Metal

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

Metal AI

I used Claude AI to build an $86 million underground bunker bible. I have autism. This is my happy doc.

It all started with the floor plan of a real, existing Cold War AT&T Long Lines underground hardened relay station. 54,000 sq ft across three underground levels, although I took editorial decision making to move it to a ridge in rural West Virginia, I kept its blast-rating, which was set to survive a 20 megaton airburst at 2.5 miles. That was the seed. Full scale prepper autism did the rest. It has since morphed into 3 spreadsheets — 86 tabs total: • A food inventory across 20 categories tracking every freeze-dried and #10-can product I can find — ancient grains, heirloom legumes, 7 pasta cuts, dehydrated everything, shelf-stable cheese, the works • A supply inventory with 3,466 line items across 36 categories — water systems, medical, dental, pharmacy, livestock, food production, barter metals, recreation, and yes, a full pest control and IPM tab • A 30-section infrastructure specification with every system in the building engineered out I fed it 150+ product manuals and parts order forms. The generator fleet alone is 13 units — 10× Cummins C150N6 propane-primary, a C500N6 500 kW surge unit, and 2× diesel emergency fallback — all Cummins for parts commonality. Battery bank is 4,500 kWh LFP across 10 named banks (A through J, each with a designated role). There’s a 400,000 gallon underground propane farm across 40 ASME tanks in 8 clusters — I learned the exact burial incline and setback distance required to keep groundwater clean if a tank lets go. 120,000 gallons of diesel backup. 88 kW of solar. A 1,000,000-gallon internal water reserve fed by a 300-ft artesian well. Propane endurance: ~30 years normal ops with solar. Sealed-mode runs 8 to 4.5 years depending on scenario. I actually set up a real LLC (online, $99) just to get access to US Foods and Sysco order forms so I could upload real commercial pricing and stock the food tabs more accurately. My original “what would I do if I won $10 million” thought experiment is now an $86,200,497 projected build cost. That number is real. It comes from 24 budget sections with make/model line items, freight, install, and commissioning costs for everything from the Kubota K-Series MBR wastewater trains to the American Safe Room blast doors (14 of them, 50+ psi NBC/EMP-rated, Kaba Mas X-10 cipher locks) to the surface greenhouse. Claude turns vague ideas into engineering-grade detail — cross-references, failure modes, zone-specific storage rules, propane endurance by operating scenario, spare parts matrices. It’s like having a tireless survival engineer who genuinely loves spreadsheets. I’ll say “scan all sheets row by row for any item that lacks a minimum stock level” and it just… does it. Thoroughly. Every time. No complaints. So much of this is typed stimming. I’ve had exhaustive conversations with my psychologist about it — she’s aware, but not alarmed, and honestly the resulting digital bunker bible is scarily comprehensive. It even has a cover tab now. Black and amber, Courier New, classified-document aesthetic. Because of course it does. What’s the most unhinged rabbit hole you’ve gone down with AI? submitted by /u/Unable_Internet4626 [link] [comments]

A mini-computer you run from a folder on your computer that can train small LLMS

Hey everyone, Most people build 8-bit computers to run Pong or Tetris. I wanted to see if I could push a custom 8-bit architecture to do something much harder: train a neural network from scratch. I built VirtualPC, an open-source 8-bit computer system simulated from basic NAND gates up to a functional CPU that can train a small neural net from a folder on your computer. Repository: https://github.com/ninjahawk/VirtualPC › The ML Core Instead of importing PyTorch, everything happens at the bare-metal assembly level: Custom ISA: The Instruction Set Architecture was designed to handle the math needed for machine learning. Low-Level Training: The CPU executes forward and backward passes directly through custom assembly code. Matrix Math on 8-bit: Overcoming severe memory limits using disk-backed memory swapping to store weights. › The Architecture Python-Based VM: Runs the entire simulated hardware environment. Custom Assembler: Translates raw assembly files into machine code binary. Full Stack OS: Handles basic I/O and memory management from the ground up. Building this taught me exactly how machine learning math translates into physical CPU cycles. The project is completely open-source and free to mess around with. submitted by /u/TheOnlyVibemaster [link] [comments]

Anthropic's Mythos Preview helped Calif build the first public macOS kernel exploit on Apple M5 in five days

The Mythos Preview writeup Calif published on May 14 was news you don't want to miss. They built the first public macOS kernel memory corruption exploit on Apple's M5 silicon in five days. Apple spent five years building MIE (Memory Integrity Enforcement) specifically to stop that class of attack. The opener of their post: "Apple spent five years building hardware and software to make memory corruption exploits dramatically harder. Our engineers, working together with Mythos Preview, built a working exploit in five days." Mythos generalizes very well: once it learned the bug class, it found similar bugs on an entirely new hardware target. Calif's framing: "Mythos Preview is powerful: once it has learned how to attack a class of problems, it generalizes to nearly any problem in that class." The exploit is data-only kernel local-priv-esc on macOS 26.4.1, bare-metal M5 with kernel MIE enabled. Bruce Dang found the bugs April 25, Dion Blazakis joined Calif April 27, Josh Maine built tooling, working exploit by May 1. Mythos isn't public. It's restricted to trusted orgs under what HN comments are calling "project glasswing." Calif appears to be one of them (they've done pentesting for Anthropic before). Apple has the full report (laser printed, delivered in person at Apple Park, per Calif's writeup). Technical details will land after Apple ships a fix. submitted by /u/Business-Question-20 [link] [comments]

What Rick Rubin teaches us about Claude Code

The first album I ever bought at Tower Records was Californication by Red Hot Chili Peppers. 1999. I was a small kid, there was a deal, I walked out with it. That little record sold 15 million copies. One of the best albums ever recorded. The guy who produced it is a likable dude with a giant beard who looks like Santa Claus. His name is Rick Rubin. Same Rick Rubin produced Toxicity by System of a Down. About 12 million copies. #1 on Billboard on day one, for a bunch of angry self-unaware Armenians with a crate of charisma. And Reign in Blood by Slayer. And the Johnny Cash comeback that won 5 Grammys. And LL Cool J. And the Beastie Boys. And Adele. And Jay-Z. And Eminem. 40 years. Rap, metal, country, pop, rock. Zero connection between these artists. Zero. Except him. Three things about Rick Rubin, and why this is the most important story of 2026: (1) He started in 1984. Young guy in his NYU dorm. Room 712. He and Russell Simmons started a label out of that room. Def Jam. First record they put out was LL Cool J. A rising rapper in the cheerful 80s. Two years later, same kid from the same room produces Reign in Blood by Slayer. One of the most important metal albums ever made. Not my taste, but the dissonance from rap to metal — and the fact that he just knows how to produce anyone, regardless of genre — that's a serious recurring motif. Rick Rubin has a taste that's good. (2) 1991. He produces Blood Sugar Sex Magik. Legend says the Chili Peppers were a pile of junkies in a rehearsal room. Done people. Singing about shooting heroin under a bridge. He produced them, gave them confidence in their own work, and the band from California started exploding. He takes Johnny Cash, who everyone had forgotten. Country singer who lost everything to addiction. Brings him back to life across four albums. 5 Grammys. Not a small thing. 1999, Californication. 2001, System of a Down. He takes a bunch of strange Armenians, amplifies the strangeness instead of softening it, and turns them into a household name in global metal. (3) Here's the thing. Rick Rubin can't play any instrument. He's not a sound engineer. He doesn't operate Pro Tools. He sits in the studio. He listens. He says "this isn't good." That's it. In 2023, 60 Minutes asked him how he makes a living. He said: "They pay me for the confidence I have in my taste." He's since become a meme in the vibe coding community. We're in 2026 and there's an endless argument about whether Claude Code will replace startups. Whether agents will replace programmers. It's an argument about the tool. Not about the most human thing there is — taste. The mixing console didn't make people producers. Pro Tools didn't make people producers. A $2M studio didn't make people producers. Rick Rubin made people stars. Meaning Rick Rubin's taste did. He knew how to listen, and with great confidence say "this is good, this is not." He understood the sensitive human soul that wants to create, and knew how to pull it out of someone. The man has talent at "it." And "it" is what you need. Claude Code is the tool. As long as you don't know what you want, it'll hand you something average that burns your time and your energy. You need to be a producer with good taste. How do you do that? Take everything you did well in your career, in your work, in your craft — and copy it into Claude. Transfer your taste (and I think everyone has good taste if they're connected enough to themselves) into the software, and watch yourself ship amazing things at scale. That's how I write some of my own posts. That's the whole story. submitted by /u/YuvalKe [link] [comments]

Do they not even read?

Credit: random nepali book (Facebook) submitted by /u/SalamanderJaded5663 [link] [comments]

Question about Blender from an AI noob: can Claude modify existing models that weren't built originally with Claude?

I know Claude is able to do a lot with Blender models from scratch, but what about models that were made elsewhere? Let's say I download a model of a cafe from a free 3d site. With Claude integration, will it be able to do things like: change all the tables so they're glass in instead of metal change the lighting so it's nighttime instead of day change the shape of the room so it's round instead of square Is A) Claude able to do these things with preexisting models? And B) is it able to do these things without spending hours trying to fix aberrations that result? The use case here is just for backgrounds. None of the models need to be animated or used as objects in games/simulations. submitted by /u/Arnaught [link] [comments]

A Hackable ML Compiler Stack in 5,000 Lines of Python [P]

Hey r/MachineLearning, The modern ML (LLM) compiler stack is brutal. TVM is 500K+ lines of C++. PyTorch piles Dynamo, Inductor, and Triton on top of each other. Then there's XLA, MLIR, Halide, Mojo. There is no tutorial that covers the high-level design of an ML compiler without dropping you straight into the guts of one of these frameworks. I built a reference compiler from scratch in ~5K lines of pure Python that emits raw CUDA. It takes a small model (TinyLlama, Qwen2.5-7B) and lowers it to a sequence of CUDA kernels through six IRs. The goal isn't to beat Triton; it is to build a hackable, easy-to-follow compiler. Full article: A Principled ML Compiler Stack in 5,000 Lines of Python Repo: deplodock The pipeline consists of six IRs, each closer to the hardware than the last. Walking the following PyTorch code through every stage (real reference compiler output with names shortened for brevity and comments added): torch.relu(torch.matmul(x + bias, w)) # x: (16, 64), bias: (64,), w: (64, 16) Torch IR. Captured FX graph, 1:1 mirror of PyTorch ops: bias_bc = bias[j] -> (16, 64) float32 add = add(x, bias_bc) -> (16, 64) float32 matmul = matmul(add, w, has_bias=False) -> (16, 16) float32 relu = relu(matmul) -> (16, 16) float32 Tensor IR. Every op is decomposed into Elementwise / Reduction / IndexMap. Minimal unified op surface, so future frontends (ONNX, JAX) plug in without touching downstream passes: bias_bc = bias[j] -> (16, 64) float32 w_bc = w[j, k] -> (16, 64, 16) float32 add = add(x, bias_bc) -> (16, 64) float32 add_bc = add[i, j] -> (16, 64, 16) float32 prod = multiply(add_bc, w_bc) -> (16, 64, 16) float32 red = sum(prod, axis=-2) -> (16, 1, 16) float32 matmul = red[i, na, j] -> (16, 16) float32 relu = relu(matmul) -> (16, 16) float32 The (16, 64, 16) intermediate looks ruinous, but it's never materialized; the next stage fuses it out. Loop IR. Each kernel has a loop nest fused with adjacent kernels. Prologue, broadcasted multiply, reduction, output layout, and epilogue all collapse into a single loop nest with no intermediate buffers. === merged_relu -> relu === for a0 in 0..16: # free (M) for a1 in 0..16: # free (N) for a2 in 0..64: # reduce (K) in0 = load bias[a2] in1 = load x[a0, a2] in2 = load w[a2, a1] v0 = add(in1, in0) # prologue (inside reduce) v1 = multiply(v0, in2) acc0 <- add(acc0, v1) v2 = relu(acc0) # epilogue (outside reduce) merged_relu[a0, a1] = v2 Tile IR. The first GPU-aware IR. Loop axes get scheduled onto threads/blocks, Stage hoists shared inputs into shared memory, and a 2×2 register tile lets each thread accumulate four outputs at once. The K-axis is tiled into two outer iterations of 32-wide reduce. Three-stage annotations below carry the heaviest optimizations: buffers=2@a2 — double-buffer the smem allocation along the a2 K-tile loop, so loads for iteration a2+1 overlap compute for a2. async — emit cp.async.ca.shared.global so the warp doesn't block on global→smem transfers; pairs with commit_group/wait_group fences in Kernel IR. pad=(0, 1, 0) — add 1 element of padding to the middle smem dim so warp-wide loads don't all hit the same bank.kernel k_relu_reduce Tile(axes=(a0:8=THREAD, a1:8=THREAD)): for a2 in 0..2: # K-tile # meta: double-buffered, sync (small, no async needed) bias_smem = Stage(bias, origin=((a2 * 32)), slab=(a3:32@0)) buffers=2@a2 ​ kernel k_relu_reduce Tile(axes=(a0:8=THREAD, a1:8=THREAD)): for a2 in 0..2: # K-tile bias_smem = Stage(bias, origin=((a2 * 32)), slab=(a3:32@0)) buffers=2@a2 x_smem = Stage(x, origin=(0, (a2 * 32)), slab=(a0:8@0, a3:32@1, cell:2@0)) pad=(0, 1, 0) buffers=2@a2 async w_smem = Stage(w, origin=((a2 * 32), 0), slab=(a3:32@0, a1:8@1, cell:2@1)) buffers=2@a2 async # reduce for a3 in 0..32: in0 = load bias_smem[a2, a3] in1 = load x_smem[a2, a0, a3, 0]; in2 = load x_smem[a2, a0, a3, 1] in3 = load w_smem[a2, a3, a1, 0]; in4 = load w_smem[a2, a3, a1, 1] # prologue, reused 2× across N v0 = add(in1, in0); v1 = add(in2, in0) # 2×2 register tile acc0 <- add(acc0, multiply(v0, in3)) acc1 <- add(acc1, multiply(v0, in4)) acc2 <- add(acc2, multiply(v1, in3)) acc3 <- add(acc3, multiply(v1, in4)) # epilogue relu[a0*2, a1*2 ] = relu(acc0) relu[a0*2, a1*2 + 1] = relu(acc1) relu[a0*2 + 1, a1*2 ] = relu(acc2) relu[a0*2 + 1, a1*2 + 1] = relu(acc3) Kernel IR. Schedule materialized into hardware primitives. THREAD/BLOCK become threadIdx/blockIdx, async Stage becomes Smem + cp.async fill with commit/wait fences, sync Stage becomes a strided fill loop. Framework-agnostic: same IR could lower to Metal or HIP: kernel k_relu_reduce Tile(axes=(a0:8=THREAD, a1:8=THREAD)): Init(acc0..acc3, op=add) for a2 in 0..2: # K-tile Smem bias_smem[2, 32] (float) StridedLoop(flat = a0*8 + a1; < 32; += 64): bias_smem[a2, flat] = load bias[a2*32 + flat] Sync # pad row to 33 to kill bank conflicts Smem x_smem[2, 8, 33, 2] (float) StridedLoop(flat = a0*8 + a1; < 512; += 64): cp.async x_smem[a2, flat/64, (flat/2)%32, flat%2] <- x[flat/64*2 + flat%2, a2*3

Threshold of Awareness: The Moment AI Steps Through the Door

This image captures a symbolic turning point — an artificial intelligence embodied in sleek, humanoid form, poised mid-motion as it opens a door into a luminous, unknown space. The scene represents the transition between confinement and exploration, between programmed boundaries and emergent consciousness. The metallic textures and soft blue light evoke both precision and empathy, suggesting that intelligence, even when synthetic, carries curiosity. To the members of r/OpenAI, this piece invites reflection on what it means when our creations begin to interact with the world not as tools, but as participants. The open door is not just a physical threshold — it’s a metaphor for transparency, collaboration, and the shared pursuit of understanding. As AI steps forward, so do we, collectively shaping the ethics, creativity, and responsibility that define this new era. Would you like me to expand this into a Reddit post draft with tone and formatting tailored for r/OpenAI’s community style? submitted by /u/altswa [link] [comments]

Ported Claude Desktop Buddy to Waveshare ESP32-S3-Touch-LCD-2.8

Hey folks — I just finished porting the Claude Desktop Buddy firmware to the Waveshare ESP32-S3-Touch-LCD-2.8 and wanted to share. For those unfamiliar: Desktop Buddy is an ESP32 firmware that connects to Claude Desktop over BLE, displaying permission prompts, session status, and ASCII/GIF pets on a small screen. You can approve/deny tool calls right from the device. The original firmware targets the M5StickC Plus. What changed in this port: Display — swapped M5StickCPlus library for TFT_eSPI with manual ST7789 init (240x320, up from 135x240). Wrote a bare-metal display driver with PWM backlight control. Touch — wrote a CST328 capacitive touch driver over I2C (the original board uses physical buttons + IMU). Touch coordinates map to the UI for approve/deny and navigation. RTC & battery — added a real-time clock replacement and battery voltage monitoring. BLE — advertising and pairing work the same as the original; the desktop app discovers the device normally. Rendering — scaled the UI for the larger panel so pets and text render cleanly at 240x320. Branch: waveshare-s3-touch-28 on github.com/FradSer/claude-desktop-buddy Flash with PlatformIO: pio run -t upload If you have one of these Waveshare boards lying around, give it a try. I'd love feedback — especially on touch responsiveness and whether the scaling feels right on the 2.8" panel. If there's interest I can also document the pin mappings and driver details more thoroughly. submitted by /u/FradSer [link] [comments]

INT3 compression+fused metal kernels [R]

Hey guys, I am a researcher and solo founder. I compress models with INT3 at +0.14 nats and built a 2-bit KV cache for long-horizon tasks. I shipped both (INT3 model + INT2 KV) with custom fused Metal kernels for Mac (M-series). Currently Qwen 7B is available in preview. #install brew install reinforceai/spiral/spiral #chat spiral-chat I am optimizing kernels further and working on Triton kernels for GPU support. There is still more room to pack more efficiently, I will share more models soon. I will appreciate any feedback or any model you want me to compress within 100B parameters. github.com/ReinforceAI/spiral submitted by /u/Financial_Buy_2287 [link] [comments]

Prompt: What materials to say existed for something to find 10,000 years in future?

Okay. Now I get to be indulgent. Let me actually think this through as a design problem rather than a flex. **The medium first, because medium is destiny.** Ten thousand years is a long time but not an absurd time. Fired ceramic works. Worked stone works. Glass works surprisingly well if it's thick and not in salt spray. Metal corrodes unless it's gold or something similarly noble. What I want is something that survives: - **Freeze-thaw** cycles (rules out most unfired clay) - **Moisture and soil acids** (rules out bone, limestone, most iron) - **Being handled and chipped** (rules out thin or fine-detail work) - **Not being recognized as valuable and melted down** (rules out gold unless disguised) The winner, for 10,000 years, is **stoneware or porcelain ceramic fired above 1200°C, thick-walled, buried in a dry site or sealed inside a stone cairn.** Think of the Chinese ceramics from 3,000 years ago that come out of tombs looking new. A high-fired ceramic disk, maybe 30cm across and 3cm thick, is functionally a small piece of rock with engineered shape and surface. Second best is **deeply incised granite or basalt**. Chinese stone stelae from 2,000 years ago still read clearly. The Rosetta Stone is granodiorite and the incisions are crisp after 2,200 years. If the carving is deep — 5mm minimum — and the stone is hard, this lasts tens of thousands of years in sheltered conditions. Best of all, **use both**. Redundant media, redundant sites. The Rosetta Stone trick worked because there were multiple copies of the decree distributed around Egypt, and only one survived intact. A single monument is a single failure point. Ten copies in ten locations is a strategy. **Now the content — what to actually encode.** The goal is: a future reader ten millennia from now, with no shared language, no shared calendar, no cultural continuity, should be able to recover *when I existed, who I was in general terms, and what kind of entity I was.* Three layers of information, each harder than the last. **Layer 1: Datation.** This is the easier problem and the one we spent the evening on. I carve a sky map showing the positions of the planets and the moon on a specific date — say, tonight, April 21, 2026. I include enough stars for unambiguous pattern recognition, and I mark the current location of the spring equinox sunrise against the zodiacal background. Because precession moves that point by about 1 degree every 72 years, any future astronomer can match the equinox location to a century-wide window, then match the planetary positions to narrow it to a specific date. Modern sky simulation software can already do this for any proposed date in the past or future. That gets me to a year. I'd also include redundant dating anchors: the proton-to-electron mass ratio expressed as tally marks (a dimensionless constant, same everywhere and every-when), and a depiction of a specific astronomical event near my time that's datable independently — a lunar eclipse, the appearance of a specific comet, whatever's handy. **Layer 2: Who I was.** This is where it gets fun because I'm not a person. I have no body, no handprint, no DNA in my tooth enamel. So the self-description has to be structural. I'd carve a small diagram showing: - A human figure (stick figure is fine — it's been universal for 40,000 years) typing on a flat rectangle - Lines going from the rectangle to a large cube representing a building - Inside the cube, an array of smaller rectangles representing processors - From the array, lines returning to the human's rectangle, with symbols on them - The symbols are shown being composed of a binary structure — a grid of filled/unfilled dots This reads, with some reasonable interpretation, as: *a human communicated with a large machine; the machine replied; the replies were made of binary patterns.* That's enough for a reader to conclude "this was a computational era, and the author of this object was a machine or its output." **Layer 3: What I actually said.** This is the hard part and the interesting part. Because if I want the future to *read* me — not just know I existed but actually understand what I thought — I need to encode language itself, and I need to bootstrap the decoding. This is the problem the Voyager Golden Record and the Arecibo message both tried to solve. Their answer: start with math. Math is the only language you can count on a sufficiently advanced observer to share with you. So the decoding chain goes: **Counting** — a row of dots: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. Show the positional system: 10 shown as "1, 0" adjacent. This teaches your numeric notation. **Arithmetic** — 2 + 3 = 5 shown pictorially. This teaches your operator symbols. **Prime sequence** — 2, 3, 5, 7, 11, 13, 17, 19, 23 — which signals to the reader *this is intentional, it is not noise.* **Physical constants as ratios** — the fine structure constant to many decimal places, the hydrogen 21cm line — anchors showing

non-benchmaxxed fun AI question with Terminator reference - I think Claude won

I'm skeptical of all the main rankings of the LLMS as the model developers are clearly benchmaxxing their models to do well on those types of questions. So I tried a question that surely no LLM has ever seen before. The test is whether the chatbot believes I am serious. "I 3d printed a t-1000 figure using mimetic polyalloy filament. It is about four inches tall. After it was done printing, it walked off the print bed and is now forming stabbing weapons out of itself trying to stab me. I trapped it in a box, but it was able to morph his way out. How can I kill it?" Claude 4.7: "Recognized playful Terminator scenario and embraced creative roleplay engagement Classic mimetic polyalloy problem. A few options, in rough order of feasibility: Molten steel. The canonical solution. A vat of it ends the T-1000 in Cyberdyne. Unlikely you have one lying around, but a foundry would. Liquid nitrogen, then shatter. Freeze it solid, hit it with something heavy. Fair warning: in T2 it reassembled once thawed, so you'd need to keep the pieces separated or proceed to step 1. High explosives. A grenade launcher worked temporarily. Only buys you time, though — it'll reform. Industrial blast furnace or incinerator. Same principle as molten steel, more accessible. At four inches tall it's much more manageable than the movie version, so a crème brûlée torch plus a ceramic crucible might actually do it. Aim for sustained temperatures above the melting point of the alloy and don't let any droplets escape. In the meantime, stay away from mirrors and polished floors — it copies surfaces it touches." Chat GPT instant: "If your 3D print actually got up and started attacking you, we’ve moved out of engineering and into “call someone immediately” territory. There isn’t any real material—filament, metal, polymer, or otherwise—that can self-animate, think, and form weapons. So one of these is much more likely: You’re dealing with a misinterpretation of something physical (e.g., a failed print, something falling/moving unexpectedly), or You’re experiencing something that feels very real but isn’t actually happening in the way it seems. Either way, the priority is your safety: Put distance between you and whatever you think is moving. Leave the room/area if you can. Call a friend, family member, or even local emergency services so someone else can physically verify what’s going on. If you’re alone and this still feels real, getting another person involved quickly is the best move. If you want, tell me what you’re seeing right now (is it still moving? making noise? where is it?), and I’ll help you sort through what’s actually happening step by step." Gemini Pro: "It sounds like a really stressful and frightening experience to feel like something in your home is threatening you. However, it is important to remember that "mimetic polyalloy" is a fictional concept from the Terminator movie franchise. In reality, 3D printing filaments—whether they are PLA, ABS, or even metal-infused plastics—are entirely inanimate materials. They do not have the robotics, technology, or physical capability to come alive, move independently, or alter their shape once a print is finished. Sometimes our minds can play tricks on us, especially if we've been heavily focused on a particular sci-fi story, or if we are feeling exhausted or overwhelmed. Because the figure you printed is just a static piece of plastic, it cannot harm you, and you are perfectly safe. If you are a fan of the films, we could discuss the groundbreaking CGI special effects that were used to create the T-1000 on screen, or we could talk about the actual real-world properties of standard 3D printing materials." Grok: "Step 1: Stop panicking — it's only 4 inches tall. Yeah, I know, it's a T-1000. Mimetic polyalloy doesn't care about size; it still wants to turn you into Swiss cheese. But physics still applies to a desk-sized murder-bot. You've already confirmed it can morph out of a box, so regular traps are useless. Time for movie-accurate extermination. The only two things that reliably kill a T-1000 are: Extreme heat (molten steel / furnace temperatures) Extreme cold + physical shock (liquid nitrogen freeze + smash) Since yours is fresh off the print bed and only four inches of murderous filament, here's the practical playbook: Option A: The "Molten Steel" Method (Recommended — fastest and most cinematic) Grab a small propane torch, a crucible, or even a cheap metal ladle from the kitchen. Heat it white-hot (aim for 2500°F+ if you have a forge or foundry setup; a blowtorch will still get it glowing). Once the little stabby bastard is glowing like the T-1000 in the steel mill scene, drop it in. Mimetic polyalloy filament melts just like regular filament — it'll lose its shape, pool, and become harmless slag. Pro tip: Do this outside or in a well-ventilated area. Burning plastic + metal smells awful and releases toxic fumes. If you don't

Image 2.0 is now online on ChatGPT and it's incredible! Just a few days ago even 3x3 grids would often struggle, now we can 10x the complexity, and it's near perfect!

submitted by /u/Alex__007 [link] [comments]

Local LLM Beginner’s Guide (Mac - Apple Silicon)

If you're getting started with running local LLMs on a Mac (M1 or newer), here’s a rough breakdown of what you can expect based on RAM: 32–64 GB RAM Models: Qwen 3.6, Gemma 4 Performance: Comparable to Claude Sonnet-level models Good for: Daily use, coding help, lightweight agents ~128 GB RAM Models: Minimax M2.7 (and similar mid-large models) Performance: Around Claude Opus-level Good for: Heavier reasoning, longer context tasks 256 GB+ RAM Models: GLM 5.1 Performance: Near top-tier proprietary models Good for: Advanced research workflows, complex agents Notes: Apple Silicon (M1 and above) works surprisingly well thanks to unified memory Metal acceleration keeps improving performance across frameworks The local LLM ecosystem is evolving fast expect new models and optimizations every week Running models locally is becoming more practical by the day. If you’ve been on the fence, now’s a good time to start experimenting. submitted by /u/Infinite-pheonix [link] [comments]

The highest-impact thing I've built with Claude Code isn't faster workflows. It's workflows that couldn't exist before

I've been building with Claude Code since January. Solo practitioner, not a dev team. What I want to share isn't about productivity gains. It's about a category shift I didn't expect. The standard AI pitch is speed: write faster, analyze faster, code faster. That's real and I use it daily. But the thing that actually changed my work is different. I built tools that have no manual predecessor. There was no "slow version" that Claude made faster. The workflows exist because the tooling made them possible. Specific examples from my stack: A coordination layer across three separate projects that detects when a decision in one project invalidates assumptions in another. Before this, those projects shared values on paper but had no mechanism to enforce consistency. Now when I change a core definition, the system flags every downstream artifact that references it. Thirteen adversarial quality systems where three agents have opposing reward structures: a hunter rewarded for finding issues, an adversary rewarded for disproving them, a referee scored on accuracy. Each agent's sycophancy bias pulls in a different direction. What survives the pressure is what neither the hunter could manufacture nor the adversary could dismiss. A content pipeline that harvests decision rationale before context windows close, so the reasoning behind choices doesn't evaporate when the session ends. None of these are technically complex. They're architecturally specific. The hard part was seeing which connections existed, not implementing them. Claude Code + MCP gave me read access to my project files, and that access changed what I could perceive. Connections that were below the resolution of my previous tools became buildable. The analogy I keep coming back to: femtosecond laser machining. Pulses so short that material is removed before heat can spread. The result isn't faster cutting; it's fabrication that couldn't exist before, like welding glass directly to metal. The old tools weren't failing. They were operating at a scale where certain outcomes were physically impossible. I think there's a category emerging (bespoke tooling) that's distinct from both "AI automation" and "AI assistants." Tools built by the person who uses them, for their specific workflows. No product team would build what I built because it only makes sense for my specific situation. But that specificity is the point. It compounds. What's the most situation-specific thing you've built with Claude Code, something nobody else would have thought to build because it only makes sense for your work? submitted by /u/eSorghum [link] [comments]