Fast forward ten years, and at the start of 2026, our home had stopped feeling smart. The problems decomposed broadly into two buckets. First, performance felt like it had degraded, where asking for our lights to turn off or on could result in delays of up to 30 seconds to a minute. Part of this can be explained by the fact that there are a variety of network delays associated with those requests, not to mention an LLM in the loop, ultimately introducing a latency that defeats the purpose of just using the physical switch.

Second, my expectations for what a smart home should be had increased. In a world in which AI is taking off, a smart home to me is one that can answer arbitrary questions about itself, or better yet, one in which a robot could come in and help out. Do I need to pick up dishwasher detergent on the way home? Where did I leave that book by Michael Pollan? Or where did I leave the keys? That would require distilling and understanding properties about our home, as well as their spatial relationships with each other for a robot helper.

Part of the reason my expectations have changed is because of the pace of technology. Even with a good understanding of deep learning and computer vision, building something like this in 2016 would have involved staffing up a medium-sized team to build a prototype, and hoping that interim results might buy time to launch and iterate towards a longer term roadmap. In 2026, with AI coding agents and significantly more powerful models, I felt confident that I could take this on on my own.

Table Stakes - Smart, Faster Lighting

The first task was to get to parity with what I already had with our existing home automation app, but making it more responsive. This was surprisingly simple with Claude Code. The first version of it was ready in 30 minutes, with millisecond latency from the app within an hour.

While next steps here could involve incorporating locks, the garage door, the dishwasher, and the washer/dryer, I felt confident enough that we could get there that I switch to the next task: creating an object-focused map of our home.

An object-focused map of our home

In a 1945 article in The Atlantic, Vannevar Bush’s “As We May Think” imagined a memory extension system for humans, and while most people looking back on that have described it as foreseeing the world wide web, the world wide web can’t tell me where I left our copy of Michael Pollan’s book.

However, I was able to design a system that could do that for our home, along with answering other relevant questions.

What about the robots?

Currently, we have Roombas on each floor of our home that sometimes crash into walls and furniture, so I’ve started thinking about what might happen if a more sophisticated robot were to present itself in our home. It could query the system above, but how would we provide it directions? To address this, I’ve started working through building a smarter map of our home. This is still a work in progress, and currently the maps are just from a handful of photos.

Because of our DGX Sparks, I’ve been able to run many models locally, paying for modest electricity costs, and without requiring special cooling systems.

Closing thoughts

A lot of this really comes down to questions about how to design and organize the information, to make sense of the physical world to simplify our lives. Part of this could come from being better organized in the physical world, but in the absence of a physical organization system, everything is in its place if you have a virtual organization system for finding it.

On that note, as I was writing this post, Andrej Karpathy posted about an LLM Wiki system for building a personal knowledge base, and if I squint, it’s touching on related ideas of how we organize knowledge and information for AI.

Have we reduced many of the problems in AI to building better search engines?

As I became a practitioner and student of improv and acting, I kept re-encountering the Stanislavsky acting system, which, when applied to written scenes, involved a lot of work deconstructing characters, arcs, etc.

As I moved from the spontaneity of the improv stage to the rigid deconstruction of the Stanislavsky system, I hit a wall: the “computational” tax of deep character analysis.

The Cumbersome Process of Analyzing Characters

During a class on Uta Hagen’s approach to Stanislavsky’s system, or a class on Shakespearean acting, I remember hours upon hours of analysis to model the characters I was playing (Jim O’Connor in Tennessee Williams’s The Glass Menagerie and Cassius in Shakespeare’s Julius Caesar).

As a writer, this computational challenge scales significantly. It’s not just analyzing one model but rather becoming the architect of an entire ecosystem of interacting systems. Now, there are multiple characters to analyze, multiple models to construct, and a more challenging computational problem.

Where AI can Help

If we view characters as multi-scale latent dynamical systems, where internal drives (latent states) shift across beats (scales) in response to external stimuli, then LLMs become the perfect tool to map these transitions. To that end, and with the help of Claude Code, I’ve built Uta: a play analysis engine that uses LLMs to extracts a play’s beats and a character’s tactic DNA from raw text, and then injects that model into a real-time improvisation loop. You can see visualizations of the extracted model in the screenshots below, or try out the code yourself (API key required to regenerate character models).

Once constructed, those models can be fed into an improvisation engine that can be used to generate novel scenes with a human, or allow for two characters to interact dynamically.

While LLMs are contaminated by existing scholarship on Shakespeare and Chekhov, Uta’s true power lies in analyzing original, unpublished scripts where the model has no prior bias, allowing a writer to stress-test their own characters. However, I did my initial testing on better known plays, and couldn’t resist trying out the engine on Hamlet (I played Jeff). Here is a lightly edited version of the generated output for flow and clarity1:

python scripts/run_improvisation.py session \

--character HAMLET --play hamlet \

--setting “An office kitchenette in Scranton” \

--stakes “Jeff has taken Hamlet’s pizza from the fridge and is pretending it’s his own”

JEFF: Hey, Hammy. I did order a pizza, and this is mine. Sorry, dude

HAMLET: Did you order it, Jeff — and if so, from where, and when, and does the box bear your name or some other inscription that I, in my apparent madness, might have overlooked?

JEFF: Relax, dude. It says “Ham” on it because I used the box for a ham sandwich last week.

HAMLET: A box, you say — a box that once held ham, and thus the name, and thus your innocence is sealed, and thus the matter ends, neat as a funeral. How convenient that the world arranges itself so tidily in your favor, Jeff, and how strange that a man who eats his lunch and takes another’s should wear that same unbothered smile.

JEFF: What are you going to do about it?

HAMLET: Do about it? Ha — a fine question, Jeff, a fine question indeed, for what does one *do* when the universe arranges itself so neatly against him — when the theft arrives gift-wrapped in ham and innocence, when every villain smiles and every grievance is made to look like madness? No, no — keep it. Eat well.

You can also have the scene be entirely AI-generated:

python scripts/run_improvisation.py \

crossplay \

--character-a LOPAKHIN --play-a cherry_orchard \

--character-b HAMLET --play-b hamlet \

--setting “A crumbling estate in an unnamed country, dusk” \

--stakes “Both men have come to say goodbye to something they cannot name”

=== Cross-Play Scene: LOPAKHIN meets HAMLET ===

Setting: A crumbling estate in an unnamed country, dusk

Stakes: Both men have come to say goodbye to something they cannot name

LOPAKHIN: I walked the whole length of it this evening — what’s left of it — and I’ll tell you something strange: I kept thinking how much money could still be made here, kept running the numbers in my head, and then I’d stop, and there’d be nothing. Just the stumps, and the smell of the cut wood still in the air, sweet as — I don’t know what.

[tactic=confess | score=4.67]

HAMLET: You stopped — that is the thing I find I cannot stop thinking about — not the numbers, not the stumps, but that you stopped, and there was nothing there to meet you: tell me, when the reckoning fails and the silence comes in, do you find it empty, or do you find it is somehow — full?

[tactic=probe | score=4.83]

LOPAKHIN: I stood there — I had the papers, I had the numbers, I had everything a man is supposed to have when he has won — and it was full, yes, God help me, it was full, but not with anything I knew how to hold: it was full of your mother’s voice, and the smell of the trees before they were mine, and something that felt — I don’t know, I don’t have the word for it — something that felt like being eight years old and not yet ashamed.

[tactic=surrender | score=5.00]

HAMLET: That fullness — yes, I know it, or have known its shadow — when the thing we have pursued arrives at last and we discover it carries in its hands not triumph but the precise shape of everything we have lost: tell me, do you think a man can act his way through such a moment, or does it merely pass through him, like light through a window, leaving the glass exactly as cold as before?

[tactic=receive | score=4.67]

LOPAKHIN: It passed through me — like you said, like light — and I stood there in the orchard, axe-men all around me, and I was eight years old again, and my father’s hands were on me, and I was weeping, and I did not know if it was grief or something worse — something closer to joy — and I have never told that to a living person, and I am telling you now, and I still don’t know what it means.

[tactic=confess | score=4.83]

HAMLET: Then we are both still eight years old, and the orchard is still standing, and nothing that came after has yet happened — and I think that is the only mercy I have understood today.

[tactic=shelter | score=5.00]

LOPAKHIN: I stood there in the dark after the axes stopped, and I wept — not for you, not for them, but because I didn’t know whose hands those were, swinging, and I still don’t know.

[tactic=receive | score=4.67]

HAMLET: Then they were no one’s hands — and no one’s hands have done no wrong — and the orchard stands yet, Lopakhin, it stands yet, in the only place where axes cannot reach it.

[tactic=consecrate | score=5.00]

LOPAKHIN: Yes. Yes, let it stand there. I won’t take that from you — or from myself.

[tactic=receive | score=4.50]

HAMLET: Then let this be the rite of it — that we two stood here, at the end of what was, and did not look away from each other; the trees are gone, yes, but the going was witnessed, and a witnessed thing is not altogether lost.

[tactic=receive | score=4.83]

LOPAKHIN: Yes. Witnessed. I think — I think that is the word I have been looking for, perhaps my whole life, without knowing it was a word at all.

[tactic=receive | score=4.50]

By automating the labor of Stanislavsky, we don’t replace the actor or writer. Instead, we give them a mirror to see their characters more clearly, even in a Scranton kitchenette.

Distill wasn’t meant to be another publication venue. It was an argument that the medium itself mattered. Instead of static equations and frozen figures, readers could touch the concepts. JavaScript was the bridge between mathematical abstraction and human intuition.

Of course, there was a caveat this kind of interaction required knowledge of Javascript.

Fast-forward: LLMs as game designers and teaching collaborators

When large language models arrived, so did the ability to generate interactive content through words rather than code it line-by-line.

Today, if you want to make a small game that illustrates e.g. Bayes’ rule, you can simply describe it to an LLM. You just say:

“Create a React app that be viewed on a mobile browser to visualize how prior and likelihood combine to form a posterior, with sliders for both.”

…and a few seconds later, you have a working frontend.

Pair that with a global CDN (e.g. Vercel, Cloudflare, etc.), and you’ve got instant worldwide distribution of a concept in pure-JavaScript. No backend, database, or DevOps. The web becomes a lightweight teaching platform.

This changes the economics of pedagogy.
Where Distill once asked: “What if explanation could be interactive?”,
we can now ask: “What if interaction itself could be authored conversationally?”

Why games are a powerful medium for learning

Games are the most natural teaching technology we’ve ever invented.
They reward exploration, expose feedback loops, and invite persistence.
When structured around an idea like a theorem, a model, or a process, they become living diagrams.

A well-designed educational game does what a textbook cannot: it builds intuition through experience. The player doesn’t memorize the rule: they feel it.
For abstract fields like machine learning or information theory, that kinesthetic dimension can make the difference between rote knowledge and deep understanding.

Building Hamming: a Wordle for coding theory

I’ve experimented with this idea myself. The result is Hamming, a browser-based game inspired by Wordle but grounded in information theory and error-correcting codes.

Each game of Hamming is a puzzle where you get 7 rounds of questions to find a mystery number, but one of the answers will be a lie, i.e. the opposite of the truthful answer to the question.

It’s playful, but it teaches a genuine concept: how redundancy allows communication to survive noise. It also allows one to play with the structure of a Hamming (7, 4) code.

I wrote the first version of Hamming describing the rules to an LLM in natural language, iterating on edge cases, and letting the model scaffold the JavaScript. Then I deployed it to a static site using Cloudflare’s CDN. From prompt to playable game, the whole process took less than an afternoon.

You can try it yourself at hamming.krisheswaran.com.

Reflection

Nearly a decade ago, in that corner of Google Brain, Chris was showing me how JavaScript could animate understanding. Now, we can do the same thing, but with the machine writing much of the code for us.

What hasn’t changed is the goal: to make abstract ideas tangible.

Play with it: hamming.krisheswaran.com

“And I remember one of the first jokes you told me: an introverted mathematician looks at their own shoes while an extroverted mathematician looks at YOUR shoes.”

We spend the first thirty minutes catching each other up on life. Outside of a group Zoom call during COVID, it’s been fifteen years since we’d spent any significant time together. When we were graduate students, Alex was the first person to teach me sleight of hand tricks. He founded a startup around the time I left Google, and he’s very open about his experiences.

“We spent a lot of time just vetting the right problem to go after with leaders in our domain. We hadn’t built a proof of concept or anything.”

Wait… what? “If you didn’t have a proof of concept, how did you get the meetings?”

“We’re mid-career, Krish. I just leveraged my experience and network, and it got meetings. You have a network and a reputation from 15 years of experience at Google. I’m pretty sure you could do the same.”

“And you go in with nothing?”

“I go in with my experience, and then I spend the meeting just listening and understanding what their problems are.”

“Of course, that builds trust.”

Both our heads are nodding now. Alex returns to something from an earlier part of our conversation. “It sounds like you had a lot of experience at Google socially engineering your environment to enable progress. That’s exactly the skill set you need to be using.”

The subtext? Sure, there are indie hackers out there, but that’s not how I can differentiate myself.

Within a few days, and with the help of my network and reputation, I’ve landed a meeting with a C-suite executive leading one of the most prestigious organizations in my target domain. Alex has taught me a new trick.

My friend and former colleague Varun has described the experience of CEO Founders as analogous to the hero’s journey popularized by Joseph Campbell. This newest trick is the Supernatural Aid propelling me forward.

Based on that discussion, we found a pathway to reassign half my reports to other managers. This change meant that I would hand off the radiology AI efforts that I’d been on since the start of 2017, a difficult decision that paved the way for a continued set of difficult decisions and led me on my current path.

Part of the team I continued to lead during my final months at Google was exploring a new research effort into building a foundation model to help on a wide range of geospatial problems including health, economics, and the environment. Those final months taught me a lot, and while I left Google before the work was ready for publication, I was really glad to learn that the team’s paper just came out. The news reached me as I was working to convert our home garage into a test site for my more recent, post-Google efforts, about as hands on as I could possibly get.

My phone vibrates. “chatgpt has given me a reasonable answer to try.”

As my dad goes off to try out the advice of his new AI child, I think through the implications. Computer-related troubleshooting has been one of the things I have done for friends and family for decades. While I have come to rely more heavily on ChatGPT for my own debugging needs, it’s another thing to realize that those who may have come to me in the past have it as an option as well. That’s both a relief and a little humbling.

I check in with my dad on how the advice has gone. He hasn’t tried it out yet, but the steps he describes sound reasonable to me. With some luck, this will address the residual errors, and it will compile before lunch.

I nod.

“That must be quite an adjustment and leave you feeling a bit vulnerable.”

I’ve just shared a recent exchange in which I told someone about my project, and how their response is still echoing in my head. Don’t tell me you’re spending all your time on that.

That response came at a moment when I was otherwise in flow. I’ve developed much stronger intuitions about how to maximize performance out of the AI systems I’ve been working with. Many of the key technical issues are being addressed. I have a plan and am executing methodically. And I am able to test my system out in the wild, on my phone.

“Yes, it did leave me feeling more vulnerable.”

“It also sounds like you’ve been getting a lot more enjoyment out of the process, and this started you on a track to thinking about the product.”

That’s a fair point. This took me out of a process that I was enjoying, and knocked me out of equilibrium into thinking about the product.

“Yeah, and maybe I should be thinking about that product, but probably with one of my business advisors.”

“Yeah, you’re working solo, but when you make yourself vulnerable, you want it to be with people you trust. Have you considered a founders cohort?”

I shake my head. I’d seen someone mention it but hadn’t pursued it.

“Maybe you should consider it. You’ll probably get better advice than the Walgreens person administering your vaccine.”

I laugh and nod.

The comment came as I was on track with a plan, and in my mind, this was one of the pieces that would fall into place later during a pilot. However, as I reflected on the side comment, it dawned on me that waiting for the pilot phase to solve this would add more risk to the pilot itself.

And so over the course of a day, rather than address any of the list of problems that were part of my plan, I built a system that allowed me to mitigate those later risks up front. As I introduced that system into my solution over the next week, it became clear that many of the problems I had been encountering and had been added to my list were automatically solved by creating this new system. Furthermore, it opened up a number of new possibilities that had not felt feasible just a week earlier.

One thing I have valued recently is how quickly I can change course when things aren’t working, or if they just feel a little off. While these changes don’t always lead to success, they’ve helped me maintain a momentum that has enabled more successes to emerge as the result of quickly being able to test a range of ideas.

That was hard to replicate in other environments I had operated in. One key difference has been size. Leading a large effort in a big company involves a lot of stakeholder buy-in. While those checks and balances are important to avoid degrading a user experience or avoiding harm, when a project is at an early stage, those kinds of constraints can significantly reduce that momentum. However, when smaller teams are able to incubate those kinds of changes within larger organizations, and those teams are able to demonstrate success, I’ve found the lack of but-in that may have created headwinds for those teams early on are often rewarded later.

As I think about how this effort will play out over the coming months, I am excited to take in other side comments, and make quick adjustments where needed.

One good thing about having a working proof of concept has been the ability to test what works and what doesn’t. It’s been a way to get quick feedback at a technical level.

Progress update? We could use this next week. Does it actually work?

One challenge about telling people you have a working prototype is that it creates expectations, and early tests have revealed the system is not ready for real world testing.

That would probably work, and I’ve known it for months. Why am I procrastinating on it?

One complication about getting technical is that it’s exciting to be working in that capacity, and exploring broadly is a way to prolong this phase of the project. I think I have a fear that I will have to give that part up again at some point, but maybe the thing I need to remind myself of is that I can always return to this phase just as I have now, and that other phases of the project will bring their own forms of enjoyment.

“Did you notice that naive linear regression performed almost as well as the state-space model?” And there it was. I had been so caught up in what Brown’s team had built that I didn’t really care how they built it. They could have demonstrated the result much earlier without relying on the state-space model, and I would have been just as impressed.

This idea ended up working to my benefit multiple times during my time at Google. In a company with a short attention span, demonstrating things quickly has its advantages, and that often meant leaving more sophisticated approaches on the table to get to something that worked.

However, those more sophisticated approaches were more exciting to work on, and I would often find myself or members of my team drawn to them. New technologies, techniques, and approaches would call to us, and we would be taken in. It required a certain discipline to prioritize what was important to get to a minimal viable solution.

And so it’s been with my life outside of Google. The pace of technology has generated a trail of methods with a strong allure, and my first couple months have included days in which I was drawn in by ideas that have introduced more challenges: difficulty debugging, complex undocumented dependencies, and expensive resources to run.

And then there are days like yesterday, in which, shortly before bed, I wrote down a technique that I would be able to validate in a day, relying primarily on textbook approaches with well documented libraries, debugging tools, and best practices. Also, it could run on my local CPU. With a few hours of effort today, I had an end to end system that mostly worked, or at least provided a baseline.

Minimal. Viable.

The first month has made me feel like a beginner in so many aspects, learning a new problem domain, understanding what that problem might be, being overwhelmed by the many sub-problems that might emerge, and trying to figure out what might be important. And then the responses come in, and I start parallel conversations with the advisor and the potential customer.

It’s a good start. Are you free to chat later today? I’d like you to try fleshing this part out. Oh, you really don’t know how this industry works, do you? From what I’ve seen, businesses in this sector may also require you to build out these other features. I don’t think some of the things you’re working on are important right now, but some others are.

It’s difficult to make oneself vulnerable and get feedback on one’s work. It opens oneself up to all kinds of comments and anxiety about how others perceive the effort you put in. However, when the responses come in, the conversations that result make me realize that people are actually taking me seriously, and with the feedback, I get a much clearer sense of what I will be doing for the next few months.

I filed my PhD disseration in Berkeley on a Thursday afternoon, which gave me a three-day gap before starting what became a fifteen-year stint at Google. When I think back to the days in between, the main event that marked the transition from graduate to tech employee was the Festival of Harps at the Chapel of the Chimes. The event included Karma Moffett’s indelible Tibetan Bell Experience, which I still reminisce about with the friends who attended with me. They have since moved to Toronto.

Instead of a Vipassana retreat, a hacker residency, or a harp festival, Aarti and I are in SFO, waiting to board a flight to Toronto. In addition to the friends with whom I attended that harp festival, Aarti’s brother has moved there, and I can’t think of a better way to reset than to spend time in a different setting catching up with loved ones.

We end up sitting next to someone who helps small businesses, and after informing them of my current status, I get a lesson on protecting myself, my intellectual property, following compliance rules, and all the mistakes new ventures make when they first get started. I am on my own.

Notifications on my inbox continue throughout the day but are mostly from recruiters. On the car ride back from Carmel, I let Aarti know that I’m feeling a little overwhelmed. She brings up the fact that it’s only my first day, and I can figure things out on my own schedule. My goal for the first week– no, the first month– should be to have fun.

And so I keep that in mind as I proceed through the week. It stays with me as I read a few papers from 2008 that I had nearly forgotten about, run a number of machine learning experiments and document my learnings, trace down a bug to an issue with ARPACK, which was written in FORTRAN 77, learn the ins and outs of structuring a Delaware C Corp and the tax implications, discover a few additional nuances of the use case I am exploring from the end user, come up with an initial plan to get to a minimum viable product (MVP) or early prototype, create my first GitHub pull request for feedback, check out a product demo of the new Apple Vision Pro glasses, discover a cafe near the Presidio, get a haircut, start drafting a new short fiction piece, start reading a new novel, receive compliments on my progress in kickboxing, discuss the prospect of coaching a new improv troupe, and ignore over 100 recruiter emails.

Having fun… check.

“15 years.”

“I’ve been at Apple for 11.”

“Do you think you’ll stay?”

“Yeah, I’m a lifer.”

Two hours earlier, I had turned in my laptop, corporate phone, and badge at Google’s San Francisco office. When I got home, Aarti told me that despite my mentions of wanting to do something new, she had just assumed that I was too attached to my team to leave Google. She was proud of me.

The final meetings I’ve had with Google colleagues over the past few weeks have centered around a related question: what caused me to change from a Google “lifer” to an entrepreneur? Indeed, there were opportunities that I had passed up before to stay at Google. As a technical person living in the Bay Area, the past 15 years have been riddled with moments in which recruiters, start ups, former colleagues, and friends have tried to entice me to join them in a compelling venture. At times, I’d even entertained these conversations, only to be offered a project at Google that made the proposed external one seem less appealing by comparison.

The answer I have given people in those final meetings starts in December of 2022 when I started playing extensively with ChatGPT and continues into the present, in which I was running experiments using a number of open source foundation models in my spare time, reading papers, and opening my eyes to possibilities outside of the company that has been my first and only full time job. Sometimes I end the answer by describing the start ups, VCs, and potential clients I have spoken with that gave me the conviction to pursue this new direction.

What that answer misses are the many other moments that came before it. There were my first 20% projects at Google, at the encouragement of my managers, which gave me the confidence to explore new ideas outside of my day-to-day. There were the technical mentors who taught me everything from distributed systems to applied machine learning, enabling me to delve deep into multiple technical areas while giving me the breadth to understand a variety of considerations when building a solution. There were the managers who taught me the nuances of working with others and many of the soft skills that enabled me to be effective in navigating a variety of hurdles. Then there were the many opportunites to lead technical efforts, in which I was working directly with clients and learning what was a must-have and what was an optimization for later. There was the time I got to participate inside an Alphabet startup and learn a lot more about the nuances of a business. There were the times I started to manage people and teams, in which I learned how to navigate a number of difficult situations, pivot where necessary, take risks along the way, and learn from both successes and failures. There were the moments in which I helped identify new external partners, and learned how to build trust with their organizations. All these moments suggest that my experiences at Google have played an outsized role in incubating the confidence and experience leading up to the present.

And for that, I am greatful for the many people I met over the years for helping me here. The coming months will no doubt involve a number of new experiences that will challenge me, but I think the foundations I have developed have given me the grit with which to engage with these. Thank you, and stay tuned for updates as I embark on this new venture.

Rewatching old episodes of Beyond 2000 offers a window into how future versions of ourselves might view the current moment. Through the episodes, one can learn about the different strategies companies took to try to introduce novel ideas into the market. Many of the designs ended up being either an accessible wedge for the later disruptive technology (e.g. Kodak photo CDs for analog cameras) or a bit too forward-looking to be practical (the bulky wearable computer that holds a CD player). In each of these cases, the disruptive technologies that would pave the way for large-scale adoption had yet to be developed.

I’ve been thinking a lot about how to find the sweet spot in technology that both shows the future while capturing value to benefit people today. Does it start by showcasing the art of the possible and then adapting those concepts into real world use cases? Who are the partners one can engage with to get early feedback on whether an idea is moving in the right direction? Here, I’ve come across another video, an interview of NVIDIA’s Jensen Huang at Stanford. Unlike old episodes of Beyond 2000, something tells me this video will still be relevant thirty years from now.

Our goal is to try to understand ways to help the city work more effectively, and which actions might have the most impact. Sifting through the data gives us an understanding of things like San Francisco’s procurement process and also a sense of frustration at being provided with so much data, but ultimately despite a number of open standards, how difficult it actually is to work with.

Data has certainly been key in driving a variety of decisions in my day job, and perhaps the most important is which problems and areas to focus on. The hope is that we can do something similar in engaging locally. I write some convenience functions to better analyze the data. I feel a tap on my shoulder. Aarti points me to a graph she’s created. I smile.

The job doubled as a business class. When I joined Google Fiber in 2012, it had a hundred or so employees, an order of magnitude smaller than the thousands at YouTube. That meant visibility into every aspect of the business. Understanding those mechanics benefited me significantly in the roles I have taken since. Here are two examples that serve as instructive use cases:

1. Connecting a home to the Internet requires an up front investment, but once established, is relatively inexpensive to operate, resulting in a high operating margin. Furthermore, customers tend to be sticky. Thus, a company that is able to handle the capital expenditures to connect homes, could, given a large enough user base, benefit from the healthy operating margins to offset that cost. It’s also worth noting that certain investments to provide connectivity, like spectrum licenses, are assets that don’t depreciate, so they effectively function as cash and could be sold at a later date.

2. Internet providers charge a lot for TV service because customers want it. Unlike the Internet connection itself, the cost of providing that content on a monthly basis is high, so the margins are very low. However, because other ISPs also offer TV service, this becomes table stakes in order to have parity with one’s competition. There are multiple ways to respond to this. One is to make up for the low margins with high volumes, which is what the largest providers are able to do. Another is to bundle the television service with higher margin offerings (triple play). A third could simply be to encourage customers to use over-the-top services instead of their own TV offering, which may reduce revenue but depending on the price point could actually increase profits.

I’ve thought back to these examples when thinking about how the AI landscape might evolve. There are currently multiple companies spending a lot of money to build large models. The bet is that there will be a way to recoup these costs by charging customers for the privilege of using these models (even so-called open models have language around this). If costs for updating and serving these models could eventually be reduced, then one could image a a business model similar to an ISP’s connection to the home.¹

What about downstream business customers of these AI services? If the costs of serving these models cut significantly into their margins, then they are in a similar position to the TV service. Options available to them may include making up for the lower margins through higher volumes, bundling their AI product with a higher margin offering, or developing a strategy in which the cost associated with using these models does not scale with the size of their business.

It’s early days for AI companies, who themselves have an interest in creating a healthy ecosystem to nurture business investment in this space. As these AI models continue to amaze me in the art of the possible, I am equally excited to see how the business models evolve.

1. There is also a world in which either some of those costs do not come down and/or some data to train these models needs to be purchased/licensed, in which case the business model may resemble something more like TV service. ↩

In the over ten years since, I have hosted a number of similar parties, both for primary and general elections, all with the goal of enabling more informed decision making. As for the motivation for these parties, the state of California, Alameda County (and San Francisco County post-move), were more than happy to oblige with a number of confusingly worded initiatives with unclear implications.

Then came COVID, which completely changed the dynamic. The lack of in-person gatherings combined with friends moving away meant that attendance slowly dwindled along with my motivation to keep hosting.

That didn’t do me any favors when it came time for the upcoming March 5 primary, which as usual has several local measures. What has changed since my first ballot party is that there are a number of web sites offering voter guides, mostly created by different special interest groups. As I looked online, though, I noticed that some of my sources, like SPUR, were only endorsing a subset of initiatives this time. While I don’t always agree with SPUR, I’ve found their analysis on the pros and cons of different ballot measures helpful, and it looked like they weren’t providing this for the measures they weren’t endorsing. This included Prop E, which contains a number of sub-proposals and would require some research.

As I compiled that research into a spreadsheet, I noticed other sheets in my drive from earlier ballot parties. These sheets included endorsement matrices from organizations like SPUR, along with free text descriptions of the initiatives, as well as their pros and cons. Could I use the past endorsements and a language model to help me predict what my votes might be on these measures?

On a whim, I started putting together a recommender system based on the data and see how well it would do in predicting endorsements from a held out set. Unfortunately, I couldn’t find a list of my past votes, so the system would respond with recommendations treating each web site as a voter persona, so I could browse the recommendations for SPUR, the League, the Chronicle, and other sites that had put together voter guides in past years. I decided to use BERT as my language model to convert the ballot-related text into embeddings.

The process was surprisingly straightforward. Within a couple hours, I had preprocessed the data, created train/test splits, written a simple Factorization Machine, got my training loss to converge, and done a basic hyperparamter sweep. That’s where the good news ended. My evals indicated that the recommendations being produced were mediocre at best. It was at this point that I took a closer look at the data and discovered one possible culprit. The text in the sheet was of varying quality, which in turn depended on who had been assigned to which ballot measure. Indeed, subtle changes to the text changed the outcome of the recommendation. Furthermore, despite years of ballot parties, we weren’t talking about a lot of data, so it wasn’t surprising to me that the model would overfit.

While I had some ideas on how to address these challenges, with the election only a few days away, implementing them felt like procrastinating on the actual task of how I would vote. Still, I was curious to see how the model had predicted SPUR’s endorsement of Prop E, the one on which SPUR had refused to take a position one way or the other. When I looked at the model’s predictions on 2024 ballot initiatives (I had held out all of 2024), the model had assigned nearly all of them a probability > 0.9 (likely YES endorsement) or < 0.1 (likely NO endorsement). Despite the noted problems with the model itself, I was nonetheless quite excited to see how it had placed its bet on SPUR and Prop E. I cross-referenced the correct row and column corresponding to SPUR’s Prop E prediction and realized then that it was the one exception for which the model had not assigned a high or low probability. Instead, the model had assigned a probability of 0.4821, refusing to take a position one way or the other.

After my efforts, I sought out a zero-shot opinion on the matter via ChatGPT because why not? It provided a few paragraphs of analysis followed by the following statement: “As of now, there hasn’t been a specific recommendation from SPUR (San Francisco Bay Area Planning and Urban Research Association) or a direct statement on their stance regarding Proposition E.” I returned to my ballot research.

So at the start of 2023, I started using OpenAI’s text-davinci-003 API to change how I consumed the news. NewsAPI.org and RSS feed libraries provided different mechanisms to access news headlines and summaries, so I created a custom news reader to filter how the news initially showed up. At the time, my prompt was to just look at the headline:

 Summarize the above in one sentence for an important person who wants to get
 the details succinctly, sticking just to the facts and removing opinions.

Over time, I evolved the prompt and API, e.g. switching to GPT3.5, and experimenting with providing few shot examples. This morning, I took a look at how my news reader was changing the underlying articles.

The first focuses on a decade-plus covert relationship between the US and Ukraine to mitigate for Russian influence. I was generally satisfied with the changes made by my reader, in which what initially read like the trailer for a movie about a covert op was adjusted into an investigative report on the same topic.

The second article focuses on an UK politics. As someone who hasn’t been following this story, it wasn’t clear to me whether some of the hedging ended up distorting some of the facts (e.g. is “controlled by” a better paraphrasing of MP Lee Anderson’s statement than “influenced by”?), but I did appreciate that the standard gotcha articles in which a politician refuses to take a stand on an issue was called out a bit more plainly.

The final article is about the Republican primary. I thought this one does present a change that distorts content, specifically in how it characterizes Haley: “Counterforce within MAGA movement.” It might be splitting hairs, but I think a more accurate characterization would be a counterforce within the Republican Party to the MAGA movement.

While the re-writing of these headlines does sometimes distort the potential content of the information contained therein, overall, I have enjoyed the more boring presentation of facts compared to the clickbait of the current news headlines. This is a news I can bear.

When it came to non-fiction, most of my books were from grad school, and a decade had passed since I’d graduated. I had started managing double-digit sized teams at that point and asked myself how much I’d realistically be able to stay technical. This was also around the time that Marie Kondo had reached peek popularity, and I took a hard look at my bookshelf to determine what still sparked joy. I kept Durrett’s Probability: Theory and Examples while saying goodbye to Billingsley’s Probability and Measure. Yes to Feller (volumes I and II). No to Emanuel Parzen’s Stochastic Processes. Goodbye Structure and Interpretation of Computer Programs. Goodbye, Gallager’s Information Theory and Reliable Communication. I’ll miss you most of all.

This was at the start of a pandemic. Places closed. Then they opened again with some restrictions. We started leaving the house again. We stepped into a nearby bookstore. As I walked to the math and science section, I was surprised by the lack of Dover or Springer books. In their place were popular math and science books. I went to other bookstores and noticed the same pattern. Occassionally, there would be a stray Dover at a used bookstore or a Six Not So Easy Pieces, but by and large bookstores had konmaried these books out.

I searched online. Some of the books I had given away were out of print. I felt a sense of loss, so I tried to find a constructive way to cope. Yes, I went on a book shopping spree. This was around the time that text to imeage models were becoming popular, and I was playing around with stable diffusion to create AI art, so I leaned into getting back some probability and statistics books including some new ones focused on machine learning. Welcome back, Casella and Berger. It’s been too long, Noga Alon.

My favorite addition to the shelf has been Kevin P. Murphy’s two-volume Probabilistic Machine Learning. It’s a tour de force through so many of the algorithms and approaches surrounding both what we experience in our software today, grounded in explanations that should be accessible to anyone with a solid foundation in probability theory and statistics. It’s the one I find myself picking up the most often, flipping to random pages based on something I had thought about on that given day. And it’s made me feel more connected technically to the work that is happening on the teams I manage as well as the broader research community. It sparks joy.