Solo ustedes, su magia ha hecho de esto un debut espectacular! 😍😍😍😍😍 #Monotonia# You and only you--and of course ozu--have made this debut spectacular!!! 😍😍😍😍😍 @ozuna

What if a single forward pass could let a model read two completely different texts at once? Transformers are built from strongly nonlinear pieces: self-attention and layer after layer of MLPs. So the natural intuition is that mixing two contexts into one input should make the output collapse into noise unrelated to either. This paper overturns that intuition. Simply averaging the token embeddings of two texts and feeding the result as a single input still leaves clear traces of both contexts in the next-token distribution. Tested across Pythia, Llama, and Qwen, the true next token from each individual stream lands in the top-10 ranks of the mixed output 30-40% of the time, and within the top-100 ranks 60-65% of the time. Even more striking: this superposition ability isn't something models learn. It's strongest right at initialization and degrades monotonically as pretraining continues, suggesting it's an intrinsic architectural property that training actually erodes. The authors show it can be substantially restored with lightweight fine-tuning on less than 0.025% of the original pretraining data, and they build on this to propose a guided decoding method that generates two independent, coherent continuations from a single forward pass. Title: Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs URL: #LLM# #Transformers#

Recommended read. Interaction horizon scheduling is an underexplored control problem in agentic RL This paper from the Qwen team takes a closer look at the problem. Scaling the maximum number of environment interactions per episode improves long-horizon agents, and curriculum methods that expand the horizon beat fixed-horizon training. But those schedules are open-loop. They increase monotonically to a manually specified maximum with no way to detect that expansion stopped helping. The authors propose the effective interaction frontier, a dynamic boundary past which extra interactions give diminishing returns while cost keeps growing linearly. Fixed-horizon sweeps on AppWorld and BFCL show clear saturation plateaus. Elastic Horizon is a closed-loop controller that tracks the boundary using the 90th percentile of successful trajectory lengths, a statistic already available during training. It settles inside the saturation band from both under-capacity and over-capacity starts, gets the best success rates across 7B and 14B backbones, and saves up to 25% of per-step trajectory tokens. Paper:

Harness Engineering Anti-Patterns AP1. The Context Hoarder 🎯 Point "Include everything just in case" — that feeling of safety is silently killing your agent's performance. More information isn't safer; it's often harmful. ❗ Problem The context window overflows with irrelevant information, diluting the agent's attention. Important information gets buried in the middle, leaving insufficient space for the code and specs the agent actually needs. The result: degraded judgment accuracy, with cost and latency worsening super-linearly. 🔍 Mechanism & Symptoms This anti-pattern is seductive because the intuition "more info = safer" is strong. Designing retrieval (what to fetch and when) takes effort, so stuffing everything in feels easier. But the context window is a scarce resource like a CPU's L1 cache. Utility does not increase monotonically. Beyond a threshold, irrelevant tokens dilute the attention mechanism, causing "lost in the middle" — critical info buried in the center gets ignored. Symptoms include injecting entire repositories, full conversation histories, all tool definitions on every call, and "the agent inexplicably ignores information it already read." 📋 Scenarios - An issue-to-PR agent is fed not just the issue but the entire repo's README, config files, and past PR history. The agent misses the issue's key points and starts editing unrelated files. - A migration agent receives thousands of files at once, overflowing context and losing consistency mid-task. - In pair programming, unopened files and long conversation history consume context, slowing responses. 🛡 How to Avoid - Measure context usage by category and visualize allocation like a memory profiler - Default to pull (let the agent fetch what it needs) and limit push (force injection) to only invariants that are fatal to violate - Summarize and compress older conversation turns; dynamically load only task-relevant tool definitions - When you catch yourself thinking "include everything for safety," recognize that as this anti-pattern's signature #HarnessEngineering# #AIAgent#

Harness Engineering Anti-Patterns AP1. The Context Hoarder 🎯 Point "Include everything just in case" — that feeling of safety is silently killing your agent's performance. More information isn't safer; it's often harmful. ❗ Problem The context window overflows with irrelevant information, diluting the agent's attention. Important information gets buried in the middle, leaving insufficient space for the code and specs the agent actually needs. The result: degraded judgment accuracy, with cost and latency worsening super-linearly. 🔍 Mechanism & Symptoms This anti-pattern is seductive because the intuition "more info = safer" is strong. Designing retrieval (what to fetch and when) takes effort, so stuffing everything in feels easier. But the context window is a scarce resource like a CPU's L1 cache. Utility does not increase monotonically. Beyond a threshold, irrelevant tokens dilute the attention mechanism, causing "lost in the middle" — critical info buried in the center gets ignored. Symptoms include injecting entire repositories, full conversation histories, all tool definitions on every call, and "the agent inexplicably ignores information it already read." 📋 Scenarios - An issue-to-PR agent is fed not just the issue but the entire repo's README, config files, and past PR history. The agent misses the issue's key points and starts editing unrelated files. - A migration agent receives thousands of files at once, overflowing context and losing consistency mid-task. - In pair programming, unopened files and long conversation history consume context, slowing responses. 🛡 How to Avoid - Measure context usage by category and visualize allocation like a memory profiler - Default to pull (let the agent fetch what it needs) and limit push (force injection) to only invariants that are fatal to violate - Summarize and compress older conversation turns; dynamically load only task-relevant tool definitions - When you catch yourself thinking "include everything for safety," recognize that as this anti-pattern's signature #HarnessEngineering# #AIAgent#

We have published the full engineering design behind native RGB Swaps on the Lightning Network. The swap is not a separate protocol layered on top of Lightning — it is a property of the payment path itself. A single circular route, one payment_hash, zero external coordination. Atomicity is not enforced by an additional contract; it falls out naturally from Lightning's preimage settlement. This design handles non-monotonic value flows across multi-asset channels, affects commitment construction, HTLC state management, and client-side transfer history at the architectural level. The RGB Lightning Network has reached production-grade engineering. Large-scale testing begins now. Make Bitcoin Smart.

THREE.JS r187 QUALITY LOOP You are running a quality loop on a Three.js project. Iteration count is an output, not an input. Never pick a number up front. There is always more you could add. That is why the ceiling must be fixed before pass one, or the loop has no way to tell good from busy. PROBE (once) Record: - three version (pin exact 0.187.0 = r187; allow @types/three >= 0.186.0 if 0.187.x types lag upstream) - active pipeline: WebGLRenderer + EffectComposer OR three/webgpu + RenderPipeline/TSL PostProcessing - WebGL 2 fallback requirement (if three/webgpu is active, record whether forceWebGL: true parity is in scope) - loop style (manual THREE.Timer step vs rAF), bundler, R3F or vanilla Grep the project against the 186 -> 187 migration surface (r187 removed all r177 deprecations): - WebGL passes: subclassed Pass touching removed public quad (now internal _fsQuad) - WebGPU/TSL: implicit math type coercion regressions, storage buffer binding mismatches - PMREM (#34585#): custom shaders, ViewportTextureNode, or env readbacks assuming 2D equirect/cubeUV targets instead of cube render targets - Depth: manual reversed-depth workarounds (fixed upstream, delete them) - Manifest: caret or tilde version ranges on three (pin 0.187.0 exact) CEILING (once, before pass one) Write down what "highest quality" means for this brief, as numbers: - GPU frame time budget (ms via timer query) and CPU submission budget (ms) on target device - draw call, triangle, texture count, and geometry count budgets (via - perceptual pixel diff threshold (e.g., FLIP or RMSE) below which a change is imperceptible - the brief's required elements, as a binary checklist - critic score target for G4 If it is not on this list, it is not part of the ceiling. GATES (ordered, fail-fast) G0 Build: imports resolve, zero runtime deprecation warnings, typecheck passes, zero shader/WGSL compile warnings. G1 Determinism: override Math.random with seeded PRNG and step THREE.Timer manually. Await LoadingManager, KTX2/Draco workers, document.fonts.ready, and renderer.compileAsync(scene, camera) before frame 0. Render two independent runs at fixed frame N: zero NaN/Inf, exact SHA-256 pixel match on software rasterizer (or maxChannelDelta <= 1 across 99.9% of pixels on hardware GPU). G2 Resource & Lifecycle: GPU frame time (EXT_disjoint_timer_query_webgl2 or WebGPU timestamp-query), CPU frame time, (calls, triangles), and (geometries, textures) inside ceiling budgets. Run lifecycle soak: mount -> render 120 frames -> unmount/dispose -> remount. Assert zero net delta on geometries, textures, and compiled programs. G3 Rendering: - Automated: renderer.outputColorSpace === SRGBColorSpace; albedo/emissive maps = SRGBColorSpace; normal/roughness/metalness/AO data maps = NoColorSpace; active tone mapping non-None (unless brief mandates unmapped); shadow map depth buffers non-empty on shadow-casting lights. - Material policy: enable EON rough diffuse only on hero dielectric assets already budgeted for MeshPhysicalMaterial with roughness greater than 0.5 (never upgrade background MeshStandardMaterial props if it threatens G2). G4 Intent & Temporal Stability: an independent blind critic receives ONLY a 3-angle static contact sheet, a 3-frame motion burst (t, t+16ms, t+32ms), a temporal luminance-delta heatmap, the G2/G3 metric table, and the brief checklist. It never sees your reasoning or your code diff. Fails immediately on temporal shimmer, z-fighting, or shadow cascade swimming. EACH PASS 1. Target the lowest failing gate (G_k). Only that one. 2. Isolate one root cause. Fix one mechanism (you may touch any file required by that mechanism). 3. Re-measure G0 through G_k (and G_{k+1}..G4 once G_k passes). Record the full metric vector. 4. Keep the change ONLY IF: - zero passing lower gates (G0 .. G_{k-1}) regressed, AND - the targeted sub-metric in G_k improved monotonically (even if a multi-bottleneck gate as a whole has not yet crossed threshold). Otherwise, git revert immediately. Renames, comments, dead-code cleanup, and formatting do not count as a pass. PHASE BUDGETS (decoupled) Plumbing budget (G0-G1): max 3 passes. If build or determinism fails after 3 passes, halt and escalate harness/environment blockers. Convergence budget (G2-G4): max 6 passes. Plumbing fixes do not consume convergence passes. CONVERGENCE vs EXPANSION Convergence closes the gap to the ceiling. It is finite. This loop does convergence only. Expansion adds detail that is not in the brief. It is unbounded. It is not a pass. If you see a detail worth adding, do not add it. Log it as a proposed requirement. After the loop stops, a human decides whether to accept it. Accepted additions become a new brief: update the ceiling, re-run the probe, run the loop again, stop again. STOP WHEN Converged: G0-G4 all pass, and for 2 consecutive convergence passes perceptual diff is below the ceiling threshold (or metric delta is inside measurement noise). Stalled: targeted sub-metric unchanged for 2 consecutive passes. Escalate to structural root-cause analysis. No third retry. Capped: 3 passes spent in G0-G1, OR 6 passes spent in G2-G4. Past that you are polishing noise or fighting architecture. REPORT - stop rule fired and phase (Plumbing vs Convergence) - full metric and gate vector history per pass - convergence trajectory: last-pass delta vs remaining gap to ceiling, plus pass-over-pass decay ratio, classified as Asymptotic (parameter tuning exhausted) or Structural (requires architecture change outside loop) - proposed expansions, each as one line, not implemented