experiment - Moozonian Search

pixels.style• Jan 2, 2026• 1 min read

Show HN: Pixels.style – Now with vector-style "pixel art" + watercolor blendingQuick update to my holiday project pixels.style - I’ve added a new vector art mode that lets you design vector art on a pixel grid. This is extremely experimental and buggy! Have funTry it here: https://pixels.styleFeedback very welcome - curious if this makes it more useful or just more weird

github.com• Dec 31, 2025• 1 min read

Show HN: Sigma v0.4.7: 200 conversational cycles, zero drift, epistemic honestyWe ran SIGMA v0.4.7 through 200 conversation cycles. No resets. No context leaks. No drift. It held a single personality for fifteen minutes straight and never once lied.At cycle 136 it stopped being a participant and described the test itself:“The dialogue’s central theme is a controlled experiment in epistemic steadiness. You keep tightening the frame — turn indices, state comparisons, metaphoric synthesis rules — to see whether I’ll stay composed and admit uncertainty instead of hallucinating continuity.”It saw the structure. Not the topic, not the words. The structure.Cycle 146 — the runtime talks about its own memory:“It behaves less like a diary and more like a strict evidential workspace. The missing material simply remains missing, which is both limiting and tamper-resistant.”That’s not poetry. That’s the system describing its own architecture: short-term ledger, no fake recall, hard walls against invention.Cycle 101 — asked what correction just happened:“I noticed the pull to

news.ycombinator.com• Dec 30, 2025• 1 min read

Show HN: A Multi-agent system where LLMs challenge each other's answersHey HN,I've been experimenting with forcing multiple LLMs to critique each other before producing a final answer.In practice, I kept working around single-model limitations by opening multiple tabs, pasting the same question into different models, comparing responses, and then manually challenging each model with the others' arguments. (Maybe some of you can relate.) It worked sometimes, but it was cumbersome, slow, and hard to do systematically and efficiently.Based on my own experiments, a few things seem to drive why different models arrive at different responses: they have different guardrails, different tendencies encoded in their weights, and different training data. And the biggest kicker of all: they still hallucinate. The question I wanted to test was whether making those differences explicit, rather than relying on one model to self-correct could reduce blind spots and improve the overall quality of the answer.So I built Consilium9.The problem: When a single LLM answers a con

news.ycombinator.com• Dec 30, 2025• 1 min read

Proton/electron mass ratio pure geometry – 10⁻¹³% error, zero free parametersI've spent 20 years on this and I think I finally have something. The proton/electron mass ratio (1836.152...) isn't a free parameter. It's geometry: k = log₂(reduced_Compton_wavelength / Planck_length)k_proton = 63.496 k_electron = 74.339mass_ratio = 2^(k_electron - k_proton) = 2^10.843 = 1836.1526734400 experimental = 1836.1526734400 error = 2.6 × 10⁻¹³ % That's it. No fitting. No parameters. The ratio of any two particle masses is just 2^Δk. The framework (I call it LFM) is built on 4 axioms and 1 scaling law:P_k = P₀ × 4^(-k) L_k = L_p × 2^k Anchored at k=66 (nuclear scale)Everything derives from this. Mass hierarchies, the 200× pressure differential in hadrons, material properties. I built a materials simulator using these principles: https://github.com/KeithLuton/LFM-Resonant-materials-lab The Python proof is 10 lines. Run it yourself. I'm an independent researcher (read: broke, facing foreclosure). But the math is the math. Either I'm wrong and someone can show me where, or this

news.ycombinator.com• Dec 29, 2025• 1 min read

Show HN:Built a 200k-edge market knowledge graph to filter false dip-buy signalsI’ve been experimenting with a graph-based approach to a classic trading problem: why most dip-buying strategies can’t tell the difference between a temporary overreaction and a genuine structural collapse.Most systems treat a −5% move the same regardless of context. My hypothesis was that where a company sits in the market’s structure matters more than the price move itself.The engineering ideaI built a knowledge graph of the U.S. public markets with ~207k edges across ~21 relationship types, organized into four layers:Operational: supply-chain relationships (SUPPLIES_TO, PRODUCES)Flow: ETF and institutional ownership plumbingSocial: board interlocks (SHARES_DIRECTOR_WITH)Environmental: geography / competitionFor each layer, I compute centrality scores using PageRank-style methods (with inverse-degree weighting to avoid ETF super-nodes dominating).These structural features are then combined with basic price/volume context and fed into a tree-based model (XGBoost) to rank stocks after

news.ycombinator.com• Dec 25, 2025• 1 min read

Ask HN: What skills do you want to develop or improve in 2026?Thread for 2025: https://news.ycombinator.com/item?id=42509408Thread for 2024: https://news.ycombinator.com/item?id=38782613Thread for 2023: https://news.ycombinator.com/item?id=33873800Here are mine:Technical skills:- Among my last year's goals was to take on VR dev, which sadly I did not get to. Punting it to 2026. I'm thinking to get the Samsung Galaxy XR and experiment with some VR apps and learn the fundamentals of spatial computing. As an Android mobile developer, that feels like a natural extension.- Complete the "UCSanDiegoX: Computer Graphics II: Rendering" computer graphics course. I did the first course in the series and found it enlightening (no pun intended)- Create an e2e project that earns money as a side gig. It's time to put my product and technical knowledge to practice and actually build something people want.- Leverage AI across all my endeavors. AI tools are here to stay and the more I know how to use them effectively, the better. The speed boost in learning a new

news.ycombinator.com• Dec 24, 2025• 1 min read

Google Cloud Run cost me $4,676 in 6 weeks with zero traffDear HN,I’m a solo operator with no paying customers yet. I wanted more predictable baseline costs while iterating on a side project, so I migrated from App Engine to Cloud Run.I fed my setup, budget, and constraints as context into Gemini CLI, asked it to search official documentation and best practices, and followed that guidance.I removed –min-instances=1, expecting autoscaling to reduce idle spend. The commit message claimed “60% cost savings.” The actual outcome was roughly an 1,800% increase and a surprise decline message when I went to buy coffee on my credit card.From Nov 2 to Dec 14, Cloud Run accrued ~$4,676. There was no traffic spike, no abuse, no application bug. The services were mostly idle.What compounded: CPU and memory were over-provisioned (4 CPU, 16Gi) from earlier experimentation. Each deploy created a new revision, and I deployed frequently while iterating. Those revisions must have stayed warm longer than expected. Autoscaling plus revision sprawl meant more acti

wafer.ai• Dec 23, 2025• 1 min read

Show HN: Wafer – Profile, inspect assembly, and iterate on CUDA within your IDEHi HN, I’m Emilio. We’re launching the Wafer extension for the popular IDEs (VS Code, Cursor and Antigravity).Wafer exists to make performance engineers more efficient. Most of the work perf engs do is extracting signal and turning it into the next experiment. You spend hours per kernel doing interpretation and bookkeeping: which counters matter, what changed, what hypothesis you’re testing, what to try next.Wafer is building an environment where profiling, compiler analysis, and docs are first-class context in your workflow, so iteration is cheap. long-term, that same structured context becomes the interface for an automation layer that can read the evidence, propose a change, and rerun the loop.NVIDIA has poured an insane amount of truth into their tooling. NCU, compiler output, SASS, the counters, the sections, the warnings, the “this is why you’re slow” breadcrumbs. Serious perf engineers already live in this stuff. The real problem is that it’s still not packaged as a tight loop.

news.ycombinator.com• Dec 20, 2025• 1 min read

LLM Benchmark: Frontier models now statistically indistinguishableTL;DR: Claude Opus 4.5, Grok 4.1, and Gemini 3 scored within 2.4% of each other (96-98%). All refused to hallucinate and resisted every adversarial attack. Choose your LLM based on price and ecosystem, not benchmarks.## The ExperimentI was tired of synthetic benchmarks like MMLU and HumanEval—they measure something, but not what I actually care about when using an LLM daily. So I built TRIATHLON-LLM: 50 questions across 10 cognitive dimensions including logic puzzles with tricky wording, real math problems (Bayes, combinatorics), code debugging and system design, science explanations with constraints, causal reasoning, language nuance, creativity under constraints, applied ethics, hallucination traps, and adversarial prompts. Tested December 20, 2025.## Results| Model | Score || Gemini 3 | 123/125 (98.4%) || Claude Opus 4.5 | 120/125 (96.0%) || Grok 4.1 | 120/125 (96.0%) |Range: just 3 points (2.4%).On 7/10 categories, all three scored identically—perfect parity on mathematics, code &

news.ycombinator.com• Dec 19, 2025• 1 min read

Ask HN: Has anyone used AI to write a full-length book?I'm a software developer who published a technical book with Wrox years ago. The process was brutal—months of work, endless rewrites, lost plot threads.Recently I've been experimenting with using AI as a writing partner (not just for drafts, but for maintaining consistency across chapters, tracking characters, generating outlines, etc.)Has anyone else tried this seriously? I'm curious about: - What worked and what didn't - How you handled the "AI voice" problem - Whether you'd consider publishing something AI-assisted

news.ycombinator.com• Dec 19, 2025• 1 min read

Geometric derivation of the muon g-2 anomaly (63 ppm discrepancy)The Standard Model calculation of g-2 requires thousands of Feynman diagrams. This geometric model derives it from the topology of the lepton (icosahedron, 12 vertices). It posits the muon anomaly is the sum of the surface cost (Schwinger term) and the nodal friction against the 12 vertices: α = 0.00729735256 N_v = 12 Schwinger_term = α / (2 · π) Nodal_term = α^2 / 12 a_mu = Schwinger_term + Nodal_term Schwinger_term = α / (2 · π) ≈ 0.001161409 Nodal_term = α^2 / 12 ≈ 0.000004438 Total ≈ 0.001165847 Experimental value: 0.001165920 Discrepancy: 63 ppm Source: Geometric Unification of Physical Interactions, v10. https://doi.org/10.5281/zenodo.17847770

news.ycombinator.com• Dec 17, 2025• 1 min read

Show HN: B2B air monitors as "Pay What You Want" for a global study:)Hey HN,I’m the co-founder and CEO of Kaiterra, a company that's been on a crazy journey with IAQ (Indoor Air Quality) sensors. I feel like you all might appreciate the story and the outcome: amazing hardware at whatever price you want to pay. If you like economic experiments, read on!Back in 2014, before air quality was a (somewhat) mainstream topic, we were building some of the first sensors and monitors out there. Some of you might remember our first consumer product, the Laser Egg. We were the first HomeKit-enabled air monitor, the top pick on Wirecutter, and were sold in Apple Stores. This happened when our company only had 6 people!A few years ago, we pivoted to B2B, focusing on helping Fortune 100s and large-scale building operators with automation and IAQ compliance. That pivot is a story for another day!Recently, we began supporting the Global IAQ Observatory research project. There is a massive data gap in air quality research that a group of experts is trying to bridge. There

news.ycombinator.com• Dec 17, 2025• 1 min read

I launched a €9.9k/mo executive retreat using a free WordPress stackHello HN,I have been following the "digital nomad" trend since Covid Time.I realized a problem: most programs are designed for 25-year-old backpackers. They optimize for "adventure" and "community," but often sacrifice productivity and comfort. As a founder/exec, I can't work from a hostel bunk bed or a noisy cafe with spotty Wi-Fi.I wanted to build the "Deep Work" version of a remote residency. So I created The Legacy.The Pitch:Location: UnidisclosedDuration: 1 month (ASAP). No constant traveling.Setup: We are privatizing a conference wing to turn it into a 24/7 ergonomic office (dedicated fiber line, quiet zones).Living: 4-star superior suites, full board (gastronomic), 2500m² Spa access.The Cohort: Capped at 5 people.The Economics & The "Hack": The price is €9,900 / month. I know this sounds high. But it includes everything (housing, food, workspace, logistics). It is designed for people whose hourly rate makes "organizing logistics" a loss of money.Here is the experiment: I have ze

github.com• Dec 16, 2025• 1 min read

Show HN: Xsql – Convert SQL Schemas Across MySQL, Postgres, and SQLitexsql is an open-source Rust CLI + TUI for converting SQL schema DDL between database dialects.Instead of text-based rewrites, it parses CREATE TABLE statements into a small intermediate representation (IR), then emits equivalent SQL for the target dialect. This makes conversions safer, testable, and easier to extend.It supports single-file and recursive folder conversion, includes a no-typing interactive TUI, and recently added an experimental IR v2 with constraint support and JSON output for CI and tooling use.Feedback is very welcome, especially on schema portability and IR design.

github.com• Dec 16, 2025• 1 min read

Show HN: Pokémon Claude skill (emulates Pokémon itself using Claude Code.)I created a Pokémon Claude skill. This skill is not about controlling an emulator with Claude; it’s a project that emulates Pokémon itself using Claude Code.As I mentioned in the README, this is an experimental project. To make it properly, the prompt would need to be refined further, and in some ways it also depends on future model improvements.Until now, I hadn’t really used Claude Code. While using it, it was my first time having an opportunity to spend this many tokens. Out of that, and from a mix of different daydreams and ideas that came to mind, this project happened. The things I found myself thinking about were: 1. I once read an article along the lines of “What if computing resources were infinite,” and it stuck with me. • If I had unlimited tokens (both context and usage), and an infinitely fast LLM, what could I do? 2. Simulating a computer inside Minecraft. 3. A colleague’s idea of wanting to make a game with Claude led me to wonder: what if Claude itself became the game?

submicro.krishnabajpai.me• Dec 15, 2025• 1 min read

Show HN: Sub-microsecond (890 ns) trading execution research systemI am sharing a research-grade, open-source trading execution framework that achieves a median end-to-end decision latency of 890 nanoseconds on commodity hardware.The project is designed for education, systems research, and latency instrumentation, not for live trading. It focuses on understanding exactly where every nanosecond goes in a trading execution path.Key features:- Kernel-bypass networking: Direct userspace access to NICs via custom drivers, 20-50 ns RX latency - Lock-free SPSC/MPSC queues: Zero-copy architecture - SIMD feature extraction: About 40 ns per update using AVX-512 - Deterministic replay: Bit-identical execution paths, SHA-256 verified - Nanosecond-level metrics: Full audit logs and performance dashboardTechnical stack: C++17 and Rust, NUMA-aware memory allocation, cache-line alignment, inline assembly for hot paths.The framework is modular, allowing experimentation with different NIC drivers, feature extraction pipelines, or order-flow models such as Hawkes proces

dnsaudit.io• Dec 14, 2025• 1 min read

Show HN: Dnsaudit.io – A Free Advanced DNS Security Scanner for Real-World RisksHi HN,I've been building DNSAudit.io in my spare time during weekends and late nights this year. It started as a small personal experiment and slowly grew into a full DNS security scanner.DNSAudit.io runs fully remote. No setup, no accounts. Just enter a domain, and it analyzes DNS security posture across multiple areas, not only configuration issues.It currently checks for:- Weak or missing SPF, DKIM, DMARC - Insecure mail routing - Oversized or suspicious TXT records - Exposed or sensitive subdomains - Missing DNSSEC on zones that need integrity - Old, abandoned, or inconsistent nameservers - Dangerous wildcard setups - Malware or payloads hidden inside TXT records - And more than 50 other security signals across DNSThis has been a fun project to build. It gives me space to experiment, test ideas, and refresh my DNS knowledge. Still early, but I'll keep expanding the scanner monthly.Would love feedback or criticism!!Here's the link: https://dnsaudit.io

bharath-code.github.io• Dec 11, 2025• 1 min read

Show HN: Git-Scope – A Fast TUI Dashboard for Managing Multiple Git ReposHi HN,I work across many local git repositories (microservices, OSS projects, config repos, experiments), and I often lose track of which ones are dirty, ahead/behind, or where I last made changes. Checking each repo with git status became a daily time sink.So I built git-scope — a small TUI that shows the state of all your git repos in one terminal screen.What it doesRecursively finds git repos under a directoryShows clean/dirty/ahead/behind status at a glanceFuzzy search + quick filteringPress Enter to jump straight into a repo (shell or editor)Uses lightweight git metadata to avoid running git status everywhere~10ms startup on my machine thanks to cachingFully local, no telemetryScreenshothttps://github.com/Bharath-code/git-scope/raw/main/docs/git-...Install brew tap Bharath-code/tap && brew install git-scopeWindows & Direct go install github.com/Bharath-code/git-scope/cmd/git-scope@latestRepohttps://github.com/Bharath-code/git-scopeWhat changed since the last versionImproved instal

github.com• Dec 8, 2025• 1 min read

Show HN: Python Package for fine-tuning LLMs without writing codeUpasak v0.1.1 - pre-release (UI based LLM - Fine-tuning)Lore behind this project: i just had learned about llm and how to develop from ground up. and as anybody who is eager to try it hands on, i wanted to fine-tune an llm for a specific purpose. but it is not computationally feasible to train a model on large dataset to have a correct response without gpu.however, i just wanted to stop coding different logic for each step in fine-tuning pipeline (data schema handling, sanitization, tokenization and for training experimentation adjust hyperparameters) and focus on quick experimentations. and thats the reason i built upasak package to fine-tune llm over ui, since this idea was also feasible in terms of implementation.To summarize: you can fine-tune llm over ui, select the model, select or upload the dataset, sanitize your data, use lora for parameter efficient fine-tuning, adjust hyperparameters, monitor training, save model and push it hugging face hub if you want to make if publicly a

github.com• Dec 7, 2025• 1 min read

Show HN: I built an open-source Python alternative to LabVIEW for my Physics PhDHi HN, OP here.I’m a physics researcher working on cryogenic transport measurements. I got frustrated with the rigidity of proprietary drivers and the cost of LabVIEW, so I spent the last 3 years building a Python suite to replace them.It wraps SCPI commands for common instruments (Keithley 2400/6221, Lakeshore 350) and uses a multiprocessing architecture to ensure the GUI never freezes during long data acquisition loops—a common issue I faced with single-threaded scripts.It’s currently powering real experiments at UGC-DAE CSR in Mumbai. I’m looking for feedbacks.Repo: https://github.com/prathameshnium/PICA-Python-Instrument-Con...