Richard Socher
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,695 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,695 words
Add missing citations, update stale details, or suggest a clearer explanation.
Richard Socher (born August 1983) is a German-born computer scientist, entrepreneur, and investor known for his work applying deep learning to natural language processing. He completed his PhD at Stanford University under Christopher Manning and Andrew Ng, where his thesis on recursive neural networks helped move the field of NLP toward neural methods [1][2]. He founded the startup MetaMind, which Salesforce acquired in 2016, then served as Salesforce's chief scientist before leaving to start the AI search engine You.com in 2020 [3][4]. In May 2026 he came out of stealth as the founder of Recursive Superintelligence, a startup pursuing self-improving AI that raised $650 million [5][6].
| Born | August 1983, Dresden, Germany [2] |
| Fields | Natural language processing, deep learning, machine learning |
| Education | University of Leipzig; Saarland University; PhD, Stanford University (2014) [2] |
| Known for | Recursive neural networks, the Stanford Sentiment Treebank, GloVe word vectors, founding MetaMind and You.com [1][7][8] |
| Current role | Founder and CEO, Recursive Superintelligence (2026); founder, You.com; founder and managing partner, AIX Ventures [5][9] |
Socher was born in Dresden in the former East Germany and attended secondary school in the city's Dresden-Plauen district [2]. He studied computational linguistics at the University of Leipzig, where he finished a degree in 2006, and then at Saarland University in Saarbrücken, completing further study in 2008 [2]. He moved to the United States for graduate school and enrolled at Stanford, where he earned a PhD in computer science in 2014 [1][2].
His doctoral committee was led by two researchers who would become central figures in modern machine learning. Christopher Manning, a computational linguist, and Andrew Ng, a machine learning researcher, served as principal co-advisors, with Percy Liang also on the committee [1]. The pairing was unusual at the time: Socher has often described championing neural networks for language tasks during a period when much of the NLP research community was skeptical that the approach would work [4][8]. His dissertation, titled "Recursive Deep Learning for Natural Language Processing and Computer Vision," received Stanford's award for the best computer science PhD thesis [1][2].
Socher's academic work centers on a family of models he called recursive deep learning, in which neural networks operate over tree structures rather than flat sequences. According to his Google Scholar profile, his publications have been cited well over 200,000 times, placing him among the most-cited researchers in NLP [7].
His most recognized paper is "Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank," presented at EMNLP in 2013 with co-authors including Manning, Ng, and the linguist Christopher Potts [10]. The paper introduced the Stanford Sentiment Treebank, a dataset of movie-review sentences annotated for sentiment at every node of their parse trees, and a model called the Recursive Neural Tensor Network that learned how sentiment composes as phrases combine. The work won the Association for Computational Linguistics Test-of-Time Award in 2023, a decade after publication [10].
In 2014 Socher co-authored "GloVe: Global Vectors for Word Representation" with Jeffrey Pennington, who was the lead author, and Manning [8]. GloVe is an unsupervised method that produces vector representations of words by factoring a matrix of word co-occurrence statistics drawn from a large corpus. It became one of the standard sets of pretrained word embeddings used across NLP research and was also recognized with an ACL Test-of-Time Award [8]. Socher's personal site describes him as having helped popularize the most widely used word vectors, though GloVe was a collaborative effort and the credit belongs to the three authors together [11].
After leaving academia for industry, Socher continued to publish. In 2016 his group released "Ask Me Anything: Dynamic Memory Networks for Natural Language Processing," which framed many language tasks as a question-answering problem handled by a network with an episodic memory component, along with a companion paper extending the idea to visual question answering [12]. In 2018 he and colleagues including Bryan McCann and Caiming Xiong published "The Natural Language Decathlon: Multitask Learning as Question Answering," known as decaNLP, which proposed casting ten different NLP tasks as question answering so that a single model could learn all of them [13]. The decaNLP work anticipated later interest in general-purpose models that handle many tasks through a common text interface.
Earlier in his career Socher was one of the authors of "ImageNet: A Large-Scale Hierarchical Image Database," the 2009 CVPR paper that introduced the ImageNet dataset under the direction of Fei-Fei Li [11]. ImageNet became the benchmark that drove the deep learning breakthroughs in computer vision later in the decade. Socher was a contributor among several authors rather than the project lead, and the dataset is most closely associated with Li's group at Princeton and Stanford [11].
Socher founded MetaMind in July 2014 [3]. The startup built a general-purpose platform for predicting outcomes on language, vision, and database tasks, and offered image recognition and text classification through the cloud [3]. It raised about $8 million from investors including the venture firm Khosla Ventures and Salesforce chief executive Marc Benioff, and Socher served as its chief executive and chief technology officer [3][9]. Salesforce acquired MetaMind in April 2016, reportedly for roughly $33 million, and folded its technology and team into the company's AI efforts [3].
Following the acquisition, Socher became chief scientist and an executive vice president at Salesforce, the enterprise software company best known for customer relationship management products [4][9]. He led the research organization that became Salesforce Research, overseeing teams working on fundamental and applied research, search, customer service automation, and a cross-product AI platform that handled both structured and unstructured data [9]. During this period his group produced work on multitask learning, question answering, and protein and code generation models. He left Salesforce in 2020 to start his own company [4].
In December 2020 Socher announced You.com, a search engine he co-founded with Bryan McCann, a former Salesforce NLP researcher [4]. The site opened a public beta in November 2021 and positioned itself as a more private, customizable alternative to Google, with results grouped into app-like sections that users could rearrange [14]. In December 2022, shortly after the release of ChatGPT, You.com added a chatbot called YouChat that returned conversational answers alongside live web results, an early example of combining a large language model with web search [14].
You.com raised money in several rounds. An initial investment was led by Benioff, followed by a $25 million Series A in 2022 led by Radical Ventures, a $50 million Series B in 2024 led by Georgian, and a $100 million Series C in 2025 led by Cox Enterprises that valued the company at about $1.5 billion [14]. Over time the company shifted its emphasis from consumer search toward AI tools and APIs aimed at businesses and at developers building their own AI products [4][14]. Socher served as chief executive.
Socher is a founder and managing partner of AIX Ventures, a venture capital fund focused exclusively on AI startups [9][15]. The fund describes itself as run by practitioners and its founding group includes Socher, his former PhD advisor Christopher Manning, and Kaggle founder Anthony Goldbloom [15]. AIX has invested early in a number of well-known AI companies, including Hugging Face, Perplexity, and Weights & Biases [15].
In May 2026 Socher launched Recursive Superintelligence, sometimes referred to simply as Recursive, a San Francisco company that emerged from stealth with $650 million in funding and a reported valuation of about $4.65 billion [5][6]. The round was co-led by Alphabet's GV and Greycroft, with participation from the venture arms of Nvidia and AMD [6]. According to TechCrunch, the founding team included AI researcher Peter Norvig, Cresta co-founder Tim Shi, and Tim Rocktäschel, who had led work on open-endedness and self-improvement at Google DeepMind [5].
The company's stated goal is to build recursively self-improving AI: systems that can identify their own weaknesses and rewrite their own code to address them, with the longer-term aim of automating scientific research [5][6]. Socher framed the work as making the full cycle of ideation, implementation, and validation of research ideas automatic, starting with a model that can improve its own codebase before expanding into fields such as physics, chemistry, and early-stage biology [5][6]. The idea connects to ongoing debates about recursive self-improvement and the risks and possibilities of highly capable AI systems.
Socher has received several honors over his career. His Stanford dissertation won the university's best computer science PhD thesis award, and two of his papers later received ACL Test-of-Time awards [1][8][10]. He was named a World Economic Forum Young Global Leader in 2016 and a WEF Technology Pioneer, and he appeared on TIME magazine's inaugural list of the 100 most influential people in AI in 2023 [9][16]. In April 2024 the Technische Universität Dresden, in his home city, awarded him an honorary doctorate for his contributions to deep learning and natural language processing and for supporting AI education in Dresden, where he helped set up a student innovation center at his former school [2].