I'm on the same page as Ilya. A top AI expert under Elon Musk has left to develop an AI with "empathy."

Right after its establishment, an AI startup is seeking funds with a valuation of $28 billion. What's its background?

According to a report by Zhidongxi on November 26th, foreign media Business Insider cited sources familiar with the matter as saying that the US AI startup Humans& is raising $1 billion (approximately RMB 7.11 billion) with a target valuation of $4 billion (approximately RMB 28.43 billion).

The founder of this startup, Eric Zelikman, left Elon Musk's large - model unicorn xAI in September this year.

At xAI, as an AI researcher, Zelikman was deeply involved in the R & D of Grok 2, Grok 3, and Grok 4 agents. He also pioneered the STaR algorithm, the first algorithm to train language models for natural language reasoning through self - generated reasoning chains.

With such little experience, what gives Humans& the confidence to aim for a high valuation?

Zelikman believes that the reinforcement learning (RL) paradigm has limitations because models often solidify existing biases rather than providing novel insights. The focus of model research should shift to assisting individuals rather than trying to replace them.

Therefore, Zelikman founded Humans&, aiming to develop models that can learn user behavior and empathize with users.

He believes that by improving human - centered models, AI is more likely to achieve lofty goals that have been difficult to reach so far, such as curing cancer.

▲ Eric Zelikman, former xAI researcher, pioneer of the STaR algorithm, co - founder and CEO of Humans& (Source: Eric Zelikman's personal website)

01 From the core R & D at xAI to the pioneer of the STaR algorithm, a genius researcher spanning academia and industry

Eric Zelikman entered Stanford University in 2016 to study Symbolic Systems. After graduating with an honor degree in Symbolic Systems in 2020, he directly started his doctoral studies. In 2024, he suspended his studies and became an AI researcher at xAI.

▲ Zelikman's educational background (Source: Eric Zelikman's personal website)

During his more than one - year tenure at xAI, as an early member, Zelikman was deeply involved in the construction of pre - training data for Grok 2, led and scaled up the reinforcement learning reasoning framework for Grok 3, and established the reinforcement learning infrastructure and training plan for the Grok 4 agent.

▲ Zelikman's work experience (Source: Eric Zelikman's personal website)

During his time at xAI, he also pioneered the STaR algorithm. His most well - known achievement is co - authoring Quiet - STaR — a method that allows language models to generate internal reasoning processes before generating output, which significantly improves the model's performance in tasks such as CommonsenseQA and GSM8K.

▲ The paper on the STaR algorithm (Source: arxiv)

In addition, he developed the Parsel framework, which enhances the algorithmic reasoning ability of language models through a combinatorial decomposition method, with a pass rate on complex programming tasks more than 75% higher than previous methods.

Zelikman's academic papers have been highly recommended (top 8%) at top conferences such as ICLR 2022, NeurIPS 2022, COLING 2022, EMNLP 2022, and ICML 2023.

He has also won the Best Reviewer Awards (top 1 - 1.5%) at ACL 2023, NeurIPS 2023, EMNLP 2023, and ICLR 2024 consecutively.

▲ Excerpts of Zelikman's paper achievements (Source: Eric Zelikman's personal website)

In September this year, Zelikman left xAI and returned to campus to continue his doctoral studies at Stanford University. He also founded Humans&.

02 "Today's language models are too cold - hearted." Zelikman wants to create AI that truly understands humans

"Humans& (AI)" means humans and AI. As the name suggests, the mission of Humans& is to create AI that can better understand humans.

▲ The homepage of Humans& (Source: Humans&)

In October this year, Eric Zelikman talked about his journey of founding Humans& on the podcast "No Priors" hosted by venture capitalist Sarah Guo.

Zelikman said that during his doctoral studies, he concluded from his research on simulating different types of students through training language models that accurately modeling user characteristics can design more suitable human service systems.

▲ Zelikman participating in the "No Priors" podcast (Source: YouTube - No Priors)

He found that the actual application depth of current top - tier models is far from reaching their capacity limits. The core obstacle is that they lack the ability to understand human goals. Zelikman believes that the existing training paradigm focuses too much on single - task scenarios and lacks consideration of long - term impacts.

"The most fundamental problem is that when you treat each round of dialogue as an independent game, the model actually cannot understand the long - term impacts of its own words and actions."

At the TED AI Los Angeles event in October this year, Eric Zelikman was invited to give a speech. He said that we are currently in a task - centered reinforcement learning (RL) paradigm. However, this training method has limitations because models often solidify existing biases rather than providing novel insights. Therefore, the focus should shift to assisting individuals rather than trying to replace them.

He believes that humans are still the source of innovation.

Therefore, Zelikman founded Humans&, aiming to develop models that can learn user behavior and empathize with users. "The core goal of the model should be to understand you," he said. "Although it's difficult to achieve perfectly, it will definitely make a huge breakthrough compared to existing models."

Zelikman believes that by improving human - centered models, AI is more likely to achieve lofty goals that have been difficult to reach so far, such as curing cancer.

"By building models that are truly good at collaborating with large groups and truly understand the goals, aspirations, and values of different people, our chances of solving these fundamental human problems will increase significantly."

Currently, the technical team of Humans& is still recruiting. From the recruitment information for technical staff on its X (formerly Twitter) homepage, we know that Zelikman offers a minimum annual salary of $350,000 (approximately RMB 2.48 million) for technical staff, and the office is located in the San Francisco Bay Area, USA.

▲ Recruitment information of Humans& (Source: X)

03 Conclusion: Making AI more human - friendly is the current development trend

When GPT - 5.1 was released, OpenAI claimed that it had made improvements in both intelligence and communication methods. It not only provided tone control options and personality options, but users could also directly adjust the conciseness, friendliness, readability of the responses, and the frequency of using emojis.

AI guru Ilya also mentioned in a recent interview the importance of value functions such as "emotion" for further improving model capabilities.

This also reveals the next competitive dimension in AI development: future models not only need higher IQ but also "EQ" for collaborating with humans. This shift in demand from "function" to "emotion" marks that AI is moving from being a tool to being more interactive and human - friendly.

This article is from the WeChat official account "Zhidongxi" (ID: zhidxcom). The author is Wang Han. It is published by 36Kr with authorization.