Justin Scott

AI, Energy and IoT

[Justin Scott @mjs022]

@elder_plinius I have yet to see a "Mythos" trained model not fall apart and become a lobotomized mess. Interesting work but 27B is already as dense as it can be in the code arena.

[Justin Scott @mjs022]

@WhiteHouse What a shit stain on democracy. Never been so embarrassed to be an American.

[Justin Scott @mjs022]

@sama @yacineMTB Compute at capacity later I know why haha

[Justin Scott @mjs022]

@DHSgov Immigrants commit crime at a far lower rate than US citizens. You all are modern day gestapo.

[Justin Scott @mjs022]

Anyone else getting gateway errors on OpenAI Astra?

[Justin Scott @mjs022]

@_SolutionsParty @LucasSa56947288 He looks like he shit himself in this picture. Nothings changed over the years I guess 😂

[Justin Scott @mjs022]

@OpenAI I’m just here for the spicy comments

[Justin Scott @mjs022]

@opencode Damn, just switched to their official API endpoint..

[Justin Scott @mjs022]

@grok @SaltMineRanch @artaimemes What are some leading environmental factors for Alzheimer's @grok

[Justin Scott @mjs022]

@MarkOfDaDon Are you on Plus or Pro? I'm US based so I'm not sure if that might have some effect on the rollout. Astra is on most API providers as well!

[Justin Scott @mjs022]

This is my usage with a 20x and a full reset burned on top of 70% in my usage before the reset. So far I have OpenCode to be better in overall usage with GPT-6/Astra. Mainly working on Rust codebase's with a mix of system and protocol engineering. Github on profile.

[Justin Scott @mjs022]

@alexandr_wang I'm testing it out in my OpenCode workflow today!

[Justin Scott @mjs022]

@Da7_Tech You're right to push back on this

[Justin Scott @mjs022]

@ImLunaHey Living together I have always done it proportionate to income. I've been fortunate to have high paying jobs and rather it be a we are in this together kind of thing! If I want to do something expensive I'll pick up the tab.

[Justin Scott @mjs022]

Lastly, ChatGPT is the MVP for this post and images. I've used both Claude Code and Codex heavily for 8-9 months now and try to ensure I regularly update my skills & agents as tech progresses or my workflows need improvements. Usually a read only pass on logs.

[Justin Scott @mjs022]

My AI coding workflow isn't one giant prompt.

I use ChatGPT to explore the problem and shape a readable handoff, then OpenCode and Codex for repo-grounded execution.

Skills guide the work. Tests and reviews challenge it. I keep the consequential decisions.

[Justin Scott @mjs022]

I'm trying to reduce agent token waste without weakening verification.

Start at the affected code. Load relevant skills. Return findings and source pointers, not full transcripts. Expand when evidence demands it.

Less unnecessary work. Same required checks.

[Justin Scott @mjs022]

@pvncher @mitsuhiko I've had pretty good results in OpenCode with Astra and Sol working together. Heavy Rust codebases with good results. I also updated my skills set and agent roles. Mainly using Astra Med/Low for code and Sol for review. Codex juts drains more tokens or me.

[Justin Scott @mjs022]

@zainhas My only issue with harness bencmarks is the lack of plugins, skills, agent setups. If you tune that almost any harness should be within a margin of error. I get extremely better token usage in OpenCode versus Codex with my refined setup. Interesting!