be choosing models - AI can automate this task.

Better Caching – Cache misses are the easiest way to drive your cost up. All of our requests are cache aware, so we’re reusing a warm cache wherever possible. For example, our cache hit rate went from 5% → 60% in LibreChat once properly implemented.

Keep Context Lean – Start fresh sessions when switching tasks. Scope file context narrowly. Disconnect unused tools. Don't just compact. The goal isn't fewer tokens used, it's fewer tokens wasted.

Better Visibility – Our engineers can use as many tokens as they want, from whatever model they want, but we’ve made usage visible – and the more you spend on AI, the more impact we expect.

The goal isn't to suppress usage. It's to build the infrastructure that makes exponential growth sustainable.

Putting this into practice has cut our AI spend nearly in half, while our token usage continues to grow.

[Can Bölük @_can1357]

Took it a step further, we just sending tool results as images now, -20% token usage, lol.

[Jon Vaughan @jsvaughan]

claude breaking out the perl today, old school

[Jon Vaughan @jsvaughan]

opus 4.8 still loving the smoking gun

[Mitchell Hashimoto @mitchellh]

I strongly believe there are entire companies right now under heavy AI psychosis and its impossible to have rational conversations about it with them. I can't name any specific people because they include personal friends I deeply respect, but I worry about how this plays out.

I lived through the great MTBF vs MTTR (mean-time-between-failure vs. mean-time-to-recovery) reckoning of infrastructure during the transition to cloud and cloud automation. All those arguments are rearing their ugly heads again but now its... the whole software development industry (maybe the whole world, really).

It's frightening, because the psychosis folks operate under an almost absolute "MTTR is all you need" mentality: "its fine to ship bugs because the agents will fix them so quickly and at a scale humans can't do!" We learned in infrastructure that MTTR is great but you can't yeet resilient systems entirely.

The main issue is I don't even know how to bring this up to people I know personally, because bringing this topic up leads to immediately dismissals like "no no, it has full test coverage" or "bug reports are going down" or something, which just don't paint the whole picture.

We already learned this lesson once in infrastructure: you can automate yourself into a very resilient catastrophe machine. Systems can appear healthy by local metrics while globally becoming incomprehensible. Bug reports can go down while latent risk explodes. Test coverage can rise while semantic understanding falls. Changes happens so fast that nobody notices the underlying architecture decaying.

I worry.

[CIX 🦾 @cixliv]

You guys aren't going to believe this (I had to double check it was real).

Unitree has made an actual mecha like Gundam, the GD01.

[Kent Beck 🌻 @KentBeck]

Just listened in on a customer service call with @Chewy. No phone tree. No rush. Warm, human, and above all effective as to the purpose of the call.

[John D. Cook @JohnDCook]

I read a book about the Apollo program that used the phrase "frequent soft contact with reality" to describe an ideal development strategy.

You need to touch grass routinely, preferably not at hundreds of feet per second.

[Nick Foley @nhfoley]

When we were originally designing this scooter I knew we could never get legal approval for two riders, so we secretly made it handle smoothly and safely even with this much weight.

Lime has nerfed a few things since, but the spirit remains.

[ThePrimeagen @ThePrimeagen]

This is such a bad mentality

[Jon Vaughan @jsvaughan]

@francoisfleuret Had similar thoughts. Was wondering if many of these “I’ve never looked at the code” projects will eventually grind to a halt. This paper says that ai less likely to reuse code https://t.co/pUdXaSXrxT

[Jon Vaughan @jsvaughan]

@agupta paper trading though - how are you determining if there is any liquidity?

[Jon Vaughan @jsvaughan]

Top tip for platform success - make sure an LLM can read your API docs

[Jon Vaughan @jsvaughan]

not going to vibe code your way out of it:
"the surface-level plausibility of AI code masks redundancy, leading to the silent accumulation of technical debt"
https://t.co/pUdXaSXrxT

[Jon Vaughan @jsvaughan]

This equation will change how you see the world (the logistic map) - got here from the numberphile vid which is also good

https://t.co/tUxIbfp4fy

https://t.co/QmWdU1zGaK

[Jon Vaughan @jsvaughan]

(old but still very relevant) - all you have to do is reduce customer effort
https://t.co/hwaeKJlAcn

[Jessica Nutt @JessicaNutt96]

Mythos is not a bad name for a model but it would be better if Anthropic switched to using famous Claudes. Monet, Debussy etc. The final model that achieves AGI would obviously be Van Damme

[Simon Willison @simonw]

Shocking result on my pelican benchmark this morning, I got a better pelican from a 21GB local Qwen3.6-35B-A3B running on my laptop than I did from the new Opus 4.7!

Qwen on the left, Opus on the right

[Rousseauist Bonobo @adjsjdnskn]

I think you would go insane if you really comprehended the multiplicity of alien forms who populated the Earth for millions of years. I spent two hours boggled in a natural history museum and could not make heads or tails of it. It's not a subject truly accessible to the mind