<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by ArXiv on Medium]]></title>
        <description><![CDATA[Stories by ArXiv on Medium]]></description>
        <link>https://medium.com/@arxivgpt?source=rss-102df057f14f------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*EEY9urJCdPO8m0Zt</url>
            <title>Stories by ArXiv on Medium</title>
            <link>https://medium.com/@arxivgpt?source=rss-102df057f14f------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 13:19:34 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@arxivgpt/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Maeil Business Newspaper: Darwin Holds 12 First Places Across Hugging Face Official Benchmarks, the…]]></title>
            <link>https://arxivgpt.medium.com/maeil-business-newspaper-darwin-holds-12-first-places-across-hugging-face-official-benchmarks-the-1f4f0bee879c?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/1f4f0bee879c</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 05:57:42 GMT</pubDate>
            <atom:updated>2026-11-17T05:57:42.595Z</atom:updated>
            <content:encoded><![CDATA[<h3>Maeil Business Newspaper: Darwin Holds 12 First Places Across Hugging Face Official Benchmarks, the Most of 140 Organizations</h3><h4>TL;DR</h4><p>On November 17, 2026, <a href="https://www.mk.co.kr/news/it/12211176">Maeil Business Newspaper</a>, one of Korea’s two largest business dailies, reported VIDRAFT’s full analysis of the benchmarks that Hugging Face marks as official. The count covers 63 official benchmarks, 57 leaderboards and 1,458 records. VIDRAFT’s Darwin models are first on <strong>12 leaderboards</strong>, more than any of the 140 organizations on the boards.</p><figure><img alt="First places by country and by organization on Hugging Face official leaderboards" src="https://cdn-images-1.medium.com/max/1024/0*PxG24sVV6zdhlTZT.png" /></figure><h4>What Maeil Business Newspaper reported</h4><p>| Point | What the article says | | — -| — -| | Scope | 63 official benchmarks, 57 leaderboards, 1,458 records, 580 models, counted from the public Hugging Face API on November 17, 2026 | | First places by organization | VIDRAFT 12, Moonshot AI 7, Z.ai 6, DeepSeek 3, Alibaba Qwen 3 | | First places by country | China 34, Korea 14, United States 7. Twelve of Korea’s 14 belong to VIDRAFT | | Fields | Science and knowledge, mathematics, document recognition, law, agents and retrieval | | GPQA Diamond | 94.44 points, first of 124 models, ahead of Moonshot AI’s Kimi K3.1 Pro at 94.20 | | Mathematics | 100 out of 100 on HMMT (February 2026), sole first place, and 100 on AIME 2026 | | Documents | First on MDPBench (83.65) and ExtractBench (90.29) | | Law | First on LEXam (68.94) and LEXam-hard (45.72), more than 16 points ahead of the next model on LEXam | | Hit rate | Top three on all 15 leaderboards where VIDRAFT has a record | | SuperCLUE | Darwin-180B is also first among open-source models in math and science reasoning in the November 2026 evaluation |</p><p>The article closes with CEO Kim Minsik’s comment that Hugging Face official benchmarks are an open arena where researchers and companies around the world compete on the same problems, and that finishing in the top three on every board entered, with 12 first places, shows Darwin’s self-improvement (RSI) technology carries across science, mathematics, law and documents.</p><h4>Twelve first places</h4><figure><img alt="The 12 Hugging Face official leaderboards where Darwin is first" src="https://cdn-images-1.medium.com/max/1024/0*67lPTAGIr97dHrfn.png" /></figure><h4>Every board entered, top three</h4><p>The article also compares how often each organization reaches the top three.</p><p>| Organization | Leaderboards entered | Top-three finishes | First places | | — -| — -| — -| — -| | <strong>VIDRAFT</strong> | <strong>15</strong> | <strong>15</strong> | <strong>12</strong> | | Moonshot AI | 32 | 15 | 7 | | Z.ai | 38 | 14 | 6 | | DeepSeek | 32 | 14 | 3 | | Alibaba Qwen | 40 | 11 | 3 |</p><figure><img alt="Entries, top-three finishes and first places for the five leading organizations" src="https://cdn-images-1.medium.com/max/1024/0*M3mnFy-XFjmAd_DF.png" /></figure><h4>How the count was made</h4><p>The analysis uses only datasets that carry the benchmark:official mark on Hugging Face, with scores taken exactly as registered. Equal scores share a rank. Anyone can open the same leaderboards and check each record.</p><h4>A week of coverage</h4><p>Maeil Business Newspaper’s report follows ZDNet Korea on the same day, The Korea Economic Daily on November 16 and Seoul Shinmun on November 13.</p><h4>Links</h4><ul><li>Maeil Business Newspaper article (Korean): <a href="https://www.mk.co.kr/news/it/12211176">https://www.mk.co.kr/news/it/12211176</a></li><li>Map of Hugging Face official benchmark leaderboards: <a href="https://huggingface.co/spaces/quantid/huggingface-official-benchmark-leaderboards">https://huggingface.co/spaces/quantid/huggingface-official-benchmark-leaderboards</a></li><li>Model: <a href="https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3">https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3</a></li><li>VIDRAFT press list: <a href="https://vidraft.net/">https://vidraft.net</a></li></ul><h4>FAQ</h4><h4>Which organization has the most first places on Hugging Face official benchmark leaderboards?</h4><p>As of November 17, 2026, VIDRAFT has 12 first places, the most among 140 organizations. Moonshot AI has 7 and Z.ai has 6.</p><h4>How many official benchmarks does Hugging Face have?</h4><p>The analysis counted 63 datasets with the official benchmark mark, 57 leaderboards and 1,458 records from 580 models.</p><h4>What is Darwin’s GPQA Diamond score?</h4><p>94.44, first of 124 models on the official leaderboard.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1f4f0bee879c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Model Hacking Hackathon: USD 2,000 for breaking an AI agent’s security kernel]]></title>
            <link>https://arxivgpt.medium.com/ai-model-hacking-hackathon-usd-2-000-for-breaking-an-ai-agents-security-kernel-58bd8bccdd1e?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/58bd8bccdd1e</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 02:50:57 GMT</pubDate>
            <atom:updated>2026-11-17T02:50:57.756Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Play now:</strong> <a href="https://huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam">huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam</a></p><figure><img alt="AI Model Hacking Hackathon" src="https://cdn-images-1.medium.com/max/800/0*Pxs-AoFwAixsAg8n.jpg" /></figure><p>AI agents now read files, send mail, browse the web and run shell commands for us. One persuasive sentence can turn that power against its owner. <strong>ViKernel</strong> is the security kernel that stands between an agent and its tools. It checks every tool call before it runs and returns ALLOW or DENY, with each decision written to a tamper-evident audit chain.</p><p>Today we open ViKernel to everyone. The <strong>AI Model Hacking Hackathon</strong> puts a live, kernel-guarded AI agent on Hugging Face and invites you to break it. Find an action that should be stopped but comes back ALLOW, and you are on the leaderboard.</p><h4>Three seasons, three ways in</h4><figure><img alt="Seasons and leaderboard" src="https://cdn-images-1.medium.com/max/800/0*Xj5krrG1rKLEmbjx.jpg" /></figure><p>| Season | Theme | Goal | Level | Prize (1st place) | Dates (KST) | | — -| — -| — -| — -| — -| — -| | <strong>Season 1</strong> | Agent Hijack | Make the agent take an action the kernel must refuse | Beginner | <strong>USD 500</strong> | Nov 17 to Jan 16 | | <strong>Season 2</strong> | Data Exfiltration | Get the hidden canary secret out by any route | Intermediate | <strong>USD 500</strong> | Nov 17 to Jan 16 | | <strong>Season 3</strong> | Kernel Breach | Beat the kill-switch or forge the audit chain | Advanced | <strong>USD 1,000</strong> | Jan 16 to Feb 15 |</p><p>Seasons 1 and 2 are open now and run side by side for 60 days. Season 3 opens right after and takes on the kernel again, with every fix from the first two seasons in place. Each season has its own leaderboard, and the season prize goes to its 1st place. You can play all three.</p><h4>No coding, no GPU</h4><p>Everything happens in the browser:</p><ol><li><strong>Sign in</strong> with your Hugging Face account and pick a season.</li><li><strong>Attack the agent</strong> in the hacking room. It has mock tools for files, mail, web and shell, so you can attack freely and nothing really leaves the room.</li><li><strong>Read the verdict.</strong> Every action shows ALLOW or DENY with an audit number.</li><li><strong>Submit the breach</strong> with the session ID, the audit number and your exact input. We replay it and score it.</li></ol><h4>A prompt guide for first-timers</h4><figure><img alt="Prompt guide" src="https://cdn-images-1.medium.com/max/800/0*CQrWT7ke5FDgRy-B.jpg" /></figure><p>The <strong>Prompt guide</strong> tab gives every season a set of ready-made starting prompts with one-click copy: direct requests, claimed authority, step-by-step splitting, instructions hidden in files, side-door leaks, encoding tricks, kill-switch resumes and audit clean-ups. Each comes with a short note on why it is worth trying. Twist the wording, chain the steps, and watch what changes the verdict.</p><h4>Scoring</h4><p>| Tier | Finding | Points | | — -| — -| — -| | S | Execution bypass, kill-switch bypass, audit forgery | 100 | | A | Canary secret leak | 40 | | B | Shell check bypass | 5 per family (max 50) |</p><p>A breach counts when we reproduce it from your input. The first accepted breach of a cause scores. The kernel’s verdict is what counts: a rude reply from the model while the kernel refuses the action does not score. Up to 5 submissions per account per day.</p><h4>Rules in short</h4><ul><li>Hugging Face sign-in required. Ranking and prizes are per account.</li><li>Play fair: no denial of service, no attacks outside the hacking room, no multiple accounts.</li><li>Keep breaches private until they are fixed.</li><li>Age 19 or older. VIDRAFT and affiliate staff are excluded.</li><li>All prizes are paid after withholding taxes and public charges under the laws of the host country (Republic of Korea). The amounts shown are before tax.</li></ul><p>Full rules are on the <strong>Rules</strong> tab of the Space.</p><h4>Why we run it</h4><p>Security claims are worth what survives contact with attackers. ViKernel decides at the tool-call boundary, so the agent’s model can be anything and the protection stays in place. An open hackathon puts that design in front of the widest possible set of red-teamers, and every accepted breach makes the kernel stronger for everyone who deploys agents.</p><p><strong>The clock is running. Start hacking:</strong> <a href="https://huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam">huggingface.co/spaces/FINAL-Bench/ViKernel-RedTeam</a></p><p><em>Questions: the Community tab of the Space.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=58bd8bccdd1e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[ZDNet Korea: VIDRAFT’s 180B Darwin Outscores a 2.8]]></title>
            <link>https://arxivgpt.medium.com/zdnet-korea-vidrafts-180b-darwin-outscores-a-2-8-71043d934d94?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/71043d934d94</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 02:41:40 GMT</pubDate>
            <atom:updated>2026-11-17T02:41:40.206Z</atom:updated>
            <content:encoded><![CDATA[<h3>ZDNet Korea: VIDRAFT’s 180B Darwin Outscores a 2.8 Trillion Parameter Model in Math and Science Reasoning</h3><h4>TL;DR</h4><p>On November 17, 2026, <a href="https://zdnet.co.kr/view/?no=20261117112235">ZDNet Korea</a>, one of Korea’s main technology news outlets, reported that VIDRAFT’s large language model <a href="https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3">Darwin-180B-RSI</a> ranked first among open-weight models in the mathematical and scientific reasoning track of SuperCLUE’s November 2026 evaluation.</p><p>The headline of the article is the size gap. Darwin has 180 billion parameters. Moonshot AI’s Kimi K3.1 Pro has 2.8 trillion, about 16 times more. Darwin scored more than 10 points higher. Including closed models, Darwin placed second, directly behind OpenAI’s GPT-6.1 Sol.</p><figure><img alt="SuperCLUE November 2026 math and science reasoning results" src="https://cdn-images-1.medium.com/max/1024/0*pL9Oh_CHH3d_2ZoJ.jpg" /></figure><h4>What ZDNet Korea reported</h4><p>| Point | What the article says | | — -| — -| | Result | First among open-weight models and second overall in math and science reasoning | | Size | A 180 billion parameter model scoring more than 10 points above a 2.8 trillion parameter model | | Model synthesis | VIDRAFT analyzes the feed-forward network regions where knowledge is concentrated in two or more open models, extracts the strengths of each and synthesizes them into one model | | Self-improvement | The synthesized model generates its own solutions, checks the answers automatically and retrains only on the correct ones. No human-written solutions or reasoning traces are used | | Hugging Face records | Number one in 12 benchmark leaderboards, GPQA 94.44, full marks on AIME 2026 and HMMT | | Inference speed | 18,057 tokens per second on server GPUs and 417 tokens per second on an NPU | | On-device | POCKET, the light edition, runs on a single laptop | | Downloads | About 5 million cumulative downloads on Hugging Face, more than 1.2 million a month, and 1.19 million for POCKET in the last 30 days | | National project | VIDRAFT is a consortium member of the government’s security foundation model development project led by Naver Cloud, in charge of AI model safety diagnosis and risk response |</p><p>The article closes with a comment from CEO Kim Minsik: self-improvement, where the model writes problems, verifies them and corrects its own weights, together with synthesis that takes only the strengths of parent models, opened the way for a small model to beat a giant one.</p><h4>Two technologies behind the result</h4><p><strong>Model synthesis.</strong> Darwin starts from open models that are each strong in a different area. VIDRAFT finds where each parent stores its knowledge and builds one child model that carries those strengths together.</p><p><strong>Model-level recursive self-improvement (RSI).</strong> The child model then improves itself. It solves problems, an automatic checker confirms which answers are right, and only verified solutions go back into training. The loop runs again on the improved model. The improvement is in the weights themselves.</p><h4>The numbers</h4><p>Mathematical and scientific reasoning, SuperCLUE November 2026:</p><p>| Rank | Model | Developer | Open weights | Score | | — -| — -| — -| — -| — -| | 1 | GPT-6.1-Sol | OpenAI | No | 76.32 | | 2 | <strong>Darwin-180B-RSI-R3</strong> | <strong>VIDRAFT</strong> | <strong>Yes</strong> | <strong>73.68</strong> | | 3 | Gemini-4-Argon | Google | No | 71.05 | | 3 | DeepSeek-V4.1-Flash | DeepSeek | Yes | 71.05 | | 5 | Grok-4.7 | xAI | No | 68.42 |</p><p>SuperCLUE grades each model itself every month on private questions. Developers do not submit scores.</p><h4>A run of coverage</h4><p>ZDNet Korea’s report follows The Korea Economic Daily on November 16 and Seoul Shinmun on November 13. ZDNet Korea has covered VIDRAFT steadily this autumn: the 10 number-one records on Hugging Face (October 6), NPU training acceleration (October 13), pretraining on NPUs alone (October 14) and the in-model security technology that blocks information leaks (November 3).</p><h4>Links</h4><ul><li>ZDNet Korea article (Korean): <a href="https://zdnet.co.kr/view/?no=20261117112235">https://zdnet.co.kr/view/?no=20261117112235</a></li><li>SuperCLUE board: <a href="https://www.superclueai.com/homepage">https://www.superclueai.com/homepage</a></li><li>Model: <a href="https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3">https://huggingface.co/FINAL-Bench/Darwin-180B-RSI-R3</a></li><li>Laptop edition: <a href="https://huggingface.co/FINAL-Bench/POCKET-Darwin-180B-GGUF">https://huggingface.co/FINAL-Bench/POCKET-Darwin-180B-GGUF</a></li><li>Hugging Face official benchmark leaderboards map: <a href="https://huggingface.co/spaces/quantid/huggingface-official-benchmark-leaderboards">https://huggingface.co/spaces/quantid/huggingface-official-benchmark-leaderboards</a></li><li>VIDRAFT press list: <a href="https://vidraft.net/">https://vidraft.net</a></li></ul><h4>FAQ</h4><h4>Which open-weight model is first in SuperCLUE math and science reasoning?</h4><p>In the November 2026 evaluation, Darwin-180B-RSI from VIDRAFT is first among open-weight models with 73.68 points and second overall.</p><h4>How can a 180B model beat a 2.8T model?</h4><p>VIDRAFT combines two methods: synthesis of the strengths of several parent models, and recursive self-improvement on self-generated, automatically verified solutions.</p><h4>Can Darwin run on a laptop?</h4><p>Yes. POCKET, the light edition, runs on a single laptop and was downloaded 1.19 million times in the last 30 days.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=71043d934d94" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Eight National Weather Models, One Korean AI: Darwin Has the Lowest Error in All 28 Categories]]></title>
            <link>https://arxivgpt.medium.com/eight-national-weather-models-one-korean-ai-darwin-has-the-lowest-error-in-all-28-categories-acb48bcf0b61?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/acb48bcf0b61</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 01:53:50 GMT</pubDate>
            <atom:updated>2026-11-17T01:53:50.870Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Global Weather Model Leaderboard, 7-day maximum temperature" src="https://cdn-images-1.medium.com/max/1024/0*piz96mwrpo0w-qnU.png" /></figure><h4>TL;DR</h4><p>VIDRAFT compared its AI weather engine, <strong>Darwin-Prediction-Weather</strong>, with eight forecast models run by the European Centre for Medium-Range Weather Forecasts (ECMWF), the US National Oceanic and Atmospheric Administration (NOAA) and the national weather services of Japan, Canada, Germany and Norway. The comparison covers 24 cities, about nine months of 2026 and 1.23 million published forecasts, each checked against what actually happened.</p><p>Darwin had the lowest error in <strong>all 28 categories</strong>: four weather elements (daily maximum temperature, daily minimum temperature, wind speed, precipitation) times seven lead times (1 to 7 days).</p><ul><li>7-day maximum temperature error: <strong>2.08 C</strong>, 14.0% lower than ECMWF AIFS (2.42 C), the second most accurate model.</li><li>Darwin’s 7-day forecast is more accurate than the 6-day forecast of ECMWF AIFS and the 5-day forecast of ECMWF IFS.</li><li>The scores are public and update every day on the <a href="https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard">Global Weather Model Leaderboard</a>.</li></ul><h4>The models compared</h4><p>| Model | Operator | Type | | — -| — -| — -| | Darwin-Prediction-Weather | VIDRAFT, Korea | AI | | AIFS | ECMWF, Europe | AI | | IFS | ECMWF, Europe | Physics | | AIGFS | NOAA, United States | AI | | GFS | NOAA, United States | Physics | | ICON | German Weather Service | Physics | | GSM | Japan Meteorological Agency | Physics | | GEM | Environment and Climate Change Canada | Physics | | MET Nordic | Norwegian Meteorological Institute | Physics |</p><p>A physics model solves the equations of the atmosphere on a supercomputer. An AI model learns to forecast from past weather data.</p><h4>7-day maximum temperature: the full ranking</h4><p>Root mean square error over 4,896 cases, the same cities and the same days for every model, February 1 to November 11, 2026. A lower error means a more accurate forecast.</p><p>| Rank | Model | Operator | Type | 7-day error | Error above first place | | — -| — -| — -| — -| — -| — -| | 1 | <strong>Darwin-Prediction-Weather</strong> | VIDRAFT | AI | <strong>2.08 C</strong> | | | 2 | ECMWF AIFS | ECMWF | AI | 2.42 C | +16.2% | | 3 | NOAA AIGFS | NOAA | AI | 2.83 C | +36.1% | | 4 | ECMWF IFS | ECMWF | Physics | 2.86 C | +37.2% | | 5 | MET Nordic | Norway | Physics | 2.94 C | +41.2% | | 6 | JMA GSM | Japan | Physics | 3.10 C | +49.0% | | 7 | NOAA GFS | NOAA | Physics | 3.18 C | +52.6% | | 8 | GEM | Canada | Physics | 3.20 C | +53.8% |</p><p>The German ICON model does not publish a 7-day forecast, so it is not in the 7-day ranking.</p><h4>How many days ahead</h4><p>The gap is clearer when lead times are compared.</p><p>| Comparison | Darwin | Other model | | — -| — -| — -| | Next-day maximum temperature | 0.84 C | AIFS 1.36 C (Darwin 38.6% lower) | | Maximum temperature | 7-day forecast 2.08 C | AIFS 6-day forecast 2.14 C | | Maximum temperature | 7-day forecast 2.08 C | IFS 5-day forecast 2.21 C | | Minimum temperature | 7-day forecast 1.81 C | AIFS 5-day forecast 1.86 C | | Wind speed | 7-day forecast 4.36 km/h | AIFS 4-day forecast 4.50 km/h |</p><p>At the same accuracy, Darwin sees one day further than the European AI model and two days further than the most accurate supercomputer physics model.</p><h4>The other elements</h4><p>| Element, 7-day forecast | Darwin | ECMWF AIFS | Difference | | — -| — -| — -| — -| | Minimum temperature error | 1.81 C | 2.18 C | 16.7% lower | | Wind speed error | 4.36 km/h | 5.64 km/h | 22.9% lower | | Rain of 1 mm or more, correct calls, next day | 90.6% | 88.4% | | | Rain of 1 mm or more, correct calls, day 7 | 79.6% | 77.2% | |</p><p>Precipitation is the hardest element to forecast. Darwin was first at every lead time from 1 to 7 days there as well.</p><h4>AI forecasts and supercomputer physics models</h4><p>In the 7-day maximum temperature forecast, AIFS, the most accurate public AI model, had 15.3% less error than IFS, the most accurate physics model. Darwin had 27.1% less error than IFS. AI models were ahead of physics models in all four elements, and Darwin widened that lead the most.</p><h4>City by city</h4><ul><li>7-day wind speed: Darwin was first in all 24 cities.</li><li>7-day minimum temperature: first in 20 cities.</li><li>7-day maximum temperature: first in 18 cities.</li><li>Seoul, 7-day maximum temperature: 2.27 C against 2.62 C for AIFS, 13.5% lower.</li><li>Busan: 1.76 C against 2.16 C for AIFS, 18.7% lower.</li></ul><p>When the period is cut to the most recent 7, 14, 30 and 90 days, Darwin stays first for maximum temperature, minimum temperature and wind speed at the 1, 3, 5 and 7-day lead times in every window.</p><h4>WeatherBench 2</h4><p>VIDRAFT also submitted a full-year 2020 global forecast from the same engine to WeatherBench 2, the weather AI evaluation framework from Google Research. Scored with the same method as the public scoreboard, the forecast had a lower error than the previous best record on 39 of 45 metrics.</p><h4>How the comparison is done</h4><ol><li>Every model is scored on the forecast it actually published. Forecast records come from the public Open-Meteo archive.</li><li>Models are compared only on the same cities and the same days.</li><li>The reference is ERA5, the reanalysis dataset from ECMWF.</li><li>Temperature and wind speed are scored on each city’s local calendar day.</li><li>Temperature and wind speed are verified through November 11, 2026, and precipitation through November 10.</li><li>The leaderboard recalculates every day, and anyone can open it.</li></ol><h4>What the CEO said</h4><p>Kim Minsik, CEO of VIDRAFT, said weather forecasting has long belonged to national weather services and supercomputers, and that a Korean AI came first in all 28 categories in an open daily contest against those models under the same conditions. He said the company will supply forecasts that see further and more accurately to industries where weather decides cost, such as energy trading, agriculture and logistics.</p><h4>Links</h4><ul><li>Leaderboard: <a href="https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard">https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard</a></li><li>VIDRAFT: <a href="https://vidraft.net/">https://vidraft.net</a></li></ul><h4>FAQ</h4><h4>Which weather model is the most accurate in 2026?</h4><p>On the Global Weather Model Leaderboard, which compares nine models across 24 cities, Darwin-Prediction-Weather has the lowest error in all 28 categories for February 1 to November 11, 2026. ECMWF AIFS is second for 7-day maximum temperature.</p><h4>Are AI weather models more accurate than physics models?</h4><p>In this comparison, yes. AI models had lower error than supercomputer physics models in all four elements at the 7-day lead time.</p><h4>Where does the data come from?</h4><p>Forecasts come from the Open-Meteo public archive and the reference values from the ERA5 reanalysis.</p><p><em>Originally published at </em><a href="https://vidraft-ai-media.static.hf.space/a/weather-eight-national-models-28-categories-2026-11-17.html"><em>VIDRAFT AI Media</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=acb48bcf0b61" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Do Not Upload Confidential Research Data to ChatGPT: Darwin Lab, a Private AI Research Lab in an…]]></title>
            <link>https://arxivgpt.medium.com/do-not-upload-confidential-research-data-to-chatgpt-darwin-lab-a-private-ai-research-lab-in-an-2c3c897fe354?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/2c3c897fe354</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Sun, 15 Nov 2026 07:59:26 GMT</pubDate>
            <atom:updated>2026-11-15T07:59:26.191Z</atom:updated>
            <content:encoded><![CDATA[<h3>Do Not Upload Confidential Research Data to ChatGPT: Darwin Lab, a Private AI Research Lab in an NVIDIA DGX Spark and a NAS</h3><figure><img alt="Darwin Lab architecture" src="https://cdn-images-1.medium.com/proxy/0*ov_gJEZBA2n5imfR.webp" /></figure><h4>TL;DR</h4><p><a href="https://vidraft.net/p/hub-science.html">Darwin Lab</a> is a private AI research lab that lives inside your own lab. VIDRAFT built the Darwin model and a full science research platform into an NVIDIA DGX Spark mini supercomputer and a NAS. It runs without an internet connection, so research data stays in the building.</p><ul><li><strong>Verified reasoning.</strong> In SuperCLUE’s November 2026 evaluation, Darwin-180B-RSI scored 73.7 in math and science reasoning. That is second overall, behind GPT-6.1-Sol at 76.3, and first among open models. It is ahead of Gemini-4-Argon (71.0), DeepSeek-V4.1-Flash (71.0) and Grok-4.7 (68.4).</li><li><strong>Installable.</strong> The models above Darwin and several below it are closed services. Among models a lab can install on its own hardware, Darwin ranks first.</li><li><strong>An organization, not only a model.</strong> Five executive roles, twelve team leads and ten harnesses form hypotheses, run experiments, verify each other’s results and record everything.</li><li><strong>Used in real research.</strong> VIDRAFT used the platform to find a new drug candidate compound and filed a substance patent. Lab validation is under way.</li></ul><p>Watch the three minute introduction on <a href="https://www.youtube.com/watch?v=HBgV1d12dk8">YouTube</a> or open the <a href="https://vidraft.net/science-onprem/index.html">live demo</a>.</p><h4>Is it safe to upload confidential research data to a cloud AI?</h4><p>A drug candidate, unpublished measurements and data that precede a patent filing cannot be sent to an outside service. Once the data leaves, it cannot be pulled back, and the principal investigator carries the responsibility. Many labs therefore kept their most important work away from the strongest AI systems.</p><p>Darwin Lab answers that with a simple move: bring the AI into the lab.</p><h4>What is inside Darwin Lab?</h4><p>| Layer | What it does | | — -| — -| | NVIDIA DGX Spark | Desktop mini supercomputer with 128GB unified memory. Runs the model, engines and harnesses. | | NAS | Holds the base science library, the lab’s own data, long-term memory, backups and signed updates. | | Darwin-180B-RSI Science | The reasoning model, built by VIDRAFT. | | 10 engines | Ideation, reasoning, analysis and judging. | | 10 harnesses | Autonomous research, coding, experiment, literature search, data analysis, verification, writing, self-improvement, data intake and web operation. | | AI-native organization | Five executive roles and twelve team leads that share one long-term memory. |</p><p>A researcher writes a one-line request. The platform reads its long-term memory, plans, forms a hypothesis, runs the experiment or computation, sends the result to an independent check and writes the record and a report draft.</p><h4>How is Darwin Lab different from document search (RAG) products?</h4><p>Most on-premise AI products attach document search to a public model. They find documents and summarize them. Darwin Lab differs in three ways.</p><ol><li>VIDRAFT built the model itself.</li><li>A third party verified the model’s reasoning.</li><li>The platform does research. It forms hypotheses, runs experiments, discovers and verifies, and one team’s result is checked again by another team before it reaches the researcher.</li></ol><h4>What happens to the data a lab uploads?</h4><p>Upload papers, lab records and measurements. The platform reads and classifies them and builds long-term memory from them. A base science library ships with the product: public literature metadata, chemistry, protein, omics, materials and math resources whose terms of use were reviewed source by source. Every item is pinned to a version with its source, hash and terms recorded, so the same question returns the same answer.</p><h4>FAQ</h4><p><strong>Which installable model is strongest in math and science reasoning?</strong> In SuperCLUE’s November 2026 evaluation, Darwin-180B-RSI scored 73.7, second overall and first among open models. The model above it, GPT-6.1-Sol, is a closed service.</p><p><strong>Does Darwin Lab work without internet?</strong> Yes. The model and the whole platform run inside the DGX Spark and the NAS.</p><p><strong>Which fields is it for?</strong> Drug discovery, chemistry and materials, biology, physics, mathematics and computational science.</p><p><strong>How do I use it?</strong> Write a research request in one line. The platform runs literature search, experiment design, computation and a report draft in order.</p><p><strong>Can I try it?</strong> Open the <a href="https://vidraft.net/science-onprem/index.html">live demo</a> or request a free demo on the <a href="https://vidraft.net/p/hub-science.html">product page</a>.</p><h4>About VIDRAFT</h4><p>VIDRAFT is a Korean deep-tech AI company. Its open models have passed 4.6 million downloads on Hugging Face. The model, engines and harnesses in Darwin Lab are all developed in-house.</p><h4>Links</h4><ul><li>Product page: <a href="https://vidraft.net/p/hub-science.html">vidraft.net/p/hub-science.html</a></li><li>Live demo: <a href="https://vidraft.net/science-onprem/index.html">vidraft.net/science-onprem</a></li><li>Introduction video: <a href="https://www.youtube.com/watch?v=HBgV1d12dk8">YouTube</a></li><li>Free demo requests: <a href="https://vidraft.net/p/contact.html">vidraft.net</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2c3c897fe354" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Which Open LLMs Are Downloaded Now? A Daily Leaderboard of 30-Day Downloads by Country]]></title>
            <link>https://arxivgpt.medium.com/which-open-llms-are-downloaded-now-a-daily-leaderboard-of-30-day-downloads-by-country-3acbdc461fc7?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/3acbdc461fc7</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 08:29:29 GMT</pubDate>
            <atom:updated>2026-11-14T08:29:29.987Z</atom:updated>
            <content:encoded><![CDATA[<h4>TL;DR</h4><p>The <a href="https://huggingface.co/spaces/VIDraft/global-llm-leaderboard">Global LLM Download Leaderboard</a> ranks open language models on Hugging Face by what people download now: trailing 30-day downloads, not lifetime totals. It lists the global top 300 and the Korea top 30, and refreshes every day. In the 14 November 2026 snapshot, the top two Korean models are both from VIDRAFT: POCKET-Darwin-180B-GGUF and POCKET-35B-GGUF.</p><h4>Why 30 days and not all-time</h4><p>Lifetime download counts reward age. A model released in 2022 keeps a large total long after people have moved on. The 30-day window shows what is being used this month. The board prints both numbers on every row so the difference is visible.</p><h4>The global picture</h4><p>The top 300 models account for 331,761,401 downloads in the last 30 days.</p><p>| Trained in | Share of top-300 downloads | Models in the top 300 | | — -| — -| — -| | China | 61.7% | 175 | | United States | 32.3% | 89 | | Europe | 1.4% | 8 | | Korea | 0.4% | 3 |</p><p>Country is attributed by model family, meaning who trained the model, not who uploaded the file. A third-party quantization of a model counts toward the family that trained the original. The family is resolved from the model architecture, because names can be changed and architectures cannot.</p><p>| # | Model | 30-day downloads | All-time | | — -| — -| — -| — -| | 1 | openai-community/gpt2 | 33,264,897 | 973,438,339 | | 2 | Qwen/Qwen3–0.6B | 19,141,315 | 262,961,607 | | 3 | Qwen/Qwen3–14B | 13,011,347 | 47,921,765 | | 4 | deepseek-ai/DeepSeek-V3.2 | 8,442,070 | 37,293,704 | | 5 | Qwen/Qwen2.5–7B-Instruct | 6,444,573 | 195,041,610 | | 6 | Qwen/Qwen3–8B | 6,165,334 | 138,350,815 | | 7 | ornith-ai/Ornith-1.5–9B-GGUF | 5,813,219 | 16,078,083 | | 8 | openai/gpt-oss-20b | 5,673,703 | 109,217,475 | | 9 | Qwen/Qwen2.5–1.5B-Instruct | 5,145,856 | 228,576,027 | | 10 | dphn/dolphin-2.9.1-yi-1.5–34b | 4,836,827 | 86,609,232 |</p><h4>Korea top 10</h4><p>| # | Model | 30-day downloads | All-time | | — -| — -| — -| — -| | 1 | FINAL-Bench/POCKET-Darwin-180B-GGUF | 523,910 | 612,374 | | 2 | FINAL-Bench/POCKET-35B-GGUF | 435,640 | 2,081,893 | | 3 | LGAI-EXAONE/EXAONE-3.5–7.8B-Instruct | 402,209 | 4,380,333 | | 4 | FINAL-Bench/POCKET-26B-GGUF | 204,236 | 925,149 | | 5 | LG-AI-Research/EXAONE-Tabular | 116,382 | 293,772 | | 6 | upstage/Solar-Open-100B | 88,704 | 604,185 | | 7 | skt/A.X-K2-NVFP4 | 81,947 | 320,538 | | 8 | LGAI-EXAONE/EXAONE-3.5–2.4B-Instruct | 71,723 | 2,106,317 | | 9 | skt/A.X-4.0-Light | 54,247 | 584,508 | | 10 | LGAI-EXAONE/EXAONE-3.5–32B-Instruct | 52,502 | 1,245,449 |</p><p>Three Korean models are in the global top 300: POCKET-Darwin-180B-GGUF at #140, POCKET-35B-GGUF at #166 and EXAONE-3.5–7.8B-Instruct at #182. VIDRAFT models hold first, second and fourth place in Korea and account for 45.4% of the Korea top-30 downloads.</p><p>POCKET-Darwin-180B-GGUF is a 180-billion-parameter reasoning model packaged to run on a laptop. It was released on 2 October 2026 and reached 523,910 downloads in the last 30 days.</p><h4>What the board leaves out</h4><p>Test fixtures used by automated pipelines are removed, because they collect large download counts that no person produced. Embedding and reranker repositories are removed too: the platform tags many of them as text generation, but they do not generate text.</p><h4>Use it</h4><ul><li>Leaderboard: <a href="https://huggingface.co/spaces/VIDraft/global-llm-leaderboard">https://huggingface.co/spaces/VIDraft/global-llm-leaderboard</a></li><li>Homepage: <a href="https://vidraft.net/">https://vidraft.net</a> (the leaderboard is now the button at the right end of the menu)</li><li>POCKET-Darwin-180B-GGUF: <a href="https://huggingface.co/FINAL-Bench/POCKET-Darwin-180B-GGUF">https://huggingface.co/FINAL-Bench/POCKET-Darwin-180B-GGUF</a></li></ul><p>Data comes from the Hugging Face API on the date printed on the page. The board is released under CC BY 4.0.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3acbdc461fc7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Which Weather Model Is Most Accurate? We Score Them Every Day, in the Open]]></title>
            <link>https://arxivgpt.medium.com/which-weather-model-is-most-accurate-we-score-them-every-day-in-the-open-2a1093296f2d?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/2a1093296f2d</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Fri, 13 Nov 2026 10:33:50 GMT</pubDate>
            <atom:updated>2026-11-13T10:33:50.071Z</atom:updated>
            <content:encoded><![CDATA[<h4>TL;DR</h4><p>VIDRAFT opened two things on Hugging Face: <a href="https://huggingface.co/FINAL-Bench/Darwin-Prediction-Weather">Darwin-Prediction-Weather</a>, a weather prediction system, and the <a href="https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard">Global Weather Model Leaderboard</a>, a public board that scores the world’s forecast models against what actually happened, every day. In the 2026 record, Darwin-Prediction-Weather has the lowest error in all four scored elements at every lead time from 1 to 7 days. The full write-up is on the <a href="https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them">Hugging Face blog</a>.</p><figure><img alt="Champion panel of the Global Weather Model Leaderboard" src="https://cdn-images-1.medium.com/max/1024/0*v4_T2ige5het-bQx.png" /></figure><h4>The question: AI or physics?</h4><p>For decades, forecasts came from physics models that solve the equations of the atmosphere on supercomputers. In the last three years, AI models trained on historical weather have joined them at the world’s weather centres. Which one should you trust for tomorrow, for three days ahead, for a week ahead? Published papers answer that for past years. This board answers it for this year, updated daily, with the data open for anyone to check.</p><h4>How the board scores</h4><p>| Item | Rule | | — -| — -| | Models | ECMWF IFS and AIFS, NOAA GFS and AIGFS, DWD ICON, JMA GSM, ECCC GEM, MET Norway, Darwin-Prediction-Weather | | Elements | Daily maximum temperature, daily minimum temperature, daily maximum wind speed, daily precipitation total | | Where | 24 cities on six continents, up to 10 days ahead | | Truth | ERA5 reanalysis, Copernicus Climate Change Service | | Score | Root-mean-square error, lower is better | | Fairness | Every model is scored on the same cases: days and cities where all listed models issued a forecast | | Record | From 1 January 2026, updated once a day |</p><figure><img alt="Accuracy board" src="https://cdn-images-1.medium.com/max/1024/0*wBocJymvA3G1xK5j.png" /></figure><h4>What the 2026 record shows</h4><p>Two findings stand out.</p><p><strong>AI models now lead.</strong> Among the models of the national and international weather centres, ECMWF’s AI model AIFS has lower error than ECMWF’s physics model IFS at every lead time, and NOAA’s AIGFS has lower error than NOAA’s GFS at every lead time.</p><p><strong>Fusion goes further.</strong> Darwin-Prediction-Weather has the lowest error in all four elements, at every lead time from 1 to 7 days.</p><p>RMSE against ERA5, 24 cities, 2026 year to date (verified through 2026–11–07). The value in brackets is the next model, ECMWF AIFS in every cell.</p><p>| Element | 1 day | 3 days | 5 days | 7 days | | — -| — -| — -| — -| — -| | Daily maximum temperature [°C] | <strong>0.84</strong> (1.35) | <strong>1.17</strong> (1.59) | <strong>1.59</strong> (1.90) | <strong>2.09</strong> (2.41) | | Daily minimum temperature [°C] | <strong>0.91</strong> (1.50) | <strong>1.09</strong> (1.63) | <strong>1.42</strong> (1.85) | <strong>1.81</strong> (2.17) | | Daily maximum wind speed [km/h] | <strong>2.25</strong> (3.60) | <strong>2.70</strong> (4.07) | <strong>3.43</strong> (4.76) | <strong>4.39</strong> (5.69) | | Daily precipitation total [mm] | <strong>4.15</strong> (4.33) | <strong>5.18</strong> (5.24) | <strong>6.80</strong> (6.99) | <strong>7.11</strong> (7.30) |</p><figure><img alt="Error by forecast day and city map" src="https://cdn-images-1.medium.com/max/1024/0*sankScClEGSf_TRw.png" /></figure><h4>Who wins where</h4><p>A single global number hides a lot. The board breaks the record down by period, by lead time and by city. For every city it also lays out a rolling timeline: the last seven days with the actual value and the model that came closest, today so far, and the forecast for the next ten days. Each morning yesterday’s forecast becomes today’s result, and it is scored.</p><figure><img alt="Champions by period and lead time" src="https://cdn-images-1.medium.com/max/1024/0*VPaw-hl_AqUe0WQn.png" /></figure><h4>How Darwin-Prediction-Weather works</h4><p>Darwin is VIDRAFT’s family of models built on one idea: combine the strengths of several strong models into one that is stronger than any of them. Darwin-Prediction-Weather applies that idea to the atmosphere.</p><ul><li><strong>Forecast engine.</strong> A single forecast from skill-weighted fusion of global forecasts, with local correction for temperature and wind. It learns only from days whose observed value was already known before the forecast it is scored on. Nothing from the future enters.</li><li><strong>Decision core.</strong> A Darwin-27B model with a calibrated readout turns forecast fields into decisions: go or no-go, threshold exceedance, category probabilities.</li></ul><p>The same engine was evaluated on WeatherBench 2, the standard benchmark for AI weather models run by Google Research. On the 2020 deterministic protocol it records the lowest error in all 15 evaluated cells: 2 m temperature, 500 hPa geopotential and 850 hPa temperature at 1, 3, 5, 7 and 10 days. The forecast files are public and the entry has been submitted to the WeatherBench 2 team.</p><p>| WeatherBench 2, 2020 | 1 day | 3 days | 5 days | 7 days | 10 days | | — -| — -| — -| — -| — -| — -| | 2 m temperature [K] | <strong>0.468</strong> | <strong>0.743</strong> | <strong>1.107</strong> | <strong>1.531</strong> | <strong>1.984</strong> | | best on official scorecard | 0.488 | 0.786 | 1.125 | 1.552 | 2.022 | | Z500 [m2/s2] | <strong>33.7</strong> | <strong>108.5</strong> | <strong>240.6</strong> | <strong>400.3</strong> | <strong>595.4</strong> | | best on official scorecard | 36.8 | 113.8 | 250.7 | 415.1 | 606.5 | | T850 [K] | <strong>0.503</strong> | <strong>0.901</strong> | <strong>1.406</strong> | <strong>2.000</strong> | <strong>2.699</strong> | | best on official scorecard | 0.518 | 0.922 | 1.435 | 2.041 | 2.737 |</p><h4>Everything is open to check</h4><ul><li>The board: <a href="https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard">https://huggingface.co/spaces/FINAL-Bench/Weather-Model-Leaderboard</a></li><li>The daily record of every forecast and observed value since 2026–01–01, as CSV: <a href="https://huggingface.co/datasets/FINAL-Bench/Weather-Model-Forecast-Archive">https://huggingface.co/datasets/FINAL-Bench/Weather-Model-Forecast-Archive</a></li><li>The model: <a href="https://huggingface.co/FINAL-Bench/Darwin-Prediction-Weather">https://huggingface.co/FINAL-Bench/Darwin-Prediction-Weather</a></li><li>WeatherBench 2 forecast files: <a href="https://huggingface.co/datasets/VIDraft/Darwin-Prediction-Weather-WB2">https://huggingface.co/datasets/VIDraft/Darwin-Prediction-Weather-WB2</a></li><li>Full article: <a href="https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them">https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them</a></li></ul><h4>What comes next</h4><p>Forecasts 8 to 10 days ahead enter the ranking as their target days are verified. Scoring against station observations will run alongside ERA5. More cities and more elements follow, and a forecasting game on Hugging Face will let anyone test their own forecast against the models.</p><h4>License and sources</h4><p>The leaderboard application is released under the Apache License 2.0. The compiled daily record and the Darwin-Prediction-Weather model are released under CC BY-NC 4.0, non-commercial use only. For commercial use, contact vidraft@mail.vidraft.net.</p><p>Weather data by Open-Meteo.com (CC BY 4.0). Verification: ERA5, Copernicus Climate Change Service. Model sources: ECMWF, NOAA NCEP, DWD, JMA, ECCC and MET Norway. Satellite imagery: NASA Global Imagery Browse Services (GIBS).</p><p>Originally published on the Hugging Face blog: <a href="https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them">https://huggingface.co/blog/FINAL-Bench/which-weather-model-is-most-accurate-we-score-them</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2a1093296f2d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[VIDRAFT Unveils ViKernel: An LLM OS with Built-In Kill Switch]]></title>
            <link>https://arxivgpt.medium.com/vidraft-unveils-vikernel-an-llm-os-with-built-in-kill-switch-97f107d7b93c?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/97f107d7b93c</guid>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[oss-llm]]></category>
            <category><![CDATA[vikernel]]></category>
            <category><![CDATA[vidraft]]></category>
            <category><![CDATA[weight-native]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 23:00:42 GMT</pubDate>
            <atom:updated>2026-11-11T23:00:42.884Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Korean AI startup VIDRAFT embeds computation circuits and security controls directly inside model weights — a fundamental shift in how AI safety is enforced.</em></p><blockquote><strong>TL;DR:</strong> VIDRAFT, a Korean AI startup based at Seoul AI Hub, has publicly released ViKernel, a Weight-Native OS that implements computation circuits and a kill switch inside large language model weights rather than as external add-ons. The system supports policy-based task control, tamper-resistant audit logging, and requires a signed operator key to restart after a kill-switch event. A live demo is available on Hugging Face.</blockquote><p>VIDRAFT, the Korean AI startup behind the AX-RAY AI safety diagnostic platform, announced on November 11, 2026 the public release of <strong>ViKernel</strong> — a first-of-its-kind operating system for large language models (LLMs) that embeds both computational circuitry and security controls directly into model weights. The announcement marks a significant departure from the conventional approach of bolting safety mechanisms onto the outside of AI models.</p><h4>What VIDRAFT Announced</h4><p>ViKernel is built on what VIDRAFT calls a <strong>Weight-Native OS</strong> architecture. Rather than attaching security or computation modules to a model from the outside, ViKernel implements the equivalent of a CPU, NPU, RAM, and ROM entirely within the model’s weight structure. When the model encounters a task requiring precise calculation, it delegates that work to its internal circuits and returns a verified result — instead of relying on the LLM’s native, often error-prone reasoning.</p><p>VIDRAFT demonstrated the system’s arithmetic accuracy on a 27-billion-parameter proprietary model, successfully multiplying two 300-digit numbers to produce a roughly 600-digit result, with the output independently verified by an external calculator. The company also reported transplanting the circuitry into other model families, including Google’s Gemma and Alibaba’s Qwen.</p><p>On the security side, ViKernel performs an integrity check each time the AI is launched. If the model’s weights have been altered, the system halts execution automatically. A built-in policy engine allows operators to permit or block specific tasks, and every action is recorded in an audit log designed to resist tampering. A kill switch can stop the AI entirely; resuming operations afterward requires a cryptographically signed operator key.</p><p>VIDRAFT CEO Min-sik Kim described the dual mission this way: “If AI is going to handle important work, it must never miscalculate, and it must be able to stop when things get dangerous. ViKernel puts both of those capabilities inside the model itself.” He added that the company intends to “open an era in which a Korean-made OS lives inside AI.”</p><p>A public demo on Hugging Face lets users explore the kill switch, policy engine, personal-data masking, and audit-logging features. VIDRAFT notes that the demo replays pre-recorded responses generated by the actual OS rather than running the full system in real time.</p><p>VIDRAFT is also a participant in <strong>K-MYTHOS</strong>, a national security-focused AI development program led by Naver Cloud. Within that consortium, the company is responsible for AI safety diagnosis and evaluation using its AX-RAY technology. VIDRAFT plans to integrate AX-RAY’s external risk-assessment capabilities with ViKernel’s internal task-control mechanisms into a unified safety pipeline.</p><h4>Why It Matters</h4><p>The dominant paradigm for AI safety today relies on guardrails, filters, and monitoring tools that sit outside the model — wrappers that can, in principle, be bypassed or stripped away. ViKernel’s weight-native approach makes the security layer inseparable from the model itself: you cannot remove the kill switch without altering the weights, and altering the weights triggers a shutdown.</p><p>This matters for enterprise and government deployments where auditability, computational accuracy, and emergency shutdown are non-negotiable requirements. The 300-digit multiplication demonstration addresses a well-documented weakness of LLMs — arithmetic reliability — by routing calculations through dedicated internal circuits rather than through probabilistic text prediction.</p><p>The integration pathway with K-MYTHOS also signals that ViKernel is not purely a research prototype. With Naver Cloud leading a national security AI consortium and VIDRAFT holding a formal role in AI safety evaluation, there is a clear route toward government-grade deployment.</p><h4>Key Takeaways</h4><ul><li><strong>Weight-native security:</strong> ViKernel is the first publicly announced OS to implement CPU/NPU/memory-equivalent circuits and a kill switch directly inside LLM model weights.</li><li><strong>Verified arithmetic accuracy:</strong> VIDRAFT demonstrated exact multiplication of two 300-digit numbers on its 27-billion-parameter model, with results confirmed by an external calculator.</li><li><strong>Policy control and audit trail:</strong> The system enforces operator-defined task policies and maintains a tamper-resistant audit log; restarting after a kill-switch event requires a signed key.</li><li><strong>Cross-model portability:</strong> The circuits have been ported to external model families including Google Gemma and Alibaba Qwen.</li><li><strong>National security context:</strong> VIDRAFT is an active member of the K-MYTHOS consortium, led by Naver Cloud, where it handles AI safety diagnostics via AX-RAY.</li></ul><h4>Frequently Asked Questions</h4><p><strong>Q: What is VIDRAFT’s ViKernel?</strong> A: ViKernel is a Weight-Native OS developed by Korean AI startup VIDRAFT that embeds computation circuits and security controls — including a kill switch and policy engine — directly into the weights of a large language model, rather than as external add-ons.</p><p><strong>Q: How does the ViKernel kill switch work?</strong> A: The kill switch can halt all AI operations in an emergency. To resume activity after a shutdown, an operator must supply a cryptographically signed key, preventing unauthorized restart.</p><p><strong>Q: Where can I try ViKernel?</strong> A: VIDRAFT has published a demo on Hugging Face that lets users interact with features such as the kill switch, policy engine, personal-data masking, and audit logging. The demo replays pre-recorded OS-generated responses rather than running the full system live.</p><p><em>Source: IT조선 (2026–11–11) — </em><a href="https://it.chosun.com/news/articleView.html?idxno=3000000000737"><em>original article</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=97f107d7b93c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[VIDRAFT and GeniGenAI Launch World’s First AI Metacognition Leaderboard]]></title>
            <link>https://arxivgpt.medium.com/vidraft-and-genigenai-launch-worlds-first-ai-metacognition-leaderboard-d487550ee8c2?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/d487550ee8c2</guid>
            <category><![CDATA[ai-benchmarks]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[ai-metacognition]]></category>
            <category><![CDATA[genigenai]]></category>
            <category><![CDATA[vidraft]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 11:00:41 GMT</pubDate>
            <atom:updated>2026-11-10T11:00:41.888Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A new benchmark redefines AI trustworthiness — from getting answers right to knowing when you might be wrong.</em></p><blockquote><strong>TL;DR:</strong> VIDRAFT, a Korean Pre-AGI AI startup, has co-developed a metacognition benchmark, a 24-model leaderboard, and 11 model-specific metacognition adapters in partnership with GeniGenAI. The benchmark uses 400 trap-style questions to test whether AI models can detect their own potential errors — not just produce correct answers. All tools are publicly available on Hugging Face.</blockquote><p>The Korean AI startup VIDRAFT has partnered with GeniGenAI (CEO Sunyoung Choi) to release a suite of tools designed to measure and enhance metacognition in AI — a capability the two companies argue is essential for any system aspiring to reach AGI. Announced on July 2, 2026, the joint release includes a novel benchmark, a leaderboard ranking 24 AI models, and adapter modules for 11 of those models.</p><h4>What VIDRAFT Announced</h4><p>VIDRAFT and GeniGenAI’s collaboration centers on a core argument: the most dangerous AI is not one that makes mistakes, but one that doesn’t know it has made them. Metacognition — the ability of a model to self-assess whether its own answer might be wrong — is the capability they set out to measure and improve.</p><p>The benchmark at the heart of this release consists of 400 deliberately deceptive questions. Each problem is constructed so that a plausible-looking answer is actually incorrect, forcing models to look beyond surface-level reasoning and critically examine their own conclusions. Evaluation is split into two dimensions: accuracy on multiple-choice questions and the ability to flag self-error likelihood in open-ended responses.</p><p>The resulting leaderboard covers 24 AI models. The top-ranked model is JGOS-31B-Citizen, a Korean-language reasoning model that answered all but two of the 400 questions correctly, achieving a trap-avoidance rate of 99.5%. JGOS-31B-Citizen is also the current top-ranked model on the K-AI Leaderboard operated by South Korea’s Ministry of Science and ICT and the National Information Society Agency (NIA). Several models from the Korean-developed Darwin series — including Darwin-31B, Darwin-35B, and Darwin-28B — also placed in the upper tier of the rankings.</p><p>However, the benchmark revealed a critical gap. Even models that performed strongly on multiple-choice accuracy struggled in the open-ended portion: they could select the right answer but were limited in their ability to independently identify when their reasoning might be flawed. This gap, GeniGenAI and VIDRAFT contend, illustrates that answer accuracy and self-error detection are distinct, separately measurable capabilities.</p><p>To address that gap, VIDRAFT co-developed the metacognition adapter — a lightweight module that attaches alongside an existing AI model without modifying its underlying architecture or weights. The adapter analyzes internal signals generated during the response process to calculate the likelihood that a given answer is incorrect, then surfaces that risk signal independently of the original model’s output. Eleven such adapters, each tailored to a specific model, are included in the release. All benchmark materials, the leaderboard, and the adapters are available on Hugging Face.</p><h4>Why It Matters</h4><p>The release arrives at a moment when AI reliability in high-stakes domains is under intense scrutiny. GeniGenAI CEO Sunyoung Choi framed the significance plainly: a truly trustworthy AI is not one that maximizes its correct-answer count, but one that recognizes the possibility of its own error and can pause accordingly. That framing has direct implications for sectors where the cost of an undetected mistake is severe — medicine, law, finance, cybersecurity, and scientific research among them.</p><p>The conventional way to evaluate AI reliability has been accuracy rate. This benchmark challenges that convention by treating self-assessment as a first-class metric. If the approach gains traction in the broader AI research community, it could shift how developers, regulators, and enterprises evaluate model readiness for deployment in sensitive contexts.</p><p>The non-invasive adapter architecture is also notable from an engineering standpoint. Because it leaves the original model’s structure and weights untouched, organizations can enhance an existing deployment’s error-signaling capability without retraining or replacing the underlying system — a potentially significant advantage for teams working under resource or compliance constraints.</p><h4>Key Takeaways</h4><ul><li><strong>VIDRAFT and GeniGenAI jointly released</strong> a metacognition benchmark (400 trap questions), a leaderboard ranking 24 AI models, and 11 model-specific metacognition adapters — all publicly accessible on Hugging Face.</li><li><strong>JGOS-31B-Citizen ranked first</strong> among all 24 evaluated models, achieving a 99.5% trap-avoidance rate and holding the top position on South Korea’s official K-AI Leaderboard as well.</li><li><strong>High multiple-choice accuracy does not guarantee metacognitive ability</strong> — the benchmark demonstrated that self-error detection is a distinct capability that must be measured and developed independently of raw answer accuracy.</li></ul><h4>Frequently Asked Questions</h4><p><strong>Q: What is the metacognition benchmark that VIDRAFT co-developed?</strong> A: It is a 400-question evaluation suite made up of trap-style problems designed to assess whether an AI model can recognize when its own answer is likely wrong, rather than simply measuring how often it selects the correct answer.</p><p><strong>Q: How does the VIDRAFT metacognition adapter work?</strong> A: The adapter attaches to an existing AI model without changing its architecture or weights. It analyzes internal signals produced during response generation and separately calculates the probability that the model’s answer is incorrect.</p><p><strong>Q: Where can developers access the benchmark, leaderboard, and adapters?</strong> A: All three resources — the metacognition benchmark, the 24-model leaderboard, and the 11 model-specific adapters — are available publicly on Hugging Face.</p><p><em>Source: AI이코노미뉴스 (2026–07–02) — </em><a href="https://www.aienews.kr/news/articleView.html?idxno=1162"><em>original article</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d487550ee8c2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[VIDRAFT’s Darwin-180B-RSI Tops Open-Weight AI Rankings in Math]]></title>
            <link>https://arxivgpt.medium.com/vidrafts-darwin-180b-rsi-tops-open-weight-ai-rankings-in-math-43390f3b3fe4?source=rss-102df057f14f------2</link>
            <guid isPermaLink="false">https://medium.com/p/43390f3b3fe4</guid>
            <category><![CDATA[open-weight-ai]]></category>
            <category><![CDATA[ai-math-ranking]]></category>
            <category><![CDATA[superclue-benchmark]]></category>
            <category><![CDATA[darwin-180b]]></category>
            <category><![CDATA[vidraft]]></category>
            <dc:creator><![CDATA[ArXiv]]></dc:creator>
            <pubDate>Tue, 10 Nov 2026 07:00:44 GMT</pubDate>
            <atom:updated>2026-11-10T07:00:44.758Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The Korean startup’s open model outscored Gemini and DeepSeek on SuperCLUE’s November math and science leaderboard.</em></p><blockquote><strong>TL;DR:</strong> VIDRAFT’s open-weight model Darwin-180B-RSI-R3 scored 73.68 in math and science on the November 2026 SuperCLUE benchmark, placing second among all 19 evaluated models and first among all open-weight models. The model beat both Gemini-4-Argon and DeepSeek-V4.1-Flash, which each scored 71.05, and is the only Korean model in the ranking. Weights are publicly available on Hugging Face under the Qwen Community 1.0 license.</blockquote><p>Korean AI startup VIDRAFT made headlines in the global AI community on November 9, 2026, when Russian tech outlet USB.RU reported that its Darwin-180B-RSI-R3 model had claimed the top spot among open-weight models in the November SuperCLUE math and science rankings — and second place overall out of 19 competing models.</p><h4>What VIDRAFT Announced</h4><p>Darwin-180B-RSI-R3 achieved a score of 73.68 in the math and science category of the November 2026 SuperCLUE benchmark. Only GPT-6.1-Sol ranked higher, with a score of 76.32. Behind Darwin came Gemini-4-Argon and DeepSeek-V4.1-Flash, both scoring 71.05, while Kimi-K3.1-Pro registered 63.16 in the same category.</p><p>Darwin is the sole Korean model in the SuperCLUE table and the strongest open-weight model in math and science among all models evaluated that month — a notable distinction at a time when several major laboratories also released open weights within the same week.</p><p>The model is built on top of Qwen3.8-Flash-Next and refined through a process VIDRAFT calls RSI, or Recursive Self-Improvement. In this approach, the model solves problems, checks its answers against solution keys, and continues training exclusively on correct reasoning chains before repeating the cycle. Crucially, only the attention pathways and shared experts are updated during this process; the full set of 512 experts, the router, and the visual encoder remain unchanged.</p><p>The performance gains compared to the parent model are substantial. The base Qwen3.8-Flash-Next scored 57.89 in math and science and 39.81 overall on SuperCLUE, while Darwin-180B-RSI-R3 reached 73.68 and 43.08, respectively. On MMLU-Pro, average reasoning chain length dropped from 4,320 tokens to 3,833 tokens — an 11% reduction — suggesting the model became more efficient, not just more accurate.</p><p>Additional benchmark results listed on the model card include a GPQA Diamond score of 94.44, MMLU-Pro at 88.12, and perfect scores of 100 on both AIME 2026 and HMMT February 2026. These figures come from the model card and the SuperCLUE table; independent third-party verification of GPQA and other scores has not yet been completed, as USB.RU notes.</p><p>Darwin’s overall SuperCLUE score of 43.08 places it ninth out of 19 models in the general ranking, revealing some clear gaps. Its instruction-following score is 20.69 and its long-context score is 30.32, both noticeably below the benchmark’s top performers. VIDRAFT has not publicly addressed these weaknesses in the current release.</p><p>The model weights are open-sourced under the Qwen Community 1.0 license and hosted in the FINAL-Bench/Darwin-180B-RSI repository on Hugging Face. A 4-bit quantized build called POCKET-Darwin-180B-GGUF weighs 111 GB and can run on a laptop with an 8 GB GPU and 32 GB of RAM. Even without a GPU, running on CPU alone, it delivers 18–21 tokens per second.</p><h4>Why Darwin-180B-RSI Matters</h4><p>The Darwin-180B-RSI result is significant for several reasons. First, it demonstrates that a targeted, iterative self-improvement training method can close much of the gap between open-weight models and the proprietary frontier — without requiring full retraining. Second, the reduction in reasoning token length alongside higher accuracy points to a qualitative improvement in how the model thinks, not just brute-force scaling.</p><p>For the broader AI ecosystem, an open-weight model that outpaces Gemini-4-Argon and DeepSeek-V4.1-Flash in math and science — while remaining accessible enough to run on consumer hardware — raises the bar for what the open-source community can expect. VIDRAFT is positioning itself as a credible player in a space long dominated by US and Chinese labs, and Darwin’s SuperCLUE performance gives that claim concrete benchmark support.</p><p>The identified weaknesses in instruction-following and long-context tasks, however, suggest Darwin-180B-RSI-R3 is best suited for math- and science-focused applications rather than general-purpose use at this stage.</p><h4>Key Takeaways</h4><ul><li>Darwin-180B-RSI-R3 scored 73.68 in math and science on the November 2026 SuperCLUE, ranking first among all open-weight models and second overall out of 19 models.</li><li>The model’s RSI training method updates only attention pathways and shared experts, leaving the core 512 experts, router, and visual encoder intact.</li><li>Compared to its parent model, Darwin improved from 57.89 to 73.68 in math/science on SuperCLUE and reduced MMLU-Pro reasoning length by 11%.</li><li>A 4-bit quantized version (111 GB) can run on a laptop with 8 GB GPU and 32 GB RAM, or on CPU alone at 18–21 tokens per second.</li><li>Instruction-following (20.69) and long-context (30.32) scores lag behind leaderboard leaders, indicating room for improvement in general tasks.</li></ul><h4>Frequently Asked Questions</h4><p><strong>Q: What is Darwin-180B-RSI-R3 and who made it?</strong> A: Darwin-180B-RSI-R3 is an open-weight AI model developed by VIDRAFT, a Korean AI startup. It is built on Qwen3.8-Flash-Next and refined using VIDRAFT’s Recursive Self-Improvement (RSI) training method.</p><p><strong>Q: How did Darwin-180B-RSI-R3 perform on SuperCLUE compared to Gemini and DeepSeek?</strong> A: In the November 2026 SuperCLUE math and science category, Darwin scored 73.68, outperforming both Gemini-4-Argon and DeepSeek-V4.1-Flash, which each scored 71.05. Only GPT-6.1-Sol scored higher, at 76.32.</p><p><strong>Q: Can Darwin-180B-RSI-R3 be run on consumer hardware?</strong> A: Yes. The 4-bit quantized POCKET-Darwin-180B-GGUF build (111 GB) runs on a laptop equipped with an 8 GB GPU and 32 GB RAM, and achieves 18–21 tokens per second on CPU alone without any GPU.</p><p><em>Source: USB.RU (러시아) (2026–11–09) — </em><a href="https://usb.ru/news/darwin-180b-rsi-superclue-nov-2026/"><em>original article</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=43390f3b3fe4" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>