Humans&

Humans& is an American artificial intelligence company founded in 2025 that describes itself as a "human-centric" frontier AI lab. The startup, whose name is stylized with a trailing ampersand (the company itself writes it in lowercase, as "humans&"), was established by researchers who had previously worked at Anthropic, xAI, Google, OpenAI and Meta. In January 2026 it announced a $480 million seed round at a $4.48 billion valuation, one of the largest seed financings on record. Rather than building autonomous agents intended to replace human labor, Humans& states that it is building foundation models aimed at social intelligence and at helping people coordinate and work together. ^[1][2][8]

The company began releasing its work in the second half of 2026. In September it published a research preview of Persimmon, a "user model" built to simulate how people behave in multi-person chats. In October it released Ampersand, which it calls a "collaborator model," through a paid API, followed two days later by Boba, a free texting assistant built on Ampersand. ^[10][13][16] It trains these models on hardware it owns, in data centers set up with the cloud provider DeepInfra. ^[13][20]

Overview

Humans& is a frontier AI research company based in San Francisco, where its job listings are located; the law firm Goodwin has described it as based in California. ^[7][32] Its chief executive, Eric Zelikman, dates his role as co-founder and CEO to September 2025, and the company emerged from stealth on January 20, 2026, when it disclosed its seed round. At that point the company was roughly three months old and had not released a product. ^[1][3][5] TechCrunch reported that it then had 20-odd employees, who, according to the company, also came from OpenAI, Meta, Reflection, AI2 and MIT. ^[1]

The company positions itself in contrast to the dominant strategy among large AI laboratories, which has centered on building increasingly capable autonomous agents to automate work. Humans& frames its thesis around coordination rather than conversation or automation: it argues that the next major frontier for AI lies in helping groups of people communicate, build trust, and act together, with AI serving as connective tissue inside organizations and communities rather than as a replacement for the people in them. ^[1][4]

The table below summarizes the company's key facts as of October 2026.

Founding and founders

Humans& was founded by a group of researchers drawn from several of the leading AI laboratories. Press coverage of the launch named five co-founders. ^[1][2]

Eric Zelikman is the co-founder and chief executive officer. He was an early employee at xAI, where, according to his website, he contributed to pretraining data for Grok 2, started and scaled reinforcement learning for reasoning for Grok 3 Thinking, built the agent reinforcement learning infrastructure and recipe for Grok 4, and developed its first multi-agent reinforcement learning recipes. His website dates his time at xAI from March 2024 to September 2025. Before joining xAI he was a Ph.D. candidate at Stanford University, where he authored the paper introducing STaR ("Self-Taught Reasoner"), an early method for training language models to reason using their own generated rationales. His doctoral advisers were Nick Haber and Noah Goodman, a fellow co-founder of Humans&. ^[4][5][31]

Andi Peng is a co-founder and former Anthropic researcher. At Anthropic she worked on reinforcement learning and on the post-training of Claude models spanning the 3.5 through 4.5 generations; before that she was a Ph.D. student at MIT CSAIL. ^[1][2][34] Peng left the company on July 29, 2026. She announced the departure on X the next day, posting the farewell note she had sent to colleagues, in which she said that day would be her last. ^[23][24]

Yuchen He is a co-founder whose previous employers include xAI, where he helped develop the Grok chatbot, and OpenAI. He has described the company's plan to train its model in a way that has humans and AI systems interacting and collaborating together, using long-horizon and multi-agent reinforcement learning. ^[1][4]

Georges Harik is a co-founder and was Google's seventh employee, where he helped build the company's first advertising systems. ^[1] Reuters reported that he also worked on the launch of Gmail, initiated Google Docs and led Google's acquisition of Android. ^[6] His later research includes probabilistic programming, such as the 2010 paper "Variational Program Inference" with Noam Shazeer, and he co-authored the 2024 Quiet-STaR paper with Zelikman and Goodman. ^[29][30] He also participated in the seed round as an investor and co-led it. ^[2][6]

Noah Goodman is a co-founder and a professor of psychology and computer science at Stanford University. He was a doctoral adviser to Zelikman during the latter's time at Stanford. ^[1][5][33]

Forbes, citing LinkedIn, also described Ray Ramadorai, who previously designed systems for Microsoft's large data centers, as a founding member of the team. ^[9]

The seed round and valuation

Before the launch, Forbes reported on October 31, 2025, citing six sources, that Zelikman was seeking to raise $1 billion for Humans& and that the company was already negotiating a valuation of $5 billion. The report said, citing one source, that the company had told investors its new training method would require more compute than current approaches, and it quoted an investor who had declined to take part as saying the amount was too large for an early-stage company. ^[9]

On January 20, 2026, Humans& announced that it had raised $480 million in a seed round at a valuation of $4.48 billion, smaller on both counts than the figures in the earlier report. ^[2][6] Co-founder Andi Peng told Crunchbase News that the company would spend the majority of the capital on compute for training models, and Crunchbase reported that the round was raised all in cash and unstructured. ^[2]

On March 16, 2026, the law firm Goodwin, which advised the company, announced that it had guided Humans& in raising more than $500 million in the seed financing, at a valuation of approximately $4.5 billion. Goodwin named the same lead investors, SV Angel and Georges Harik, with participation from Nvidia, Jeff Bezos and GV. ^[7]

The financing drew attention because of its scale relative to the company's stage. At three months old and with no released product, the company raised what Crunchbase News called one of the largest seed rounds ever raised, and both Crunchbase News and TechCrunch identified Thinking Machines Lab's 2025 round as the largest seed round on record. ^[1][2]

Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, raised a $2 billion seed round in 2025 that Crunchbase described as by far the largest in its dataset; TechCrunch reported its valuation as $12 billion. Humans& has been grouped with this small set of so-called mega-seed rounds, in which experienced AI researchers raise unusually large amounts of early capital on the strength of their track records. ^[1][2]

Mission and focus

Humans& describes its mission in terms of human connection and collaboration. The company argues that progress for people comes from understanding one another, building trust, making connections, and working together, and it states that AI can be reimagined to center on people and their relationships rather than on individual task automation. In its framing, AI at its best should act as connective tissue that strengthens organizations and communities. ^[1][4]

Technically, the company says it is building a foundation model architecture optimized for social intelligence rather than for information retrieval or code generation. Its stated research areas include long-horizon and multi-agent reinforcement learning, memory systems, and modeling user understanding, all aimed at enabling planning and coordination among groups of people and AI systems over extended periods. The company has said it is designing its product alongside the model, and reporting has described its ambition as a coordination layer for teams that could eventually overlap with collaboration tools such as group messaging and shared documents. ^[1][4]

At launch, Reuters reported that the company expected to release a product early in 2026, and Zelikman told the agency: "The model will coordinate with people, and other AIs where appropriate, in order to allow people to do more and to bring them together." ^[6] The company also told TechCrunch that it had already turned away interested parties and was not interested in being acquired. ^[4]

Compute infrastructure

Humans& chose to own its training hardware rather than rent it. In a September 2026 post on X, Zelikman wrote that the company had built a cluster and owned the hardware, thanking DeepInfra, Nvidia and Supermicro for helping to make it happen. In a follow-up post he wrote that renting the compute of the company's current clusters "would likely have cost us over $1 billion on a 5-year contract," adding that more was on the way. ^[20][21] DeepInfra said it supported the Persimmon launch with DeepCluster, a dedicated NVIDIA Blackwell cluster that it deploys and operates. ^[22] The company's Ampersand announcement says training runs on hardware Humans& owns, in data centers set up in collaboration with DeepInfra, and the Ampersand model card credits B300 GPUs with "the machine work." ^[13][14]

The company has also published engineering research on low-precision training and inference. "The 4-bitter Lesson" (July 10, 2026), by Ziang Li and colleagues, describes a recipe for stable reinforcement learning that uses NVIDIA's 4-bit NVFP4 format for both weights and activations. The post credits collaborators at RadixArk and NVIDIA, and the company open-sourced the implementation in projects including NVIDIA's TransformerEngine, FlashInfer and SGLang. ^[18] A follow-up post on October 23, 2026 described a decoding pipeline that keeps weights in NVFP4 but leaves activations in 16-bit precision; the company said this made text generation up to 10% faster while being more accurate. It said the technique was used in Ampersand's reinforcement learning and in serving Boba, and that it had contributed the work to FlashInfer, SGLang and Miles. ^[19][27][28]

Models and products

Persimmon

Persimmon, released as a research preview on September 10, 2026, is what the company calls a "user model." Rather than acting as an assistant, it generates the next turn of a conversation from the perspective of simulated people, conditioned on a description of the setting, profiles of the users, and the conversation so far. ^[10][11] Humans& argued that frontier assistant models prompted to play users produce far less varied behavior than real people, which narrows what evaluations and training environments built on them can show. ^[10]

The model card lists Persimmon as initialized from NVIDIA's Nemotron 3 Ultra base model, which has 550 billion total parameters (55 billion active) in a hybrid of Mamba-2, mixture of experts and attention layers, with a context length of up to 1 million tokens. ^[11] Persimmon was mid-trained on conversations between people from the public internet and then post-trained with reinforcement learning on its own generations against an adaptive discriminator, which the company said improved coherence over long conversations. ^[10][11] Humans& said it used thousands of Blackwell-generation GPUs on a cluster it had recently built. ^[12]

The company evaluated Persimmon by "distribution matching," asking whether the model's responses look as if they were drawn from the same distribution as real human conversations rather than checking for a fixed list of traits. One test adapts the Turing test to group chats. ^[10] All of the results below come from the company's own evaluations.

Because the model is meant to reproduce human behavior, including disagreement, mistakes and sometimes adversarial behavior, the company said it does not use assistant-style guardrail and refusal criteria as the sole standard for it. Access is limited to a research and evaluation preview by application, with the stated aim of preventing misuse such as impersonation, fabricated testimony and consequential decisions about real people; academic applicants were encouraged to request API credit grants. ^[10][11] The model card credits Manya Bansal and Alexis Ross with co-leading the data, training and evaluation work, and Joey Hong with leading post-training. ^[11]

The launch post also reported Persimmon's scores on the User-Sim Index (USI), a benchmark proposed by Zhou et al. in 2026, while listing its limitations. ^[10] In October 2026 Zelikman argued on X that user models probably should not be evaluated on the USI, posting code for a trivial "user model" that he said scored 95% across the benchmark's four behavioral dimensions, and added that regex-based signals "shouldn't be headline results." ^[25][26]

Ampersand

Ampersand v0.1, released on October 20, 2026, is described by Humans& as "the first model trained for people to achieve more together," a collaborator model meant to help groups of people reach individual and shared goals over long periods. ^[13][14] It is post-trained from Z.ai's open-weight GLM-5.3, which the Ampersand model card describes as a mixture-of-experts transformer with about 744 billion total parameters (about 40 billion active) and a context length of 1,048,576 tokens. ^[14]

Ampersand was trained with reinforcement learning in long-horizon, multi-party social simulations in which Persimmon plays the simulated users, with rollouts spanning hundreds of thousands of tokens. ^[14] Because a single episode can span many trainer updates, the company proposed a new off-policy learning algorithm, Contrastive Off-policy Reward Optimization (CORO), which compares trajectories with one another instead of calibrating against an estimate of how well the model performs from each simulation seed. Training used a multi-agent framework built on Miles, the company's NVFP4 recipe, and a meta-optimizer that adjusts settings such as the reward and batch composition during a run so that unintended behaviors can be corrected without restarting. ^[13] In the Boba announcement the company called its method Reinforcement Learning from Simulated Feedback (RLSF), carried out "over millions of simulated days interacting with thousands of simulated people." ^[16]

The company's reported evaluations, which the model card describes as tentative, compare Ampersand with its GLM-5.3 base:

Sources: ^[13][14]

Humans& acknowledged that in its main simulation Ampersand incurred a higher interaction cost than other leading models, much of it from coaching people who simply wanted an answer or preferred to work things out themselves. ^[13] The model card reported that accuracy on the cybersecurity portion of the WMDP benchmark rose from 80.6% for GLM-5.3 to 89.0% for Ampersand, and that prompt-injection susceptibility stayed similar to GLM-5.3's, with a successful injection rate "substantially higher than frontier models." It also reported that conversations simulated by Persimmon looked more like real conversations with a user to several frontier assistant models, including GPT-6 Astra and GPT-6.1 Sol, than conversations simulated by Claude Opus 5.5, lowering those models' self-reported probability that they were being evaluated. ^[14]

Ampersand is offered without a waitlist through an Agents API that follows OpenAI's Agents API, so developers can use the official OpenAI SDK. In a group chat, an application gives the model two tools, send_message and pass, and after each message the model either replies or stays silent. ^[13] Launch pricing was $1.00 per million input tokens, $0.20 per million cached input tokens and $5.20 per million output tokens, including reasoning tokens. ^[15]

Boba

Boba, unveiled on October 22, 2026, is the company's first consumer product: an assistant built on Ampersand that people can text one-on-one or add to a group chat. ^[16] It works over iMessage, RCS and SMS or in a web browser, can make reminders and calendar invitations, handle email and research the web, and is offered free, with no subscription or ads. ^[16][17] At launch Humans& said Boba was available in several countries, including the United States, Canada and Australia, and that early users had already sent it more than a million messages. ^[16] In late October 2026 the Boba website cited more than 3 million messages exchanged by early users, more than 300,000 of them in group chats. ^[17] In a company survey of early users (290 completed responses, a 29% response rate), 70% said they loved Boba overall and 25% said they liked it. ^[16] The site says messages are stored to give the assistant memory and may be used to improve the company's AI and future products, but are not sold to advertisers or used to train third-party models. ^[17]

Investors

The seed round was led by Ron Conway's venture firm SV Angel together with co-founder Georges Harik. A range of additional investors participated. ^[2][6]

Emerson Collective is the firm founded by Laurene Powell Jobs. Reporting noted that the round also included participation from many other investors beyond those named. ^[1][2] The company's website lists further backers, including Conviction, Bloomberg Beta and AME Cloud Ventures (Jerry Yang), and individual investors such as Anne Wojcicki, Marissa Mayer, Bill Maris, Igor Babuschkin, Thomas Wolf and Logan Kilpatrick. ^[8]

Significance

Humans& attracted notice for two reasons that reinforced each other: the size of its seed round and the pedigree of its founding team. The combination of a roughly half-billion-dollar raise, a multibillion-dollar valuation, and founders with experience building models such as Claude and Grok placed the company among the most heavily capitalized early-stage AI ventures, despite having no product at announcement. ^[1][2]

The launch was also discussed as a notable example of a contrarian product thesis within the AI industry. While many leading laboratories framed their work around autonomous agents that operate with little human involvement, Humans& publicly committed to the opposite premise, that AI should augment human collaboration rather than replace human labor. Coverage treated the company as a test of whether "human-centric" or coordination-focused AI could compete with the prevailing agent-centric approach. ^[1][4]

The departure of co-founder Andi Peng in July 2026 drew coverage; RuntimeWire wrote that her exit removed "a key Anthropic-trained researcher" before the company's bet had produced a public product. ^[24] Between September and October 2026 the company released its first models and a consumer app. The benchmark results it published for Persimmon and Ampersand come from its own evaluations. ^[10][13][14]

Context

The Humans& financing came during a period of very large early-stage venture capital fundraising by AI startups led by well-known researchers. The most prominent precedent was Thinking Machines Lab, founded by Mira Murati, which raised a $2 billion seed in 2025. TechCrunch listed other so-called mega-seed rounds, including Unconventional AI's $475 million raise at a $4.5 billion valuation in December 2025 and Lila Sciences' $200 million seed round in March 2025. ^[1] Forbes compared Humans& with Thinking Machines Lab and Safe Superintelligence, labs that had raised large sums before launching any product. ^[9] Crunchbase data showed that more than 41% of the $38.4 billion invested in global seed funding in 2025 went to companies in AI-focused categories, and Crunchbase described 2025 as a record-setting year for U.S. seed rounds of $100 million or more. ^[2]

See also

• List of neolabs
• Neolab
• Thinking Machines Lab
• STaR (Self-Taught Reasoner)
• Quiet-STaR
• Nemotron 3
• GLM-5.3
• Multi-agent system

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20Humans&, a 'human-centric' AI startup founded by Anthropic, xAI, Google alums, raised $480M seed round, TechCrunch, January 20, 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Humans& Raises Huge $480M Seed Round At $4.48B Valuation For 'Human-Centric AI Lab', Crunchbase News, January 2026.
3. ^New AI Startup Humans& Raises $480M Seed at $4.8B Valuation, Built In, January 21, 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7Humans& thinks coordination is the next frontier for AI, and they're building a model to prove it, TechCrunch, January 25, 2026.
5. ^1 ^2 ^3 ^4Eric Zelikman, personal website, accessed June 2026 and October 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6Akash Sriram, AI startup Humans& raises $480 million at $4.5 billion valuation in seed round, Reuters (via Yahoo Finance), January 20, 2026.
7. ^1 ^2 ^3 ^4Goodwin Guides humans& in Raising Over $500 Million at $4.5 Billion Valuation in Seed Round, Goodwin Procter press release, March 16, 2026.
8. ^1 ^2 ^3AI for humans, company website (launch announcement dated January 20, 2026), accessed October 2026.
9. ^1 ^2 ^3Quién es Eric Zelikman, el ex xAI que busca mil millones para crear una "IA que colabore con humanos", Forbes Argentina (Spanish-language edition of the Forbes report), October 31, 2025.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Persimmon, humans& blog, September 10, 2026.
11. ^1 ^2 ^3 ^4 ^5 ^6Persimmon v0.1 Model Card, humans&, September 10, 2026.
12. ^humans& (@humansand), post on X about training Persimmon, September 10, 2026.
13. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11Ampersand, humans& blog, October 20, 2026.
14. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Ampersand v0.1 Model Card, humans&, October 20, 2026.
15. ^Pricing and limits, Ampersand developer documentation, humans&, accessed October 2026.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Boba, humans& blog, October 22, 2026.
17. ^1 ^2 ^3 ^4Boba · The AI that gets your goals, Boba website, accessed October 27, 2026.
18. ^1 ^2Ziang Li and colleagues, The 4-bitter Lesson: Balancing Stability and Performance in NVFP4 RL, humans& blog, July 10, 2026.
19. ^1 ^2Ziang Li and colleagues, Faster-than-light NVFP4 decode, with precision, humans& blog, October 23, 2026.
20. ^1 ^2Eric Zelikman (@ericzelikman), post on X about the humans& cluster, September 27, 2026.
21. ^Eric Zelikman (@ericzelikman), post on X about the cost of renting equivalent compute, September 28, 2026.
22. ^DeepInfra (@DeepInfra), post on X about DeepCluster and Persimmon, September 11, 2026.
23. ^1 ^2Andi Peng (@TheAndiPenguin), post on X announcing her departure from humans&, July 30, 2026.
24. ^1 ^2Andi Peng leaves humans& six months after its $480 million seed, RuntimeWire, July 30, 2026.
25. ^Eric Zelikman (@ericzelikman), post on X about the User-Sim Index, October 6, 2026.
26. ^Eric Zelikman (@ericzelikman), follow-up post on X about regex-based evaluation signals, October 6, 2026.
27. ^humans& (@humansand), post on X about the NVFP4 decode pipeline, October 23, 2026.
28. ^humans& (@humansand), post on X about open-source contributions, October 23, 2026.
29. ^Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber and Noah D. Goodman, Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking, arXiv:2403.09629, March 2024.
30. ^Georges Harik and Noam Shazeer, Variational Program Inference, arXiv:1006.0991, June 2010.
31. ^Eric Zelikman, Yuhuai Wu, Jesse Mu and Noah D. Goodman, STaR: Bootstrapping Reasoning With Reasoning, arXiv:2203.14465, March 2022.
32. ^1 ^2Jobs at humans&, humans& careers page (Ashby), accessed October 2026.
33. ^Noah Goodman, Stanford Profiles, accessed October 2026.
34. ^Andi Peng, personal website, accessed October 2026.