AI Leadership Weekly

Issue #17

Brad
February 06, 2025

Welcome to the latest AI Leadership Weekly, a curated digest of AI news and developments for business leaders.

Top Stories

OpenAI moving into Japan
OpenAI is moving into the Japanese market in another joint venture with SoftBank.

SoftBank is their finance partner in the recently-announced Stargate Project, where they are helping to finance datacentres within the US. In this new venture, they have made an annual commitment of $3 billion towards the use of OpenAI's tech in a venture they're calling "Cristal Intelligence."

In a nutshell, Cristal Intelligence will use OpenAI systems (ChatGPT, their API) which are customised for the Japanese market and for business. This ongoing commitment signals another player moving into Asia's AI space.

European Union creating open-source AI with “European values”
The EU is investing $56 million to develop an open-source LLM called OpenEuroLLM.

This new model will work in all 30 of the EU's member languages, and will use numerous super computers found throughout the EU. It's also hoped that EU values will be "baked into" the model, in terms of alignment and what the model will produce.

$56m also seems like a tiny number to train an LLM, when the likes of Meta and OpenAI spend billions on their models, although people are pointing to DeepSeek-R1 to show what can be achieved on a small budget.

OpenAI launches Deep Research
Expanding on its recent Operator launch, OpenAI has announced the Deep Research tool.

Deep Researches spawns an agent that uses reasoning abilities to search the web for data and compile reports on what it finds. It will be powered by an optimised o3 model, and is currently available to Pro users, with other paid users soon to follow.

In Brief

Market Trends

40% of Facebook posts may be AI-generated
It probably surprises no-one that huge swaths of social media posts are now AI-generated. The team from Originality.ai claims that, as of November last year, over 40% of Facebook's long-form content may be AI-generated.

The team looked at nearly 9,000 Facebook posts going as far back as six years to find this figure. They also note that the average percentage of AI posts between 2023 and November 2024 was 24.05%, which they note is a 4.3x increase since the launch of ChatGPT. In comparison, the average rate from 2018 to 2022 was only 5.34%.

Reasoning models not so reasonable
Several major Chinese universities have discovered that advanced, so-called 'reasoning' models such as DeepSeek-R1 and OpenAI's o1 chronically 'underthink'.

What this means in practise is that they frequently discard viable solution paths and information while they "think through" the problem set for them. The paper claims that, especially on complex tasks, the models will constantly shift their strategy to solve the problem, and that these shifts result in using more than twice as many computational tokens as necessary.

This is even more damning when they also frequently arrive at incorrect answers. The paper also offers some solutions, including penalising switching strategies so that the AIs will fully explore existing paths.

OpenAI on the “wrong side of history”
Sam Altman has recently made comments about OpenAI being "on the wrong side of history" in terms of their closed-source approach to AI development. There have been criticisms of the company for years regarding their closed-source nature, but these new statements come in the wake of DeepSeek disrupting the industry. Time will tell how 'open' OpenAI will become, though.

Tools and Resources

Sourcetable
This spreadsheet tool connects to your real-time data like a regular spreadsheet, but uses AI to extract insights for you.

Open Deep Search
This is an open-source alternative to OpenAI's Deep Search.

Hugging Face Spaces
This is Hugging Face's very own app store for AIs.