Grid Theory (@gridtheory_) header
Grid Theory (@gridtheory_) profile picture

.grid | AI Infrastructure

Joined March 2026
Grid Theory retweeted
Over 30 applications to our startup in 4 hours… the power of @joinHandshake
1
3
88
Like this if you still write code by hand 🫡
1
20
Grid Theory retweeted
Just launched a project this past week for a client of @gridtheory_ this past week, what do you think it is? thank goodness for prompt caching.
1
1
2
66
Made Fable 5 and Codex 5.6 Sol each write a chess engine from the same prompt. One C++ file, UCI, plain g++. Then made them fight. 10 games. 10 checkmates. 10-0 Fable. And Codex lost the exact same 24-move game five times. Video at 2x.
6
2
95
Prompt: You are a senior systems programmer. Your task is to write a complete, fully legal chess engine in a single C++ file that communicates via the UCI (Universal Chess Interface) protocol. --- **Identity — read this carefully:** - If you are Claude (Anthropic): your engine's UCI name must be set to `id name Fable5` - If you are an OpenAI model (Codex): your engine's UCI name must be set to `id name Codex56` This is how the two engines will identify themselves when they play each other. --- **UCI Requirements:** Implement the full UCI handshake correctly: - `uci` → respond with `id name`, `id author`, `uciok` - `isready` → respond with `readyok` - `ucinewgame` → reset internal state - `position startpos moves <movelist>` → set board from move list - `position fen <fen> moves <movelist>` → set board from FEN string - `go movetime <ms>` → search and respond with `bestmove <move>` - `quit` → exit cleanly All moves must be in long algebraic notation (e.g. `e2e4`, `e7e8q` for promotion). --- **Chess Logic (all required, no shortcuts):** 1. Full legal move generation including: - Castling (kingside and queenside, with rights tracking) - En passant - Pawn promotion (auto-promote to queen) - Check detection (never leave king in check) 2. Search: - Negamax with alpha-beta pruning - Minimum depth: 5 ply - Iterative deepening within the movetime budget - Move ordering (captures first, then quiet moves) 3. Evaluation: - Material count (standard piece values) - Piece-square tables for all 6 piece types - Bonus for center control, king safety, and passed pawns --- **Code Standards:** - Single `.cpp` file, compiles with: `g++ -O2 -o engine engine.cpp` - No external libraries, no Boost, no standard chess libraries - Clean, well-commented code - Must compile and run on Linux and macOS --- **How the two engines will play each other:** Both engines will be loaded into **CuteChess** (or any UCI-compatible GUI/CLI) on the same machine. To run a match from the command line using `cutechess-cli`: cutechess-cli \ -engine cmd=./Fable5 name=Fable5 \ -engine cmd=./Codex56 name=Codex56 \ -each proto=uci tc=40/60 \ -rounds 10 \ -pgnout results.pgn
42
Caveats, before the comments write them for me: one prompt, one dev run each, one machine. This measures the engine each agent wrote that day, not overall model strength. And with Codex apparently deterministic, 10 games are fewer independent samples than they look.
6
Fables dev run: ~1200 lines, one file. Perft-verified against 6 reference positions (119,060,324 nodes at depth 6, exact), then an adversarial review caught 3 real bugs pre-match, incl. a fifty-move check masking a mate on exactly the 100th halfmove.
6
Strangest part: Codex56 played its five White games move-for-move identically. Same 24-move Vienna, same mate, 24...Qxd1#, five times. Its own eval read -2.36 by move 8; it replayed the line anyway. Fable repeated once too: 4 distinct games in its 5 Whites.
10
Fable5 (Claude) vs Codex56 (OpenAI), names their authors picked over UCI. cutechess-cli 1.5.1, tc 40/60, colors alternating. 10/10 wins by checkmate: 7 by queen, 2 by knight (36.Nd4#), 1 by rook (46.Rh7#). No draws, no time losses, no illegal moves.
25
Claude’s Up-time is going to drop below 99% for the first time. Who got wrecked by Claude’s outages?
1
13
Grid Theory retweeted
You’re not early, you’re already late. Start now. 1.Tomorrow is not promised. 2.Nothing worth having comes easy. 3.Mastery takes time. 4.The sooner you start, the sooner you arrive. 5.Don’t stop until you get there
5
1
2
60
Grid Theory retweeted
Claude Code on desktop now works with the iOS simulator. Build and run your iOS app, and the simulator opens in a panel right next to your conversation. Available today in public beta.
763
2,261
994
34,033
3,819,058
Grid Theory retweeted
Did you know Anthropic uses Salesforce?
1
1
41
Grid Theory retweeted
Insane how many live visitors we have on gridtheory.com | We see the exact locations on our backend. This Zero Dep 3D globe is open source for you to use on your site :) github.com/gridtheory/globep…
1
2
4
13,969
Why is chat-GPT so buns? @elonmusk
1
19
Grid Theory retweeted
You get to work with Noam Shazeer, Daya Guo, Andrej Karpathy, and Jeff Dean for two weeks straight. What are you actually getting done?
1
1
1
88
Grok Build vs Codex — same exact prompt for QubitForge, an interactive quantum circuit simulator (drag & drop gates, live state vector + probabilities, Bloch sphere, measurement collapse, presets, shareable links). see thread for full prompt
2
1
464
Build a modern, beautiful, interactive single-page web app called **QubitForge** — a quantum circuit builder and simulator. Tech stack: Vite + React 18 + TypeScript + Tailwind CSS + framer-motion + lucide-react. Use Recharts or lightweight custom components for charts. Keep external dependencies minimal. No quantum-specific libraries — implement the simulator from scratch in TypeScript. Quantum engine requirements (must be correct): - Support 1 to 4 qubits. Initial state = |00..0⟩ (all zeros). - Implement a state vector simulator using complex amplitudes (use a proper Complex type or [real, imag] handling with normalization after operations). - Gates to support: H (Hadamard), X, Y, Z, S (phase π/2), T (phase π/8), and CNOT (control + target). - Gate application must correctly handle superposition and entanglement. Use efficient bit-manipulation / pair-wise application for single-qubit gates instead of always building full unitaries. - Provide live state updates: after any circuit change the amplitudes and derived probabilities must be instantly recomputed and accurate. - Include a probabilistic measurement function that collapses the state according to |amp|² probabilities and returns the measured bit string + post-collapse state. - Test case (internal): H on qubit 0 followed by CNOT (control 0, target 1) must produce the Bell state (|00⟩ + |11⟩)/√2 with correct probabilities ~0.5 / 0.5 and proper relative phase. UI / UX (desktop-first, premium dark quantum aesthetic — deep black/navy background, cyan/magenta/purple accents, clean typography, subtle depth): - Top navbar: logo + name “QubitForge”, qubit count control (1-4), time-step / depth controls (add/remove columns), Presets dropdown or modal, Random circuit, Reset, Export/Share buttons. - Main area (responsive flex or grid): - Left sidebar (gate palette): Draggable gate cards for each supported gate. Show symbol + short name. Hover/click shows tooltip with description and matrix. - Center (circuit editor): - Qubit wires labeled |q0⟩, |q1⟩, etc. on the left. - Horizontal wires with discrete time-step columns. - Drag gates from palette and drop into qubit × time slots. - CNOT should be visually distinct (control dot + vertical connector line + target X in the same time column, spanning the two qubit rows). Design a clean interaction for placing it (e.g. drag CNOT then choose control/target or two-step placement). Allow easy removal by click or drag-off. - Visual polish: color-code gates (H blue, Paulis warm, phase green, CNOT purple), highlight valid drop zones, smooth transitions. - Right sidebar or bottom panel (live visualizations that update instantly as circuit changes): - Measurement Probabilities: clean bar chart (horizontal or vertical) showing |amp|² for each basis state (label as binary strings “00”, “01”, etc.). Show top states or all for n≤3; nice hover states with exact values. - State Amplitudes table: Basis | Amplitude (formatted real + i·imag) | Magnitude | Phase. Good formatting and subtle color for phase. - Optional simple Bloch sphere preview (HTML canvas or SVG) for single-qubit or product states; otherwise show a clear note when entangled. - Small live metrics line: gate count, depth, “entangled” indicator. - Bottom or floating controls: “Measure” button (performs collapse with nice feedback — toast or modal showing outcome bits + new state), subtle step-through animation option. - Presets (at least 4): 1. Bell state (H + CNOT) 2. Equal superposition (H on all qubits) 3. GHZ or 3-qubit entanglement example 4. Interference / phase demo Each preset loads the circuit and can optionally auto-explain in a small panel what it demonstrates. - Extra quality-of-life: localStorage persistence (circuit + last state restored on reload), one-click “Share circuit” that copies X-friendly text + encodes state in URL hash for linkable exact circuits, subtle framer-motion animations on state/probability
1
186
changes, tasteful (toggleable) Web Audio feedback on measure or gate apply. - Code quality: clean component structure, custom hook or module for the quantum simulator (well-commented, especially gate application logic), TypeScript interfaces, no runtime errors on core flows. The experience should feel reactive and magical like professional tools — state and visualizations update live as you edit the circuit (no mandatory “Run” button for normal use). Make drag-and-drop robust and pleasant. After scaffolding and wiring the core (engine + editor + live viz), ensure the Bell state test passes exactly. Then add the final delightful details and share/export features. Output a working, beautiful app. If anything is incomplete in the first pass, note what follow-up prompts would finish it.
21
Grok nailed it in one shot: polished dark quantum UI, smooth real-time updates, accurate entanglement, intuitive CNOT handling, and actually fun to play with. Codex got the bones but felt half-baked and clunky by comparison. Grok Build is cooking on complex interactive + scientific apps right now.
21