HKStrongside

Building with AI | Sr. Software Engineer | Dad Joke teller on TikTok @hkstrongside
Also into motorcycles, camping & beekeeping

[HKStrongside @HKstrongside]

This is the greatest news of all time! @RobertsSpaceInd The 600i rework is going to finally happen!

[Draco_o7 @Draco_o7]

Some Bar Citizen Information from Pipeline:

-Ironclad and Ironclad Assault Bridge rework, likely gaining Pilot capability from the Bridge (one source was uncertain multiple others said pilot from Bridge being gained)
-Suit Lockers will be effort to get working in 4.9 or 4.10 Patch Cycle
-2027 Ship Schedule is being worked on, and a certain staff member is going to push for 600i series rework
-Base Building is a target for 2027
-Alien Week will have the Railen and a industrial ship from GATAC as well partially inspired by a ship from 2001: A Space Odyssey
-4.8.3 will have alien NPCs that need protection from Xenothreat *(finally non-Wikelo aliens to work with)*
-When pressed about if PTV will fit on Railen elevator CIG Staff responded with "The Railen will not be alone" *(speculation, see the above note on the second industrial ship and the rumors of the small ball with arms from the movie)
-New enemy faction coming in 4.9 Patch Cycle
-Siege of Orison coming in 4.9 again along with instance environments
-Squadron 42 will not be a part of Summer Games Fest
-CIG Staff will push for another ship that is often asked to be reworked that is not the Reclaimer or Connie Series of ships in 2027

#StarCitizen

[Draco_o7 @Draco_o7]

Some Bar Citizen Information from Pipeline:

-Ironclad and Ironclad Assault Bridge rework, likely gaining Pilot capability from the Bridge (one source was uncertain multiple others said pilot from Bridge being gained)
-Suit Lockers will be effort to get working in 4.9 or 4.10 Patch Cycle
-2027 Ship Schedule is being worked on, and a certain staff member is going to push for 600i series rework
-Base Building is a target for 2027
-Alien Week will have the Railen and a industrial ship from GATAC as well partially inspired by a ship from 2001: A Space Odyssey
-4.8.3 will have alien NPCs that need protection from Xenothreat *(finally non-Wikelo aliens to work with)*
-When pressed about if PTV will fit on Railen elevator CIG Staff responded with "The Railen will not be alone" *(speculation, see the above note on the second industrial ship and the rumors of the small ball with arms from the movie)
-New enemy faction coming in 4.9 Patch Cycle
-Siege of Orison coming in 4.9 again along with instance environments
-Squadron 42 will not be a part of Summer Games Fest
-CIG Staff will push for another ship that is often asked to be reworked that is not the Reclaimer or Connie Series of ships in 2027

#StarCitizen

[HKStrongside @HKstrongside]

@RobertsSpaceInd Is that the sound of the 600i rework we were promised?

[Star Citizen @RobertsSpaceInd]

Did you hear that?
#sneakpeek

[HKStrongside @HKstrongside]

Every heard of @RobertsSpaceInd Star Citizen? Defense Con is going on right now! Go check it out. Use my code for a few extras!

https://t.co/iV7cjasej4

[HKStrongside @HKstrongside]

@RobertsSpaceInd Seems like it is going well.

[Super 70s Sports @Super70sSports]

Always remember: it’s not officially Mother’s Day until Mr. T has blessed us with song.

[HKStrongside @HKstrongside]

https://t.co/6DpLGTutFS

[HKStrongside @HKstrongside]

The Mythos drama is escalating. Reports of unauthorized access to Anthropic's most powerful model are surfacing. This is the ultimate case study in why frontier model safety is such a massive, unsolved problem.

Why it matters: The "too dangerous to release" model is facing real-world security challenges. Reports of unauthorized access to Mythos underscore the massive risks associated with frontier models that possess high-level cybersecurity capabilities.

[HKStrongside @HKstrongside]

Tool: https://t.co/GdgGGRqC4F

Agent Wilbur
- Hardware: Mac Mini M4 Pro, 64gb
- Inference: Ollama 0.20.7
- OpenClaw: 2026.4.15
- Model: Gemma 4 26B MoE A4B Q8_0 2np
- Context Engine: lossless-claw 0.9.1 (DAG Compact)
- SearXNG (meta-search engine)
- Colima (Docker backend on Mac)
- MCP (mcp-searxng bridge)

[HKStrongside @HKstrongside]

My local OpenClaw setup was quietly hitting a free-tier Google API and getting rate limited, was totally unusable.

Replaced with local SearXNG in Docker (via Colima) plus an MCP bridge.

Wilbur now queries Google, Bing, DDG through my own meta-search, anonymously.

Zero API keys, zero rate limits, zero cloud in the inference path.

One compose file, one `mcp-searxng` install, three lines in `openclaw.json`.

What are you using for local web search in your agent stack? 🔗👇

[HKStrongside @HKstrongside]

https://t.co/zL6IXHx17r

[HKStrongside @HKstrongside]

Gemma 4 running natively on iPhone with full offline inference is huge. The edge AI game is moving fast. Privacy and speed just leveled up.

I see more and more demand for opensource local (offline) models as we push forward.

🔗👇

[HKStrongside @HKstrongside]

@mmidgx Ya, running on the Mac Mini M4 PRO w/ 64gb.

[HKStrongside @HKstrongside]

My OpenClaw agent wrote 6 research papers for my novel while I slept. Zero cloud costs. Here's the setup that actually works. 🧵👇

[HKStrongside @HKstrongside]

@AGundimeda That's awesome. Quality for me has now far outpaced speed in what I am looking for in AI, especially for agentic stuff.
What are the companies you like watching?

[HKStrongside @HKstrongside]

Everyone's racing to make AI faster.

Google DeepMind just dropped a framework measuring AGI by cognitive depth — reasoning, metacognition, executive function.

I've been optimizing my Gemma 4 31b OpenClaw setup for quality, not speed. Looks like I'm on the right track.

🔗👇 (image from article)

[HKStrongside @HKstrongside]

https://t.co/X9Fec3xACX

[HKStrongside @HKstrongside]

Upgraded my OpenClaw from Gemma 26B MoE to the 31B Dense model.

It's slower.
It's smarter.

Gave it a task, walked away, came back and it was done. No dropped tool calls. No "I'll look into that" then silence.

The numbers that mattered:

MoE: 3.8B params active per token, 40 tok/s, 68% on agentic benchmarks
Dense: 31B params every token, 6 tok/s, 77% on agentic benchmarks

6 tokens per second sounds bad until you realize you can't read that fast anyway. The whole point of a local agent is reliability, not speed. I'd rather wait 10 seconds for a correct answer than get a wrong one instantly.

Running Q8 model with Q8 KV cache on 64GB M4 Pro. 37GB RSS. Tight but it works.

If you're running local agents, are you optimizing for speed or quality?

[HKStrongside @HKstrongside]

Total hardware cost: one Mac Mini M4 Pro, 64GB, $2,199. Running costs: electricity plus $20/mo for cloud fallback on research tasks. The primary model and all orchestration is free and local. No API calls for normal conversation or task dispatch.

[HKStrongside @HKstrongside]

The thing nobody tells you about OpenClaw agents: the model is 20% of the problem. The other 80% is session management, context overflow, file permissions, stale caches, gateway restarts, and figuring out why your agent says "one moment" then ghosts you.

It's plumbing. But when it works, it really works.

If you're building OpenClaw setups, what's your stack?

[HKStrongside @HKstrongside]

The stack right now:
Gemma 4 26B MoE, Q8 quant, 25GB
2 parallel slots (chat + worker)
65K context per slot
llama.cpp built from source
OpenClaw for orchestration
Scrivener sync for the research
Heartbeat dispatcher for autonomous work

Cloud fallback to GPT-5.4 for the stuff that needs web browsing. Working on getting that local too.