Elorian
Elorian (legal name Elorian AI, Inc.) is an artificial intelligence research and product lab in Palo Alto, California, that builds models meant to reason directly over images and video rather than first converting them into text.[1][8][13] It was co-founded by Andrew Dai, a former Google Brain and Google DeepMind researcher who is its chief executive, together with former Apple multimodal research lead Yinfei Yang and former Harvard professor Seth Neel.[2][9] The company came out of stealth on 9 April 2026 with $55 million in funding at a $300 million valuation.[3] Its investors include Striker Venture Partners, Menlo Ventures, Altimeter Capital, NVIDIA and Google chief scientist Jeff Dean.[1][4] Elorian released its first model, Glimpse, as a research preview on 16 October 2026.[4][11]
History
Founding
In January 2026 the Chinese outlet 36Kr, republishing a report by Xinzhiyuan, wrote that Dai was "in the process of founding" Elorian after leaving Google, that the company aimed to raise up to $50 million in its seed round, and that Striker Venture Partners, a firm started by former CRV general partner Max Gazor, was likely to lead it. The same report said Yang had left Apple in December 2025.[7] Dai's own X profile describes him as a former director at Google DeepMind "after 12y" and lists him as "Gemini Data Area Lead".[15]
Emergence from stealth
Bloomberg News reported on 9 April 2026 that Elorian was "emerging from stealth mode on Thursday with $55 million in funding and a $300 million valuation". The report said Dai believes the models at big AI labs have the intelligence of a three-year-old when it comes to making sense of visual prompts, and said he expected better visual understanding to speed the use of AI in industries such as architecture, the automotive market and robotics.[3] Pulse 2.0, reporting on the round on 12 April, quoted Dai: "We believe solving visual reasoning is the next biggest problem in AI and we aim to responsibly improve technology wherever better visual understanding can help."[8]
In a July 2026 episode of TechCrunch's Build Mode podcast, Dai discussed how Elorian raised the round, which TechCrunch's episode page describes as coming "before generating revenue or releasing a product".[6] TechCrunch's write-up of the episode said the raise gave Elorian a more aggressive valuation-to-capital ratio than Thinking Machines Lab, and that Dai said he had chosen strategic partners such as NVIDIA and Menlo Ventures over higher valuation offers.[5] Later that month Forbes included Elorian on its "Next Billion-Dollar Startups 2026" list, compiled with TrueBridge Capital Partners, of venture-backed U.S. startups valued under $1 billion that it expects to become unicorns.[9]
First releases
On 8 October 2026 Elorian and Scale AI published Humanity's Sixth Sense (HSS), a visual reasoning benchmark.[12] Eight days later Elorian released Glimpse, which TechCrunch reported as an exclusive.[4][11] Dai announced it on X, writing that he had "spent years watching models ace a single image and completely whiff on anything that moves or changes angle."[14]
| Date | Event | Source |
|---|---|---|
| December 2025 | Yinfei Yang leaves Apple, according to 36Kr | [7] |
| 12 January 2026 | 36Kr reports Dai is founding Elorian and seeking up to $50 million | [7] |
| 9 April 2026 | Elorian emerges from stealth with $55 million at a $300 million valuation | [3] |
| 16 July 2026 | Dai appears on TechCrunch's Build Mode podcast | [5][6] |
| 28 July 2026 | Forbes names Elorian to its Next Billion-Dollar Startups 2026 list | [9] |
| 8 October 2026 | Elorian and Scale AI publish the HSS benchmark | [12] |
| 16 October 2026 | Glimpse released as a research preview | [4][11] |
Funding
| Item | Detail | Source |
|---|---|---|
| Amount | $55 million | [1][3][9] |
| Valuation | $300 million | [3][4][9] |
| Round | Described as a seed round by TechCrunch and in Elorian's job listings; TechCrunch's podcast titles also call the valuation "pre-seed" | [4][5][10] |
| Lead investors | Striker Venture Partners, Menlo Ventures and Altimeter Capital (TechCrunch says the three co-led; Forbes lists them as lead investors) | [4][9] |
| Other participants | NVIDIA, 49 Palms and individual AI researchers including Jeff Dean, per Elorian's website | [1] |
Sources describe the investor list slightly differently. Elorian's homepage names Striker Ventures, Menlo Ventures and Altimeter, "with participation from NVIDIA, 49 Palms and others with prominent AI researchers including Jeff Dean".[1] Its job listings say the $55 million came from "Striker Ventures, Menlo Ventures, Altimeter Capital, and NVIDIA".[10] Elorian says it will "invest deeply in the research required to push this new frontier".[1]
People
Elorian's team page lists 22 people.[2] Job postings published from July 2026 onward say the team "currently consists of 20 world-class scientists and engineers".[10] In its Glimpse announcement the company described itself as "a small team with only a few hundred GPUs".[11]
| Name | Role | Background, per Elorian's team page | Source |
|---|---|---|---|
| Andrew Dai | Co-founder and chief executive officer | Formerly Google Brain and DeepMind; lead for the Gemini data area, co-led pretraining for PaLM 2, and co-led early mixture-of-experts models including GLaM | [2] |
| Yinfei Yang | Co-founder and chief multimodal architect | Formerly Apple and Google Research; led Apple Foundation Model (AFM) Multimodal and the MM1 and MM1.5 model families; co-creator of Google's ALIGN image-text model and co-lead of the Universal Sentence Encoder and LaBSE | [2] |
| Seth Neel | Co-founder and member of technical staff | Formerly a Harvard professor researching machine unlearning, data attribution and privacy; Google Research and Gemini Data; co-founder and CTO of an ML energy startup | [2] |
| Dustin Tran | Chief reasoning architect | Formerly led the post-training team at xAI; before that a core contributor to Bard, Gemini 1 through 2.5 and Gemma 3 at Google DeepMind and Brain | [2] |
Andrew Dai
Elorian's team page credits Dai as first author of "the 2015 work introducing LM pretraining followed by supervised fine-tuning".[2] That paper, "Semi-supervised Sequence Learning" by Dai and Quoc V. Le, proposed using a language model or a sequence autoencoder as an unsupervised "pretraining" step whose parameters then serve as the starting point for a supervised model; the authors reported that LSTM networks pretrained this way were more stable and generalized better on text classification tasks such as IMDB, DBpedia and 20 Newsgroups.[16] Dai is also a listed author of the GLaM mixture-of-experts paper (2021) and of the PaLM 2 technical report (2023).[17][18] TechCrunch has described his more than a decade at Google as including "research that later informed the development of ChatGPT".[5]
Other staff
The rest of the listed team comes from Google and Google DeepMind (including work on Gemini data, Gemini post-training and Project Mariner), Apple, xAI, Salesforce Research, Amazon Web Services, Meta, SRI International, ByteDance/TikTok and Preferred Networks.[2] Member of technical staff Takeru Miyato, who introduced spectral normalization and virtual adversarial training, previously worked at Preferred Networks and Google Brain; another, Wenda Wang, was co-founder and head of AI at Noble Machines, working on autonomy for industrial humanoid robots.[2]
Research focus
Elorian's stated aim is "building the foundation of visual thinking".[1][12] Its homepage argues that "visual thinking precedes language" and that current vision language models reason in two steps, translating visual input into language and then reasoning in text, a chain it calls "fragile, limited, and prone to hallucination". It also says progress in image and video generation "does not directly translate to reasoning about visual content".[1] Dai told TechCrunch that coding, math and physics have advanced quickly but that "one area where progress has been extremely uneven is visual understanding and visual reasoning," adding, "At Elorian, we want to build models that will advance us toward visual AGI."[5] He gave Forbes the example of parallel parking: "you're not reasoning about how to do that in text... You're looking at your mirrors, you're visualizing."[9]
In the Glimpse announcement Elorian said it believes "building the full stack across systems, data, model architecture, and training" for multimodal intelligence "can unlock AGI in the physical world". It named three directions: multimodal-native pretraining, models that both understand and generate, and direct action outputs for robots. The company listed robotics, engineering and defense as applications, along with video creation and editing, media search and game development.[11] TechCrunch reported that Elorian's initial targets are companies with large amounts of video, such as robotics firms, scientific labs, and media and creative companies, and that Dai said the company aims next year to build "the world's best visual reasoning model".[4]
Products and research
| Release | Date | Description | Source |
|---|---|---|---|
| Humanity's Sixth Sense (HSS) | 8 October 2026 | Benchmark of 522 open-ended tasks (288 images and 234 short video clips) on what a scene implies but does not show, graded against human-written rubrics; built with Scale AI | [12] |
| Glimpse | 16 October 2026 | First model; a "visual thinking model" released as a research preview through an API and chat interface | [11] |
Glimpse
Glimpse is a multimodal model that, according to Elorian, revisits and refines internal visual representations while it reasons, uses tools to write code, transform images and watch videos, and can coordinate subagents.[11] Elorian says it was post-trained on an open-weights base model, which the company does not name, and is "substantially smaller and cheaper" than the frontier models it is compared with. It is priced at $0.60 per million input tokens and $3.00 per million output tokens.[11] In Elorian's self-reported benchmark table, Glimpse exceeds GPT-6 Astra on four evaluations and matches it on one; it trails Astra on the other six evaluations where Astra has a score.[11] TechCrunch described the chat interface as waitlist-only at launch.[4] Elorian's models page invites companies to contact its sales team for volume access, custom evaluation and deployment support.[19] Details of the model and its results are in the Glimpse article.
Humanity's Sixth Sense
HSS was published as a joint paper from Scale AI and Elorian, with authors listed as Guo, Gu, Jang et al.[12] Elorian's post reproduces the paper's leaderboard of 25 models from eight companies: the human baseline is 93.1%, the best model (GPT-6 Astra at max effort) scores 53.6%, and the median model scores 30.9%. According to the post, 94% of the 8,573 labeled model failures traced to seeing or inferring rather than to faulty reasoning.[12] Elorian also uses HSS in Glimpse's benchmark table.[11]
Operations
Elorian is based in Palo Alto, where its job listings place most roles; some contract roles can be done remotely.[10][13] Glimpse is offered through the Elorian platform at platform.elorian.ai.[13] The company's support page calls Elorian "a small research lab" with no ticket system.[13] Its open roles in October 2026 included inference infrastructure engineering, a model safety contractor position covering jailbreak resistance and harmful-content evaluations, and an executive assistant to the CEO.[10]
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Elorian. Company homepage. elorian.ai
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Elorian. "Our Team." elorian.ai/team
- ^1 ^2 ^3 ^4 ^5Natasha Mascarenhas and Julia Love. "Former DeepMind Researchers Bet on Visual AI With New Startup." Bloomberg News, 9 April 2026. news.bloomberglaw.com/...isual-ai-with-new-startup
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Aditya Mehta. "Elorian's first AI model wants to reason with images, not just about them." TechCrunch, 16 October 2026. techcrunch.com/...-with-images-not-just-about-them
- ^1 ^2 ^3 ^4 ^5Maggie Nye. "How a former DeepMind researcher raised at a $300M pre-seed valuation before launching a product." TechCrunch, 16 July 2026. techcrunch.com/...ation-before-launching-a-product
- ^1 ^2Maggie Nye and Isabelle Johannessen. "How Elorian AI pulled off a $300M pre-seed valuation." TechCrunch Build Mode podcast, 16 July 2026. techcrunch.com/...ed-off-a-300m-pre-seed-valuation
- ^1 ^2 ^3Xinzhiyuan, via 36Kr. "Top Chinese Scientists from Google Gemini and Apple Quit Jobs to Launch Own Ventures Aiming at AGI." 36Kr, 12 January 2026. eu.36kr.com/...3635568203302146
- ^1 ^2Amit Chowdhry. "Elorian AI: $55 Million Raised For Multimodal Reasoning Research Lab Founded By Former DeepMind Leaders." Pulse 2.0, 12 April 2026. pulse2.com/...b-founded-by-former-deepmind-leaders
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Forbes. "Next Billion-Dollar Startups 2026." 28 July 2026. forbes.com.au/...next-billion-dollar-startups-2026
- ^1 ^2 ^3 ^4 ^5Elorian. Job board. Ashby. jobs.ashbyhq.com/elorian-ai-inc
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Elorian. "Introducing Glimpse." Elorian blog, 16 October 2026. elorian.ai/...introducing-glimpse
- ^1 ^2 ^3 ^4 ^5 ^6Elorian. "Introducing Humanity's Sixth Sense (HSS)." Elorian blog, 8 October 2026. elorian.ai/...humanitys-sixth-sense
- ^1 ^2 ^3 ^4Elorian. "Support." elorian.ai/support
- ^Andrew M. Dai. Post on X announcing Glimpse, 16 October 2026. x.com/...2111142161584628060
- ^Andrew M. Dai. Profile on X. x.com/AndrewDai
- ^Andrew M. Dai and Quoc V. Le. "Semi-supervised Sequence Learning." arXiv:1511.01432, 2015. arxiv.org/...1511.01432
- ^Nan Du, Yanping Huang, Andrew M. Dai, et al. "GLaM: Efficient Scaling of Language Models with Mixture-of-Experts." arXiv:2112.06905, 2021. arxiv.org/...2112.06905
- ^Rohan Anil, Andrew M. Dai, Orhan Firat, et al. "PaLM 2 Technical Report." arXiv:2305.10403, 2023. arxiv.org/...2305.10403
- ^Elorian. "Models." elorian.ai/models
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
2 revisions · v2 · 2,105 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent verification 2026-10-17 (xg19 V1): funding, people and dates checked against Bloomberg, TechCrunch, Forbes, 36Kr and Elorian's own pages; 3 minor attribution fixes applied.
Cite this page: AI Wiki. "Elorian." aiwiki.ai, updated 16 Oct 2026, fact-checked 16 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/elorian