fallthrough

Where are you going today? Well, we're going to go over down route 53, then head down bedrock ave, and in no time we'll be going down the highway. But where are you going? We expect to hit 85 within the first 20 seconds of hitting the highway! But where are you going? I just told you!

From the outside, complicated problems seem simple to solve. Have a gun violence problem? Simple! Get rid of the guns. What's hidden is the path dependence. For example, how exactly do you suggest we get rid of half a billion guns? @skriptble.me @matthewsanabria.com #commonsense #problemsolving

At some point, we can't place the blame on frontier AI lab CEOs for saying insane things. We keep giving them a platform. And more importantly, rarely do we take the history of claims they've made to weigh new ones. That's the first change we need to make. @skriptble.me @matthewsanabria.com

Wealth management is a weird trap that makes the entire system worse while making sense at each step of the way. It's a weird twist of Simpson's Paradox. We won't fix problems like this without fixing the whole system. @skriptble.me @matthewsanabria.com #business #wealthmanagement #tech

Some believe that large companies are well run, @skriptble.me questions whether they've actually peered at the insides. There is mess everywhere, seemingly unlimited chaos, and a bunch of people barely holding the entire thing together. Elegant this is not. @matthewsanabria.com #tech #business

We love to focus on metrics as a proxy for accomplishing real things. "Look how fast we're shipping!" "Look at how many lines of code I wrote!" "Look at how many tokens I used!" But did you actually do anything useful with that stuff? No? Oh, okay then. @matthewsanabria.com @skriptble.me #tech

Just install OpenClaw! Wait, no, I mean just install Hermes Agent. Wait, never mind, scratch that, just install Pi! The whiplash of these recommendations is giving @matthewsanabria.com a headache. There must be a better way. #tech #llms #codingagents

For many, not having to write the code means they can move faster. But domain knowledge is tricky, because it means that not knowing how to implement one small component of a larger project can stall the entire thing. Coding agents fix that. @skriptble.me @matthewsanabria.com #tech #codingagents

Cryptocurrencies were poised to change the world, and then they didn't. Many believe it was because they were overhyped, @skriptble.me has a different theory: they matched people's mental model of money even though that mental model is incompatible with actual finance. @matthewsanabria.com #crypto

@adamhjk.me thinks that we should just let the LLMs go wild, but @matthewsanabria.com sees the value in continuing to do tasks ourselves that LLMs might automate. While those with decades of experience benefit from the speed, those with less experience benefit from the friction and the struggle.

Holding informed opinions today is foolish. We had web search before, but they had friction to learning something. With LLMs the barrier to deeper knowledge has been removed. It'll get things wrong, but a nuanced wrong opinion is better than an uninformed one. @skriptble.me @matthewsanabria.com

@skriptble.me wanted to optimize a ZFS dataset for the data it'd be storing, so he set the recordsize to 4MB. When he started actually using the dataset he was greeted with minute long stalls. The problem? An SQLite database causing 1,000x write amplification 😬 #softwareengineering #zfs #filesystem

Move Fast and Break Things wasn't the first phrase but it was the most blunt and honest. Tech as an industry loves to prioritize speed over all else, including any sense of direction. With AI tools, we've doubled down on this and it's made things much worse. @skriptble.me @matthewsanabria.com #tech

Epistemic markers are phrases like "I think", "I guess", "maybe", and "certainly". They tell us how confident the speaker is in what they are saying. Their usage has been declining for decades, but AI leaders have taken it to an entirely new level. @skriptble.me @matthewsanabria.com #language #tech

Collective action often requires collective suffering to achieve collective gains. But if we allow people to leave the collective without consequence, or worse, allow those who would participate in the gains to undermine the work of the collective, we will not achieve progress. @skriptble.me #unions

Coding agents are great. We can build projects we never thought possible. But it also reduces the hands on keyboard time we all have writing code. This can lead to atrophy of those programming skills. We need to find a balance between prompting and programming. @matthewsanabria.com #tech #llms

In some ways, the way people are acting around AI makes sense: the business is incentivized by shareholder pressure to invest heavily in it and employees are incentivized by the business to invest their energy into it. If want this to change, we need systemic fixes. @matthewsanabria.com #tech

We often treat load-bearing infrastructure in software ecosystems an unserious way. When catastrophic things happen in software ecosystems we rarely make the systemic changes requires to ensure it won't happen again. @matthewsanabria.com #infrastructure #software #softwareengineering

We've been told the world is run by incredibly strategic multi-billion and multi-trillion dollar companies, but for those who work inside them, it feels like nothing could be further from the truth. Would these companies function if all the EAs stopped working? @skriptble.me @matthewsanabria.com

The open source world was designed for a world where there was more friction for doing things. With LLMs a lot of that friction has been removed. Sometimes that means we have to ask a potential contributor "are you a human?". @www.jvt.me.web.brid.gy #opensource #tech #llms

It's far too easy for open source maintainers to step out of the joy economy and into the capitalist economy without noticing it's happened. All of the sudden that fun side project is now filled with issues, sometimes from big companies who would like you to fix their problem for free. @skriptble.me

Do LLMs hold all of the world's knowledge? Absolutely not, and relatively they hold about as much knowledge as you do. But our brains are bad at understanding relative scales as non-linear. @skriptble.me #tech #llms #softwareengineering

It usually only takes 1 project for an open source maintainer to learn they need to set boundaries. Unfortunately, that 1 project often becomes load bearing infrastructure before we've even realized what's happened, putting everyone in quite a predicament. @www.jvt.me.web.brid.gy #opensource

Software folks are not the main characters of their employers. And yet, for almost our entire existence as a discipline, we have acted like we are. The consequences? We've become inscrutable the rest of the company. That's made open source difficult for a business to understand. @skriptble.me

"You've built a concept in the tone of Git, but it's off key" @skriptble.me Pull Requests created a faulty foundation. It made sense when GitHub was created, but at scale it doesn't work as well as alternatives. GitHub Actions compounded this problem. Also, YAML is *awful*. #git #pullrequests

People make projects for joy. And then others start depending on those projects. You wind up being buried in work and the project no longer generates joy for you. Instead it generates exhaustion, and eventually burnout. We need a better way to handle this. @skriptble.me #opensource #software

GitHub and GitLab do too much. @skriptble.me thinks they should actually be hubs instead of being the everything platform. We already went through this with social media companies and the solutions were more open platforms like the Fediverse and Bluesky. #tech #software #opensource #github #gitlab

Are software engineers suffering from main character syndrome? @skriptble.me argues that they have been and there's a reason to actually fix that. If software engineers can't argue for things they support (e.g. open source) to the business, they won't get buy-in. #tech #software #business

Does more hours mean we'll get more work done? Absolutely not. C. Northcote Parkinson wrote (satirically) about this: "Work expands so as to fill the time available for its completion." @skriptble.me expands it a little: we'd get more work done if we worked fewer hours. #tech #workinghours

Benchmarks keep going up but the models don't feel better. @steveklabnik.com and @skriptble.com on why "10% improvement" means nothing if using it still feels bad, and how some open models post frontier numbers by overtraining on the benchmarks. fallthrough.fm/ep/70

@steveklabnik.com doesn't notice 4.7 regressions because he kicks off a task, walks away for an hour, and checks the result. If it takes 75 minutes instead of 60, he never feels it. The subjective experience of your workflow shapes your whole take on whether a model got worse. fallthrough.fm/ep/70

"If Claude goes down I still want to be able to work" is where the solution era ends and the toolbox era begins. fallthrough.fm/ep/70 @skriptble.me

Anthropic is telling customers to fuck off, not because it doesn't want their money, but because it can't handle the growth. We don't have a product strategy here, we have a capacity crisis. fallthrough.fm/ep/70 @steveklabnik.com

OpenAI overbuilt capacity and in hindsight, that looks prescient. Anthropic was conservative and is visibly struggling to serve the traffic they actually have. fallthrough.fm/ep/70 @steveklabnik.com

@skriptble.me doesn't like Opus 4.7. @steveklabnik.com doesn't have a problem with it. The difference? Steve is mostly dispatching long running tasks. fallthrough.fm/ep/70

GitHub's traffic went up 45x in one year and it feels like their service uptime went down proportionally. Not because they're incompetent, but because of architectural decisions. SaaS means one person's outage is everyone's outage. The solution? BYOC. @steveklabnik.com fallthrough.fm/ep/70

"I have read the constitution and it is a little bit wild. It feels like a religious document." And it better not call itself we, that'll scare @ianlopshire.bsky.social. fallthrough.fm/ep/64

AI agents push you toward monorepos sooner, because you want the full codebase on disk so Claude has context. And five agents running simultaneously means a lot more commit traffic than you're used to. @steveklabnik.com on the architectural ripple effects: fallthrough.fm/ep/70

Gemini is depressed, Codex is sycophantic, and Claude is made by a quasi-religious org. @steveklabnik.com's conclusion? The weirdos who are weirdest about it are also making the best one. fallthrough.fm/ep/64

Git was designed for the Linux kernel, which most people would describe as large. If you've worked at Google or Meta, you'd describe it as small. @steveklabnik.com fallthrough.fm/ep/70

JJ has no staging area. At least, that's how it's presented. As @steveklabnik.com discovered, the staging area is a workflow pattern in JJ, you can opt in or you can opt out. fallthrough.fm/ep/70

"It's easier to scale a big thing down than to scale a small thing up." -@steveklabnik.com Listen to the full conversation about JJ, Git, and scaling software at fallthrough.fm/ep/70 #git #systemdesign #scale

"What if GitHub was not just one host, but a bunch of them?" Tangled is doing that, built on AT Proto, so your Bluesky social graph carries over. fallthrough.fm/ep/70 #OpenSource

@steveklabnik.com: "One of the most beloved features of JJ is jj undo and it's much easier in JJ to see what a command does: just run it and see what happens. I know that if something goes wrong, I can immediately run jj undo and get back to the state that I was just in." fallthrough.fm/ep/70 #jj

A couple years ago, it was unthinkable to build a project not on GitHub. But now, it's healthy to have more hosts and other ways of doing things. @steveklabnik.com on the current state of GitHub and the rise of alternatives. fallthrough.fm/ep/70 #OpenSource

@steveklabnik.com on JJ's approach to handling conflicts: "JJ allows commits to stay in a conflicted state. It's chill with that happening." @steveklabnik.com explains why this is important for developers working on multiple branches simultaneously. #OpenSource #Git fallthrough.fm/ep/70

GitHub raised the popularity of Git, but its Pull Request model has always been at odds with how git works. Review per branch isn't how the Linux mailing list works and now tools like JJ push developers to work closer to the way Git was meant to be used. @steveklabnik.com fallthrough.fm/ep/70

@steveklabnik.com could debate 600,000 babies and he argues this means that Claude is intelligent. Confused? This clip from episode 64 should clear things up.

There's so much debate about AGI and superintelligence, but @skriptble.me argues that we should instead be more focused on who's doing the debating, not on what they're debating. @ianlopshire.bsky.social disagrees. Who do you agree with?

Fighting revolutions does suck, as @steveklabnik.com. But as @skriptble.me counters, that doesn't mean the lives of those fighting is only despair. #History #Power

The DEI conversation got exhausting because it kept getting trotted out by companies that didn't feel bad about who wasn't there. They felt bad about where they landed in the distribution. fallthrough.fm/65 #Diversity #TechCulture

"If you can't be honest about the capabilities of your technology, we don't want to hear it." The AI industry right now is obsessed with the appearance of disruption. fallthrough.fm/65 #AI #AIHype

We are not powerless. Collectively we wield tremendous power. We just have to decide to actually wield it. fallthrough.fm/ep/64 #Labor #Unions

We love to overindex, especially during hype-cycles. Currently, we're overvaluing LLMs and undervaluing humans. We must not lose sight of the reality that LLMs need us, while they are a convenience for us. fallthrough.fm/65 #AI #LLM

@matthewsanabria.com has been clinging to Linux on the desktop for ages now. He knew it was only a matter of time, though. A MacBook Air, iPhone Air, and AidPods Max later and he's all in on the walled garden ecosystem. fallthrough.fm/65 #Apple #Linux

If someone keeps fathering children but never sticking around, we have a word for that and it's not flattering. Why do we use a different word when they do it with companies? @skriptble.me fallthrough.fm/65 #StartupCulture #Founders

Like social media feeds, we should worry about how AI will pull us away from each other by pulling us toward it instead. "It's incentivizing humans to be isolated, when we're naturally a communal group of people" --@matthewsanabria.com fallthrough.fm/64 #AI

The founders celebrated on stage are rarely the ones who did the late-night booth setup the day before the event. "You singular are not. You plural collective are." @matthewsanabria.com fallthrough.fm/65 #StartupCulture

The AI hyper tells us to worry about LLMs taking our jobs, but the AI that's actually controlling our lives is much quieter. It's far more subtle. fallthrough.fm/ep/64

The social media companies had this problem. Now the AI companies have this problem. "We're a big family" is the warning sign. Families that demand ideological conformity are cults. fallthrough.fm/65 #TechCulture #Diversity

"You literally cannot DMCA it if it's not real." Anthropic issued DMCA notices for the Claude Code leak and in doing so settled the debate about whether it was genuine. fallthrough.fm/65 #Anthropic #ClaudeCode

The AI overlord isn't a Terminator. It's the algorithm you're scrolling right now. @matthewsanabria.dev on Fallthrough Ep 64. Full episode: fallthrough.fm/ep/64

"Humans are collectively good but individually bad." @ianlopshire.bsky.social on Fallthrough Ep 64 framing superintelligence as a stats problem: 9 billion humans, one training run, normal distribution. How do you like those odds? Full episode: fallthrough.fm/ep/64

The Linux holdout is gone. MacBook Air, iPhone, AirPods Max, all in one week. Calling customer support by dialing from the computer was the final unlock. fallthrough.fm/65 #Apple #Linux

The scary AI isn't god-tier superintelligence. It's something smarter than humans but still dumb, still doing what its boss tells it. Ian on Fallthrough Ep 64. Full episode: fallthrough.fm/ep/64

Every "AI is making us dumber" panic you've heard? Plato had the same one about writing 2,400 years ago. @steveklabnik.com on Fallthrough Ep 64. Full episode: fallthrough.fm/ep/64

We only call ourselves the "most intelligent" species because we wrote the test. Dogs smell better. Birds see better. Bacteria evolve faster. Beavers reshape ecosystems more efficiently than we do. @skriptble.me on Fallthrough Ep 64. Full episode: fallthrough.fm/ep/64

Terminal emulation is everywhere, and it kind of sucks in most places (e.g. IDEs, editors, web browsers, etc...). Mitchell Hashimoto has a solution for that!

For a while now, people have assumed #rustaceans wanted to rewrite all of the things in Rust. But is that actually true? Here's @steveklabnik.com's take!

@steveklabnik.com thinks bringing new people into open source is like feeding cats to coyotes. What do you think? 🤔