81 results for parameters

news.ycombinator.com
We trained a real-time world model for $2k with Minecraft mod revenuehttps://news.ycombinator.com/item?id=48661355Hey HN, this is Daniel and Artur from DreamForge. We’re two friends and students from Kazakhstan who trained a real-time controllable world model (first-person + third-person) end-to-end for roughly $2k by standing on open models and limited revenue from our game mods. You can see early demos here: https://trydreamforge.com.The slightly strange part is our small budget and how it was even formed. We come from several years of indie gamedev and game modding background. We’ve got mods with millions of downloads that have been covering a meaningful chunk of our cloud GPU costs. So, the budget constraint was the thing we’ve been building around this whole time.Our first model (DF-World 0.1 Preview) is 1.5B parameters in size, generates video at 480p, supports real-time mouse+keyboard control in both first-person and third-person view, accepts mixed multimodal inputs (text, image, video), and lets you reprompt mid-stream to change the world on the fly. It runs in ~10-12 FPS on a single RTX
Jun 24, 2026 3:22 PM
ebookaloud.com
Show HN: eBook to audiobook narration with realistic AI voiceshttps://ebookaloud.comFor a while I've wanted to try out the new AI voices for long-form narration, but everything I found required a subscription that didn't justify my limited usage. I came across the open Kokoro model [0] and the voices are very good -- good enough to listen to for hours without the fatigue I got from legacy, robotic TTS voices. The model is 82m parameters and designed to run fast, but I still struggled to get reasonable times from CPU inference on my 12-core laptop. I thought a cloud-based GPU service would let me generate audiobooks fast enough to feed my own self-hosted library, and that same pipeline could become a product other people could use.I had two goals in building this: get some exposure to AI multi-agent coding workflows, and build a TTS product targeting ebook to audiobook conversion specifically. 99% of ebookaloud was written by DeepSeek v4 in OpenCode. I've used about 750 million tokens costing $12 in credits over the course of a month, and I'm very pleased with the resu
Jun 24, 2026 3:04 PM
modeloop.app
Show HN: Modeloop – A modern model-based design toolhttps://www.modeloop.app/Hi Developers, I'm Luca, the creator of Modeloop. I've spent the last 18 months building, from scratch, a model-based design tool and today I'm finally opening the Open Beta.Modeloop is a tool designed to generate code and deploy on embedded systems. The core idea behind this product is to create a vertical solution to design, test and iterate on the model faster, leveraging the modern software practices like Git. For this reason, Modeloop avoids vendor lock-in by saving model data as readable JSON rather than proprietary formats, making model reviews significantly easier.At its core, Modeloop is conceived with a developer-first mindset, focusing on delivering an optimal user experience. This approach comes from my personal experience working in the robotics and automotive industries where I've experienced firsthand the frustration of fighting with tool configurations. The software we use should be built to accelerate innovation, not spend all day setting the right parameters to make t
Jun 2, 2026 2:30 PM
news.ycombinator.com
Ask HN: What's your experience with AI in hiring?https://news.ycombinator.com/item?id=48111921"AI-native hiring" is just faster keyword/vector matching. Every atomic hiring step such as filtering applicants, moving to different stages involves, we will be dealing with 100s of judgement parameters before we take any action forward. What are we offloading to AI today? - just admin tasks or human judgements?- Sourcing and filtering are intensive human judgements operations. - Accepting applications, managing, organising, collaborating, outreach are most time admin consuming tasks in hiring pipeline.Simple ATS sass automates most of these things. AI is a great tool, but before fitting into hiring user flows where judgements >>> automations, we need to resolve on fundamental principle, AI-augmented judgment or AI-replaced judgment? Which you think is actually working well today for HR professionals? Has AI helped, or just added noise?
May 12, 2026 6:04 PM
news.ycombinator.com
Show HN: BNNR – a closed-loop pipeline for improving vision modelshttps://news.ycombinator.com/item?id=47719542Hi HN,we’ve been working on computer vision models for a while, and one thing kept coming up: improving them is surprisingly unstructured.You train a model, try a few augmentations, tweak some hyperparameters, run it again - and if the metric goes up, you keep it. But it’s often unclear why it improved, or whether the model is actually learning something better.We kept running into questions like:- Did this change actually help, or is it just noise?- Is the model focusing on the right features?- Are we improving generalization, or just overfitting differently?So we built BNNR (Bulletproof Neural Network Recipe) - an open-source PyTorch toolkit that turns model improvement into a closed loop:- Train a controlled baseline- Explain what the model actually learned (via XAI methods like OptiCAM / GradCAM)- Improve by testing candidate strategies in parallel- Prove the result with structured comparisonsOne thing we focused on is avoiding “blind” changes. Instead of committing to a single ide
Apr 10, 2026 3:24 PM
raw.githubusercontent.com
Show HN: Linear RNN/Reservoir hybrid generative model, one C file (no deps.)https://raw.githubusercontent.com/bggb7781-collab/lrnnsmdds/refs/heads/main/lrnnsmddsI just noticed it takes literally ~5 minutes to train millions parameters on slow CPU...but before you call Yudkowsky that "it's over", an important note: the main bottleneck is the corpus size, params are just 'cleverness' but given limited info it's powerless.Anyway, here is the project:https://github.com/bggb7781-collab/lrnnsmdds/tree/maincouple of notes:1. single C file, no dependencies. Below are literally all the "dependencies", not even custom header (copy paste from the top of the single c file):#define _POSIX_C_SOURCE 200809L#include #include #include #include #include #include #include #include #include #include 4136 lines of code in one file at the moment, that's all.2. easiest way to compile on Windows: download Cygwin (https://www.cygwin.com/), then navigate to the directory where your lrnnsmdds.c file is and just run gcc on it with some optimizations, such as:gcc -std=c17 -O3 -march=native --fast-math -o lrnn lrnnsmdds.c -lmOn Linux just run gcc, if for whatever reason yo
Apr 9, 2026 9:55 PM
news.ycombinator.com
Hybrid Attentionhttps://news.ycombinator.com/item?id=47674749TLDR: Forked pytorch and triton internals . Changed attention so its linear first layer , middle quadratic layer, last linear layer Inference got much faster with a low perplexity hit in tests .Full attention O(n²): 17.96s / 5.6 tok/sHybridAttention O(n·W + n·D): 0.35s / 286.6 tok/sI have been building a small Rust focused language model from scratch in PyTorch. This is not a finetune. It is byte level, trained from random initialization on a Rust heavy corpus assembled here: https://codeberg.org/JohannaJuntos/SisyphusModel and training setupThe model has 25.6M parameters with a 512 context length. It uses a byte level vocabulary of 256, with 8 layers, 8 heads, and 512 dimensional embeddings. Positional embeddings are learned and the embedding and LM head weights are tied.Training ran for 30k steps on a 173.5M byte Rust corpus using a single RTX 4060 Ti 8GB.Final metrics were a train loss of 0.5834, validation loss of 0.8217, and perplexity of 2.15. The best validation loss occurred ar
Apr 7, 2026 1:06 PM
github.com
Show HN: Robust LLM extractor for websites in TypeScripthttps://github.com/lightfeed/extractorWe've been building data pipelines that scrape websites and extract structured data for a while now. If you've done this, you know the drill: you write CSS selectors, the site changes its layout, everything breaks at 2am, and you spend your morning rewriting parsers.LLMs seemed like the obvious fix — just throw the HTML at GPT and ask for JSON. Except in practice, it's more painful than that:- Raw HTML is full of nav bars, footers, and tracking junk that eats your token budget. A typical product page is 80% noise. - LLMs return malformed JSON more often than you'd expect, especially with nested arrays and complex schemas. One bad bracket and your pipeline crashes. - Relative URLs, markdown-escaped links, tracking parameters — the "small" URL issues compound fast when you're processing thousands of pages. - You end up writing the same boilerplate: HTML cleanup → markdown conversion → LLM call → JSON parsing → error recovery → schema validation. Over and over.We got tired of rebuilding t
Mar 26, 2026 3:55 AM
github.com
Show HN: Three new Kitten TTS models – smallest less than 25MBhttps://github.com/KittenML/KittenTTSKitten TTS (https://github.com/KittenML/KittenTTS) is an open-source series of tiny and expressive text-to-speech models for on-device applications. We had a thread last year here: https://news.ycombinator.com/item?id=44807868.Today we're releasing three new models with 80M, 40M and 14M parameters.The largest model (80M) has the highest quality. The 14M variant reaches new SOTA in expressivity among similar sized models, despite being
Mar 19, 2026 3:56 PM
github.com
Show HN: Investing Withous Bias or Emotionhttps://github.com/kachurovskiy/stratcraft/I've spent years actively trading and losing money. It affected my health and relationships negatively. I also realized that my opinion on what's worth buying is more than worthless, what matters is the opinion of the market.But I didn't want to settle for index funds that hold a lot of under-performing stocks so I've built this tool that allows trading 12000 instruments (stocks, ETFs, metals, real estate, crypto) via statistically proven strategies.Project is open source, tool is self-hosted, at the very minimum it requires an Alpaca brokerage account, they also provide the necessary market data, alternatively you can use Tiingo or EODHD.I'm currently running 2 paper trading accounts, fine-tuning parameters like slippage, planning to start ADX on real money soon.I'm not trying to beat hedge funds and other big players here, instead I'm looking to join them by waiting for a trend to be established (daily time frame, wide stops, long duration trades) and naturally diversifying over all
Mar 16, 2026 8:36 PM
news.ycombinator.com
Is anyone compressing AI models for the 4B people without GPUs or internet?https://news.ycombinator.com/item?id=47213965Hey HN,I'm a 20yo solo builder from India,I got frustrated that every capable AI model assumes you have a GPU, a credit card, or reliable internet. None of those are true for most of the world — including me.So I started digging into the compression literature and ways through which i can solve this problemWhat I found: - DeepSeek distilled 671B reasoning into 1.5B that runs on a laptop - TRM (Samsung, 2025) beat DeepSeek R1 on ARC-AGI with 7M parameters by iterating instead of scaling - RWKV runs in constant memory with no quadratic attention cost - GRPO lets you specialize a tiny model on a narrow domain in hours on CPUThe techniques exist. What doesn't exist: a systematic effort to apply all of them together, specifically for low-resource languages and low-end hardware, and give the results away free.I'm building this. Calling it KIRO.The goal is simple: take every major open source frontier model, compress it into domain-specific versions under 500MB, and deploy them offline on the
Mar 2, 2026 4:42 AM
morsel.modelscript.org
Show HN: Morsel – an open-source, web-based visual Modelica editorhttps://morsel.modelscript.orgHey HN,I’m excited to share Morsel, a modern, web-based visual editor for Modelica that I've been working on. It's part of the broader ModelScript compilation, analysis, and visualization framework.Modelica is a powerful, object-oriented language for equation-based modeling of complex physical systems (mechanical, electrical, electronic, hydraulic, thermal, control, etc.). However, tooling in the Modelica ecosystem has historically been heavy, proprietary, or desktop-bound. I built Morsel to bring a fast, beautiful, and accessible Modelica modeling experience right to the browser.Key Features ============- Interactive Diagrams: High-performance visualization of Modelica connections and topologies using AntV X6.- Synchronized Code Editor: Real-time, two-way synchronization between the visual diagram and the underlying Modelica code (powered by Monaco Editor).- Intelligent Properties Panel: Easily edit component parameters, modifiers, and annotations through a responsive UI that directly
Feb 26, 2026 8:21 PM
news.ycombinator.com
Is it possible to build ithttps://news.ycombinator.com/item?id=47163195● Core: Vertically aligned monolayer graphene (electron velocity > 150\,000\text{ cm}^2/\text{Vs}). ● Interconnect layer: Nanolayer silicon (Si) to create an energy barrier (bandgap). ● Shield/Gate: Coaxial gold (Au) coating (noise shielding and instant heat dissipation). ● Base (Substrate): Synthetic single-crystal diamond (thermal conductivity > 2200\text{ W/mK}). B. Performance parameters (Estimated) ● Operating frequency: 500\text{ GHz} - 1.2\text{ THz} (vs. today's 5\text{ GHz}). ● Energy efficiency: 90\ \% power reduction compared to traditional FET design. ● Thermal Management: Capable of passive operation without the need for active cooling
Feb 26, 2026 7:55 AM
github.com
Show HN: Sleeping LLM – A language model that remembers by sleepinghttps://github.com/vbario/sleeping-llmI built a system that gives LLMs persistent memory from conversations — not through RAG or databases, but by editing the model's actual weights. The knowledge lives in the parameters. The context window is empty.During wake, facts from conversation are injected directly into MLP weights via MEMIT (a single forward pass, instant recall). During sleep, the system audits which memories degraded, refreshes them with null-space constraints (guaranteeing orthogonality to working memories), then progressively transfers knowledge into LoRA — like biological memory consolidation from hippocampus to neocortex.The key problem was a hard capacity ceiling: the 8B model sustains 0.92 recall up to 13 facts, then crashes to 0.57 at fact 14 — a sharp phase transition, not gradual decay. And LoRA consolidation was blocked by what I call the "alignment tax": RLHF training fights back against injected knowledge (37% recall loss on 8B from a single LoRA pass).The fix: per-fact graduated consolidation. Each
Feb 26, 2026 6:09 AM
cuto.video
Show HN: Cuto – Turn raw footage into viral shorts in 30s with an AI prompthttps://cuto.videoI built Cuto because I was tired of spending hours on the repetitive parts of video editing: splitting long footage, syncing audio and captions, filtering out bad takes, and adjusting the pacing. I just wanted to focus on the storytelling.Cuto is an AI video editing workspace that automates the grunt work. You drag and drop up to 12 video clips, write a quick prompt detailing what you want (e.g., "Find the best 30s product highlight and add a strong CTA"), and the engine does the rest.Within 30 seconds, it hands back an editable visual timeline with: - Auto-split shots and filtered out silences/bad takes - Synchronized captions with rhythm alignment - Selected highlight effects and custom watermark branding - Auto-generated social-ready copy for TikTok, Shorts, and ReelsUnlike traditional NLEs (Non-Linear Editors), we wanted an interface where the visual parameters are controlled intuitively by natural language first, giving you a solid 80% draft instantly. From there, you can still tw
Feb 22, 2026 10:11 AM
news.ycombinator.com
Ask HN: How do you test your MyBatis dynamic SQL queries?https://news.ycombinator.com/item?id=47104032(Full Disclosure: I am developing this plugin and drafting this post with the assistance of AI agents. It’s been an eye-opening experiment in how AI can help "scratch an itch" faster than ever.)I’m currently working on a project using MyBatis, and one of the biggest friction points for me is verifying dynamic SQL. Whenever I have a complex query with , , or OGNL expressions, the feedback loop is painfully slow.My current "workflow" (which I find incredibly tedious) is: 1. Copy the raw SQL from the XML mapper. 2. Manually scrub the XML tags like and . 3. Replace all #{param} placeholders with actual values, one by one. (or :param) 4. Finally, paste the cleaned-up SQL into the IDE’s database console to see if it even runs.I was surprised to find that, despite many MyBatis plugins existing, none of them seem to offer a "direct execution with parameters" feature within the IDE. I expected this to be a standard feature by now, but it feels like we’re still stuck with manual labor for query
Feb 21, 2026 7:52 PM
spacedaily.com
New axis grid links complex earth data in space and timehttps://www.spacedaily.com/reports/New_axis_grid_links_complex_earth_data_in_space_and_time_999.htmlBerlin, Germany (SPX) Feb 04, 2026 A new axis based grid model illustrates how independent spatial, temporal, and parametric axes can be combined to describe complex earth observation data sets in a single coherent framework. The representation separates space, time, and measured parameters into distinct axes, then aligns them in a unified grid that can accommodate a wide range of satellite and in situ measurements collecte
Feb 11, 2026 11:57 AM
spacedaily.com
Quantum interferometer boosts sensitivity to tiny laser beam shiftshttps://www.spacedaily.com/reports/Quantum_interferometer_boosts_sensitivity_to_tiny_laser_beam_shifts_999.htmlBerlin, Germany (SPX) Jan 09, 2026 Researchers at Ludwig-Maximilians-Universitat Munchen have demonstrated a quantum interferometric method that can register extremely small tilts and displacements of a laser beam. The approach is based on weak value amplification, a concept from quantum mechanics that enhances the response of an interferometer when internal beam parameters change only slightly. In many experiments and tech
Jan 12, 2026 9:57 AM
news.ycombinator.com
New Year 2026: Fusion Updates from Helion and Commonwealth Fusionhttps://news.ycombinator.com/item?id=46527550Helion Energy (backed by Sam Altman and Thiel) has a short years end blog post:https://mailchi.mp/helionenergy/2025-at-helion?e=05425b940c> In 2025, Polaris transitioned fully into operations. Our team is now conducting thermonuclear fusion operations daily, refining test parameters, and using real-time plasma, neutron, and electrical diagnostics to increase the amount of fusion energy we recover from the system. Polaris consistently exceeds both the plasma temperatures and FRC size achieved on Trenta.> We also received final approval from the Washington State Department of Health to operate with deuterium–tritium (D-T) fuel, marking the first time a private fusion energy company has been licensed to perform D-T fusion.> One of our biggest milestones of the year was receiving approvals and beginning construction on the world’s first commercial fusion power plant in Malaga, Washington.What is not said: Achieving net energy from the Polaris test machine. Three years ago they announced th
Jan 7, 2026 3:33 PM
depthsight.pro
Show HN: Solo trader built algo trading platform via AI – no dev experiencehttps://www.depthsight.pro/Hey HN,I'm Artem, a trader from Ukraine. I tried learning to code 15 years ago and failed—I wanted to build trading systems, not fight with syntax.8 months ago, I decided to try again using Claude/GPT/Gemini as my "pair programmers." I built DepthSight by describing what I needed in plain language. No prior dev experience.What it does: • Native AI strategy generation from plain text • Visual logic builder with weighted conditions • Dual backtesting: vector engine for speed, event-driven for accuracy • Genetic algorithms to evolve strategy parameters • Order book + tape analysis (market depth, open interest) • Full mobile PWAThe Twist: Instead of a signup page, I built a terminal ARG.Type "help" in the terminal at the bottom of depthsight.pro to start. You'll need to: - Decrypt hidden shards (Base64, HEX) - Find files in fake logs - Crack sudo access - Solve the final puzzleRewards: First solve: Lifetime Pro (unlimited compute) Top 10: 3 months free Everyone else: Beta access + 30% life
Dec 30, 2025 5:07 PM