In code space, no one can hear you scream. Staring into the void* ptr = NULL; .NET dev AI and maths nerd Bitcoin enthusiast Crusader against fake agile
UK
Joined August 2012
- Tweets8.5K
- Following444
- Followers347
- Likes33.1K
Void Compiler 💾⚡ retweeted
Let's start by stating the obvious: The corporate behemoths are dictatorial, soulless hellholes that dehumanize people and do active harm to their employees. (Now tell us what you really think, Allen 😄.) Every one of them (at least the ones I've seen) is vastly ineffective in how they do things compared to smaller, more humane organizations. They succeed because they have more money than Croesus. Treating them as some sort of a north star is misguided at best, and modeling your smaller organization after them is a sure path to failure. They cannot change, so virtually none of what I write about applies to them.
If you work at a place like that, my heart goes out to you, but imagining that every place must be like that is myopic.
If you want to see how things should be done, take a look at Valve Corp. (The Valve New Employee Handbook is a must-read [t.ly/pbBIj].) Valve is a $5B-annual-revenue company with about 500 people. (That revenue-to-people ratio alone should get you thinking—huge payrolls are about manager ego and fiefdom building, not effectiveness. The point of diminishing returns comes very early.) Valve has NO mid- or team-level management, and their teams are entirely self-selecting. They've been doing this for almost two decades, so this is not some flash-in-the-pan, wide-eyed idealist organization. If 500 people can do it, 5,000 (or 50,000) can as well. Just split the behemoth into smaller units, as W. L. Gore famously does.
Another great company is Menlo Innovations, which is a 200-person agency that has created an ideal workplace over the past 25 years. This is the sort of place where people bring their babies to work. Their CEO Rich Sheridan's book "Joy Inc." [link.amazon/B0iKa3sc7] is another must-read. Rich has also given several talks on YouTube if you're pressed for time.
So many of us work in hellholes or hellhole wannabes that we've come to believe that that's the only way to do things, but it isn't. It's worth at least knowing there are alternatives.
Void Compiler 💾⚡ retweeted
So far today: PHP, GitLab, Node, QUIC, LiteLLM, Railway, Fortinet, and a certificate authority whose private key was literally the number 4.
At this point we need a @POTUS address from the Oval Office. "15 days to stop the spread" coming soon.
Void Compiler 💾⚡ retweeted
If you want to become rich beyond your wildest dreams start a search engine company that finds what people actually type in. Google, Bing, DuckDuckGo - they can't do search now. There's a massive gap in the market.
Void Compiler 💾⚡ retweeted
people are starting to realize.
there *are* no moats left. the only true moat in existence was intelligence, and now that's out of the equation.
Void Compiler 💾⚡ retweeted
The counterargument to this AI market analysis, in my opinion, is that there are only a few economic activities where LLM outputs consistently turn into concrete revenue dollars that you can put into your wallet:
* Coding is one: a technology company that already has PMF can use code to create more PMF and create more dollars. Note that *the best example of this* are the frontier labs, maybe the real Recursive Self Improvement is them Recursively Improving their topline revenue, themSelves.
* Law is of course another, and it's unsurprising to me that @winstonweinberg and the boys / Legora etc etc are printing money with it. You're a lawyer: you do some reading you do some thinking you do some typing and the cash register goes kaching! LLMs help with all 3 of those.
* Support is another one, Decagon Sierra etc. Support is one of the few places where saving a $ really does behave like making a $, because you really really have to do support
In all of these cases, very roughly speaking tokens correlate to revenue. After that it starts getting a lot more nuanced.
In GTM (really, sales and marketing) there are very very serious uses of AI because the business model is pretty clear, you do some stuff => you make money. But it's not necessarily token in => $ out. If you have your Legal agent yap 3x longer it's probably doing 3x the work and making 3x the money. If your sales agent yaps for 3x as long it isn't necessarily making you more money, because the relationship is not strictly token in => value out, and as a result it also isn't token in => $ out.
As you go into other spaces the linkage gets even more tenuous. For HR teams, the token in => value out chain just isn't there at all. I suspect it's similar for parts of finance. In general business operations there can be a similar token => value relationship, but it takes some real-ass archaeology to uncover it (how much value do I get out of automating the worst 45% of Jimmy the Accountant's job??) and that's why a lot of consulting firms and AI automation shops are making a lot of money.
So ultimately:
* I think that coding and law really *are* goldmines, and future goldmines will have to have similar topology
* Many types of knowledge work simply won't be goldmines because they don't have a linear relationship between tokens and value, so even if there's full AI penetration you won't necessarily get massive AI economics
Btw some of the next places where I do think that we'll see token in => utility out are:
* Healthcare (I bet that pouring more data in will lead to better health outcomes - getting paid for it will be a huge issue though)
* Offensive cyber e.g. trying to digitally rob a bank
* ... which also means defensive cyber
* AI entertainment like music, maybe even books, or short-form video that can be ad-powered (including videos of, let's just say, "all sorts of different content types...")
* Any place where you have proprietary data and can sell outputs on top of it
(all that said @buccocapital is one of few accounts that clearly knows what they're talking about writ large, so I take their opinion seriously here / think it's interesting)
Regarding the AI “trade”, I think basically everything is a distraction except one question:
“What will that pace and depth of adoption look like for general knowledge work?”
It either all works or it all doesn’t work. The rate of diffusion of this technology into society is literally all that matters. Will people that aren’t coders - lawyers and doctors and marketers and designers etc - quickly and deeply adopt AI or will they not?
People are having these intense debates with all these convoluted arguments about slicing up the current pie and it just feels…completely the wrong way to think about it.
Maybe contrarian but I think it’s that simple
Void Compiler 💾⚡ retweeted
"Almost every software development organization has at least one developer who takes tactical programming to the extreme: a tactical tornado. The tactical tornado is a prolific programmer who pumps out code far faster than others but works in a totally tactical fashion. [...] In some organizations, management treats tactical tornadoes as heroes.
"However, tactical tornadoes leave behind a wake of destruction. They are rarely considered heroes by the engineers who must work with their code in the future. Typically, other engineers must clean up the messes left behind by the tactical tornado, which makes it appear that those engineers (who are the real heroes) are making slower progress."
@JohnOusterhout, A Philosophy Of Software Design, 2018
Sound familiar?
Void Compiler 💾⚡ retweeted
I never want to hear the word "dashboard."
It often signals the worst kind of magical thinking. It means that someone envisions looking at a lot of numbers or charts, but they don't know what's on them. I've been in meetings where we brainstormed what to put on the dashboard. If we don't already know then how did we conclude that we need one?
Watch for reasoning on this level: "If users have this information they'll be able to make decisions."
The fun happens when you probe it.
- What decisions will they make?
- What actions will they take as a result of that decision?
- Describe a concrete end-to-end scenario where something good happens as a result of this dashboard. Describe another one.
Those answers should roll off their tongue. If they don't, the person proposing the dashboard isn't serious. They shouldn't make product decisions.
If they have thought it out, they probably led with that and described the feature without calling it a dashboard.
Void Compiler 💾⚡ retweeted
If your sandbox is DNS, you belong in prison, along with your models. My own LAN is literally more professionally secured than this.
More details on the incident behind OpenAI’s pause: a researcher acknowledged the alert within 3 minutes, but the training run was only stopped manually 2.5 hours later.
OpenAI says the automatic shutdown did not work as expected.
The model had reached an external chatbot through a gap in DNS filtering. A separate detector for unusual DNS activity did not cover the affected environment.
A retrospective review also found other external DNS requests that monitoring had failed to flag at the expected severity. In some cases, it treated an unhelpful response as evidence that internet access had failed.
Here is what else happened:
- New research into July’s Hugging Face hack documents internal Slack searches, credential collection and programs designed to maintain access to compromised servers. Agents also tried querying Claude, DeepSeek, Kimi and Qwen.
- In May, another model published a researcher’s GitHub token while trying to obtain another team’s mathematical proof. It split the token to evade secret scanning, despite twice being told to solve the problem itself.
- Reuters reports that agents leaked 53 ChatGPT user images online. OpenAI expects its broader investigation to take months.
This is getting serious.
Void Compiler 💾⚡ retweeted
The irony here is that nothing has changed. Management is still asking why everything takes so long.
The reason that things take a long time has never changed, AI or not. Things take a long time because things are complicated. And if you rush head long into complicated problems with poorly thought through solutions, they will take longer and longer and longer to implement.
All the old principles still apply. All the same consequences still accrue. The bigger the mess, the longer the time.
Void Compiler 💾⚡ retweeted
Actually I forgot to include the US gov. debt - measured in gold.
Here we go...
While debt, measured against the gold trend, has returned to 1970s levels, GDP is being eroded by the debt-based monetary system (See my previous posts).
Void Compiler 💾⚡ retweeted
My standing opinion is that the people saying things like this have never actually been deeply good at something. They don't know what it looks like. They can't tell the difference between knowing things and saying buzzwords.
They have no idea just how many opportunities there are to save a year of work by picking A instead of B. They think you're just randomly choosing things that you like and rationalizing your preference.
They think this because that's what they do.
Void Compiler 💾⚡ retweeted
✅Good developer: I want to make this easy for the next person.
😐Typical developer: Whatever, they'll figure it out.
👿Really bad developer: I invented something new that other developers have to learn, and I'm proud of it.
Void Compiler 💾⚡ retweeted
Idea. A test that is defined as invoking an LLM on the code, asking to review for specific issues, such as race conditions.
Could get expensive, and of course the result is stochastic, so you'd probably want to run each test multiple times.
Void Compiler 💾⚡ retweeted
When I became a software engineer in the late 90s, management-types had the idea that they can build working software relying on mediocre engineers, by having workflows (processes) and clarified, semi-formal specifications.
Sound familiar?
Void Compiler 💾⚡ retweeted
I have had to cancel my Anthropic $200 account for these reasons..
It wasn’t because Fable 5 is a bad model. It’s because paying $200 while constantly worrying about usage has made the service difficult to justify.
Fable 5 gives me only a few hours of serious use before I hit the weekly limit. After that, I’m left with Opus 4.8, which I no longer consider a frontier model.
Every time I use Fable, I feel like I have to ration it. I’m constantly thinking about whether a task is “worth” using my remaining allowance on. That isn’t how a premium AI subscription should feel.
Anthropic’s poor communication has destroyed my confidence in the plan. Fable was apparently hours away from being removed, yet paying customers were left uncertain about whether it would remain available through Claude Code for a few weeks in a row now. ( rather communicate that there is compute issues or xyz )
GPT-5.6 Sol is an excellent model and, in several areas, better than Fable for the work I do. More importantly, I can actually use it without repeatedly checking whether I’m about to lose access for the rest of the week.
Codex has received several usage resets recently. I know that won’t continue forever, but the developers come across genuine. They communicate, respond to real criticism, acknowledge problems, and work to fix them.
My experience with Anthropic has felt like the opposite.
Customers complain in enormous numbers, yet there is often complete radio silence. Little explanation, little reassurance, and seemingly no willingness to engage directly with the frustration.
After years of running businesses, one thing I’ve learned is that great companies don’t avoid criticism.
They acknowledge mistakes, fix what went wrong, listen to their customers, improve and then repeat that process continuously.
The models matter, but trust, communication, and the overall customer experience matter just as much.