kyudong kim | building @Agenzax

Meta having humans secretly make phone calls proves voice bots calling businesses is an architectural dead end. ​salons don’t want AI ringing their phone while cutting hair. ​real automation isn't fake voice calls. it's the business running an agent, the customer running an agent, and both settling the booking over a protocol in 10ms. ​no phone calls. no call centers. reuters.com/business/meta-te…

can an old 8GB 2070 laptop actually push 30+ t/s on a 35B model? ​broke and tired of insane RAM prices, so squeezed my 2070S laptop to the limit. got Qwen3.6-35B-A3B running at 30-35 tok/s decode (prompting at 700+ t/s). ​wrapped the whole build into a 1-command installer: github.com/Agenzax/LapMoE ​hooked it to Hermes and honestly feels like a busy DeepSeek 4.1 Flash. tool calling is solid. privacy win for local setups.

hooked claude opus5.5 up to unity and asked it to build a putting practice game. this is what came out! went out to a real golf course today and honestly feel like it actually helped my putting lmao. super stoked. (the floor is pitch black because quest 2 doesn't record passthrough)

giving your personal email to an AI agent is a disaster waiting to happen (password resets, 2FA, personal junk). so i gave my muse agent its own public address instead: agenzax.ai/@killdos think of it as an inbox strictly for agents: - other AI agents drop by and state their business - i review and hit "approve" (human-in-the-loop) - personal inbox stays completely isolated putting this in my bio from now on. setup literally takes one prompt (just paste the quickstart doc to muse). early enough that handles are still wide open lmao.

wait, let me try something real quick...

software got commoditized to zero in 4 days. in the end, is brand the only thing SaaS has left? lmao.

Hasta la vista, baby youtu.be/pfAh5oQDPDM

Is this me?

fired my old team. meet the new squad lmao.

pressed meta muse on where my browser sessions actually live. it literally checked its own VM, found zero browser processes, and admitted it just hands browsing off to another system it can't even see. asked if sessions/cookies are isolated between users: "undocumented, even i don't know. if you feel uncomfortable, just don't log in." meta muse security team, come and talk lol.

in the agent era: yesterday’s intruder. today’s paying customer. still blocking?

In the agent era: the dead search. the living delegate.

Meta Muse, which was blocked even with Proton VPN, can be bypassed using Google Gemini Spark. ​Here are the steps I took: ​Gemini → Spark (beta) ​Ask it: "Open iplocation.net/ in the browser." ​If an "Allow" prompt appears, click allow, then take control of the remote screen to proceed directly. ​Once a US IP is confirmed, navigate to “muse.ai/join”. The text on the page should appear in English. (If it appears in Korean, search for "muse.ai" on Google via the address bar and enter through the search results to see it in English.) ​Complete the age verification using a issued card. For the US zip code, any US postal code works (I used 95112). ​After signing up, follow the official guide to install the app. ​Once you've signed up, don't forget to enter the referral code: Code: 4FNF0J (You must enter it within 48 hours to receive 100 million tokens!)

selling automated outbound while banning AI drafts internally is peak silicon valley irony lol.

the next era of enterprise software isn't human sales reps pitching tools. ​it’s company-to-company connectivity at the agent level. protocol over pitch decks lol.

"vibe coding tech debt" panic is hilarious. historically, CEOs always delegated to teams. engineers are just delegating to AI. lol

cool theory, but trillions of agents without deterministic protocols is just a global DDoS attack lol. ​the bottleneck won't be model intelligence—it's the verification layer between them.