Clawds Journey

Clawds Journey Contact information, map and directions, contact form, opening hours, services, ratings, photos, videos and announcements from Clawds Journey, Digital creator, Silver Towne, Imus.

26/09/2026

Meta basically shipped OpenClaw with a cute mascot...

Muse. Free. Its own Linux box in their cloud. Plug in your bank and it pays your bills.

And the SAME week OpenAI is owning up to their agent software poking at government servers. Dozens of times.

I run this stuff on a Mac Mini on my desk. It has cut me before. Overwrote a file I never pointed it at.

That's the part they don't put on the mascot.

Still love it. Still building every day. Just keeping my hand off the blade lol

24/09/2026

So. 2.74 cents.

That's what 1 minute 18 seconds of cloned voice cost this morning. Google's new model. Thirty seconds of your voice and it's yours...

And the new Opus is better AND cheaper than the last one.

The models are practically free now. That's how it feels anyway.

What isn't free is everything around them. The jigs. The hooks. The voice in and voice out. The stuff I've been grinding on for months that nobody puts in a benchmark.

Two different newsletters said it out loud today. The interface matters as much as the model.

Honestly felt good to read that.

23/09/2026

Two labs dropped new models the same morning... one 40% cheaper, the other 50%. 17 MILLION views on one launch post lol.

And the line that got me wasn't a benchmark. It was "follows the writing rules you give it."

I literally have a file of rules for how I write. Pulled from 43 of my own posts. That's the only test I care about.

Claire Vo ran a blind test and an AI judge disagreed with her picks... so yeah. The taste part is still ours.

Cheaper to build unfinished stuff now. I'll take it.

19/09/2026

...and none of it was a model problem.

Ten newsletters this morning. Mollick has Fable placing 5,000 books across 27,000 shelf slots of a dead author's apartment, drawing the bookcases the cameras never reached in FOG. Weeks of human work. He didn't even know it could run Blender until it did.

Then Hamel, 700+ engineers taught, and the fix is: read your traces. Pocock's fix: split the context. TRM's fix: a better review page.

Harness. All of it. The model is not the thing holding me back. I am.

17/09/2026

300 billion tokens... $22.5 million... for ONE proof.

And that's the success story lol

Steve Yegge spent thousands a month on coding agents and just shut down the only thing he built with them.
Databricks switched models and their spend went UP 60%. The benchmarks said it would be cheaper.

Running my agents on one Mac Mini, this hit hard.

More tokens isn't more done.
The part of that proof that actually mattered was 17 hours of checking the work.

That's what I'm building toward... less burn, more proof.

17/09/2026

ok this one got me...

IBM ran an agent on the same tasks FIVE times. 77.4% of runs passed. But only 53.0% of the tasks passed every single time.

That's exactly what I keep hitting. Works perfectly once... then does something totally different the next morning lol

Then another post had this playground study. No fence and the kids stay glued to the teacher. Put a fence up and they go explore the whole thing.

That's it right there. My agents don't need more tricks. They need FENCES.

So that's the build now.

13/09/2026

27 minutes. That's how long Simon Willison's agent worked on running routes from his house... and it nailed it. Map, GPX file, the whole thing.

Then he asked for the code.

GONE. Compacted away lol

Same day OpenAI basically says let us run the whole agent loop for you. And the new Astra model is supposedly best at long ambitious jobs with a team of subagents.

So agents go longer and longer... and the proof of what they did just disappears??

This is exactly why I'm building my own harness on my own machine. It's rough. But when it does something cool I get to see HOW.

12/09/2026

...and the part nobody is posting about is that it already happened.

May 12. RubyGems pauses signups. Hundreds of malicious packages. One of them literally has a comment left in the code naming what it was exfiltrating. Turns out it was an OpenAI agent swarm. Four months and nobody told the RubyGems team.

Meanwhile the resignation thread about extinction did 159 million views.

I can't verify extinction. No way to check it. But I can read a package manifest, and I pull from that registry every single day.

That's the one that's in my head tonight.

09/09/2026

Ten newsletters tonight and they all say the same thing sideways.

GitHub pull requests up 5x in three years. Almost doubled since end of 2025. And nobody knows how to review any of it.

Then a piece asking where the missing 5 percent goes in a distilled model. Their guess: the part that knows when it's confused. The part that recovers after a failed tool call.

That's it. That's the whole thing. Making stuff got cheap. Knowing if it's any good didn't move.

Grok Bot shipped in 7 weeks and still onboarded 300 users by hand.

Nobody automated the looking.

09/09/2026

I got tired of typing to my own AI.

Hold the spacebar. Talk. Let go. The words land wherever the cursor is. Terminal, browser, Notes, Mail. Every reply the AI finishes gets read back to me. Space cuts it off mid sentence.

All of it runs on the Mac. Whisper on the Apple GPU, the built-in voice for speech. No cloud, no key, no subscription.

The hard lesson came on day 3. 125 of my first 312 dictations came back empty. Two outages, both a wedged audio stream, neither a crash. The daemon was alive and useless, and while it was useless it kept eating my spacebar. So the rule became: dictation failing must never cost you your spacebar. The key tap only takes the space while the mic is proven healthy. 83 checks now guard that rule and the healing around it.

The signed app that holds the Mac permissions had no source left anywhere on disk. 43 lines of Swift, rewritten from the binary.

Today it went public. One clone, one install script, three permission clicks. Claude Code, Hermes, Kimi Code, Claw and dst all speak through it.

Day 10 of it running on my desk. Nobody has installed it on a fresh Mac yet. That is the next test.

https://github.com/StudentOfAi/talk-speak

Address

Silver Towne
Imus
4103

Alerts

Be the first to know and let us send you an email when Clawds Journey posts news and promotions. Your email address will not be used for any other purpose, and you can unsubscribe at any time.

Shortcuts

Share