# Standard Intelligence > Source: https://aiwiki.ai/wiki/standard_intelligence > Updated: 2026-10-28 > Fact-checked: 2026-10-27 > Categories: AI Agents, AI Companies, AI Research, Speech & Audio AI > License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)" > Cite as: AI Wiki. "Standard Intelligence." aiwiki.ai, 28 Oct 2026. https://aiwiki.ai/wiki/standard_intelligence > From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution. **Standard Intelligence** (legal name Standard Intelligence PBC) is an artificial intelligence research company in San Francisco that trains models to operate computers by learning from video of people using them.[1][2] Galen Mead and Devansh Pandey founded it in 2024.[3] The company says its long-term mission is "building aligned AGI", which it also describes as building "aligned general learners".[3][4] Its main model, FDM-1, was announced in February 2026. FDM-1 is a [computer use](https://aiwiki.ai/wiki/computer_use) model trained on screen-recording video that a separate model had labeled with keyboard and mouse actions, and its outputs are key presses and mouse movements.[1] Earlier, in November 2024, the company released hertz-dev, an open 8.5-billion-parameter base model for full-duplex conversational audio.[5]
Standard Intelligence logo
Standard Intelligence logo, from the official website si.inc. Trademark of its owner.
On 30 April 2026 the company announced that it had raised $75 million from [Sequoia Capital](https://aiwiki.ai/wiki/sequoia_capital) and Spark Capital. It also named [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy), Milan Kovac and Stanley Druckenmiller among its angel investors and advisors.[4] Crunchbase News reported that the round was raised at a $425 million pre-money valuation.[6] The Information reported a $500 million post-money valuation.[7] The company had six people at the time.[4] | Standard Intelligence | | | --- | --- | | Legal name | Standard Intelligence PBC; previously Standard Intelligence, Inc.[2] | | Incorporated | 2024, in Delaware[2] | | Founders | Galen Mead, Devansh Pandey[3][8] | | Headquarters | San Francisco, California[4][2] | | Focus | Computer-use models pretrained on screen-recording video; earlier work on conversational audio[1][5] | | Models | hertz-dev (November 2024), FDM-1 (February 2026)[5][1] | | Team size | 6 (April 2026)[4] | | Latest funding | $75 million, announced 30 April 2026; Sequoia Capital and Spark Capital[4] | | Reported valuation | $425 million pre-money (Crunchbase News); $500 million post-money (The Information), April 2026[6][7] | | Website | [si.inc](https://si.inc) | ## History ### Founding Mead and Pandey met as teenagers at the Atlas Fellowship in 2022. Sequoia Capital, which later invested in the company, describes the program as "a selective fellowship for high-school students interested in AI alignment and AGI".[8] According to Sequoia, both founders left their undergraduate programs "out of a sense of urgency" to work on AGI, and in April 2026 they were 21 and 20 years old respectively.[8] Sequoia lists 2024 as the company's founding year. Its SEC filings give 2024 as the year of incorporation, name Delaware as the state of incorporation, and list "Standard Intelligence, Inc." as a previous name of the issuer.[3][2] ### hertz-dev (2024) The company's first public release was hertz-dev, published on 3 November 2024 with code and checkpoints. Standard Intelligence called it "the first publicly released base model for conversational audio".[5] In the announcement the company said it was doing "fundamental research with the goal of building aligned general intelligence" and was "currently a team of 4 in San Francisco". The post credited eleven collaborators, including both founders.[5] ### The heap (2025) On 30 September 2025 the company described "the heap", a 30-petabyte storage cluster in a colocation facility in downtown San Francisco. It said it built the cluster for under $500,000 to store video for pretraining computer-use models.[9] The company said that storing the same data on [Amazon Web Services](https://aiwiki.ai/wiki/amazon_web_services) would have cost about $12 million a year, and that the heap cut this "~40x to $354k per year, including depreciation".[9] It racked all 30 petabytes of hardware during an event it called "Storage Stacking Saturday", which ran for 36 hours with a break to sleep. Friends who helped were offered food and custom-engraved hard drives, and contractors joined later in the event.[9] ### FDM-1 (2026) On 23 February 2026 the company published FDM-1, which it called "the first fully general computer action model". Demonstrations showed the model building a gear in Blender, driving a car around a San Francisco block, and finding a bug in a mock banking app.[1] The research team listed on the post was Neel Redkar, Yudhister Kumar, Devansh Pandey and Galen Mead.[1] The Information's later report on the company opened with a Toyota RAV4 driving around San Francisco's South Park neighborhood "with a laptop propped on the dashboard", the laptop running what it called "a new breed of AI model".[10] ### Thiel Fellowship and Series A (2026) In April 2026 the Thiel Foundation named Mead, listed with the hometown Chapel Hill, North Carolina, to its 2026 class of Thiel Fellows. The announcement described Standard Intelligence as "building aligned general learners, pretraining large models to actively explore and learn from the Internet". Each fellow in the class receives $250,000 over two years. The foundation is chaired by [Peter Thiel](https://aiwiki.ai/wiki/peter_thiel).[11] On 30 April 2026 the company announced the $75 million round from Sequoia and Spark Capital, working with Sonya Huang of Sequoia and Mikowai Ashwill and Yasmin Razavi of Spark.[4][8] Sequoia called it a Series A that it led "alongside Miko and Yasmin from Spark Capital".[8] The company said the money "unlocks several orders of magnitude of compute scaling" for the FDM work and lets it fund "blue-sky research".[4] ## Founders and team | Person | Role | Background | | --- | --- | --- | | Galen Mead | Co-founder; executive officer and director in SEC filings[2] | Met Pandey at the Atlas Fellowship in 2022; left an undergraduate program; 2026 Thiel Fellow[8][11] | | Devansh Pandey | Co-founder; executive officer and director; signed the June 2026 Form D filings as President (as Devansh Ratna Pandey)[2] | Met Mead at the Atlas Fellowship in 2022; left an undergraduate program[8] | | Sonya Huang | Director[2] | Sequoia Capital partner listed on Sequoia's page for the company[3] | Besides the two founders, Sequoia's April 2026 announcement named Neel, Yudhister, Ulisse and Ryan as the rest of the six-person team. Sequoia wrote that they had turned down "the conventional path (fancy degrees and offers from big token)".[8] The FDM-1 post lists Neel Redkar and Yudhister Kumar on its four-person research team alongside the two founders.[1] ## Research approach Standard Intelligence argues that the usual way of building computer-use agents does not scale. That method fine-tunes a [vision-language model](https://aiwiki.ai/wiki/vision_language_model) on screenshots annotated by contractors and then builds [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) environments for each task. The company says agents trained this way cannot act on more than a few seconds of context or process high-framerate video. Because contractor labels are expensive, it adds, "the largest open dataset is less than 20 hours of 30 FPS video". The post links that statement to the ScaleCUA paper.[1][12] Its alternative is to pretrain on internet video of people using computers. It writes that "building a general computer agent requires an internet-scale video corpus, just as building GPT-3 required an internet-scale text corpus".[1] The key technique is an inverse dynamics model (IDM): a model that looks at a sequence of screen frames and infers which actions produced the changes.[1] The FDM-1 post cites two earlier works. One is "Behavioral Cloning from Observation" (Torabi, Warnell and Stone, 2018). As the post describes it, the researchers taught an IDM to label the action taken between before and after states in simulated environments.[1][13] The other is OpenAI's Video PreTraining (VPT) paper (2022). VPT trained an IDM on a small amount of labeled data, used it to label online videos of people playing Minecraft, and trained a behavioral prior from the result. Its authors describe the method as a form of semi-supervised [imitation learning](https://aiwiki.ai/wiki/imitation_learning).[14] The company calls VPT "the first to apply this method at scale". It argues that design, finance and general computer work need "minutes to hours of context" rather than VPT's six seconds.[1] It also mentions VideoAgentTrek, an October 2025 paper that trained a computer-use IDM and mined 1.52 million interaction steps from 39,000 YouTube tutorial videos.[15] Standard Intelligence says that approach lacks video context and depends on "screenshot-action-CoT triplets".[1] Sequoia partner Sonya Huang summarized the bet as "the Tesla FSD approach applied to knowledge work on computer screens". She called it "deeply 'bitter lesson'-pilled", a reference to [The Bitter Lesson](https://aiwiki.ai/wiki/bitter_lesson), because it scales raw data rather than hand-engineering workflows. She noted that the team was "emphatically 'not video people'".[8] ## Models ### hertz-dev hertz-dev is an 8.5-billion-parameter, audio-only, full-duplex [autoregressive](https://aiwiki.ai/wiki/autoregressive_model) model for spoken conversation. It handles two speakers at once and can parse and generate overlapping speech.[5] The model card says it was trained on "20 million unique hours of high-quality audio data".[16] It is a base model "without fine-tuning, RLHF, or instruction-following behavior", released for researchers to fine-tune for other audio tasks.[16] The code and checkpoints are released under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license) on GitHub and [Hugging Face](https://aiwiki.ai/wiki/hugging_face).[17][16] | Component | Description | | --- | --- | | hertz-codec | Convolutional audio [variational autoencoder](https://aiwiki.ai/wiki/variational_autoencoder). It encodes mono 16 kHz speech into an 8 Hz latent at a KL-regularized 1 kbps, one 32-dimensional latent per 125 ms frame. Causal convolutions allow streaming. 5 million encoder and 95 million decoder parameters.[5] | | hertz-ar | 40-layer, 8.4-billion-parameter decoder-only [transformer](https://aiwiki.ai/wiki/transformers) with a context of 2,048 input tokens (about 4.5 minutes). The first 32 layers ("hertz-lm") predict a 15-bit quantized projection of the next latent; the last 8 layers predict the next audio latent.[5] | | hertz-lm checkpoints | One version initialized from a language model trained on 2 trillion tokens, one randomly initialized and trained only on audio latents. The company said the recipe learned linguistics either way.[5] | | Duplex | Handled as a post-training task, with two projection heads.[5] | The company reported a theoretical average latency of 80 ms and a real-world latency of 120 ms on a single RTX 4090. The blog post called this "2x lower than the previous state of the art"; the Hugging Face model card says "1.5-2x lower".[5][16] The company also said that in subjective evaluations hertz-codec outperformed SoundStream and EnCodec at 6 kbps and was on par with DAC at 8 kbps.[5] These are the company's own measurements. As of October 2026 the GitHub repository had about 1,800 stars; its code was last updated in January 2025.[17] ### FDM-1 FDM-1 is a model that takes screen video in and produces keyboard and mouse actions out. In the post, "FDM" stands for "forward dynamics model", the final stage of the training recipe.[1] The company says FDM-1 "trains and infers directly on video instead of screenshots". It also says the model "consistently improves with scale".[1] The post does not give a parameter count.[1] | Stage | What the company says it did | | --- | --- | | 1. Inverse dynamics model | Trained an IDM on 40,000 hours of contractor-labeled screen recordings[1] | | 2. Labeling | Used the IDM to label its 11-million-hour video corpus with predicted actions[1] | | 3. Forward dynamics model | Trained the FDM autoregressively on next-action prediction over interleaved video frames and actions; output tokens are key presses and mouse-movement deltas[1] | The body of the post says FDM-1 was trained on videos from "a portion" of the 11-million-hour dataset.[1] **Video encoder.** The company says its encoder fits "almost 2 hours of 30 FPS video in only 1M tokens", while VLMs spend about a million tokens on one minute of such video. It calls this 50 times more token-efficient than the previous state of the art and 100 times more efficient than OpenAI's encoder.[1] The encoder is trained on a masked compression objective. The company said the idea was similar to [V-JEPA](https://aiwiki.ai/wiki/v_jepa) but not the same method.[1][18] It reported that 32,000 tokens hold about 3.5 minutes of video, 200,000 tokens about 20 minutes, and 1 million tokens about 1 hour 40 minutes. It also reported about 100 times faster convergence than a basic [Vision Transformer](https://aiwiki.ai/wiki/vision_transformer) baseline on a text-transcription task.[1] Its frame-count comparison with [Gemini](https://aiwiki.ai/wiki/gemini), ChatGPT computer use, [Claude](https://aiwiki.ai/wiki/claude) and the [NVIDIA Cosmos](https://aiwiki.ai/wiki/nvidia_cosmos) tokenizer was computed from those vendors' published documentation.[1] **Inverse dynamics model.** Labeling is non-causal: as the post puts it, "you can't label a Cmd+C until you see the resulting pasted sequence". The company therefore used a masked [diffusion](https://aiwiki.ai/wiki/diffusion_model) architecture that predicts masked action tokens conditioned on all frames. It unmasks the most confident predictions first and repeats the process until the whole sequence is labeled. The company said this worked better than CTC loss or ordinary cross-entropy.[1] **Action tokens.** Each key press, key release and scroll event is its own token. Mouse movement per frame is split into X and Y components, normalized to the screen size, and placed in one of 49 exponentially sized bins, so small movements get finer bins. The model also predicts the next click position.[1] FDM-1 does not use chain-of-thought, byte-pair encoding or tool use. The company says it was trained "with no language model transfer", though about 1.25 trillion transcript tokens were included for instruction tuning and language grounding.[1] **Results.** On the company's internal evaluation suite, IDM-labeled data beat contractor data on mouse movement and action skills, measured as target accuracy, symbolic memory and UI manipulation. Typing and verbal understanding improved more slowly on IDM-labeled data, which the company attributed to noise from the IDM.[1] The three demonstrations were: - **CAD:** extruding faces on an n-gon to make a gear in Blender. The company saved OS checkpoints at successful steps, which it said "unlocks test-time compute for computer use tasks".[1] - **Driving:** after fine-tuning on less than one hour of driving data, the model drove turns around a San Francisco block using arrow-key presses. Control went through a fork of openpilot's "joystick mode" and a remote-steering website. openpilot is open-source driver-assistance software from [comma.ai](https://aiwiki.ai/wiki/comma_ai).[19] The company said FDM-1 started at 50% accuracy on key-press prediction (no action, left or right), "significantly higher than the baseline model" that had only its video encoder and no internet-video pretraining.[1] - **Fuzzing:** in a toy banking app, the model explored states and found that a "Submit Wire Transfer" button stayed clickable after a transfer had completed, letting the balance go negative.[1] The post reported results from the internal suite and these demonstrations. It did not report scores on public computer-use benchmarks such as [OSWorld](https://aiwiki.ai/wiki/osworld), and it did not announce public weights or an API.[1] ### Infrastructure The heap uses 2,400 hard drives, mostly used 12 TB enterprise drives, in 100 NetApp DS4246 chassis. These are attached to 10 CPU head nodes bought on eBay and connected by a 100 Gbps dedicated internet line from Zayo.[9] The company listed one-time costs of $426,500, of which $300,000 was drives, and recurring costs of $17,500 a month for internet and electricity. That came to $29,500 a month with three-year depreciation. It compared this with about $1.13 million a month on AWS, including estimated egress, and about $270,000 a month at estimated bulk-discounted [Cloudflare](https://aiwiki.ai/wiki/cloudflare) R2 pricing.[9] The company said it built the cluster "to store 90 million hours worth of video data". The company's reasoning was that training data can tolerate losses: "we don't need AWS's 13 nines of reliability; 2 is more than enough".[9] Its software is about 200 lines of Rust for writes, an nginx server for reads, and an SQLite database for metadata. The company said it avoided Ceph and MinIO to keep the system simple.[9] It also said its training runs had previously loaded Cloudflare R2 heavily enough to be rate-limited.[9] Sequoia's announcement described the cluster as roughly 20 times cheaper than hyperscaler alternatives, a lower figure than the company's own estimate of about 40 times cheaper than AWS.[8][9] For evaluation, the company built infrastructure that it says runs more than 1 million rollouts an hour across 80,000 forking virtual machines. Each VM is a minimal Ubuntu desktop with 1 vCPU and 8 GB of RAM, and a single [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) can control 42 of them.[1] Forking copies a full memory snapshot of an operating-system state, so one starting state can be reused across thousands of rollouts. The company reported an 11 ms round trip from screen capture to action.[1] ## Funding | Date | Round | Amount | Investors | Valuation | | --- | --- | --- | --- | --- | | Announced 30 April 2026 | Series A, as described by Sequoia and Crunchbase News[8][6] | $75 million[4] | Co-led by Sequoia Capital and Spark Capital; angels and advisors included Andrej Karpathy, Milan Kovac and Stanley Druckenmiller[4][6] | $425 million pre-money (Crunchbase News); $500 million post-money (The Information)[6][7] | The company's blog post named Kovac, Druckenmiller and Karpathy as angels and advisors.[4] Kovac led Tesla's [Optimus](https://aiwiki.ai/wiki/tesla_optimus) humanoid robot program until he left Tesla in June 2025.[20] The company's announcement on X tagged @karpathy, @tszzl and @_milankovac_ as examples of its angels and advisors.[21] The Information's report, by Rocket Drew, ran under the headline "Standard Intelligence Rides Neolab Fervor with Computer Use Model".[10] Two Form D notices filed with the U.S. Securities and Exchange Commission on 5 June 2026 report equity sales under the Rule 506(b) exemption. One reports $35,613,499 sold to 36 investors, with a first sale on 2 April 2026. The other reports $44,199,974 sold to 10 investors, with a first sale on 13 April 2026.[2][22] Together the two filings total about $79.8 million. They do not say how the amounts relate to the $75 million the company announced. In both filings the company declined to disclose its revenue range.[2][22] SiliconANGLE reported that the company would use the money to buy more computing capacity and to develop safety guardrails for computer-use models.[23] ## Safety and alignment aims The company presents its computer-use work as a step toward general learning systems and says it is also studying how to align them. In the funding announcement it wrote that "current alignment techniques are insufficient to predictably and safely steer a model with human-level learning capabilities". It said it was studying "small versions of this problem in controlled environments to develop a science of alignment for general learners".[4] The FDM-1 post says the company believes "artificial general intelligence will be created within our lifetimes, and likely within the next decade". It adds that "there are still a lot of technical problems to be solved before aligned general learners can exist".[1] The heap post gave the long-term goal as "general models that can learn in-context and do arbitrary tasks while aligned with human values".[9] See [AI alignment](https://aiwiki.ai/wiki/ai_alignment). ## Reception Sequoia's Sonya Huang, whose firm co-led the Series A, called the 11-million-hour dataset "the largest in the industry" and described FDM-1 as "an early glimpse of what this paradigm could become".[8] SiliconANGLE's report on the round attributed the dataset, encoder and demonstration claims to the company.[23] The Information's headline placed the company within what it called "neolab fervor".[10] Debarghya Das's "The Ultimate Neolab List" also includes the company, classed as a frontier lab working on computer use.[24] See [neolab](https://aiwiki.ai/wiki/neolab). The FDM-1 figures, including the dataset size, encoder efficiency, evaluation results and demonstrations, are the company's own and come from its internal evaluations. As of 27 October 2026 the company's website listed four posts, the most recent being the April 2026 funding announcement, and none of them announced a release of FDM-1's weights or an API.[25][1] ## See also - [Computer use](https://aiwiki.ai/wiki/computer_use) - [OpenAI Operator](https://aiwiki.ai/wiki/openai_operator) - [OSWorld](https://aiwiki.ai/wiki/osworld) - [Imitation learning](https://aiwiki.ai/wiki/imitation_learning) - [H Company](https://aiwiki.ai/wiki/h_company) - [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs) - [Neolab](https://aiwiki.ai/wiki/neolab) ## References 1. Standard Intelligence. ["The First Fully General Computer Action Model"](https://si.inc/posts/fdm1/). 23 February 2026. 2. U.S. Securities and Exchange Commission. ["Form D: Standard Intelligence PBC" (accession no. 0002135700-26-000001)](https://www.sec.gov/Archives/edgar/data/2135700/000213570026000001/xslFormDX01/primary_doc.xml). Filed 5 June 2026. 3. Sequoia Capital. ["Standard Intelligence"](https://sequoiacap.com/companies/standard-intelligence/). Sequoia Capital company page. Retrieved 27 October 2026. 4. Standard Intelligence. ["A Brief Fundraising Update"](https://si.inc/posts/a/). 30 April 2026. 5. Standard Intelligence. ["An Open Base Model for Conversational Speech"](https://si.inc/posts/hertz-dev/). 3 November 2024. 6. Marlize van Romburgh. ["The Week's 10 Biggest Funding Rounds: Defense Tech Leads With Multiple Large Deals, Topped By $600M For Space Security Startup True Anomaly"](https://news.crunchbase.com/venture/biggest-funding-rounds-defense-aerospace-ai-fintech/). Crunchbase News, 1 May 2026. 7. Techmeme. ["Standard Intelligence, which is developing computer use AI models, raised $75M led by Sequoia and Spark at a $500M post-money valuation (The Information)"](https://x.com/Techmeme/status/2049906820760912053). Techmeme on X, 30 April 2026. 8. Sonya Huang. ["Standard Intelligence: Training General Intelligence in Pixel Space"](https://sequoiacap.com/article/standard-intelligence-training-general-intelligence-in-pixel-space/). Sequoia Capital, 30 April 2026. 9. Standard Intelligence. ["How to Rack 30 Petabytes of Storage"](https://si.inc/posts/the-heap/). 30 September 2025. 10. Rocket Drew. ["Standard Intelligence Rides Neolab Fervor with Computer Use Model"](https://www.theinformation.com/articles/standard-intelligence-rides-neolab-fervor-computer-use-model). The Information, 30 April 2026 (subscription required; headline and opening paragraph are public). 11. Thiel Foundation. ["Thiel Foundation Announces 2026 Class of Thiel Fellows"](https://www.businesswire.com/news/home/20260420984007/en/Thiel-Foundation-Announces-2026-Class-of-Thiel-Fellows). Business Wire, 21 April 2026. 12. Zhaoyang Liu, Jingjing Xie, Zichen Ding, et al. ["ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data"](https://arxiv.org/abs/2509.15221). arXiv:2509.15221, September 2025. 13. Faraz Torabi, Garrett Warnell, Peter Stone. ["Behavioral Cloning from Observation"](https://arxiv.org/abs/1805.01954). arXiv:1805.01954, May 2018. 14. Bowen Baker, Ilge Akkaya, Peter Zhokhov, et al. ["Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos"](https://arxiv.org/abs/2206.11795). arXiv:2206.11795, June 2022. 15. Dunjie Lu, Yiheng Xu, Junli Wang, et al. ["VideoAgentTrek: Computer Use Pretraining from Unlabeled Videos"](https://arxiv.org/abs/2510.19488). arXiv:2510.19488, October 2025. 16. Standard Intelligence. ["si-pbc/hertz-dev"](https://huggingface.co/si-pbc/hertz-dev). Hugging Face model card. Retrieved 27 October 2026. 17. Standard Intelligence. ["hertz-dev: first base model for full-duplex conversational audio"](https://github.com/Standard-Intelligence/hertz-dev). GitHub repository. Retrieved 27 October 2026. 18. Adrien Bardes, Quentin Garrido, Jean Ponce, et al. ["Revisiting Feature Prediction for Learning Visual Representations from Video"](https://arxiv.org/abs/2404.08471). arXiv:2404.08471, February 2024. 19. comma.ai. ["openpilot"](https://github.com/commaai/openpilot). GitHub repository. Retrieved 27 October 2026. 20. Fred Lambert. ["Tesla's head of Optimus humanoid robot leaves the '$25 trillion' product behind"](https://electrek.co/2025/06/06/tesla-head-optimus-humanoid-robot-leaves-trillion-product/). Electrek, 6 June 2025. 21. Standard Intelligence (@si_pbc). ["We've raised 75m in new funding from Sequoia and Spark Capital"](https://x.com/si_pbc/status/2049894469122818383). X, 30 April 2026. 22. U.S. Securities and Exchange Commission. ["Form D: Standard Intelligence PBC" (accession no. 0002135700-26-000002)](https://www.sec.gov/Archives/edgar/data/2135700/000213570026000002/xslFormDX01/primary_doc.xml). Filed 5 June 2026. 23. Maria Deutscher. ["Standard Intelligence raises $75M to develop efficient computer use models"](https://siliconangle.com/2026/04/30/standard-intelligence-raises-75m-develop-efficient-computer-use-models/). SiliconANGLE, 30 April 2026. 24. Debarghya Das. ["The Ultimate Neolab List"](https://debarghyadas.com/neolabs/). Retrieved 27 October 2026. 25. Standard Intelligence. ["Standard Intelligence" (company home page)](https://si.inc/). Retrieved 27 October 2026.