deepseek v4 flash jul 31 q2 on a macbook pro m5: 35-40 generation tokens/sec really good local models are here

progressive web apps are underrated.

the robot is not the factory

btw i think slopware (ai generated software where you care little about the code) is very useful for personal software, one-off software, internal tooling, prototypes, and even mvps.

Lars Grammel@lars.grammel.cloud · 2w ago

Slopware

Slopware

factory engineering

thoughts on measuring the effectiveness of a software factory there are 2 kinds of metrics: existing, standard metrics that should improve as the software factory takes on more and more work, and metrics specific to the software factory

some progress on the AI SDK software factory UI:

Lars Grammel@lars.grammel.cloud · 4w ago

Put Sol to the test and created a desktop app for managing the AI SDK software factory today.

sol medium works pretty well for most of my coding use cases so far; when does it make sense to switch to high or max? larger projects?

Why is a dedicated software factory UI needed? once you set up the initial pieces like sandboxing, several agents, and basic orchestration, you want to run a lot of work in parallel while ensuring the processes are followed (and supporting human-in-the-loop supervision and manual work)

Put Sol to the test and created a desktop app for managing the AI SDK software factory today.

what is the evidence that pre-built agent harnesses such as claude code or codex are superior (in terms of outcomes)? e.g. compared to using an open source harness like pi or opencode? or compared to creating a custom harness for a specific task/project?

starting to build a software factory for ai sdk. very exciting.

i think open source community PRs do not make sense any longer (issues are still very valuable though). why? because the cost/benefit ratio has shifted, especially as we move towards software factories for open source packages.

Use local docker sandboxes with AI SDK on Mac You can run tools against the sandbox, manipulate it directly, or use the AI SDK harness package to run e.g. Codex or Claude in the sandbox

part of my new project setup in the last couple of months: " add a REQUIREMENTS.md add instructions to AGENTS.md to update the requirements whenever there are code changes that would have requirements impact. the requirements must be accurate, consise, non-contradictory "

Make your own Mac app launcher. I published a small base project that handles the boring parts: global shortcut, launcher UI, app scanning, and launching. Fork it, then use Codex/Claude/Cursor to add custom shortcuts, prioritization, scripts, workflows, and design. github.com/lgrammel/app...

GitHub - lgrammel/app-launcher: Hackable app launcher base for MacOS

Hackable app launcher base for MacOS. Contribute to lgrammel/app-launcher development by creating an account on GitHub.

github.com

After the last Cursor update, I got fed up with their constant changes and increasing push towards their own models. I removed Cursor and switched to Codex plus VSCodium.

When I struggle to structure my thoughts about what's happening I turn to writing. Today about the recent US Anthropic ban news, what it says about power and dependency, and what it should mean for Europeans and citizens of the world. It's a long one. lucumr.pocoo.org/2026/6/13/am...

Dangerous Technology For Americans Only

AI nationalism, safety and European weakness.

lucumr.pocoo.org

experimenting with subagents for ai sdk, would love feedback

Added step performance stats to AI SDK 7 canary: - step time - llm response time - time to first token - tool execution times - tokens per second (several variants)

okay this is amazing

i created a llama.cpp provider for vercel ai sdk that runs directly in the Node process (no separate server required, it uses llama.cpp bindings) the llama.cpp provider supports reasoning, tool calling, image inputs, prompt caching (for a single conversation)

We revamped our (previously experimental) telemetry in AI SDK: - no OTel dependency in AI SDK (if you want OTel you can register our OTel integration, genai standard comaptible) - automatic emitting to the node diagnostics channel - telemetry integrations can be more custom and powerful

okay i am getting excited about local ai

switching to this on all my personal repos

built a basic tui for ai sdk 7 agents

is fine-tuning a dead end? just read some news that openai is shutting down their finetuning service and have not seen too many fine-tuned openweight models lately

Humans have been writing software for about 75 years. AIs will write software for hundreds and thousands of years. Eventually everything - hardware, OS, libraries, APIS, services, applications - will be designed to facilitate AI software development.

one of the best articles on the future of agentic coding that i've read so far: openai.com/index/harnes...

Harness engineering: leveraging Codex in an agent-first world

By Ryan Lopopolo, Member of the Technical Staff

openai.com