The week in AI: Trump announces $500B Stargate Project; Musk denounces $500B Stargate Project

In partnership with

Welcome to The Dispatch! We are the newsletter that keeps you informed about AI. Each Thursday, we aggregate the major developments in artificial intelligence - we pass along the news, useful resources, tools and services; we highlight the top research in the field as well as exciting developments in open source. Even if you aren’t a machine learning engineer, we’ll keep you in touch with the most important developments in AI.

NEWS & OPINION

Stargate Project announced; billionaire cafeteria food fight swiftly follows

-------------------------

Some of the biggest names in technology - from Elon Musk and Sam Altman to Microsoft CEO Satya Nadella - are trading barbs after President Donald Trump unveiled an unprecedented $500B investment in building AI infrastructure across the United States.

At the White House on Tuesday President Trump, Altman and Oracle CTO Larry Ellison were beaming as SoftBank CEO Masayoshi Son confidently predicted that artificial superintelligence will kick off America’s “golden age.” The joint venture from OpenAI, Softbank and Oracle, called “Stargate,” will commit $500B in AI infrastructure investments, including “colossal data centers” and energy development in the U.S.

It was an epochal declaration that, according to Trump, would be "the largest AI infrastructure project by far in history" and would help keep "the future of technology" in the US. Everything seemed grand - until the world’s richest man and First Buddy made his presence known.

“They don’t actually have the money,” Musk wrote in a series of posts on X on Tuesday. “SoftBank has well under $10 billion secured. I have that on good authority.” He reposted an image of a crack pipe with the joking allegation that Altman and his associates were smoking it, and later declared “Sam is a swindler.”

Sam’s initial response was respectful and measured, but he shifted stance barely an hour later and shot back - implying Musk knows he is lying and asking if the world’s richest man would like to visit the first site already under way. He then stated, “I realize what is great for the country isn't always what's optimal for your companies, but in your new role I hope you'll mostly put (America) first.”

When prompted about the issue in an interview with CNBC, Nadella, whose company is still tied to OpenAI, said with a sly grin, “All I know is I’m good for my $80 billion.” Microsoft is not an investment partner for the Stargate Project - though they are a technology partner along with Nvidia and British chip designer Arm. Nadella’s comments rather aimed to highlight Microsoft’s own investments, and that these are not outlandish numbers (well, they are) given the amount of infrastructure being planned broadly for AI.

Wharton Professor Ethan Mollick (who we frequently highlight as one of the few emerging voices of reason in AI) sounded a different note of skepticism about the project on X, wondering what all this competition is speeding us toward. “For those convinced they are making AGI soon,” he asked, “what does daily life look like 5-10 years later?” He went on to note that universal basic income is a policy decision, not a description of what life will look like in a world of highly advanced AI.

OpenAI claims they are deploying $100B for the Stargate Project immediately, and the venture is expected to draw additional funders. Altman will brief the White House on ‘PhD level SuperAgents’ on Jan. 30th.

A Chinese lab shocks the AI community with DeepSeek R1

-------------------------

Hedge-fund backed Chinese AI lab DeepSeek just released DeepSeek-R1, an open-source family of reasoning models that reportedly matches or exceeds OpenAI's advanced reasoning model, o1, on certain benchmarks while being MIT-licensed for commercial use. There is a lot to unpack here, as this is probably the most important (or at least shocking) development in open source AI to date:

• Unlike traditional GPT models, R1 uses a “reasoning approach” similar to OpenAI’s o1 that takes longer, but produces more reliable results in domains like physics, science, and math. You can see R1’s line of thinking before it comes up with the final result, but there’s some debate around the overly anthropomorphized/people pleasing on display in the thinking window. The model, coming from China, is also heavily censored on some topics and focused on math/coding/logic problems. It might refuse to attempt more creative or abstract requests outside of those domains.

• R1’s 671B parameter model is the largest in the family (click the DeepThink button when using their chatbot to try), and that model is exceptional - by far the most powerful available in open source. Benchmarks aside, all you have to do is use the model for a few advanced inquires to test its mettle - you’ll likely come away impressed.

• Not only did DeepSeek open source the model weights and training methodology, the R1 chatbot linked above is free to use and the API is, astoudingly, up to 535 times cheaper than OpenAI’s o1 API ($0.14 vs $7.5 per million input tokens).

• In May 2025, Nvidia’s DIGITS will be released at $3000. It's estimated that a cluster of four of these linkable Grace Blackwell supercomputers would be powerful enough to run the 671B parameter DeepSeek-R1 model locally. That’s still a steep investment for a consumer, but just two months ago, the idea of running such a powerful reasoning model on a $12,000 personal supercomputer setup would have seemed beyond impossible. The democratization of this level of AI power is accelerating faster than anyone predicted.

• A number of the distilled models in the family are much smaller size, with the lowest at 1.5B parameters (which you could easily run locally on a laptop). Those models do well on benchmarks, but in practice get tripped up easily and hallucinate frequently. These models are an interesting proof of concept, though: DeepSeek distilled them based on different models (Qwen 2.5 and Llama 3), and found that when they supplemented those models with R1’s reasoning outputs (referred to above) - guess what? They got better at reasoning. Not much else, though.

• DeepSeek’s team does not include any well-known AI personalities or members returning from overseas to China. They headhunted (see part 4) “fresh graduates from top universities, PhD candidates in their fourth or fifth year, and some young people who graduated just a few years ago” from the Chinese university system. They also weren’t an existing AI lab, they were created by a hedge fund with a specific mission in mind.

So, what are the implications? Hard to say at this point, but potentially profound. And for American tech giants, what is the moat? The US has bet heavily on sanctioning advanced AI chips from going to China, but that bet continues to show holes. Let’s also not forget that it isn’t “OpenAI” (which abandoned its original open-source mission at the jump), but a Chinese AI lab openly sharing the most advanced models and training methodology in the space.

Google’s still here, guys

-------------------------

Behind the mainstream AI news, Google very quietly released yet another updated version of their outstanding Gemini 2.0 Flash Thinking model (make sure to check the model - it’s gemini-2.0-flash-thinking-exp-01-21). Unlike either o1 or DeepSeek, the “reasoning” portion from Google’s thinking model is remarkably clear and concise - which is a great extra layer of information tied to the response and makes it easier to identify hallucinations.

The best part? G2.0FT is free on Google’s AI studio, their hub for devs, with very generous usage limits for such a powerful model. You can also control the model’s temperature (lower temperature = less creative), system instructions and other useful settings directly within the UI to enhance/optimize your desired outputs quickly.

It also has a 1 million token context window (the limited amount of text from the conversation or document that the model keeps "in mind" while processing) - about 8x as large OpenAI’s o1. G2.0FT has already topped the LMarena leaderboard, and is in our opinion the best overall reasoning model currently accessible anywhere. Google’s TPU’s are extremely efficient for test-time compute (which is how reasoning models differ from non-reasoning), so it’s possible or even likely the platform is free because the “reasoning” data you’re giving to Google by using the model is more valuable than whatever it costs them to run your query, going forward - same with OpenAI making o3-mini available for ChatGPT’s free tier in ~two weeks.

Give it a shot, and don’t sleep on Alphabet.

MORE IN AI THIS WEEK

• Trump revokes Biden executive order on addressing AI risks

• OpenAI reportedly needs so much compute that Microsoft can’t keep up; Microsoft relinquishes role as company’s exclusive cloud provider

• Changing popular narratives in AI to empower workers

• LinkedIn accused of using private messages to train AI

• Google agrees to new $1 billion investment in OpenAI rival Anthropic

• Databricks closes $15.3B financing at $62B valuation, Meta joins as ‘strategic investor’

• Busy week for Perplexity: acquires networking platform read.cv and makes a bid to merge with TikTok

• What would a world with AGI look like?

• OpenAI quietly funded independent math benchmark before setting record with o3

The future of presentations, powered by AI

Gamma is a modern alternative to slides, powered by AI. Create beautiful and engaging presentations in minutes. Try it free today.

TECHNICAL NEWS, DEVELOPMENT, RESEARCH & OPEN SOURCE

• University of Washington and Microsoft research: Agents are not enough

• Liquid AI introduces LFM-7B: a new 7 billion parameter model based on an efficient liquid foundation model (non-transformer) architecture

• Moonshot AI introduces Kimi k1.5: a new multimodal AI model with reported o1 level joint reasoning over text and vision

• PAT-tastrophe: how we hacked Virtuals' $4.6B agentic AI & cryptocurrency ecosystem

• OpenLIT 2.0: open source dashboard for AI engineering & LLM data

• Perplexity launches Sonar, an API for AI search

• How to use Anthropic MCP Server with open LLMs, OpenAI or Google Gemini

That’s all for this week! We’ll see you next Thursday.