ABAB Pedia

Will DePue

Will DePue (born 2003) is an AI researcher and the founder of Unemployment Labs, a small independent nonprofit research group. DeepResearch, the community-analytics platform he founded in high school, was acquired by Commsor; from July 2023 to April 2026 he worked at OpenAIOpenAIOpenAI is an artificial intelligence research and product organization founded in 2015 that develops the GPT model series, ChatGPT, Codex and a developer API, and conducts research on image, speech, video and intelligent agents. After its 2025 restructuring, the nonprofit OpenAI Foundation continues to control the for-profit OpenAI Group PBC. Its growth has involved large-scale fundraising and compute partnerships, and has also been accompanied by controversies over governance, copyright, privacy and AI safety.Open the full entry, where he was a core contributor to the text-to-video models Sora and Sora 2Sora (OpenAI)Sora is a text-to-video model and app of the same name developed by OpenAI; it debuted as a research preview in February 2024, was followed by Sora 2 and a social video app in 2025, and its web and app services were discontinued in April 2026.Open the full entry and worked on post-training models including o3, o4-mini and GPT-5. On November 12, 2026 he published a checkpoint forensics report finding NVIDIANVIDIANVIDIA is an American semiconductor and computing technology company that provides graphics processing units, artificial intelligence infrastructure, networking interconnects, and computing software, and is listed on Nasdaq under the ticker NVDA.Open the full entry Blackwell GPU fingerprints in six open-weight models from DeepSeekDeepSeekDeepSeek is a Chinese AI company developing large language models and providing generative AI services through websites, applications and APIs.Open the full entry, Moonshot AIMoonshot AIMoonshot AI develops the Kimi AI assistant, researches foundation models and offers agent products.Open the full entry, MiniMaxMiniMaxMiniMax is a Chinese artificial intelligence company that develops multimodal foundation models covering text, speech, images, video and music, and runs AI-native apps for individual users and an open platform for businesses and developers.Open the full entry and Ant GroupAnt GroupAnt Group is a digital technology and fintech company headquartered in Hangzhou, Zhejiang, China. It operates Alipay, runs digital payment and merchant services, credit, investment and insurance technology businesses, and develops large AI models; Alibaba holds about one third of its shares.Open the full entry.

Contents12 sections
Key facts

Founding Thrive.fyi in High School

Born in 2003, he founded Thrive.fyi in Santa Monica, CaliforniaCaliforniaCalifornia is a U.S. state on the Pacific coast; its capital is Sacramento and its largest city is Los Angeles. California was admitted to the Union on September 9, 1850, as the 31st state. Its population at the 2020 Census (April 1, 2020) was 39,538,223, the largest of any state.Open the full entry, while still in high school, aiming to build a chat platform made for high-school students so that students and teachers could communicate directly and quickly instead of by email. The project raised a friends-and-family round and shut down in June 2021.⁠[1][2]

Founding the Community-Analytics Platform DeepResearch

In his last year of high school he began experimenting with Discord’s API and downloaded server message data, rendering graphs of community members’ relationships, then recruited three friends to turn a local Python script into a product, DeepResearch, which became the leading analytics platform on Discord. In its first six months it analysed 36 million unique community users and stored more than 100 million data points; he taught himself front-end and back-end development in a month and designed every screen in Figma.⁠[1][2]

That year he graduated in the first graduating class of Geffen Academy at UCLA as salutatorian, then studied computer science at the University of MichiganUniversity of MichiganThe University of Michigan is a university founded in Detroit in 1817 that moved to Ann Arbor, Michigan, in 1837; it is governed by a statewide-elected Board of Regents and also has campuses in Dearborn and Flint.Open the full entry, from which he later took leave and dropped out. He is also an Eagle Scout.⁠[1][3]

DeepResearch Acquired by Commsor

Less than six months after incorporation DeepResearch received an acquisition offer from Commsor; the deal closed that month, and he then worked at Commsor as a software engineer for a year, bringing the product to enterprise customers.⁠[1][2]

He helped manage Gen Z Mafia, a community of young builders where he found his DeepResearch co-founders, and later helped restart and expand its mission as The Human Colossus. From September 2022 to March 2023 he was engineering lead at the nonprofit InternetActivism, leading its Hyperlocal project, which restores communication when the internet goes down from war, disaster or government shutdowns by forming a distributed network of Bluetooth-connected phones.⁠[1][2]

Joining OpenAI

He joined OpenAI as a resident member of technical staff, working on applied research including jailbreak and prompt-injection mitigations, custom models, model capability evaluations and API fine-tuning. From January 2024 he was a full member of technical staff.⁠[2][3]

A Core Contributor to Sora

He was a core contributor to Sora, OpenAI’s first text-to-video model, and one of the authors of its technical report, Video generation models as world simulators; Sora was unveiled that month and publicly launched that December. He was later also a core contributor to Sora 2, working mainly on text conditioning, controllability and world understanding.⁠[1][2][3]

Back at OpenAI to Start a New Team

He announced on X that after a long break he was back at OpenAI to start a new team with Troy Luhman and Eric Luhman, fellow members of the Sora team, focused on an incredibly high-risk bet with a small but significant chance of leading to artificial superintelligence; the team would stay tight but was open to researchers and engineers wanting to make progress in uncharted territory. His LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry says that at OpenAI he led a research team working on highly exploratory research into novel architectures.⁠[3][1][2]

Leaving OpenAI

He worked at OpenAI for nearly three years; his LinkedIn lists the position as ending that month. On his personal website he said he planned to take a long break, travel and explore many different things, and was happy to give talks and advise impactful projects.⁠[2][1]

He then founded Unemployment Labs, a small independent nonprofit research group of former frontier-lab researchers and engineers, operating as part of the California nonprofit macrogrants, with 501(c)(3) status pending.⁠[4][5][6]

Publishing the Checkpoint Forensics Report

Under the Unemployment Labs name he published the report Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs, finding that six open-weight models released since April 2026 by DeepSeek, Moonshot AI, MiniMax and Ant Group carry a hardware fingerprint matching NVIDIA’s Blackwell GPUs, while earlier models from the same companies match the older Hopper generation. The six are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash. Exporting Blackwell chips to China requires a U.S.United StatesThe United States is a federal country in North America made up of 50 states, the District of Columbia as the seat of government, and territories such as Puerto Rico and Guam; its capital is Washington, D.C. The Continental Congress adopted the Declaration of Independence on July 4, 1776. The 2020 census (April 1, 2020) counted a resident population of 331,449,281, and the 435 seats of the House of Representatives are apportioned among the states on the basis of the census.Open the full entry licence, and no such licence has come to light.⁠[7]

Assuming each model was trained on the hardware that initialized it, the report estimates that the two largest runs, Kimi K3 and DeepSeek-V4-Pro, needed about 4,500 and 3,000 Blackwell GPUs for two to three months, about 12,000 across five runs, with the largest single cluster costing roughly $200 million to $300 million and the five together more than half a billion dollars; it says it is not confident in these estimates. The first version, which gave 2,000 to 3,000 GPUs for the largest runs and $150 million to $200 million for the largest cluster, was replaced in a same-day revision.⁠[7]

He said on X that several senior researchers, both independent and at frontier labs including OpenAI and AnthropicAnthropicAnthropic is an artificial intelligence research and product company founded in 2021 that develops the Claude models, assistant, and tools such as Claude Code, and operates as a public benefit corporation. Dario Amodei and Daniela Amodei are its co-founders, and its research covers model safety, interpretability, and steerability.Open the full entry, had confirmed the result. The report published the initial seeds linking DeepSeek-V3 to Hopper and DeepSeek-V4-Flash to Blackwell, and he released verification code on GitHub so that others could reproduce the work.⁠[5][6]

Post-Training Work at OpenAI

Before o3’s large training run he started an internal channel to add the public ARC-AGI training set to the reinforcement-learning data. Once a colleague added the plain-text ARC dataset, the score quickly rose to about 45%; the dataset then went into o3’s large training run, and the finished model reached 88% on ARC. He has rejected criticism that it cost $3,000 per task, arguing that progress comes from pushing top-line performance first, with efficiency following.⁠[1]

The Checkpoint Forensics Method

The method starts from the fact that before training, PyTorch fills billions of weights with initial values from a seed using the Philox pseudorandom algorithm, dividing the work among many GPU threads through NVIDIA’s cuRAND library. Each streaming multiprocessor (SM) uses up to 2,048 threads, so a chip’s SM count determines where the numbers land: Hopper’s H100, H200 and H800 have 132 SMs, Blackwell’s B200/B300 have 148 and GB200/GB300 have 152. Only the right seed together with the right thread count reproduces the initial arrangement.⁠[7][6]

After training, weights for tokens never used keep almost their starting values, and the rest retain a faint statistical trace. He searched about six trillion combinations of seed and chip layout for the one that best reproduces these traces, then checked it on weights not used in the search. Running unmodified PyTorch with the recovered seeds, a physical B300 reproduced DeepSeek-V4-Flash’s preserved starting weights bit for bit and a rented GB200 reproduced Kimi K3’s, while an H100 did not. For 24 models whose developers have disclosed their training hardware, the method’s readings all agreed with the public information.⁠[7][5]

The report sets out the method’s limits: the fingerprint identifies only the hardware that created the starting weights and supports, without proving, that Blackwell also performed the training; it cannot show where the chips were or who owned them, and since Blackwell systems can be rented in data centres outside China it does not establish smuggling or an export-control violation; specially configured AMDAMDAMD (Advanced Micro Devices, Inc.) is a publicly traded U.S. semiconductor company founded in 1969 and headquartered in Santa Clara, California; its stock trades on Nasdaq under the ticker AMD. The company develops CPUs, GPUs, FPGAs, adaptive systems-on-chip and related software, mainly for the data center, personal computer, gaming and embedded markets; its product brands include Ryzen, EPYC, Radeon and Instinct.Open the full entry hardware is the only alternative that could produce the same pattern, which he excluded for DeepSeek-V4-Flash. The newest checkable models from AlibabaAlibabaAlibaba is a technology group focused on commerce, cloud computing, and AI. Founded in 1999 by 18 people led by Jack Ma, it received a major investment from Yahoo in 2005.Open the full entry, Zhipu, TencentTencentTencent is a Chinese internet and technology company that runs the QQ and Weixin (WeChat) communication and social platforms, with businesses spanning online games, marketing services, FinTech and business services, and AI products such as its HY (Hunyuan) models and the Yuanbao app.Open the full entry, XiaomiXiaomiXiaomi Corporation is a Chinese consumer electronics and smart manufacturing company with businesses in smartphones, IoT and lifestyle products, internet services and smart electric vehicles; it also develops its own mobile chips and the MiMo series of AI models.Open the full entry and StepFun all read as Hopper.⁠[7][6]

Personal Projects

He has said most of his projects come from building what he wished already existed. WebGPT runs GPT models natively in the browser with the new WebGPU API in under two thousand lines of plain JavaScript; he built it in under two weeks from no prior knowledge, and it gained about 2,700 GitHub stars. FIGMA-OS, which he built over late nights in June 2023 after Figma released prototyping tools with conditionals and variables, was the first Turing-complete 8-bit computer built in Figma.⁠[1]

He also made the Alexandria Index, free and up-to-date open embedding datasets, starting with every paper on arXiv, more than four million items; and Ephemeral Notes, a notepad in which every word fades away within 60 seconds, meant to extend working memory while thinking.⁠[1]

Sources7 sources

  1. Will DePue (opens in a new window)depue.notion.site↑
  2. Will DePue | LinkedIn (opens in a new window)linkedin.com↑
  3. A star OpenAI engineer says he's starting a new team to chase artificial superintelligence (opens in a new window)tech.yahoo.com↑
  4. Unemployment Labs (opens in a new window)unemploymentlabs.com↑
  5. Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer (opens in a new window)theregister.com↑
  6. Former OpenAI researcher says he found Nvidia Blackwell fingerprints hidden inside six Chinese AI models (opens in a new window)the-decoder.com↑
  7. Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs (opens in a new window)unemploymentlabs.com↑