Jimmy Cerone

Link of the Day: How Zed's Open-Source Edit Predictions Work

2025-11-19T03:01:40Z

How Zed's Open-Source Edit Predictions Work

There's a few interesting things in this video:

- Recorded 9 months ago, feels like a lifetime ago
- Nobody shares their secrets sauce, these folks might be the only ones
- Fine tuning (using bigger models to create data, good vs terrible output rather than single outcome training)
- speculative decoding (better guesses for next token, wider range with higher accuracy)
- Speculative vs deterministic software (find link for AI product article)
- Model hosting (so different than hosting "apps", capacity vs latency)
- Evals (unit test for LLM, only runs on a fine tuning run)

So much has changed in the last 9 months it is wild. This is before any IDE had any agent panel and before Claude Code and Codex launched for the CLI. It is hard for me to believe that all that has happened in 9 months, but here we are. While many really cool open source models have launched, very few folks show off their secret sauce like in this video. There is a ton of complexity in translating a "normal" LLM into something that can materially impact coding speed. Cursor would not have gone through the time to develop their own model if it was not worth it. So far, the cursor sauce is special. Their agents are fast and fairly smart, though I hear that Codex and Claude are still smarter.

Since this video, I've stopped using Zed as my daily driver. At the time of their video, their next edit prediction was incredible, much better than GitHub Copilot which was the leader then. Now though, it seems that Cursor and others have kind of stolen their tricks. I've slowly started to feel that the consistent UI of Cursor outweighs the snappiness of Zed as their agent / edit prediction advantage starts to fade away. Using Zed used to feel magical. Now it feels normal and a little bit outside of my comfort zone due to the differences from VS Code.

While Zed seems to have lost their advantage, their use of fine tuning and Cursor's subsequent replication of that strategy lends real credence to the idea that general models are not enough in specialized fields like coding. I wrote about the importance of training data before, but now it does seem that IDE adoption will be a compounding advantage going forward. If the smartest folks are using your IDE to write the best code, all of a sudden you have the most and best training data. I do think this is a weak spot for OpenAI. Most folks I know are using a mixture of Claude and Cursor as daily drivers with Codex sprinkled in. That usage pattern should concern OpenAI.

I don't have much to say about speculative decoding other than to say it's now on my list of things to read more about. I think this is the "special" sauce of the Zed model that allows it to be as fast as it is. I think we are going to see the primary innovation from here on out in model latency and token minimization. I don't personally think there is much juice left to squeeze on the model improvement front. We will see incremental innovation there until something comes up with a novel approach that does not use transformers.

Perhaps the most interesting part of the discussion here was model hosting. I've never heard anyone (outside of the hyperscalers) talking publicly about model hosting. I know folks who call the LLM APIs and I know people who use LM Studio to run things locally but I know nothing about the in between. The scarcity of chips for running things locally is mind boggling. There really isn't an equivalent to the "cloud" in LLMs yet. Google and Amazon have their own custom chips, but those are constrained. Apple has their private compute. Everyone else is renting or trying to build their own shit. I have no idea how folks like Cursor are managing to scale. There is no "Lambda" for LLMs yet where you can scale from 0 to 1 million requests overnight.

Finally, a subject near to my heart: testing. I fell in love with Test Driven Development right out of college and to this day my happiest days are spent in the flow loop that is writing a new feature with TDD. LLMs are really interesting though because they are not deterministic. I wrote about this property before, but once again hats off to Zed for actually talking through what they are doing here. The closest I've ever seen to an in depth discussion for testing probabilistic software is this banger from Matter about their approach to parsing. The key insight here is that you need multiple "test cases." You need to teach your LLM what good and bad are, not just what you expect. Because it's probabilistic, you need to test with probabilities. The goal is to shift the distribution of results closer to the "good" than the "bad" provided case. That's a whole new approach to testing that my brain still struggles to grasp.

Link of the Day: The Infrastructure Behind ATMs

2025-11-13T12:03:56Z

The Infrastructure Behind ATMs

In reading this article, I mostly learned how little I know about the actual infrastructure behind payments. That said, there were a couple of interesting tidbits that broke through for me:

- a bank account is a form of loan to the bank
- ATM rails power the “debit” option you see at the grocery store

I found the backstory development of the ATM here instructive as well. Often innovation is accidental. The ATM was first created for internal use, to allow banks to take care of rather menial tasks without needing to hire more tellers, which were expensive. Originally, these only existed for a particular bank in a particular branch. However, the usefulness of the technology quickly became apparent. After a few years of figuring out how to get multiple banks to agree to be available in a single ATM (which requires come complex settling logic), the ATM as we know it was born.

That’s the first business pivot. The tricky part here is that ATMs themselves are not a great business. They require lots of up front capital, not for the machines themselves but because they technically loan people the money first and then the bank pays them back later. The real financial windfall here was then using the “rails” they built for interbank transfers to power the debit transactions at grocery stores.

When you choose credit vs debit at the grocery store, you are choosing between ATM and credit card rails. What is interesting here is that the ATM rails are much cheaper, yet no retailer has figured out how to incentivize customers to use them at scale. Apparently it would be super complex?

The last fun tidbit is that in Japan, most people are paid on the same day (25th) and take their paychecks out in the form of cash. Thus ATMs there are just stuffed with cash on the days leading up to the 25th. Sounds like a new heist movie just waiting to be made.

Link of the Day: Thoughts, Observations, and Links Regarding ChatGPT Atlas

2025-11-12T19:19:20Z

Thoughts, Observations, and Links Regarding ChatGPT Atlas

While reading this article, I found myself returning to my writing on the platform shift of AI. More specifically, the question, what will the UI of AI be? I don't think we know yet. I am writing this in Dia which feels...close? Yet as Gruber points out:

After giving it a try over the last week, to me Atlas feels like … Chrome with a chat button bolted on.

Dia is the same. Atlas seems to be Dia + it can do things for you. Which I don't really want? I don't have much to do these days outside of sending emails to friends, which I prefer to do myself. Further, Similar to Gruber, I do not find myself using my web browser for all of the things (this is probably a minority view). That seems to be the requirement for a useful agentic experience. I'm much closer to using my phone for all of the things and I really wish Apple would get their shit together there so I could have Siri do things like schedule meetings.

Browser Are Probably Not the UI for AI

There are a couple of real problems here that I'm still iffy about as well. First, "computer use" mode is God awful. I resonate with Simon Williams thoughts here:

I tried out agent mode and it was like watching a first-time computer user painstakingly learn to use a mouse for the first time. I have yet to find my own use-cases for when this kind of interaction feels useful to me, though I’m not ruling that out.

I would say that this is going to get better, but I'm not sure web 2.0 companies want that to be the case. Amazon is suing Perplexity for evening try to buy anything on behalf of their clients and that feels like a trend that will continue. Benedict Evans put it well in his weekly column:

Amazon sent a cease and desist to Perplexity, telling it not to configure its ‘Comet’ browser to make automated purchases on the Amazon website. Amazon sold $65bn of ads in the last 12 months, much of it on its core e-commerce site (there’s Prime TV as well), and it does upsells and recommendations, and it doesn’t want someone else to control or disintermediate that experience. This is a systemic issue with consumer agents using the web ‘for you’, which I pointed out last year when people first suggested this model: companies want to own and control their own user experience.

So we are in a weird place with these "agentic" browsers and agentic AI in general. They are slow and not good at things (yet). But they may not be allowed to get better at things because it threatens powerful incumbents. In Personal AI in the Rugpull Economy, Don Marti explains why this is not the most promising path for agentic AI UI:

Doc Searls writes, in Personal Agentic AI,

Wouldn’t it be good for corporate AI agents to have customer hands to shake that are also equipped with agentic AI? Wouldn’t those customers be better than ones whose agency is merely human, and limited to only what corporate AI agents allow?

The obvious answer for business decision-makers today is: lol, no, a locked-in customer is worth more. If, as a person who likes to watch TV, you had an AI agent, then the agent could keep track of sports seasons and the availability of movies and TV shows, and turn your streaming subscriptions on and off. In the streaming business, like many others, the management consensus is to make things as hard and manual as possible on the customer side, and save the automation for the company side.

So it is unclear how well AI agents can get, even if the technology improves. It will require partnerships between entities that are fundamentally at odds, which seems...unlikely?

The scarier issue here is privacy. You can see this live in real time as I try out Dia. It is living in the uncanny valley where it knows both too much and too little about me... Very weird feelings. To be honest, the example in my video (it knowing what car I drive) is not that big of a deal. The real problem is:

it is not hard to steal knowledge about me from the AI model
it is not hard to get the model to take actions that aren't good for me

Simon Williams sums it up well:

The security and privacy risks involved here still feel insurmountably high to me — I certainly won’t be trusting any of these products until a bunch of security researchers have given them a very thorough beating. [...]

It was 5-10 years before people trusted payments on the web. I don't yet see an equivalent to the protections we have in web 2.0 for agentic AI. And I'm not sure what the UI for this would look like. I am not sure I trust an AI to go buy all my groceries for me, nor do I want to be consulted every time it wants to buy something for me. Where is the line there? I have no clue.

So there are 2 clear things that need to be addressed:

- the automation war between existing platforms and agents
- the privacy issue

One of the things we are missing in the hype is that many existing UIs are already very good. DoorDash is incredible. Google is an amazing way to find a quick answer (though it's gotten worse). We've got some great UIs out there for many important tasks. The hardest work in AI might be finding new ones. They've got their work cut out for them.

One area that has not seen much impact is in tasks that already have specialized apps. I'll focus on two examples with abundant MCP implementations: email and food ordering. AI Doordash agents and AI movie producers face the same challenge: the bar for a new product to make an impact is already very high:
https://elroy.bot/blog/2025/07/29/ai-is-a-floor-raiser-not-a-ceiling-raiser.html#creative-works-not-coming-to-a-theater-near-you

Link of the Day: Search Engines as Leeches on the Web

2025-11-11T13:24:08Z

Search Engines as Leeches on the Web

The reason I include this article is not necessarily for its content, but when it was written. Jakob Nielsen wrote this in 2006, before Google was the monster it is today. He wrote this before the invention of the LLM and it is eery how similar things sound then to now:

There's no doubt that search engines provide a valuable service to users. The issue here is what search engines do to the companies they feed on — the companies that fund the creation of original information. Search engines mainly build their business on other websites' content. The traditional analysis has been that search engines amply return the favor by directing traffic to these sites. While there's still some truth to that, the scenario is changing. [Loc]

I struggle to wrap my around what this means about the era of 2006 - 2022. Does that mean search companies were choking out the web even then? Or does it mean that this article was hyperbole and thus that the impact of LLMs will likewise be hyperbole?

The “open web” of blogs and websites has survived a lot and today LLMs feel like an existential threat. But apparently search engines themselves felt like an open threat as well. And maybe they were? Maybe the web of 2006 was more lively than the one of 2020. I was sadly too young to know then.

In some ways, the author predicted the future, though I find it hard to believe sites ever relied purely on organic search to be found. Their imagined future came to pass, which gives more credence to the rest of their article:

I predict that liberation from search engines will be one of the biggest strategic issues for websites in the coming years. The question is: How can websites devote more of their budgets to keeping customers, rather than simply advertising for new visitors? Here are some ideas, ranging from the proven (newsletters) to the speculative (mobile services): [Loc]

The best sites have been diversifying away from Google for years, using a blend of email and social media (its own can of worms) to drive traffic. Even so, the advent of LLMs is hurting publishers, though not in an evenly distributed way. What I found most interesting is this quote:

Recently, however, people have begun using search engines as answer engines to directly access what they want — often without truly engaging with the websites that provide (and pay for) the services. [Loc]

This felt especially relevant to today’s world of chat bots, where even link attribution is a thing of the past. This suggests a new way of using websites. If chat bots and Google can answer your question, what is the point of a website? I would argue it needs to change. We need to think of our web properties as something more than knowledge. People need to come for your voice, the vibe, or the experience.

We need to think harder about the design of our websites and the type of our content. We need to adapt. I’m not quite sure how yet, but I do know for my own habits, I go places that are niche and have a voice. And I go directly there by URL. A couple examples are:

- semafor.com (really incredible and unbiased world news)
- xxlmag.com (super niche rap content)

Link of the Day: The Next Great Distribution Shift

2025-11-05T07:59:08Z

The Next Great Distribution Shift

Before I dive in, I want to break the key claim here into a few parts:

- AI is a platform shift
- Which will entail a corresponding distribution shift
- Which will hinge on personal context as the moat

AI is a platform shift

This is fairly out of step with my tech compatriots, but I am not sold on the first assumption, that AI is a platform shift. The reason I'm not sold is best explained by Scott Galloway in his article A New AI World:

Yesterday, I skirted along the edge of the atmosphere at four-fifths the speed of sound, traveling from London to New York in seven hours. The least expensive tickets were $400. Jet transport technology has changed the world. Sixty years ago, my mother crossed the Atlantic in a steamship: It took seven days and cost 4x what flying does today. Commercial aviation has created enormous value. However, the vast majority of that value has been captured by consumers and society, vs. airlines. Since 1945 the industry has experienced years of low-margin profitability only to have its gains wiped out by periods of huge losses (e.g., $128b in 2020).

Airplanes were not a true platform shift, but cars were. Cars reshaped our entire world. They are platforms that hundreds and hundreds of companies are built on. Airplanes, in comparison, are tucked away on the edges of cities and while they are transformative, they did not have the same far reaching impact that cars did.

My question here is whether AI is a car or an airplane. Will AI reshape everything we do? Will we rebuild our cities around AI the way we did cars? I'm not so sure. I know AI is useful and I already see myself using it in many parts of my day. But it often fits alongside my existing routines and tools, rather than replacing them.

There's a few things going on here though. One, I could be totally fucking up my use of AI and failing to adopt it fast enough. That's probably true to some degree. Two, we are very early. This could be a fairly premature evaluation. If I could go back in time to the launch of mobile, I wonder if my writing about that platform shift would be similar.

Yet the article, AI Will Not Make You Rich gives me pause. In that article, Jerry Neumann compares AI not to planes, but to shipping containers. His argument is that AI will make everything faster and easier, but it won't be particularly profitable. In many ways, that argument is identical to Scott Galloway's argument that AI will be more like planes than cars.

Yet some technological innovations, though societally transformative, generate little in the way of new wealth; instead, they reinforce the status quo. Fifteen years before the microprocessor, another revolutionary idea, shipping containerization, arrived at a less propitious time, when technological advancement was a Red Queen’s race, and inventors and investors were left no better off for non-stop running.

While I'm partial to that argument, I recognize that it is hard to judge what technological innovations will end up being wealth creating looking forward. What I find more interesting than that general claim is a couple of specific ones:

And because there is no economic profit during perfect competition, there is no money to be made by innovators during maturity. Like containerization, the introduction of AI did not lead to a period of protected profits for its innovators. It led to an immediate competitive free-for-all.

A couple of years ago, we all thought models were the moats. Until Deep Seek blew that assumption out of the water. In The Next Great Distribution Shift, the argument is made that personal context will be the moat of AI. We will get to that argument in a second, but I find it interesting because that's very different than we assumed the moat would be just 2 years ago.

The most convincing argument against AI as a platform shift is the following:

While Steve Jobs was telling investors that every household would someday have a personal computer (a wild underestimate, as it turned out), others questioned the need for personal computers at all. As late as 1979, Apple’s ads didn’t tell you what a personal computer could do—it asked what you did with it.

The personal computer was surprising. So was the car. We were not sure what it meant, how to use it, what it was for. AI is... less so? Everyone seems to know exactly how to use it and where. Every is sure it will change everything. All of it reminds me of the blockchain hype from a few years ago. To be clear, I think the blockchain is incredibly useful and I'm excited about a wide variety of applications of the technology. In many ways, I'm bullish on the blockchain. In 25-30 years, we will be building most of our financial infrastructure using that technology and it will result in lower costs and more secure payments.

I view AI in the same way. It's going to make everything 10-20% over a long time horizon. LLMs are a stunning technology. But I'm not sure they are a platform shift. As the Neuron said in their latest issue:

Scientists can already do incredible things with the augmented intelligence gains from today's AI models. These gains will compound over a decade through better scaffolding (software that directs the AI to be more useful). So we'll still get incredibly valuable leverage from augmented intelligence. But it won't be “AGI” like it’s been sold to us so far; according to Karpathy, there’s nothing “general” or “intelligent” about today’s language models, really. So reset your expectations: no AGI until 2035. - 😺 Karpathy: No AGI til 2035?

The logical extension of the fact AI is not unexpected is that there is not much room for investment.

The high capex of AI companies will primarily be spent with the infrastructure companies. These companies are already valued with this expectation, so there won’t be an upside surprise. But consider that shipbuilding benefited from containerization from 1965 until demand collapsed after about 1973. If AI companies consolidate or otherwise act in concert, even a slight downturn that forces them to conserve cash could turn into a serious, sudden, and long-lasting decline in infrastructure spending. This would leave companies like Nvidia and its emerging competitors—who must all make long-term commitments to suppliers and for capacity expansion—unable to lower costs to match the new, smaller market size. Companies priced for an s-curve are overpriced if there’s a peak and decline.

A platform shift needs to be surprising! AI is not surprising and thus does not present investable opportunities. Instead, it will result in a broad based increase in utility across a wide range of industries that will be difficult to effectively monetize. There are no moats.

Which will entail a corresponding distribution shift

Ok now forget everything I just said and let's assume that AI is a platform shift. The authors next claim is that a platform shift always results in a corresponding distribution shift. This might be the most interesting and insightful part of the article. The idea the technology shifts faster than people can adopt it sounds obvious. We even have a phrase for it: "the future is here, it's just not evenly distributed." But it gets at the core of how to react well to AI.

AI is the platform shift, but folks haven't figured out the distribution of it yet. The App Store made "mobile" ready for the prime time. Google made the tangled mess that was the web useful. What will make AI easier to deploy and understand? I have no freaking clue. I'm not sure we've seen anything close to it yet. But I think the key is related to context.

Which will hinge on personal context as the moat

Imagine a protocol, similar to MCP, that stores the context in your device (ie apple secure enclave) and allows apps or models to tap into that to personalize your experience. Then the moat is probably not the processing context anymore and shifts to the storage and the ability to execute over it. - Tomás Soracco

Whoever can figure out how to make context portable (and useful) will nail distribution. The web browser (and Google) became the hubs of our digital lives. I don't think AI chat is the hub of our daily lives. Phones, maybe? I think Apple is really well positioned here, but they keep fumbling the bag. VR headsets, while I wanted to like them, are not it. AR glasses like the Ray Bans made by Meta are interesting, but probably not scalable to the level of a phone (not everyone likes glasses). AirPods present an interesting case. Scott Galloway argues they are a form of VR we live with already. I would welcome a smarter Siri paired with AirPods to ask questions about the world. But I'm not sure that is coming anytime soon.

We've seen the pendants (probably too creepy, not everyone likes necklaces either) and the weird AI pin. Those aren't it. In some ways we've seen a lot of nos and some maybes about the next distribution platform. But nothing I've seen so far feels like "it."

I almost feel like the answer is either robotics or voice or some combination of the two. I wrote about our under appreciation of the power of AI in robotics a few months ago and Kara Swisher is starting to beat the drum as well.

I always loved this company Misty Robotics and they seem closer to it than most. Maybe it's sentimentality (Misty is one of the first "cool" startups I ever saw up close), but the idea of a cute robot at home with a developer platform for tinkerers sounds promising. In some ways, it could be what the Amazon Echo never could. We are a long ways from "home robots" despite the sizzling hype videos. But we are not so far from a fun helpful assistant with a wide open app store full of useful things.

The more I write about it, hackable home robot + LLM is a fun combination. Further, it's the first form factor I've heard of that sounds like a toy, which for some reason is how these big breakthroughs tend to come about. A little robot like these (maybe combined with a Roomba?) would have context about my day (in a less creepy way than a pendant), connect to the internet, and converse with me.

Well shit now I want one. So there's my prediction after all. The distribution channel for AI will be hackable friendly home robots.

Thought of the Day: AI is Best Where Knowledge = Money

2025-10-11T15:13:07Z

I've written quite a lot about AI over the last few weeks. Whatever the new technology is, I try to think through it slowly. My goal is to understand the strategic impact of tech before I start to adopt it. Otherwise, you end up in a hammer and nail type situation. You use the technology in the crudest possible way since you don't know what it's best for.

AI = Vigilance (Extension) + Knowledge -> Money

After a lot of writing and thinking and experimenting, I think I've distilled down the best use case for AI: areas where knowledge can be converted into money. Before diving into this theory, let's start with some practical examples:

- Bug bounties - Bug bounties allow you to convert knowledge (identifying a bug) into money (in the form of a reward).

The core of the idea here is that humans are bad at vigilance tasks and computers are good at them. Previously, computers were limited to deterministic vigilance tasks. Using statistics and math, you could occasionally get away with using a computer for less deterministic vigilance tasks (monitoring and alerting with anomaly detection comes to mind). However, for the most part, you were limited to tasks with easy to define rules (think IFTT or Zapier).

At the end of the day, that is all of programming. Translating an amorphous process into something discrete and replicable for a computer. As laid out in Building AI Products In The Probabilistic Era, we are entering a new era of software. We can now create software that is probabilistic instead of deterministic.

At first blush, and in many use cases, this is not particularly useful. For most of our tasks, we wanted things to be replicable. If you ask your bank to transfer $30, you want it to do so the same way every time. You do not want it to "probably" transfer $30 with a 10% chance of transferring $100.

There is progress being made here. Hallucinations, according to OpenAI, are "solved."^1 MCPs make it easier to define what exactly a probabilistic agent can do to your data. But even so, I do not trust agents and their probabilistic nature with anything of consequence.

The only area I trust AI is in extending, not replacing, my capabilities. I trust AI when it can be more vigilant than me. Take security, for example. I cannot realistically monitor all of the security breaches the world over. Any incremental security breach that AI makes me aware of is a huge net benefit. Even if it misses some breaches, I am incrementally ahead with AI compared to without it.

So AI is best fit for cases where scale and reach is more important than precision. Or put another way, AI is best fit for tasks that humans cannot do past a certain scale.

AI is best fit then for areas where it can be vigilant in a type of knowledge space that is easily verified. Security, code, and even weather are all perfect fits. Each of these use cases have different types of benefits. AI in the security field acts more like insurance (avoiding loss) than a revenue driver. AI in trading (surfacing asymmetry) acts more like a revenue driver.

The goal then is to figure out:

1. What fields allow you to translate incremental knowledge into money
2. What that knowledge then does (protect you from downside or create upside)

So far I'm looking at the following examples as a bell weather for where and how to apply:

In a three-month trial, CRESt ran 3,500 experiments across 900 chemistries and discovered a fuel cell catalyst that reduced reliance on expensive palladium while hitting record power density. That’s notable because rare metals remain a big bottleneck in fuel-cell economics. - Weekly Dose of Optimism #164

R&D is a perfect example of the power of extending human capabilities. In R&D, AI lowers the cost of experimentation in a domain where there is tremendous upside to new knowledge. Thanks to the patent system, you can actually capture the value of that knowledge, translating knowledge almost directly into money.

The European Center for Medium-Range Weather Forecasts is the largest weather forecaster on the continent and collects unfathomable amounts of meteorological data in its quest to model the weather. It also sells that data to private companies that use it for in-house forecasting needs. Customers include energy traders, shipping lines and insurers. Interest in those commercial licenses has spiked as AI weather forecasting is seen as an increasingly viable and possibly profitable venture. The number of firms buying commercial licenses has increased to over 800, paying an average of 36,000 euros (US$42,200) per year for the data. - Numlock News: October 2, 2025 • Drive Thru, AOL, Broadway Musical

Weather is an interesting application of AI. Long outside of the purview of human knowledge, it's a domain that benefits from more and more data. The sheer amount of data means that humans don't feel the same need to understand the outcome of the AI, unlike software development. Weather is a near perfect application of The Bitter Lesson. A surprising variety of industries, from transportation to oil drilling, can benefit from improved weather data.

When writing about these applications, where knowledge can be translated into money, I find myself surprisingly giddy about AI. I am not hugely optimistic about agents for getting work done, but I am so excited about the potential for agents to find and discover incremental knowledge that makes the world better.

In the short term, I'm bullish on AI as a finder of unusual high quality information. In the long term, I'm bullish on the use of AI in robotics to automate real human work. In the meantime, I'm doubtful about the usefulness of non-deterministic agents to replace humans outside of low quality busy work that probably shouldn't be done anyway.

1 Given Hallucination is Inevitable: An Innate Limitation of Large Language Models I have my doubts about how solved this is.

Link of the Day: Wisereads Vol. 108 — Alex McCann on the death of the corporate job, Wuthering Heights, and more

2025-09-20T09:17:15Z

Link of the Day: Wisereads Vol. 108 — Alex McCann on the death of the corporate job, Wuthering Heights, and more

The title of this link is a little misleading. The real lede here is buried in the “Twitter Thread of the Week: WHY RETENTION IS SO HARD FOR NEW TECH PRODUCTS" section by Andrew Chen. I am always looking for shortcuts, key principles that can help me cut through the clutter and see things others miss. In this article, Chen lays out what I think most people (including me) miss about tech products in general and networked products in particular: retention is everything.

We understand this concept when we see it physically. The flex seal guy showed us all how important it is that a boat retains water. Yet we struggle to see it in tech, especially viral products. It does not matter how many new users you add if you cannot retain them. What I find interesting about this concept is that it means some of the nefarious tactics marketers use are not only annoying, they are ineffective. Spamming people with emails and texts if your product is not sticky does nothing. You may briefly see a surge of users but if your product isn’t great, it’s game over.

Yet retention is not synonymous with product quality, especially with networked products. Most of us do not use Facebook because it is an incredible app, but because we were required as a result of its ubiquity.

So there’s a couple of core truths here:
- Viral growth is less useful than we imagine it to be, at least in the absence of good retention
- Good retention does not require a good product. Many of us use Jira daily. That does not mean Jira is good.

These core truths remind me of Good Strategy, Bad Strategy. Chen here basically lays out the key strategic pieces of retention. There appear to be certain laws to retention that determine what does and doesn’t work.

They are counterintuitive. Who would guess that it’s easier to acquire a new user than to re-activate an old one? Who would guess that your best users are your first ones and that retention falls as you add new users?

Without knowing these core facts you can either freak out about an inevitability (falling retention) or ignore an emergency (high growth, bad retention). You may also try to chase retention where it will never appear. A dating app will never have great retention, you need to rely on other strategies to achieve profitability.

It feels like knowing these laws at the outset can save you a lot of pain and stress. Without further ado, here is the high level of the laws below. If you want to learn more, dig into the thread or jump into the Cold Start Problem.

- You can’t fix bad retention. No, adding more notifications will not fix your retention curve. You can’t A/B test your way to good retention
- Retention goes down, it doesn’t go up. And weirdly, it decays (oh, does it decay) at a predictable half life. Early retention predicts later retention
- Revenue retention expands, while usage retention shrinks. Good news: You lose people over over time, but the ones that remain sometimes spend more more money!
- Retention is relative to your product category. There’s nature, and there’s nurture. Sorry, you’ll never make a hotel booking app a daily use product
- Retention gets worse as users expand and grow. The best users are early and organic. The worst users come after that
- Churn is asymmetric. It’s far easier to lose a user forever than to re-win them back
- Retention is weirdly hard to measure. Seasonality is a real thing. New tests throw things off. Bugs happen. D365 is a real metric but you can’t wait
- Crazy viral growth with shitty retention fails. We’ve run this experiment many many times already, across multiple platforms and categories

Link of the Day: Field Notes from Shipping Real Code with Claude

2025-09-17T13:14:42Z

Link of the Day: Field Notes from Shipping Real Code with Claude

I think I am finally ready to dip my toes into “vibe coding.” In reading this article, I saw the first “real” production grade setup for an LLM. This author “gets it.” Most vibe coding articles are interesting, but not at all applicable for anyone building an application for real users. They are fun test apps that can be used for your personal life, nothing ready for the prime time.

The author here clearly uses AI day to day and gave actionable steps for incorporating things into your workflow. In many ways, it reminds me of an article version of the Kiro approach to development. I’d heard rumblings that iterative development (small chunks) and markdown rulesets were the way to code at scale with AI, but I’d never seen it well laid out until this one.

I’m going to start adding rulesets for my AI after this article. I might even use AI to help me draft the rule set based on what our code base looks like. That might end up being enlightening in and of itself, to see what AI thinks the rules of our code are.

My favorite part of this article is the author’s acknowledge of the limits of AI and the clear guardrails they lay out in the article. There were some I’d thought of (no changing test files) and some that were new to me (no database migrations) that are critical.

The key that most beginners miss is the following:

Most importantly, you’ll understand why writing your own tests remains absolutely sacred, even (especially) in the age of AI. This single principle will save you from many a midnight debugging sessions.

The danger of coding with an AI is the risk of backsliding. We see this with human developers, where one person changes legacy code without properly updating the tests and it breaks everything. The risk is elevated in the world of AI, especially when the “developer” does not code.

The 3 use cases for LLMs felt particularly accurate. Most folks fall on some range of the goldilocks scale here: all AI or none. The author lays out what I think are the most sane ways to use AI:

AI as first drafter (for throwaway code)
AI as Pair-Programming (I live here most of the time)
AI as validator (code reviews, CoPilot is great for this)

Overall, I see this as article as an indication that AI assisted coding is ready for the prime time. We finally have best practices to use that will evolve over time. This was my signal that it’s time to start implementing these best practices and changing my workflow.

I tend to be slow to adopt tech, waiting until it is derisked. It feels like the time is now here.

Link of the Day: The Visions of Neil Mehta

2025-09-16T08:59:40Z

While this article is nominally about a venture capitalist (and an interesting one at that), what I find most fascinating here are the ideas about AI. First, I think AI will massively scale our ability to learn more than it will do things for us. The best performers in the markets will use AI to surface data from previously untapped sources and dominate. And second, LLMs are airlines.

What Mehta was willing to discuss is why Greenoaks has largely stayed out of the model war, in which each player insists, almost daily, that AGI is around the corner. “They may evolve to become great businesses, like ChatGPT is, but in their first incarnation they are all kind of bad business models,” he said. “Huge capital investments up front to create this asset, the asset is worth some amount of money, which then depreciates over the course of 12 months, so you have to reinvest again 12 months later. It’s like the airline business in the 1980s; you invest in the best fleet, but then 12 months later the other airline has the newer models, and you don’t pay back the cost of your initial capital investment because the unit economics don’t work. That’s the AI model companies. They have no competitive advantage. If you create a brand like ChatGPT, or if you achieve so much scale that you capture all the capital and no one else can compete, maybe you can escape that. But it’s not obvious that everyone does.”

Scott Galloway also made this point, but I think Mehta does a better job laying out the case here. LLMs require giant capital expenditures that are not able to be amortized over time. Is anyone still using Chat GPT 3? The shelf life of that billions of dollars of investment is near zero.

In a word, they’re merely scaling the horsepower of transformer models, rather than innovating on the underlying model. Insofar as this is a fair characterization, it would explain Mehta’s skepticism of the potential for returns, competitive advantage, or anything else of classical business interest, which powers the Greenoaks machine.

In a previous post, I linked to an article that laid out the significant advances in AI. We have a loooong way to go before we squeeze the juice out of transformer models, but I'm less clear about how much better they can get. Most people overestimate how much better a technology can get and they underestimate the new ways in which it can be applied.

A long while ago, Sam Altman said he wasn’t scared of any other LLM competitor. He was scared of someone working on something in the basement that was a totally novel approach. Deep Seek was a peak at this, but the real thing that keep Sam up at night is described below:

For that, one would need to build a model that could learn a lot from very limited amounts of data, and generalize from first principles. Such is the threshold for superintelligence, which could make the kinds of connections and discoveries that would elude AGI.

If something can truly reason, the whole paradigm of sucking up tons of training data goes out the window and the economics of the business change. Allegedly, that’s what Ilya Sutskever is working at with SSI. Mehta is allegedly invested in that company. If I were looking for the next thing in AI, that’s where I’d be watching.

Link of the Day: Comparing Energy Consumption of React Framework Versions

2025-09-14T13:13:32Z

Comparing Energy Consumption of React Framework Versions

Changes in React can cause huge changes downstream, with millions of websites being affected. One of these changes is power consumption. More efficient code in the React framework can result in drastic cuts in carbon emissions. Research, however, shows that a significant amount of websites run outdated versions of the React Framework. 6 This could be because of a myriad of issues that could arise from upgrading to a newer version. Most notably breaking changes that break backwards compatibility, in which case the developer has to go in and manually change the code.

As someone who loves up to date software and yet dreads upgrades, I found this a heartening read. Tagging on potential climate impact gives me all the more reason to work towards software upgrades.

What I find most interesting about this though is the potential of scaled impact in climate. Upgrading a single site, which runs on hundreds of millions of devices, can result in real energy savings.

One of the really interesting things about tech, that we struggle to grok because it's extremistan^1, is that impact scales exponentially rather than linearly. If you taught one other person a better farming technique in an agrarian economy, it would at best scale geometrically and it would take a long time to diffuse.

If you taught a single website how to load more efficiently, you could instantly improve the carbon footprint of millions. Those types of impacts are hard to wrap our heads around, yet they are going to become more and more common given Everything is Technology.

What's kind of wild is that as software becomes more embedded in our everyday lives, the real world impact of software changes is growing. What if Nest finds a new algorithm that's 1% more efficient for temperature control? They can scale that improvement out to 1 million homes overnight.

Of course there is great risk to this type of technological centralization, but there is great potential as well. One of the areas of AI I'm excited about is the ability to find new little optimizations. These are hard to find and it costs real resources to dig them up. When we put our minds to it, we can make it happen. Google reduced the energy required for LLM inference by 33% and AWS quietly improved almost all its services over the years, saving customers money and the earth carbon.

We need more of these scaled, detailed improvements. And AI is really good at surfacing opportunities for them and making them happen faster. Whether it's upgrading an outdated package or finding a memory leak, AI can scale human impact at the scale of technology.

This underscores my core thesis about AI: the most exciting applications lie in its ability to extend, not replace human capabilities. To be honest, I don't find myself all that excited about agents. Most small businesses are not ready for agents. To unlock real economic growth there, we need process and people change. I find it unlikely that a company using paper billing is going to jump over Zapier and move right to an AI agent (though I might be wrong here!).

The market for agents is really the market for Zapier automations and we've got a long ways to go yet there. But the ability to scale humans, scanning huge volumes of information and scanning things we should, but can't, is where the real unlock lies.

Now that's not all sunshine. Just as we can find cool things because of human extension, we can find not so cool things like security vulnerabilities.

1 - Shoutout to Black Swan by Nassim Taleb for this term to describe our newly exponential world.

Link of the Day: Why LLMs Can't Really Build Software

2025-09-11T12:51:07Z

Why LLMs Can't Really Build Software

Ironically, the model of software development laid out here may end up helping LLMs build software better:

Build a mental model of the requirements
Write code that (hopefully?!) does that
Build a mental model of what the code actually does
Identify the differences, and update the code (or the requirements).

In particular, Kiro from Amazon is taking steps to apply this approach to LLM based software development. I see the future of software LLMs as less Bolt.new or Codex, where you one shot an app, and more Kiro, where you iteratively step through features one at a time.

Many of the LLMs out there and LLM coding approaches (think vibe coding) are effectively sped up waterfall development. The trouble with waterfall is that we do not have perfect knowledge about what our customers want. We never will have that perfect knowledge because prediction is hard and humans are no good at it^1.

We use agile and "lean" because we need to test software bit by bit to figure out which bits work and which bits don't. The other day I heard about someone using "daily" sprints to get software out faster than ever. There are a few edge cases where this can work (adding enterprise support for a consumer grade app being one). But for the most part, it just means you are building software you don't know people want.

For LLMs to build software well, they ironically need to adopt more of the iterative development principles from software development.

That said, there is progress being made here. Hallucinations, according to OpenAI, are "solved." Honestly, I need to read that paper alongside Hallucination is Inevitable: An Innate Limitation of Large Language Models to see who is bullshitting.

The space is moving so fast it is hard to keep up! But for now, one thing I can say with certainty is that "vibe coding" is another word for waterfall and we've proven that doesn't work (outside of certain specialized cases).

1. For more on this, check out Superforecasting and The Black Swan

Link of the Day: Building AI Products In The Probabilistic Era

2025-09-05T03:46:14Z

Building AI Products In The Probabilistic Era

I am spending a lot of time recently thinking about how AI systems fail and what use cases are best suited for these randomized prediction machines. In my own personal projects, I find myself reaching for AI when I want to extend, not replace, my capabilities. The prime example is Deeper Research, which is a simplified version of Deep Research that runs locally and does not summarize, but returns diverse search results for broad queries.

Even with all my thinking and tinkering, I've struggled to frame the killer use case of AI. I feel like I catch fleeting glimpses of it. I can see the edges of its strengths and weaknesses without being able to put my finger on the whole picture. This article, from someone out there building on the front lines, nails it. There are two key ideas here:

- Old software is deterministic, AI is probabilistic
- Old software is engineering, AI is science

⁠It's ontologically different. We're moving away from deterministic mechanicism, a world of perfect information and perfect knowledge, and walking into one made of emergent unknown behaviors, where instead of planning and engineering we observe and hypothesize.

The key to incredible products in the age of AI is finding out which parts of your product should be deterministic and which should be probabilistic^1.

While that's a powerful insight, I find myself on a bit of an SRE / monitoring kick of late and I found the implications for how we measure and change our systems the most interesting bit. We can no longer rely on tried and true RED and USE methods for understanding system health. Those only work for deterministic systems. Instead, we need to take a page from the book of scientific research. As best as I can tell, we need to run experiments and then conduct literature reviews to gather insights. The author laid out the "why" for this new approach here:

Knowing that users acquired through TikTok are more likely to build games, which are more expensive to generate on a per-token basis and therefore impact the margin calculus, is incredibly valuable across the entire company: from engineers making sure that games are efficiently generated, to marketers shifting their top-of-the-funnel strategy to a more sustainable channel, to the finance team appropriately segmenting their CAC and LTV analysis. A 20% shift from game-building users to professional web apps might mean the difference between sustainable unit economics and bleeding money on every free user — yet this insight only emerges from analyzing the actual content of AI interactions, not traditional funnel metrics.

The part that I am thinking about the most is their approach to gathering this type of data:

The easiest way of approaching it is by segmenting user inputs. You use smaller models to classify user requests to larger models, which allows you to segment your data in “regions of usage”. It’s a crude way of clustering user journeys. For Replit’s coding agent, this could be coding use cases: “what’s the likelihood of getting a positive message from the user after 3 chat interactions, for all users that submitted a prompt about React web apps?” To push things further, you can use the same approach to define milestones to achieve across different paths, which might mean classifying model internal states.

In some ways, this feels like the most interesting area in tech right now. I've yet to hear anyone else talk about this, even Replit's approach feels naive and unpolished. There is real opportunity to step in and build the "SRE for AI" role and best practices here because as far as I can tell, they do not yet exist. The best we have so far is: "use an LLM to group output together and make more guesses."

1 - Credit given to my coworker Ravi Starzl for the framing of "what should be probabilistic vs deterministic"

Link of the Day: There are no new ideas in AI — only new datasets

2025-08-27T21:27:24Z

There are no new ideas in AI — only new datasets

Yesterday I wrote about the applications of AI in robotics. Frankly, it was more theory than substance. Today I'm bringing a little bit more substance. First, this article tracks the progression of AI development better than I did, laying it out like so:

Deep neural networks: Deep neural networks first took off after the AlexNet model won an image recognition competition in 2012.
Transformers + LLMs: In 2017, Google proposed transformers in Attention Is All You Need, which led to BERT (Google, 2018) and the original GPT (OpenAI, 2018).
RLHF: This was first proposed (to my knowledge) in the InstructGPT paper from OpenAI in 2022.
Reasoning: In 2024, OpenAI released O1, which led to DeepSeek R1.

Better yet, the author lays out a strong thesis for my "commoditization" of LLMs, though they have some theory to back it up. The theory is based on an AI paper called Bitter Lesson which says:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin...Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation...the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation.

What's interesting about the implications of this thesis is that it implies we will see mostly incremental improvement in LLMs rather than step changes. As noted in the article, we've mostly used up the text available for training:

Transformers unlocked training on “The Internet” and a race to download, categorize, and parse all the text on the web (which it seems we’ve mostly done by now).

Outside of the Reinforcement Learning from Human Feedback (RLHF) and reasoning breakthroughs in 2022 and 2024, we've mostly seen incremental improvements in our ability to run systems. These breakthroughs are important for scaling these discoveries, but they are not step changes. An example is the 33x reduction in energy usage for inference achieved by Google. We need these breakthroughs to scale and apply the technology. But they are not going to get us to AGI.

For that, this article argues, we need new data sets. My prediction is we are going to see a bunch of specialized AI applications with proprietary data (Facebook is doing this with ad targeting, Waymo with cars). The new data will likely come from / for robotics and 3D world mapping, as I argued yesterday. That said, the article does make a strong case for Google. Though you could argue they are already making use of that data well given their latest video model breakthrough.

A couple of links come to mind that I can't connect well here, but still feel relevant:

- Elon Dreams and Bitter Lessons
- Of Flying Cars and the Declining Rate of Profit

Link of the Day: D.I.Y. Artificial Intelligence Comes to a Japanese Family Farm

2025-08-27T02:26:17Z

D.I.Y. Artificial Intelligence Comes to a Japanese Family Farm | The New Yorker

Koike completed his machine last year, and it works—to some degree. It sorts cucumbers with an accuracy of seventy per cent, which is low enough that they must subsequently be checked by hand. What’s more, the vegetables still need to be placed on the photo stand one by one. Koike’s mother, in other words, is in no immediate danger of being replaced, and thus far, she and her husband are none too impressed. “They are quite severe,” Koike said. “ ‘Oh, it’s not useful yet,’ ” they tell him.

The first interesting thing about this article was that it was written in 2017.

That feels like light years away in the world of AI. Back in that time, I was also using TensorFlow as part of a class at Hope College. In those days, "AI" was little more than linear regression. It could handle more variables than a human, but it was basically doing complicated calculus. Every percentage increase in predictive power took incredible amounts of tuning. The hardware wasn't there yet either. I remember kicking off a neural network training set (the new hotness that now powers LLMs) and waiting hours for terrible results.

Even so, it was a heady world back then. TensorFlow from Google was released in 2015 and it's AI Go Player beat the first human around the same time. People soon calmed down. Partly because Google sat on the technology which would go on to produce LLMs as we know them now, knowing it could cannibalize their Search Dominance^1. Partly because it is hard to commercialize these technologies.

Most of the above is an aside, though an interesting one. The real key here is what I think the true potential of AI is. More and more, I think the LLMs are a distraction. They are cool and useful for some specific tasks (like coding and marketing and image generation). But they are going to be a footnote at the end of the day.

Agents are not going to be the real use case here either. Probabilistic models have a clear upper bound in their usefulness for many business applications. No the real winner here is robotics. We are going to see a new industrial age spurred by these models. They are finally allowing us to generalize robotic movement, sensor interpretation, and memory.

Back in college, I almost worked at a manufacturing company that built the robots which built Teslas. Each machine was a huge undertaking. Why? Because they all required bespoke programming, bespoke sensors, bespoke designs. What if you could create a general purpose machine at scale that could flexibly manufacture anything?

I think the real power of LLMs is their potential to bring the promise of 3D printing to life for large scale industrial production. Well, actually not quite. That would probably require more precision than these can handle. Maybe I mean "automation" here. I'm talking autonomous cars, drone delivery, warehouse management, and more. That shit is going to get taken over by these general purpose robots.

Nvidia and many other companies are hard at work on this problem. This is where the real money is. The only way LLMs survive as anything other than a commodity is if they nail Agents. I'm doubtful. These things are incredibly useful, but the reality is there is no differentiation (outside of the artificial friend angle, which is sticky). The models are only getting smaller and faster and it's a matter of time before we can run them all locally. Between hardware and software improvements, that is inevitable.

That means the only real meat and margin here is robots. Hardware is hard and the training data for robots is much harder to come by. There are real moats in robotics. All right that's all I've got for this evening, rant over.

^1 - I forgot where I heard this argument, but there's another argument to be made that Google was just retooling everything from the ground up to be AI first. That's interesting and Sundar Pichai is on record saying just this in 2016. If you look at their bets after that, this actually grows more compelling as laid out here.

Link of the Day: The Nonwriter's Guide to Writing A Lot

2025-08-26T02:50:24Z

The Nonwriter's Guide to Writing A Lot

I am writing this as I am listening to Denver by Jack Harlow. I am trying to start writing again, which is ironic given I used to write about my prolific writing habits. It's funny that in this article, the author points out that all of us are now prolific writers if you count our various digital scribbles:

Let’s start with an observation: You are already a prolific writer...if you look at your own correspondence — every email, text, and post, every tweet you’ve cast into the soulless void that is Twitter — you will likely find that you have written more, by this point in your life, than Gandhi’s entire life’s work. You’re an internet user. A netizen; you write more text as an afterthought each day than most people pre-1980 did on purpose.

When I was young, I wrote things I probably shouldn't have shared with the public. It wasn't overly revealing, just not all that good. There's a Jack Harlow line that comes to mind there as well, where he questions all the shit his younger self said in his early work.

These days I share almost nothing with the public, even as I write often. It is weird to say that most of this writing is for me. When I wrote as a younger man, I always imagined how my words would land with an imaginary audience. Now I seem to weigh the words within myself, measuring their resonance within me.

I am proud of that shift, even as I find myself yearning to write more than I do. I am actually happy with how much I write about my persona life. But I find myself yearning to write more about the topics that matter most to me, even if it's only to get them out of my mind.

So here I am writing on Hey World, a platform no one will likely ever read. I like it that way. The pressure is lower here and yet it still feels like I'm publishing. I thought about writing on Beehiiv but I feel pressure to put out something worthwhile there since people actually get notified there. I don't want to annoy people. Nor do I want to create expectations. I would love to do this daily, but I likely won't.

To connect this back to the article of the day, since I'm mostly rambling, I want to lower the friction I feel for writing. I want to write more, in more steps. I want some of my writing to never leave my computer, some to land here, and some to go somewhere more meaningful.

Finding Roots by Leaving Them

2021-09-22T04:42:47Z

The advent of COVID and remote work forced me to reconsider where I want to live in the future and how I might make that determination.

When I was in high school, I didn't give much thought to where I'd live after college. I didn't spare a second thought for what might happen after that.

Yet as college continued, I saw my hometown change drastically. They followed Carmel and started building for growth tomorrow with money today. Suddenly, the place I thought was home was home no longer. The city knocked down buildings, built apartments that stand unfilled, and flipped parks into parking garages.

Then I moved to college in Michigan. While there, I found a city I liked, but didn't fit into.

What I realized in contemplating this is that many of my peers have made 1 of 3 choices:

1. Stay in hometown.
2. Stay in college town.
3. Stay in place of 1st job.

After seeing my friend Brian move out of NYC to use Airbnb to stay in many cities across the US, I realized that living in many cities before choosing is critical.

Why is it that we look at hundreds of cars before we buy one yet only try a few places to spend 90% of our lives?

That question is driving my thinking for the next 5-10 years. How can I "test" as many places to live as possible? How can I best choose the place to put down roots?

Just to be clear, I'm not aiming to be a "digital nomad" or any such nonsense. I'm only trying to find a good place to settle down.

3 Things #1: An Idea, An Article, An Audio Book

2021-05-29T20:25:06Z

Jason Fried is doing something like this and since I've started writing 750 words daily thanks to 750words.com (I'm now using an Obsidian hack for this, but it all started with them), I have wanted a place to put my words.

Why write about an idea, article, and book? Because that is mostly what I write about everyday. As part of a previous post, I explained how I started to limit my inputs to learn better and be less stressed. Part of adding limits was asking myself to reflect on what I read and why.

After a few weeks of doing that, I realized my thoughts about articles and the articles I read could be interesting to others. So here I am.

1. An Idea

I cheated this week (my first week - off to a great start!). I can't decided between two ideas. Both feel similar, which is why I allowed both to stay:

1. Perceived Effort is not linear
2. The size of a problem and its solution are asymmetrical.

Misconceptions about both of these ideas have caused me great stress so it was nice to finally pin down the ideas I want to remember going forward. Perceived effort is not linear is comforting because it says hard things will probably get easier over time, but in a linear way. It's nice to know that when you've been applying effort for a long time without visible progress. The size of a problem and its solution are asymmetrical is comforting because it's nice to know that big problems can be solved in small ways. The converse of that is not so fun (small problems require big solutions) but I'm trying to ignore that for now.

2. An Article

This weeks article surprised me. It's called, "Why two Clockwise designers built a video game that gives humans empathy for machines". At first, I figured this article would be fluffy bull shit pumped out by some startups' marketing machine, but after reading it I couldn't have been more wrong.

There is fascinating stuff here about teaching humans to empathize with computers via games as a way to decrease the opacity of machine learning. Even the idea that you can teach empathy via games is an interesting one in and of itself.

Perhaps the most interesting thing of all was the human machine interface created by Clockwise, which operates at the limits of human capacity. Clockwise is an interesting example of helping humans make better decisions when the number of factors involved outpaces our brains.

Oh and I highly recommend you play the game they made, it's fascinating.

3. A Book

This week my favorite book has been far and away Punished by Rewards by Alfie Kohn. The title really says it all and given the fact I haven't comprehended it yet, I'm going to leave off trying to summarize it and just urge you to read it.

The End

Social Life - It's a Thing Again

2021-05-28T16:40:39Z

As the state of Michigan begins opening back up, I'm faced with social choices again.

There are a lot of things I'm excited about with the world opening up again (food being the most obvious one), but the thing I'm least excited about is making social choices.

As an introvert and people pleaser, I'm constantly at war between pleasing people by saying yes to everything and fleeing to a mountain with no cell service.

During the pandemic, this dichotomy fell away.

"Wanna hang out?" - COVID
"Wanna go to X?" - COVID

None of the above questions were even asked thanks to COVID. I had to make ZERO social decisions because there were none to make.

With things opening back up though, the requests are once again trickling in. And I'm left floundering, my social skills weakened like my immune system as they were trapped inside all year.

As I tend to do, I've started to create systems that protect my time.

Yet I'm wondering if they will be effective or even a good idea.

Is social life something that can be engineered? What's the difference between a boundary and a wall?

As things open up, I will have to begin considering these scary questions once again.

Wish me luck.

Improving My Outputs

2021-05-14T18:45:20Z

It's come to my attention, based on one of my past articles, that I am not super satisfied with my outputs. At the moment, I'm writing 2 articles a week, one with a friend and one solo on my personal blog. For some reason, I want to write more.

In my past article, I wrote a Goldilocks of my output:

Too Little: No writing

"Just Right": 1-2 articles a week

"Too Much": An article a day

Upon reflection, I realized the above dichotomy is too simplistic. In my daily writing experiment, I publish daily on Medium, which ended up being quite stressful. My initial reflection on that experiment was that it was too stressful to write daily, yet as I hunger for more writing I think I made an error.

I think the error I made was that I assumed all daily publishing is equal. Yet, it's not.

Medium, since it's my home base for content, is where I expect high quality content to be. I want Medium to be the place where my most highly curated writing goes. With that expectation comes stress.

What I realized is that I want a layer between writing a Medium post and processing my thoughts in Obsidian.

Originally, I hoped to build out my own website for the purpose. That was a terrible failure and I still think it gives me the fear of publishing that comes with putting something on a website. The internet feels permanent and authoritative and I'm looking for quick and easy.

The two most promising things to come through are 750 words and Hey World, which is what I'm using right now. Both give me the option of writing daily without feeling too much pressure. Hey World publishes it and 750 words doesn't. Maybe it's a 3 part funnel...

Maybe, what I need to do for output is Obsidian -> 750 Words -> Hey World -> Medium.

I must say, I do quite like that.

A Return to Community

2021-05-11T15:42:02Z

This post is not about what you'd think it is.

While it's true that in my area of the world things are slowly starting to open up, I'm not writing about that.

Instead, I'm writing about my friend Neil and I's return to the community of ideas.

Almost a year ago now and two years after our failed startup, we started having weekly bonfires. I'd arrive late, as I always do and Neil would break a few matches before we got the fire roaring.

Much has changed since then. We started with a book club, reading the hard stuff that neither of us could make it through alone (like Growth by Vaclov Smil and Black Swan by Nassim Nicholas Taleb).

From there, we were inexorably drawn back into our habit of bouncing startup ideas off of each other. One thing led to another and there we were, thinking about starting a new kind of venture.

We called it, "The Idea Explorers". Our goal was to give away ideas to the community, for FREE. We were excited to share our ideas, but saddened by how we would have to do it.

Neither of us loves social media nor spends much time on it (though I do have a weakness for Twitter). We were spooked by the requirement of spending hours and hours online, building an audience.

Plus, thanks to Paul Graham's magnificent, "How to Get Startup Ideas" we realized we were coming at it all wrong.

Ideas weren't valuable.

Problems were.

So we started from scratch, this time keeping it stupid simple.

We collected problems and solutions, first in Trello, then in DevonThink, and finally in Obsidian.

After a few weeks of that, we decided to try an ambitious experiment. We'd release a product a month for a year.

Then Neil went on a loooong hike, trekking 800 miles across dry, dry Arizona.

Upon his return, we realized we were missing something.

We couldn't put our finger on it at first, but slowly it dawned on us.

It had been almost 3 years since we'd done our ideas with people.

We missed it.

So as we return our project with new energy, we hope to return to community as well.

We aren't sure yet how this will look, but we know it will not look us alone together.

Our hope is to become a hub, where people come to connect with one another over ideas.

We cannot wait to find out what that looks like with you.

Information Processing

2021-05-08T21:12:10Z

Monday Musings #1

Playing Goldilocks: A New Way to Stave Off Information Overload

When information is unlimited, processing is everything

I’m on my fourth attempt to read the Innovators by Walter Isaacson. By now, I have the first few chapters, all about the early computing theory of Ada Lovelace and the 1800s, all but committed to memory.

Back in those days, information was scant. Books were not new, but not old either. They were still expensive as fuck and reserved for the upper crust.

We live in a profoundly different world. As Kevin Kelly pointed on the Tim Ferris Podcast, no one alive 30 years ago would believe there was a free service that could find any information you can imagine (thanks Google).

Slowly, then quickly, we moved from a society where a library contained at most 300 books to one where I can hold 3000 books in one hand (thanks Amazon).

The hunger for information that led might once have led to success can just as easily lead to insanity today. I know it well.

Today I’m going to muse about how to deal with information overload, dealing with each part of my information process separately:

Input
Processing
Output

Input

I referenced input obliquely in my article about my friend Neil, but went to work in earnest today during a therapy session. My first order of business in think about input was to get a feel for what amount of input makes sense. What can I reasonably expect to do?

Too Little: No reading at all. Rarely happens, but usually when I’m obsessively doing something else. Not a good state of being.

Just Right: 3 books at once, plus 7 articles a week. This seems to be a pace I can keep up.

Too Much: 1–2 books a day, plus articles. I can only do this for 2 weeks at a time. If I’m doing this without a reason (research paper, book, etc) then I know I’m not doing well.

As I looked over my above list and thought about it, I realized with great shame that I will not read every book in existence. Even if I did read a book a day for every year of my life, I’d put nary a dent in all the world’s books.

It was this realization that led me to realize that I’ve been straining at the wrong lever most of my life.

The lever to push on in reading is not the amount of information consumed, but the type of reading and the type of processing you do to it.

I’ll get to processing in a minute, but first a quick note on the type of reading lever. Ever since I read The Wright Brothers, who were first talked about in “Gleanings in Bee Culture” (all the major outlets ignored them) and heard Balaji say on the Tim Ferriss Show that, “The future is here, it just isn’t evenly distributed,” I’ve been considering how to find weird sources others ignore.

My little hunch on the value of this type of input was confirmed when listening to Kevin Kelly on, you guessed it, the Tim Ferriss Show. On the show, Kelly said something like, “The future is unbelievable now and what’s believable now isn’t the future, it’s a projection of the past.”

The gist of all this is that interesting ideas don’t come from the mainstream. By the time they’ve reached the mainstream (via a Seth Godin book), those ideas are widely distributed (which someone like Balaji might argue indicates low quality).

The really interesting ideas require digging in obscure places. Whatever you think of him, Nassim Nicholas Taleb does this well in his book Black Swan, which is part of what makes it a worthwhile read. Taleb digs up old, obscure ideas, combines them with new ones from psychology, and repackages them into something interesting.

So, given all this, my new input strategy is to limit my inputs to a sustainable pace (3 books, 7 articles) and find original stuff at the margins.

What’s your input strategy? Let me know on Twitter! I’d love to hear about it.

Processing

In the past, I wouldn’t have even considered processing. Before I picked up writing, my sole goal was to consume as much information as possible.

Part of my movement to information health is in the recognition of a need for a pause, to process the information. The first step to towards this realization came from increasing my output, back when I wrote every for more than 70 days.

What I realized in increasing my output was that I did my best if I had material on hand to write about. Not an article, but the start of my thoughts on a topic. Most days, that was all I needed to get going, a wisp of a thought.

Later, once I’d stopped writing daily, I switched gears to doing something with my inputs after discovering Obsidian and the Zettelkasten Method. All of a sudden, a whole new world opened up to me. I’d never considered doing anything with what I learned while reading.

Sure, I would write it up from time to time but I found that boring and a slog. What I learned over time is that the work required to write an idea from my inputs was hard because I hadn’t processed it.

So I began tentatively finding my way, finding yet another answer to the question, “What can I reasonably do?”

Too Little: No processing whatsoever. Information goes in, never to emerge again. My thoughts on the topic are lost forever.

Just Right: I pull out relevant ideas from my reading and leave the chaff (random factoids) by the wayside.

Too Much: I highlight the whole book and comb through it a second time trolling for information I missed.

The key difference between each of these levels of processing is whether or not I’m focused on important ideas.

What I realized about processing is that the key lever here is the number of interesting ideas I pull from the reading.

When I first started processing, I tried to grab every tidbit in the book. After doing this for a few books, I found myself exhausted and dreaded even looking at the book I finished, knowing how much work remained.

Then, I went on a processing fast, doing none of it for a time. I’m just emerging from my fast, hopefully wiser in my ways. The key difference is that I want to be focused on important ideas, not fleeting details.

The fleeting details are fun and there are low cost ways of keeping track of those (like Memory Palaces). But, when I process information, I want to keep the useful bits because if I’m going to go to all this trouble to find interesting inputs, I’d better get the best from them.

So my goal in processing information is to keep it interesting, focusing on key ideas from the uncommon sources I surface.

Output

Of all the parts of the process, I feel the most comfortable with output. My output is most like my canary in the mines. When the rest of the process gets out of control, my sweet singing output, which keeps my spirits up, disappears.

That said, I’ve done my Goldilocks treatment on output as well.

Too Little: No writing at all.

Just Right: 1–2 articles a week.

Too Much: Writing daily. Worst case scenario. Not only does my quality take a dive, I start to hate writing in this zone.

I’m not sure why keeping writing under control is so much easier than keeping inputs or processing in check. It could be that writing is something I’ve grown to love, whereas I was born loving reading and just found processing.

It could be that writing is so hard that it’s hard to abuse. I’ve never known writing to be an easy escape. Reading makes a perfect escape, as does processing. Reading because it takes you into other worlds and processing because it allows you to explore your own.

Writing has this gritty quality to it that’s missing in the more ethereal reading and processing. Don’t get me wrong, reading can be gritty and so can processing a big idea, but it’s not a requirement.

Writing always has an edge to it for me, I think in part because it’s hard to write without facing the truth. In the end, I think it’s the connection to the real world that makes output hard and hard to abuse.

By putting something out into the world, you must weigh it’s truth, which is no easy task.

But enough about me.

What do you think of the 3 categories, input, processing, and output?

What does your process look like?

What’s the hardest part for you?

Jimmy Cerone

I ask tough questions because I'd rather know hard truths than comfortable lies. https://mailchi.mp/e1b0dc1230df/thgispeed

Daily Ramblings

2021-05-08T20:37:22Z

Topic of the Day: Promoting Writing

I keep running into a dilemma when it comes to when to post and promote writing and personal projects. The optimal time for every platform other than Twitter, which is just weird, is smack in the middle of the work day.

I will sheepishly admit that I'm not above posting during the work day, but now that I've been at my current job for a few weeks, too many of my coworkers follow me. It would not be a good look to see a well thought out post popping up on their TL when I was supposed to be working away.

One could say, why not just use scheduling software?

There is a simple answer to that question: I'm small time and cheap.

If my personal projects and writing brought in the big bucks, I'd have paid for scheduling software long ago. Until then, I guess I will just keep posting on Saturdays and hoping the algorithm is kind to me.