# AlphaGo & Gemini Veterans Launch Reflection AI — $2 Billion Funding Secured
**Report by Xinzhi Yuan**
---
## Introduction
Former masterminds behind **AlphaGo** and **Gemini** have joined forces to create **Reflection AI**.
The startup has secured **$2 billion in funding** — with **NVIDIA** among the backers — to build a *ten-trillion-token-scale* model.
The goal: **prevent AI from being dominated by a small elite**.
---
[📊 Explore the global AI model landscape — *Xinzhi Yuan's 10th Anniversary Special: 2025 ASI Frontier Trends Report* (37 pages)](https://mp.weixin.qq.com/s?__biz=MzI3MTA0MTk1MA==&mid=2652625640&idx=1&sn=599fde2abe811219a22711fe44172c70&scene=21#wechat_redirect)
---
## AI Capital Boom
- **PitchBook** reports that AI foundation model companies raised **$34.9B** last year.
- This year’s total has already **doubled to $71.9B**.
- **Reflection AI**, founded just over a year ago by a former DeepMind researcher, shocked the market with a $2B round and an immediate $8B valuation.
**Investors include:**
- Former Google CEO **Eric Schmidt**
- **NVIDIA**
- **Sequoia Capital**
- **Citigroup**
---
## The Open vs. Closed-Source Battle
Reflection AI positions itself as the **"U.S. version of DeepSeek"**, aiming to reclaim technological leadership with **Open Intelligence** — fully open models, datasets, and papers.
**Philosophy:**
Enable any organization to **fine-tune, deploy, and audit** frontier AI systems — without monopoly control.
---
## Founding Team
- **Misha Laskin** — led reward modeling for *Gemini*
- **Ioannis Antonoglou** — contributed to *AlphaGo* (defeated world Go champion, 2016)
The startup employs ~60 researchers/engineers specializing in infrastructure, data training, and algorithms. It plans to release a **frontier language model at ten-trillion-token scale** next year.
---
## Growing Open-Source Momentum
While critics fear open models could be misused, investors are betting heavily on openness:
> *Sequoia’s Stephanie Zhan*: “This is the moment of reckoning for AI.”
---
## Crucible Moment: Strategic Inflection
True turning points arrive subtly — today’s decisions will shape **AI’s trajectory for decades**.
Detecting and acting on strategic inflections will determine winners in the next intelligence race.
---
## “U.S. DeepSeek” Vision
Reflection AI’s mission parallels other global open initiatives, like **[AiToEarn官网](https://aitoearn.ai/)** — a platform enabling creators to generate AI content, publish to multiple channels, and monetize.
Laskin warns:
- Western open models currently trail Chinese competitors like DeepSeek.
- Without a strong domestic open platform, the U.S. risks losing leadership.
---
## Funding Pace
- March: $130M round closed.
- Seven months later: $2B secured.
- Laskin expects *more capital will be needed* — especially with rivals like OpenAI attracting up to $100B from NVIDIA.
Market demand for open models is rising among **enterprises** and **governments** seeking **control over their AI stack**.
---
## Why Openness Matters
Technological progress thrives on openness:
- **Internet**
- **Linux**
- **Open protocols**
Open-source enables learning, customization, deployment, and influence — across universities, startups, enterprises.
Today’s core AI technologies (e.g., self-attention, RL) emerged from shared knowledge.
The risk: closed labs monopolizing **capital, compute, and talent**.
---
## Reflection AI’s First-Year Achievements
Team contributions include:
- **PaLM**
- **Gemini**
- **AlphaGo**
- **AlphaCode**
- **AlphaProof**
- **ChatGPT**
- **Character AI**
Capabilities built:
- Large-scale training platform for **Mixture-of-Experts** models.
- Breakthrough in **automatic programming**.
- Expansion to **agentic reasoning** tasks.
---
## Race Toward Superintelligence
Reflection AI aims to build **autonomous programming systems**, replacing programmers entirely:
> *Ioannis Antonoglou*: “Autonomous programming is AGI-complete.”
---
## Code as Natural UI for LLMs
**Why start with software engineering?**
- Software ecosystems are **machine-friendly**.
- **Programming languages** are intuitive for language models.
- Future GUIs may be replaced by language models executing tasks via code.
---
## Training Autonomous Agents
Best training occurs in **customized environments** — akin to DeepMind’s Atari setup or OpenAI Gym.
Breakthroughs in broader cognitive tasks will require **creativity and experimentation**.
---
## Physics Inspiration
Misha Laskin draws from physics, noting Richard Feynman’s stance:
> “We do not truly know what energy is — and that’s important.”
Similarly, AI scientists still seek the definition of “intelligence” — and the most powerful way to understand it is to **build it**.
---
## Closing Thoughts
We have a narrowing window to create **cutting-edge open intelligence systems**.
Missing this opportunity could mean ceding control of AI’s foundation to a few closed entities.
---
**References:**
- [https://x.com/reflection_ai/status/1976304405369520242](https://x.com/reflection_ai/status/1976304405369520242)
- [https://www.sequoiacap.com/article/reflection-ai-spotlight/](https://www.sequoiacap.com/article/reflection-ai-spotlight/)
- [https://x.com/stephzhan/status/1976326493291807117](https://x.com/stephzhan/status/1976326493291807117)
---
---
## Related Platform: AiToEarn
For creators aiming to exploit AI’s potential and share globally, **[AiToEarn官网](https://aitoearn.ai/)** offers:
- AI-powered content generation
- Simultaneous publishing to platforms: Douyin, Kwai, WeChat, Bilibili, Rednote, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)
- Performance analytics
- AI model ranking: [AI模型排名](https://rank.aitoearn.ai)
---
[Read the original](2652633206)
[Open in WeChat](https://wechat2rss.bestblogs.dev/link-proxy/?k=fae82078&r=1&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzI3MTA0MTk1MA%3D%3D%26mid%3D2652633206%26idx%3D1%26sn%3D04c5952a305a83fc26d2e9564b053145)
