Google, OpenAI, DeepSeek Battle it out in AI

Good morning,

To get access to the best of my work, consider subscribing for less than $2 a week.

While DeepSeek felt like an existential threat to U.S. AI a few days ago, Google has released Gemini 2.0 Pro Experimental. Gemini 2.0 Flash Thinking Experimental is getting good reviews (shows its work + one million context). We can expect reasoning models to improve fairly quickly in 2025.

Gemini 2.0 Pro Experimental finally arrives about one year after Gemini 1.5 Pro model released back in February, 2024. Google WeatherNext, Gemini 2.0 Flash, PaliGemma 2, DataGemma, parts of Project Astra going live, and many other projects means Google actually had a pretty good year in 2024 that is likely to spill over into 2025.

It's All in the Math

OpenAI, scared of the DeepSeek hype, on Friday released o3-mini, an AI model that focuses on improved reasoning capabilities at a more modest cost than its predecessor. They promise faster response times compared to o1, reduced computational requirements and the ability to handle simpler queries with ease, all with a lower cost. OpenAI even released a Deep Research product, a name of a product already by Google. OpenAI sounds like a fairly desperate AI startup at this point.

Google’s Late 2024 Surge

While Meta had its year of efficiency to cut costs, Google had its year of innovation to respond to OpenAI. For the record, I consider Amazon, Apple and even Microsoft total failures in Generative AI, at least up until February, 2025.

Bleeding edge models mean Google, OpenAI and Anthropic are battling DeepSeek, Qwen and ByteDance out in 2025. While there are many other players both in the U.S. and China, AI infrastructure spending could continue to lower that number of model builders able to keep up. My condolences to Mistral, Cohere, 01.AI (now with Alibaba) and many smaller others. I reckon China’s ability to scale apps and products globally with Generative AI features is going to increase substantially this year.

While OpenAI has trippled its paid ChatGPT users from 6 million in 2023 to nearly 16 million in 2024, DeepSeek R1 only shows Chinese models are going to get better, cheaper and more accessible. DeepSeek is able to iterate on models much faster without commercial and product fixations. They are a free untethered research lab, with another Chinese genius getting his due.

The Man Behind DeepSeek (Liang Wenfeng)

• I really enjoyed this documentary on Liang Wengfeng, the Math prodigy behind DeepSeek, by Easy Money (a YouTube Channel). Watch the entire video here.

?? 2:24 Video: link (?? recommended)

The blogger that moved Nvidia’s Stock

A blogger called Jeffrey Emanuel sparked panic in Wall Street by his bear thesis on Nvidia after DeepSeek was trending. On January 24th, Jeffrey Emanuel sat down in his Brooklyn apartment and started writing a blog post. For hours, he pounded away on his keyboard while his wife kept their young children occupied and brought him food. Emanuel worked late into the night, and by early Saturday morning he had written nearly 12,000 words.

• The post was widely shared on X,
• Jeff’s main point shook Silicon Valley because it one of his main points simply suggested that they were nowhere near as smart and efficient as Wall Street was touting them as.

Read the viral post

At the time of writing Nvidia’s stock is only down 13% in 2025. But that still represents several tens of Billion dollars. While Meta AI and Mistral look to be very impacted by DeepSeek’s open-source models, Anthropic’s CEO looks to be on damage control.

Dario, the CEO of Anthropic, has now blamed all his failures on the lack of chip export controls on China after seeing DeepSeek’s rise. It's reasonable—DeepSeek is bad news for OpenAI, but an absolute nightmare for tier-2 LLM players like Anthropic. - JW on X.

Meanwhile Google is reportedly developing an interactive mind map feature for its AI-powered note-taking tool, NotebookLM (copying genspark), as Claude is about to be getting a web-search feature (copying OpenAI). While Silicon Valley is copying itself, Qwen, Mistral and ByteDance also released new models.

Here is one of the mind maps that already existed on Genspark on this topic about NotebookLM doing the same:

It’s “Four Seasons in one day” in AI’s pace of model release

Models from xAI, Meta AI and Claude’s reasoning model are likely to be rolled out soon. GPT-5 has to be close as well.

DeepSeek’s impact and the renewed competitiveness of China’s AI is an ongoing story that Silicon Valley has so far tried to discredit, call a hack, and bring back techno-historical theories to try and prove why BigTech is spending a quarter of a trillion dollars on AI Infrastructure and energy infrastructure to power huge AI datacenter campuses in 2025 which we don’t even have the power for yet.

The Jevons Paradox, as first formulated in 1865 by English economist William Stanley Jevons, states that greater efficiency in the use of any given resource can result in increased demand for that resource.

McKinsey’s new 2025 AI report :

Superagency in the Workplace

Read the Report

McKinsey research sizes the long-term AI opportunity at $4.4 trillion in added productivity growth potential from corporate use cases.

The report, while appearing to be sponsored by Reid Hoffman, does summarize a few things fairly well:

We’ve come a long ways in 3 years

OpenAI reaching Peak State and National Adoption

OpenAI were in damage control scrambling to do a Reddit AMA, release mini-o3 and pretend like they have a moat to DeepSeek and other open-source model makers. Meanwhile OpenAI is getting dangerously chummy with the U.S. state. OpenAI says it plans to let U.S. National Laboratories, the Department of Energy’s network of R&D labs, use its AI models for nuclear weapons security and other scientific projects. What could possibly go wrong?

Not only has the U.S. and Microsoft positioned OpenAI to be a monopoly in B2C AI, but is giving it increasing access to some of the most sensitive parts of Government and the Department of Defense. Yikes. Meanwhile OpenAI says it will make its own AI-specific hardware and robots (along with their partnerships with Figure AI whom they are likely to acquire), and is likely to become the most valuable Unicorn startup soon.

The entire article is a 16-min read to read it go here.