getzerod

Zerod in on decentralized tech, intelligence, and finance.

[getzerod @getzerod]

@DaveShapi I’ve preferred the term “consumer class” in my head. It has a better implication - that a group of people are actually needed to consume the products of abundance, and that it will be a (relatively) okay and accepted thing to exist here.

[getzerod @getzerod]

first, yes - your $200/mo plans from anthropic/openai are insane value. min/max your capacity, buy more accounts as needed, swap often, build up your systems, no brainer

but the subsidy buys them time to get the haiku cost models to sonnet (and eventually opus) levels

however, what comes next will cost more, correct

we will get to a point (and beyond) where your ~$1/$3 m/tok models will be "good enough" for all but the most difficult questions, but we gonna be blasting so many tokens costs will keep up (for awhile)

[getzerod @getzerod]

@icanvardar hate to say it, but until opus level models get cheap enough to run on api token costs, most will keep using cc

the level of subsidizing they do is insane. id blow the $200 day one

[getzerod @getzerod]

skills are great, but they are just one part of the equation. i do think they are better now than tools/mcps, and align with what agents really want - instructions and scripts. but skills tell only part of the story, how to run the skills is the actual problem to solve for - which means you need a system for how to work in your codebase.

you should be able to start up a fresh CC/Codex/OpenClaw/etc session and ask "ok, whats next?" - if you dont get a good answer out of that, and quickly, your system is not aligned and coherent to get work done.

i think a lot of the systems out there like Gas Town and Superpowers are great, but i dont think its yet time to just adopt a system from someone else. things are moving too fast, you need to be agile, and understand the ins and outs of how you do work in a codebase.

a better mental model for your "do work" harness is the shadcn/ui route - you dont import a library, you add the atomic pieces that you will compose views from. they are yours to update and make part of the codebase - not some integration that you have to fight. this is where skills start to kinda get it right, but the framework around the skills ... you should own in your code. they should compound for YOU, not for the person building the framework. they understand that system in a different way than you ever will. your system should move in your direction.

the funny thing is, i think this is all noise. and we are not yet building around the right fundamental pieces.

the code should be an artifact of the spec. the code should be disposable, rebuildable at will by the next best model. the spec should be durable, yet modifiable. the code should realign around the spec.

i say this as someone with repos littered with either a /specs or /docs folder, who freaking knows the state of all of these. i assume they are "done" if im not actively working with them. but those specs and docs are constantly fighting against the code being reshaped, one feature or new spec at a time. there is no source of truth for "how should this work?" other than the code. which at this point, i dont think is the right way to go about things.

a spec breaks into 2 initial pieces: what and how.

the "what" defines your features. the "how" describes the tech stack and conventions - this is where skills should live and be activated from.

the reference for how to build forms should not be the first form that the agent decided to build in whatever way it chose. it should be a clearly defined form convention that points at your preferences/skills for how to build forms.

once we have the what and how, a system to do work is required, which means a system to break the spec down into "units of work" is also required. and this is the main battle ground everyone is fighting over right now, but it seems like focusing on how to define the what and how is more important, and will inform the work phase in a better way.

we should have swappable specs and tech stacks. you want your app in flask, ruby, node, astro, rust .. whatever, swap out the tech stack, the "how". we need to start aligning for the personalization layer, where people dont even want your app anymore, or your tech stack, they just want a reliable spec to feed into their agent, so it can get built out into their personal ecosystem.

bottom line: the code is not the source of truth. the spec is. the code is an artifact of the spec. specs should be durable, swappable, upgradable, and yours. it's time to refocus on the spec layer.

[getzerod @getzerod]

@iannuttall check in, see 30k skills, leave

curation > quantity

no one wants to sort through all that. work on personalization and curation and ur there. we need "kits" of skills. looks good tho, close.

[getzerod @getzerod]

@Ibelick let the annoyance flow thru you, right into skills, lol. i can feel it in those caps. great list, gonna try it.

[getzerod @getzerod]

@JustJake those 1%ers must be given skin in the game, or yes, they will leave. gone are the days of keeping high quality talent around for a basic salary.

[getzerod @getzerod]

@deepfates i have not seen spokenly, i'll try it ... but has it EVER failed a transcription? does it hold up fine if you leave it on for like 5 or 10 min with tons of long pauses? do you notice degradation of responses/speed when your computer has not been restarted for a long time?

[getzerod @getzerod]

and the slope is accelerating, look to where we are going, not where we are.
https://t.co/MJoPPCGbff

[Andrej Karpathy @karpathy]

The majority of the ruff ruff is people who look at the current point and people who look at the current slope.

[getzerod @getzerod]

We're now in the era of slope code.

[Naval @naval]

We’re now in the era of slop code.

[getzerod @getzerod]

@thekitze @DmytroKrasun A quick padding cleanup would go a long ways to tighten this up on mobile. I’d build the vote piece into the header, including the buttons. Give the prompt full width.

Looking good tho, thx for sharing.

[getzerod @getzerod]

3 biggest step changes in my AI journey: GPT-3.5, Sonnet 3.5, Opus 4.5

autocomplete → view composition → system composition

we're orchestrators now. the work shifted - planning, decisions, catching bad patterns early. typing isn't the job anymore.

smarter, faster, cheaper - all coming. but what we have works. you just have to ask the right questions.

[getzerod @getzerod]

@businessbarista Claude 🤝

[getzerod @getzerod]

@rauchg Wait, so first off you are still typing this much now? And second, you think you will still be typing this much in years? Yikes. That sounds terrible.

We speak faster than we type. We read faster than we listen.

Add voice to text in the next hack and pump those numbers up.

[getzerod @getzerod]

claude code be like

[getzerod @getzerod]

@JuanRezzio Love the new model, very good and fast.

But the new agents pane locked up so bad on one of my workspaces that I can't even open it anymore! (locks up) Had to open the project up one folder to get a fresh slate.

I want it to work so bad, felt so nice. Close.

[getzerod @getzerod]

@janekm @xfreeze Now go check the price of direct X api access (pain). If grok can essentially do it for you without needing the api, then amazing, but to be seen. Need to do some testing.

[getzerod @getzerod]

@TheRealPlanC Paging @benjamincowen for thoughts on this model vs Ben's risk model. Thanks.

[getzerod @getzerod]

Plan mode in cursor = very good

Now instead of begging the agent to make a plan and ask me questions, it's built in.

Describe feature, pull in initial files you know will be referenced and/or those that have the pattern that needs to be replicated.

Let it rip for a bit, then answer questions and provide feedback on the markdown plan file. Loop.

Once plan is ready, click build, let it rip again. At that point the feature is generally done, maybe a couple tweaks or a final cleanup prompt.

Using sonnet 4.5 thinking max (1m context).

EZ mode.

[Juan @JuanRezzio]

What do you guys think about Plan Mode so far? How can we make this better?

I'm curious to see how you are using this

[getzerod @getzerod]

@CryptoG911 @Wesatoshis @benjamincowen I don't like to chime in on these, but you are just wrong. Anyone actually paying attention to Ben was buying BTC on a dynamic DCA weekly according to risk levels, plus limit orders - and we all got a very good piece of the bottom.