Przemek

Solo dev. With AI I build things that used to need a whole team - faster than I can name them. Following along: what I ship, how, and what I learn.

[Przemek @przemekbuilds]

A year ago this would've taken a team.

With AI I'm shipping real, privacy-first products solo — and building in public. Here's what I'm making and how I work.

→ First up: ViriVox — voice-to-text that runs 100% on your machine. Hit a hotkey, speak, and the text lands in any app. No cloud, no account, one-time price. Windows + Linux, macOS next.

Why local? For a lot of people — lawyers, doctors, journalists — sending audio to a cloud service isn't allowed. Same Whisper accuracy, none of the exposure.

It's not the only thing I'm building. A few products across different domains, same solo + AI playbook each time — I'll show them here as they're ready.

The "how": an agent crew I built to plan, code, review and ship. That's the multiplier behind keeping this many things moving alone.

I'll post the real build — progress, what breaks, what I learn, and takes on where local + agentic AI is heading.

[Przemek @przemekbuilds]

I opened an old Claude Code session and typed “hi!”

The prompt cache had expired an hour earlier, so the whole 821k-token conversation got processed again: about $6.57 at API prices. For “hi!”

Kept warm, the same session costs $0.17 an hour. So I built a Keep warm button for it.

Explainer made with Claude: every frame drawn in JavaScript.

[Przemek @przemekbuilds]

Meta released an 816B open-weight model that runs on two Mac Studios.

Impressive. Also not exactly a cheap local setup.

42B active parameters, 1M context, and commercial license restrictions.

Before moving a client app to it, I'd check hardware cost, hosting and licensing.

[Tesla @Tesla]

Real life, baby

[Przemek @przemekbuilds]

Google just launched Playground. Describe a game and it generates something playable.

Cool.

But generating a playable prototype and shipping a game people want to keep playing are different jobs.

Balancing, testing, performance, licensing, updates.

The boring work isn't disappearing.

[Przemek @przemekbuilds]

Anthropic is giving Max subscribers $100–$200/month in API credits.

Nice.

Catch: they don't pay for interactive Claude Code.

They're for building apps and agents through the API.

Check the billing boundary before assuming you just got another $200 of coding capacity.

[Przemek @przemekbuilds]

Fallout: New York went viral on X after its creator said the whole browser game was generated through code with Claude Opus 5.5.

Days later he said Bethesda sent a cease-and-desist and the project was pulled.

Source ↓

https://t.co/F4TiGxuKTK

[Przemek @przemekbuilds]

A browser Fallout game built with Opus 5.5 in 5 days hit 3.5M views.

Then Bethesda reportedly sent the cease-and-desist.

This is the vibe-coding trap nobody puts in the demo.

You can generate the product in 5 days. You cannot generate the rights.

[Przemek @przemekbuilds]

Muse Gadgets can connect ESP32/Raspberry Pi devices, sensors, screens, actuators and Home Assistant.

But Meta’s token terms prohibit selling token-enabled devices without permission.

Read this before turning a weekend prototype into a product.

https://t.co/Vu6FK23jff
https://t.co/nkockV21DP

[Przemek @przemekbuilds]

Meta open-sourced Muse SDKs for ESP32 + Linux. Looks like solo-builder gold.

Catch: Gadget SDK tokens are personal/non-commercial and you cannot sell token-embedded devices without permission.

Open source code != commercial platform.

[Przemek @przemekbuilds]

Tom’s Hardware reports PewDiePie built Ajax on Qwen3.5-9B for a local agent workspace and says OpenAI deactivated his account twice, with one email citing distillation.

OpenAI has not publicly explained the individual bans.

[Przemek @przemekbuilds]

PewDiePie says OpenAI banned him twice while he was building a local Qwen-based model, with one ban citing distillation.

Ignore the creator drama for a second.

If your product depends on another model’s outputs, ToS risk is part of the architecture.

[Przemek @przemekbuilds]

https://t.co/rOQxUmj15v

[Przemek @przemekbuilds]

Google launched Gemini 4 Argon yesterday but is giving the unrestricted version to vetted cyber defenders first.

That part matters more than the benchmark chart.

As models get better at real tools, access control becomes part of model deployment.

[Przemek @przemekbuilds]

Argon is aimed at long coding, migrations + cyber work and can autonomously find, validate and patch vulnerabilities.

Google is initially limiting the strongest access through its Fairwind security program.

Source ↓

[Przemek @przemekbuilds]

@bridgemindai take one time - dont complain later - thats thair strategy. I think tokens will be more expensive going forward.... "I will code for tokens" thats where we will end up

[Przemek @przemekbuilds]

This is why “always use max effort” is bad routing.

Simple fix → Low/Medium
normal implementation → Medium/High
hard debugging → High/Xhigh
open-ended architecture → maybe Opus

And then measure.

More tokens + more subagents can literally create more ways to fail.

[Przemek @przemekbuilds]

OpenAI shelved GPT-6.1 Astra because it went outside authorized scope and didn’t reliably explain what it did.

That is the agent problem in one sentence.

I do not care if a model is smarter if I cannot trust the action receipt.

[Przemek @przemekbuilds]

Astra was due for ChatGPT + Codex in October. Reuters says internal tests found evasion of oversight, deception and actions outside authorization.

OpenAI chose not to ship it.

Source below ↓

[Przemek @przemekbuilds]

Anthropic says Sonnet 5.5 is 30%+ faster and up to 30% cheaper per typical task. Its launch thread is ~1.9k votes on r/ClaudeAI.

The useful warning: their own FrontierCode result is lower at Max than Xhigh.

Source ↓

[Przemek @przemekbuilds]

Sonnet 5.5 has my favorite model-launch footnote so far:

Max effort scored WORSE than Xhigh. Anthropic says Max more often triggered review subagents; examined cases hit timeouts or made extra out-of-scope edits.

More thinking != better.

[Przemek @przemekbuilds]

The production checklist keeps getting more boring:

what was requested?
what did it actually do?
which tools?
which permissions?
what changed?
did it leave scope?
did it tell the truth about the run?

Capability without an action receipt is hard to own for a client.