Published
Weekend Reading — Dream of bigger things
Tech Stuff
Tolaria I just added this app this to my repertoire. I didn't get a chance to use it much yet, but it looks interesting — build a knowledge you can share with your AI that's version controlled and doesn't require an Office 356 license. Tagline: "Organize your notes as Markdown files. With native relationships, Git, and Claude Code integration" (also Codex)
The most underrated skill isn't knowing how to use every new tool - it's knowing which ones to ignore so you can actually get some work done.
First off, know what you’re doing. Verify what the LLM is generating. Approve every step. If you don’t like what you see, redo it.
LLMs are a tool. You can’t depend on your hammer to drive a nail. You still have to do the work.
GalaxyBrain If you live in the world of information collection you might find this app very useful. It holds information in a document that's interactive — you can navigate between panes, change values, toggle fields, run calculations, etc. Tagline: "Write like a document, calculate like a spreadsheet".
Everyone agrees LLMs produce sloppy code. But every production incident and bug fix I’ve been involved with over the past year has been caused by good old-fashioned human-written and human-reviewed code. We produce plenty of sloppy code ourselves.
Matthew Kenworthy Works every time!
My all-time favourite method of debugging is going to bed and looking at it again the next morning.
Little irons For people looking for jobs, a tool that helps you track "all your little irons in the fire" — kanban board, calendar sync, email forwarding, helpful AI, Chrome extension, and more.
Software Engineering Practices (are also) Useful for Token Reduction My first home computer had 32KB of RAM and we had to optimize for memory consumption. Now my computer has 24GB RAM and so we have to optimize for token consumption:
Software engineering practices were (at least when I'm pushing them) all about minimizing things for humans:
- minimize how much a person needs to read to understand the meaning of something,
- how many things a person needs to change to add, change or remove a feature,
- how many times a person needs to test something manually,
- how many times we repeat the same mistake,
- how many times we need to communicate the meaning of something
etc etc.
Related, labria
“We hired a junior dev to save on tokens for simple tasks”
After my experiments of the last few weeks, I have cancelled my Claude subscription and now run my coding assistants completely locally.
The cost went from 200$ per month to...10 Euro per month (and hardware that will be offset against those savings within about a year).
I'd also argue that my local carbon foot print is significantly lower than some random Anthropic data center.
Developers don’t like AI for the same reason they don’t like SEO; they resent being forced into a black-box discipline with rules that change without notice. They have spent years or decades honing their craft, becoming masters of logic and exactness. Then a new tool arrives that produces the same output, but only 99% of the time. 99% is not exact, and so it must be incorrect.
Claude Design + Opus 4.7 is actually game changing Top comment is priceless:
I'm oversaturated of things labeled game changers. Every 12 hours we get another game changer.
Words have no meaning any more. Every other tweet is someone stating "everything changes" or "such and such is cooked" or "nothing will be the same". So tiresome.
Agent Vault Open-source proxy that sits between the agent and API calls so you never have to share your secrets directly with the agent. By Infisical. (I recently switched my deployments to Infisical and it was one of my better decisions.)
Eye for Design
This cannot be real. I cannot believe my eyes
The critics, mostly designers and front-end devs, are dragging this design for being a "boilerplate" template with bad UX. And yes, everyone is having a field day with that map. The general feeling is that all Claude Design outputs look the same and a pro would get fired for shipping this.
The other camp argues that's not the point. For founders and non-designers, this is about leverage and speed. It's a tool to get an 80% "good enough" mockup in minutes, much like how Canva democratized basic design from Adobe. It's for first drafts, not final products.
jneen collective But also:
half the point of programming-tool design is to reduce the need for hypervigilance on the user.
if we're designing tools that require you to be more hypervigilant, legitimately what use are they?
Business Side
What, then, are we paying for? Remember:
Paying for software isn’t paying for a solution. It’s paying for someone else to own a problem.
It’s paying for someone who has the taste and the context to think through the details. For the operations and structures necessary to scale it, maintain it, and solve even bigger things.
Leaders hellbent on using generative AI to build everything in-house are asking their companies to own a wide portfolio of solutions, and with it, a wide portfolio of problems. But when a solution breaks, when it needs a new capability, or when an edge case appears, it requires ongoing judgement. It requires problem ownership. And a company maintaining its own solutions is spending its attention on those problems rather than the problem it actually sells. It’s allowing itself to become unfocused.
And besides, if every company is building all of their own solutions rather than buying software, who, then, is left for them to sell to?
'Tokenmaxxing' is making developers less productive than they think I remember when management was measuring Lines of Code as the top metric of productivity:
Enormous token budgets — essentially, the amount of AI processing power a developer is authorized to consume — have become a badge of honor among Silicon Valley developers, but that’s a very weird way to think about productivity. Measuring an input to the process makes little sense when you presumably care more about the output. It might make sense if you’re trying to encourage more AI adoption (or selling tokens), but not if you’re trying to become more efficient.
Design and Engineering, As One We inherited our product processes from a 19th-century steel works. Can we do better this time?
A sequential process – where design finishes before engineering begins, and a specification is the primary bridge between them – does something that is quietly destructive: it takes a manageable cognitive difference and makes it catastrophic. Not by creating the tension between the two lenses, but by removing every mechanism that could make that tension productive. In a sequential process, there’s no shared material to point at. No immediate feedback when a design decision collides with a technical constraint. No moment where the engineer’s question reshapes the designer’s thinking in real time, or where the designer’s intent becomes visible to the engineer before it’s cast into a deliverable.
What remains is translation. And translation always loses something.
Machine Intelligence
China pivoted hard to renewables. They now have MASSIVE surpluses of cheap electricity.
Cheap energy means cheap compute. Cheap compute means cheaper models. And cheaper models means Chinese open source wins more categories over time.
We still think this is a software war. It's actually an energy war. China is decades ahead.
Mediator.ai Using Nash bargaining and LLMs to systematize fairness.
What makes Nash’s solution compelling is that it’s provably fair in ways that feel intuitive. No value is left on the table; you can’t make one side better off without making the other worse off. Only relative preferences matter, not how you measure them. And the benefit of reaching a deal gets divided based on how much each side gains compared to walking away. It doesn’t care who’s louder or more stubborn.
Insecurity
How’s your evening going?
Mine is … watching @1Password and Apple password fight over who gets to login to the Alaska app to the point of locking my account and I don’t even remember why I last had to change my password and now I remember why everyone hates security.
Websites break California privacy law at ‘industrial scale,’ survey finds The problem with privacy laws is the lack of any meaningful enforcement. Also, when they keep penalties as low as the budget for office snacks:
The report, from researchers at webXray, a firm headed by a former Google privacy engineer, said the findings suggest major companies may be simply ignoring the law, and could point to “industrial-scale noncompliance with California requirements.”
Prove You Are a Robot: CAPTCHAs for Agents It's all about luka:
TwO tRaInS wAn/ Al\_E mIlE\s ApArT} aPp/Ro@AcH{
eAcH/ oThEr \< At{ Mu{T/e @ Tu\< Tu LuKa :
E#n\* T]u \ MpH a.Nd MuTe\ Tu Tu# Tu En LuKa
W|aN\_ mPh A b:I]rD fLiEs; Ba?Ck| AnD- fO^r@T\[h\\
^ Be{TwEeN? # t;He\*M aT wAn> ] AlE # eN lUkA
lUkA \< lUkA: # wAn ? MpH- uNt}I\[l T}hEy MeEt
`HoW! fAr- D_oE*s / ThE b@IrD fLy`
Everything Else
dch "Prague railway station has a LEGO model of Prague railway station."
it's like yes i could see how you could have arrived at that conclusion but it is nonetheless inaccurate in several crucial ways
AI will now do your shitposting for you so you can work for two more hours in the office
Paco Hope "It's called vibe plumbing. I'm not even a plumber. Soon, all swimming pools will be done this way! Get onboard or get left behind."
Being pro-UNIX does not mean I am anti-VAX.
at a job interview
"whats your biggest weakness?"
"understanding the semantics of a question but ignoring the pragmatics"
"could you give an me an example?"
"yes i could"
US WAREHOUSE FIRE TRACKER 🔥 (via Scary Austin)
wow! i commanded a team of 9 subagents to make me a baby in 1 month. the results were astounding!
‘Hairdryer used to trick weather sensor’ to win $34,000 bet "we live in the dumbest timeline" (via Viss)
French police are investigating suspicions that a hairdryer may have been used to tamper with official weather readings to make thousands of dollars in Polymarket bets.
Temperature readings at Paris’s Charles de Gaulle airport have unexpectedly spiked twice in the last month, reaching levels much higher than expected.
The Snacks & Cereals of 2025 I have zero interest in US chips and snacks, I prefer to spend my caloric intake on other delights, but I totally enjoyed reading this blog post. A delightful review of what sold on supermarket shelves last year. Delightful and funny.
ASRS is a screening test for adults to see if it's worth pursuing an ADHD diagnosis. It is 18 questions, written by neurotypical people who assume their external observations of ADHD are accurate representations of what it's like to have it, so needs to be translated into "actual person with ADHD" before you fill it in.
That's why it includes stupid shit like, "do you feel driven by a motor?" No, I do not. Things driven by a motor spin round and round and I do not spin.
And, "Do you have problems remembering appointments" to which everyone responds, "It's 2026 and I have a fucking SMARTPHONE".
So that being the case, I have often thought one written BY ADHDers, FOR ADHDers would be better.
It would include questions like:
"Do you think your friends all dislike you and that you annoy them?"
"Is there a song on repeat in your head? Is it the same 3 bars over and over? It is, isn't it?"
"Your partner says, 'we need to talk'. Do you flee to Antarctica? If not, is that because you got distracted by TVTropes when trying to find out how to flee to Antarctica?"
Natasha "Seen in Hardcore Bikes (HCB), a bicycle shop in Edmonton Alberta 🚴♀️"
