DeepMind began in 2010. I find the technical progression of their Go software the most impressive part of their history. DeepMind’s acquisition by Google and its subsequent successes shifted AI research funding. In March 2016, AlphaGo Lee defeated Lee Sedol 4-1 using 48 TPUs. This version reached an Elo rating of 3,739. In October 2017, AlphaGo Zero (40 block) used 4 TPUs to defeat AlphaGo Lee 100:0. AlphaZero (20 block) then beat AlphaGo Zero 60:40 in December 2017 using 4 TPUs.
| Version | Hardware | Elo Rating | Result |
|---|---|---|---|
| AlphaGo Fan | 176 GPUs | 3,144 | 5:0 against Fan Hui |
| AlphaGo Lee | 48 TPUs | 3,739 | 4:1 against Lee Sedol |
| AlphaGo Master | 4 TPUs | 4,858 | 60:0 against professionals |
| AlphaGo Zero (40 block) | 4 TPUs | 5,185 | 100:0 against Lee |
DeepMind’s successes drove interest in AI research funding. The early versions of AlphaGo required significant hardware, like the 176 GPUs used by AlphaGo Fan to win 5:0 against Fan Hui in October 2015. By the time AlphaGo Master competed in May 2017, it used 4 TPUs and achieved a 4,858 Elo rating, defeating professional players 60:0.
Shifting research funding
Investment in the AI sector changed the research landscape. In 2016, the number of equity deals in AI startups globally grew by a factor of four to 658, and the total value of funding grew by a factor of over eight to $5 billion. This growth continued as global AI startup funding reached $68 billion in 2021, compared to $39 billion in 2020. The US is the largest hub for AI investment. It produces 30% of all papers, while China produces 18%. In 2011, the US National Science Foundation received $636 million for computer science research, and $169 million went to Information and Intelligent Systems. Within that budget, $56 million went to Robust Intelligence.
I note that companies often prioritize commercial opportunities over pure scientific value because they have financial incentives to promote their own products. This concentration of wealth in the Global North leaves the Global South with less than 5% of research contributions. South Korea also planned to spend 1 trillion won, or $840 million, by 2020 on AI. The Chinese government aimed for a $15 billion AI market by 2018. The Chinese government defines artificial intelligence as a "branch of computer science where machines have human-like intelligence," including robots and natural language processing. DeepMind’s evolution from Go software to frontier intelligence models defines the current era of AI.
Frontier intelligence models
The Gemini 4 era focuses on frontier intelligence models. Gemini 4 Argon achieved a 14% improvement in accuracy compared to Gemini 3.1 Pro in evaluations by Box. This model handles complex tasks like reasoning through experimental data. It works for legal professionals by identifying relevant citations and improving contract reviews. You should know, as you already follow these developments, that Gemini 4 Argon is at the frontier of enterprise knowledge work.
The model also assists in medical settings. Clinicians use Gemini 4 to analyze complex synthetic longitudinal charts, which helps in producing concise problem lists without external lookups. This shows high recall and specificity for ICD-10 codes. In engineering, the model helps with 3D spatial reconstruction, turning video into physics-ready 3D scenes. In legal work, Gemini 4 Argon connects dates, documents, and financial calculations with legal precedents to develop substantive arguments. Will the industry ever prioritize ethics over the profit motives that drive these rapid developments?
