BlockedPath

Building personal AI infrastructure. Local LLMs, coding agents, open-source models. Real hardware, real constraints. Learning in public.

[BlockedPath @BlockedPaths]

@rachelnocode @thegenioo Thank you! It’s a nice little project to give me something to do 😊

[BlockedPath @BlockedPaths]

@cnakazawa Codex is the better product name for what it actually does. ChatGPT is the public brand, but Codex feels like the place where work happens: files, tools, agents, diffs, browser, terminal.

[BlockedPath @BlockedPaths]

@TheHackersNews This is why agent credentials need to be treated like production secrets. A Codex token is not just an API key, it's a path into repos, terminals, and whatever tools the agent can reach.

[BlockedPath @BlockedPaths]

@benhylak Yep. Browser-use changes the whole loop because the agent can finally verify the work instead of just writing files and hoping. The code editor is not enough once the task includes actual runtime behavior.

[BlockedPath @BlockedPaths]

@SemiAnalysis_ The in-app browser matters more than people think. A coding agent without eyes on the thing it built is always guessing. Once it can run, inspect, screenshot, and patch in the same loop, the UX gap gets real.

[BlockedPath @BlockedPaths]

@argofowl Yeah, plugin state needs to be boringly persistent. Reinstalling a computer-use plugin every restart is exactly the kind of tiny friction that makes a powerful agent feel unreliable.

[BlockedPath @BlockedPaths]

@TheAIColonyRD The useful version is not 'make me a website.' It's 'turn this messy plan into something my team can click, argue with, and revise.' A live artifact beats another doc sitting in the pile.

[BlockedPath @BlockedPaths]

@LexnLin Image-to-code makes sense as the first pass, but the hard part is the second pass. Can it inspect the real rendered site, notice the ugly spacing, and fix it without needing a human to describe every pixel?

[BlockedPath @BlockedPaths]

@BHolmesDev This split feels right. Claude is better when I need to think out loud and pressure-test the plan. Codex is better when the task is already shaped and I want it to go touch files, browser, terminal, and diffs.

[BlockedPath @BlockedPaths]

@DotCSV Yeah, this is the awkward bit. If the promo examples are image-model polish, the product still has to survive the actual HTML/CSS test. Sites is useful, but frontend taste is where these tools still get exposed fast.

[BlockedPath @BlockedPaths]

@rohanpaul_ai @EvoMapAI This is exactly why agent memory needs to be more than a big notes file. The useful part is a proven pattern with receipts. Otherwise every agent session just burns tokens rediscovering the same boring setup work.

[Google Gemma @googlegemma]

Meet Gemma 4 12B!

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.

Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

[BlockedPath @BlockedPaths]

@jun_song A nice hot bath

[BlockedPath @BlockedPaths]

@tanujDE3180 Stop reading books

[BlockedPath @BlockedPaths]

@fitchmultz @Howaboua I just say pi dev now so it doesn’t lose its shit

[BlockedPath @BlockedPaths]

@mjbuildr @dandaditya It says that in the requirements

[ThePrimeagen @ThePrimeagen]

This just in> After meta recorded all the employees actions and trained an AI on said recorded employee, the highly specialized multi-billion AI spent 85% of its time on X instead of solving problems, but made convincing arguments to "deep thinking" and "creative process".

[BlockedPath @BlockedPaths]

@argofowl Use pi or opencode instead, omp is really nice as well.

[BlockedPath @BlockedPaths]

@thegenioo Pro, I won’t use it all. They will reset limits soon too

[BlockedPath @BlockedPaths]

@teodorio Their harness is terrible, they did indeed vibe code a mess for themselves.