AI News AI资讯 6h ago Updated 2h ago 更新于 2小时前 46

Deep Cogito Announces $43M Series A to Advance the Post-Training Engine for Frontier Intelligence Deep Cogito宣布4300万美元A轮融资,推进前沿智能的后训练引擎

Deep Cogito raised a $43 million Series A led by TQ Ventures, bringing total funding to over $56 million, with participation from Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler The company focuses on large-scale reinforcement learning and Iterated Distillation and Amplification (IDA) research aimed at recursive self-improvement in AI models Founded by Drishan Arora and Dhruv Malrana, former leaders of Google's AI Search products, Deep Cogito operates on t Deep Cogito完成4300万美元A轮融资,由TQ Ventures领投,Benchmark、Zscaler等参投,总融资超5600万美元 公司由前Google AI Search负责人Drishan Arora和Dhruv Malrana创立,专注后训练(post-training)与强化学习技术 核心研究方向为迭代蒸馏与放大(IDA),旨在实现模型的递归自我改进,突破人类训练数据限制 已发布3B至600B+参数的开源Cogito模型系列,并为企业客户提供基于专有数据的定制化模型服务 资金将用于扩展研发团队、扩大训练基础设施,并推进下一代Cogito模型及企业级应用

65
Hot 热度
68
Quality 质量
62
Impact 影响力

Analysis 深度分析

TL;DR

  • Deep Cogito raised a $43 million Series A led by TQ Ventures, bringing total funding to over $56 million, with participation from Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler
  • The company focuses on large-scale reinforcement learning and Iterated Distillation and Amplification (IDA) research aimed at recursive self-improvement in AI models
  • Founded by Drishan Arora and Dhruv Malrana, former leaders of Google's AI Search products, Deep Cogito operates on the thesis that post-training—not pre-training—will define the next frontier of AI capability
  • The company has demonstrated post-training at scale through its open-weight Cogito model family (3B to 600B+ parameters) and is now offering enterprise customers the ability to build specialized models on proprietary data
  • Zscaler serves as both a strategic customer and investor, having worked with Deep Cogito to train intelligence directly into models rather than relying on lightweight customization

Why It Matters

Deep Cogito's focus on post-training as the critical differentiator in AI capability challenges the industry's heavy emphasis on pre-training scale, signaling a strategic pivot toward making existing models significantly more capable through reinforcement learning and self-improvement loops. The company's work with enterprise customers like Zscaler demonstrates a practical pathway for organizations to build proprietary, specialized intelligence rather than relying solely on generic frontier models, which has direct implications for how companies approach AI integration and competitive advantage.

Technical Details

  • Deep Cogito's core research direction is Iterated Distillation and Amplification (IDA), a process where a model uses additional computation to produce answers beyond its direct generation capacity, then distills those improvements back into its weights, enabling recursive self-improvement
  • The company has demonstrated post-training at scale through its open-weight Cogito model family, spanning sizes from 3B to 600B+ parameters, showing measurable capability improvements through reinforcement learning
  • Their approach goes beyond lightweight customization, working closely with enterprise customers to understand product-specific metrics and training intelligence directly into model weights using proprietary data, decisions, and outcomes
  • The long-term research goal is to build models that progressively improve their own capabilities and eventually move beyond the limits of human-generated training data
  • The same post-training engine powering the Cogito model family is being productized into a platform for enterprises seeking specialized intelligence for their own products

Industry Insight

  • Post-training is emerging as a critical competitive layer in AI, and companies that master large-scale reinforcement learning and self-improvement methodologies will likely differentiate themselves as pre-training advantages plateau among well-funded labs
  • The Deep Cogito-Zscaler partnership model—where an enterprise customer becomes a strategic investor—illustrates a growing trend of vertical integration between AI infrastructure providers and industry adopters, suggesting that domain-specific post-training will become a key value proposition
  • The thesis that "post-training determines what a model can actually become" implies that organizations should invest in post-training capabilities and proprietary data pipelines as a sustainable moat, rather than competing solely on access to frontier base models

TL;DR

  • Deep Cogito完成4300万美元A轮融资,由TQ Ventures领投,Benchmark、Zscaler等参投,总融资超5600万美元
  • 公司由前Google AI Search负责人Drishan Arora和Dhruv Malrana创立,专注后训练(post-training)与强化学习技术
  • 核心研究方向为迭代蒸馏与放大(IDA),旨在实现模型的递归自我改进,突破人类训练数据限制
  • 已发布3B至600B+参数的开源Cogito模型系列,并为企业客户提供基于专有数据的定制化模型服务
  • 资金将用于扩展研发团队、扩大训练基础设施,并推进下一代Cogito模型及企业级应用

为什么值得看

Deep Cogito代表了AI行业从"预训练竞争"向"后训练竞争"转移的关键趋势,其递归自我改进技术路线可能重塑模型能力边界。对于从业者而言,该公司将前沿研究与企业落地结合的模式,为后训练赛道的商业化提供了可参考的范本。

技术解析

  • 核心研究方向:大规模强化学习与递归自我改进(recursive self-improvement),通过IDA(Iterated Distillation and Amplification)技术让模型利用额外计算生成超越自身能力的输出,再将改进蒸馏回模型权重
  • 模型规模:已发布3B至600B+参数的开源Cogito模型系列,展示了在后训练阶段显著提升模型能力的能力
  • 企业级应用:为Zscaler等客户提供基于专有数据、决策和结果的定制化模型训练,超越轻量级微调,实现深度领域知识内化
  • 技术团队背景:创始团队来自Google AI Search,曾主导Gemini后训练及AI Mode、AI Overviews等产品开发

行业启示

  • 后训练成为新战场:随着预训练资源日益集中,后训练(强化学习、自我改进)正成为拉开模型能力差距的关键层,非头部实验室也开始展现规模化后训练能力
  • 企业专有数据价值凸显:通用前沿模型难以满足深度专业化需求,基于企业私有数据的后训练定制将成为B端AI落地的重要路径
  • 递归自我改进的长期意义:IDA等技术路线若成功,将推动模型突破人类训练数据瓶颈,实现持续自主进化,这可能重新定义AI能力发展的天花板

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Funding 融资 Research 科学研究 Training 训练 RLHF RLHF Alignment 对齐