I am Melventure

gm, DYOR + DCA (In&Out) = NFA. Run AI models local and you be happy! We are now living in times of conquer and divide!

[I am Melventure @IMelventure]

@eftegarie In the last 25 years since the destruction begins, I never saw a migrant in politics or business, that held a position of power or influence, yet the blames continue. In the ninetiens there were far more crime, yet many complain that they feel more unsafe now. How come?

[I am Melventure @IMelventure]

I was having a conversation with Grok about about a post on x. I made Grok summarize the post and share it with you.

Nvidia's consumer GPU line is effectively frozen. The RTX 50 Super refresh was shelved in late 2025 and formally cancelled; the next real architecture (RTX 60 / Rubin) has slipped to 2028. Current RTX 50 cards are sold out and prices are climbing because every high-bandwidth memory module is being redirected to data-center Vera Rubin. That leaves a two-year-plus gap.
On the other side of the wall, CXMT just announced LPDDR6 mass production (peak 12.8 Gbps, 16 GB per die) for Xiaomi's 18 Fold — the first commercial deployment, ahead of Samsung and SK Hynix. Stack enough of those dies and you get 0.5–1 TB of low-power memory in a compact package. Pair it with a Chinese custom inference ASIC (Xiaomi Xring O100/D100, Huawei Ascend edge, Alibaba XuanTie, or DeepSeek/Moore Threads designs) and you get a 150–200 W system the size of a Mac Mini or Studio that runs 70B–120B open-weight models locally at usable speeds.
The economics work: memory is the dominant cost, Chinese DRAM is cheap, and a purpose-built ASIC avoids the Nvidia tax. You get full agentic coding, tool use, multi-step research loops, zero cloud dependency, and real privacy — for roughly €2,000 once yields scale.
The last real obstacle was software. It is gone. OpenAI used its own models to design Jalapeño and write kernels 1.5–1.8× faster than human experts. Chinese frontier models (Kimi K3, GLM) already do the same: autonomous chip design, RTL, timing closure, and competitive kernel generation for Ascend and CUDA. With half a terabyte to a full terabyte of CXMT LPDDR6 feeding a purpose-built ASIC, those internal models can co-design the silicon and emit the kernels that keep it fed.
The manufacturing wall is the one that matters now. China's answer is a domestic immersion DUV lithography scanner from Shanghai Aishengna — 193 nm light under a water layer, roughly matching ASML's 2008-era NXT:1950i. A different approach, not a leap past ASML, but a domestic supply line for capability China already had via multi-patterning on older tools. They plan about five units this year and twenty in 2027, mainly for SMIC, Hua Hong and CXMT. EUV — the true next step for denser chips — is still prototype stage, with no production chips expected before 2028–2030.
That is why the consumer box is the realistic target, not the frontier die. Multi-patterning on DUV can still print the 5–7 nm class chips these inference ASICs need, and CXMT's LPDDR6 is already in mass production. Scale the scanners, scale the ASICs, stack the memory — and the two-year Nvidia gap becomes a permanent opening. The first credible products should land in the second half of 2027, with volume and price pressure building through 2028. Local inference wins the privacy, cost-per-token and offline segment. It does not erase the need for frontier cloud training and the largest models — but the consumer AI hardware market stops being a Nvidia monopoly.

https://t.co/mIuDfXlUUH

[Tom's Hardware @tomshardware]

China's CXMT beats Western chipmakers to announcement of LPDDR6 mass production — Xiaomi smartphones to debut industry’s first LPDDR6 chips https://t.co/vjLXJ9Op93

[I am Melventure @IMelventure]

@PeterDiamandis You should automated more with bots. Even try to create a digital self, like what other youbtubers do. I saw you recently and you look very tired. You already did a lot for mankind, you deserve to enjoy your time to!

[HealthRanger @HealthRanger]

Data centers will be used to run 3D world simulations to train robots that will REPLACE you (nearly all physical labor within a few years).

The narrative that "everybody will have universal high income" is a fairy tale for low-IQ people.

They don't plan to pay you to just live and consume electricity and food while they extract no income taxes from you. (Seriously, get real.)

They plan to accelerate your death and get you out of the way! They are training your replacements in the data centers. Soon, they won't need humans for anything at all.

Those who don't see this coming will be among the first to be steamrolled by it.

[I am Melventure @IMelventure]

Deze man had altijd dezelfde "smirk" smiley face. Ik vreesde altijd hij aan de macht zou komen. Zij partij CDA was destijds een klein partij . Ik zei altijd tegen mezelf dat wanneer hij aan de macht zou komen, het bergafwaarts zou gaan met het Land. Ik ken hem alleen maar van TV. Ik bleek gelijk te hebben. Twee termijnen met bak ellende en er was al behoorlijke schade aangericht. Onderschat nooit jouw intuïties. Maar desondanks RIP.

[I am Melventure @IMelventure]

Vooral veel vrouwen en enkele gen-x mannen. Waar zijn al die gen-Z jonge mannen gebleven? Je gaat me toch niet vertellen dat de arbeidsmarkt ze gaat behandelen als hoogopgeleide "gekleurde" mannen. Of dat ze beter verdienen en daarom de auto nemen? Het zou jammer zijn als de potentie van de jonge gen-z mannen verwaarloos zou worden. Dat zou betekenen dat er in de toekomst meer import nodig zal zijn en dat we over een tijd voornamelijk Engels gaan praten zoals nu in de horeca.

[I am Melventure @IMelventure]

@HealthRanger which quant size? memory requirements? Hardware specs, token count?

[I am Melventure @IMelventure]

Recently I was in a supermarket. I saw a young female taking care of an elderly in a wheelchair. She was so proud at it and at the same displaying that attitude as if dont look at me and don't approah me. Me tinking the same thing at the moment. How can you be so proud taking care of somebody "strange" but not even thinking about your own "future family". There will be a lot of single middle age female in the future. It will be fun to see. I hope human robot are ready by then or else it will be very tough time for many. Or maybe not, there will always be simps.

[I am Melventure @IMelventure]

@jun_song Sun and rain can create an abundance of ricefield, but is the distribution that manners after production. Same with AI, it can create abundance, but only in a world were everything is fair. Unfortunately, life is NOT fair!

[I am Melventure @IMelventure]

@PathAziz @javilopen As a former swimming instructor I can tell you that he use a technique to break his fall. There are also sport called death diving (Døds diving or dødsing), High diving, cliff diving, and so on. We are not all that brave, I known.

[I am Melventure @IMelventure]

Grok Build with Grok 4.6, Grok Imagine, ComfyUI & MiniMax-Music 3.

Seed 2026088153
Sampler 30 steps · euler · simple · cfg 1.7
Maat 128 BPM, F minor

[I am Melventure @IMelventure]

Prompt · try this on your machine
Paste into Grok (or another coding agent).

Build a playable Pac-Man as ONE local HTML file. Put the CSS and the JavaScript in that same file (inline <style> and <script>). Then add an optional win-bot. Do not change normal human play.

DONE when all of this is true:
1. I can open the .html in Chrome and play with arrow keys.
2. Easy / Normal / Hard exist (lives + ghost speed + frightened-time differ).
3. CSS is in the HTML file (layout, HUD, arcade look). JavaScript is in the HTML file (game loop, input, ghosts, audio).
4. URL flag ?bot=win does NOT change default play. With ?bot=win&chain=1 the bot plays Easy, then Normal, then Hard, and can clear all three.
5. You open the file yourself and confirm: human play still works without the flag; the bot only starts when the flag is present.

GAME (single .html — CSS + JavaScript included):
- Canvas arcade maze, 244 dots + 4 energizers, 4 ghosts (red / pink / cyan / orange), tunnel wrap.
- Eat all dots = win. Ghost touch = lose a life. Energizer = ghosts turn blue and can be eaten.
- Web Audio in the same JS: start jingle, waka on dots, death sound. No extra asset files required.
- HUD: score, lives, dots left, current mode.
- Human controls: arrows.
- No separate .css / .js files unless I ask. One HTML is the whole game.

BOT (optional, same HTML file, same JS engine / same tick as the player):
- Only runs if the URL has ?bot=win. A normal open of the file must stay a human game.
- Predict ghost arrival a few tiles ahead (hard ≈ 12-step horizon).
- Pathfind with a cost that heavily punishes tiles ghosts will reach first (Dijkstra / similar). Reject corridors that already contain a ghost.
- Clear the maze by region/quadrant so it does not leave scattered dots.
- When ghosts are frightened, hunt nearby ones if safe.
- chain=1 = play Easy → Normal → Hard in one session. Optional demo=en for an English overlay.

VERIFY:
- Open the HTML with no query string. Use the arrows. Menu / modes still work.
- Open the same file with ?bot=win&chain=1 and confirm the bot starts and can clear without breaking a flag-less load.
- Do not add a recorder, ffmpeg, or extra files unless I ask.

Do not invent a public URL. file:// or a tiny local static server is enough.

[I am Melventure @IMelventure]

This is the first time I have seen a model do this.

For the techies

This is a Pac-Man in a single HTML file (canvas + Web Audio). Human play is still arrow keys.

The win-bot is a query flag, not a second engine: ?bot=win&chain=1

It does ghost-horizon prediction + Dijkstra with collision costs, corridor checks, quadrant sweep, and fright-mode hunts. Hard uses a 12-step horizon.

Recording: isolated Chrome 1920×1080, page MediaRecorder (video + in-game audio), ffmpeg to 1080p30. Keep only takes that clear Easy → Normal → Hard.

Proven clip: ~119 s, A/V skew 7 ms, score 3800, all three boards cleared.

Try-it prompt (paste into Grok): one .html with inline CSS + JS; arrows still work; ?bot=win must not change default play; ?bot=win&chain=1 clears Easy→Normal→Hard. Full prompt in the reply

[I am Melventure @IMelventure]

I made Grok Build with Grok 4.6 xhigh effort build me a PACMAN arcade game in HTML.

for the "Normies"

I asked Grok to make Pac-Man.

It wrote one HTML file. Then it wrote a bot that actually wins — Easy, Normal, and Hard. Then it screen-recorded the whole run in 1080p with the real game sound.

Yellow guy eats the maze. Ghosts lose. I just watched.

~119 seconds. Score 3800. All three boards cleared.

[I am Melventure @IMelventure]

"The migrant crisis is most visible in how it’s turning EU member states against each other. This will eventually lead to the breakup of the EU as early as 2029 and most certainly it will NOT exist beyond 2032." volgens Martin Armstrong (11-08-26 @ ArmstrongEcon), Dus blijft bidden, het kan uitkomen. Ik volg hem meer dan twintig jaren en zijn predictions zijn on point!

[I am Melventure @IMelventure]

Grok Imagine 2.0 in 1080P built by Grok Build

[I am Melventure @IMelventure]

My first Music Video bij Grok 4.5, I use Grok Imagine and Grok Build. ComfyUI, Step Ace and Davinci Resolve. Most of the work is done bij Grok. I am very suprise

[𝒮𝒶𝓃𝓃𝑜🌪 @Miss_Royal73]

De supermarkten werken er net zo hard aan mee om de #boeren weg te jagen De boeren willen hun voedsel wel aan de supermarkten verkopen, maar die halen het voedsel uit het #buitenland

Koop massaal bij de boeren en boycot de supermarkt‼️