AI Leadership Weekly

Issue #16

Brad
January 30, 2025

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

Top Stories

DeepSeek-R1 Rocks the AI Industry
Chinese startup DeepSeek has released their open-source LLM 'R1', which the quant company has called their 'side-project'.

The big news overshadows the release of OpenAI’s Operator, given that the model (which beats closed-source models, including o1) was trained for only $5.6m, where as the likes of OpenAI have spent hundreds of millions of dollars training their own models. They also say they required only 2,000 AI chips for their training, versus the 16,000+ used to train other models.

There is some controversy around these claims, though. First, it's difficult to verify any of their claims regarding cost and hardware used. The second is their use of illegally-imported Nvidia AI chips. And, recently, OpenAI has made claims that they used OpenAI models to train DeepSeek-R1. DeepSeek is also currently experiencing a cyberattack.

Source: Youtube

OpenAI Releases Operator
OpenAI has finally released their long-awaited 'Operator' feature, where their AI can autonomously perform tasks.

In a recent YouTube livestream, Sam Altman and team demonstrate Operator, which, in a nutshell, lets their AI system spawn a web browser and follow instructions to perform tasks. The examples given are to order pizza and buy concert tickets. The demonstration also shows some chain of thought from the Operator AI, as well as how it asks questions when stuck, and how humans can take over at any time before handing things back to the AI.

Source: Pexels

Nvidia Stocks in Freefall
DeepSeek-R1 has had an even bigger impact on the industry: driving down AI stock prices.

The most visible casualty has been Nvidia, whose stock, at one point, had dropped 17%. The main reason for this is that if powerful models can be trained on minimal hardware, chip manufacturers such as Nvidia will see significantly less demand than is currently predicted.

In Brief

Market Trends

ChatGPT For Governments
OpenAI has announced a version of ChatGPT made specifically for government. What this means in practise is that government agencies can install a 'local' version of ChatGPT onto a private Azure instance, which the agency is more closely able to control.

This is important for governments, as copy-pasting classified or otherwise sensitive information into ChatGPT (or any other online AI provider, for that matter) would potentially violate secrecy laws (not to mention potentially leak state secrets for foreign actors).

Hugging Face Releases Smaller Models
In a world of ever-growing parameter counts, Hugging Face has been working on creating sub-1 billion-parameter models.

The latest releases of their SmolVLM have 250M and 500M parameters, and their goal is to have models that work well on as many devices as possible, including older laptops and, eventually, even in web browsers.

AI Models Fail at World History
It's well known that LLMs hallucinate, and DeepSeek-R1 has been criticised for censoring topics deemed off-limits by the Chinese government, but the Complexity Science Hub (CSH) in Austria has quantified just how bad these LLMs understand history.

Their research says that, when asked a series of yes or no questions, only 46% of their answers were correct. They also point out that the models were particularly bad at questions for certain world regions, including sub-Saharan Africa.

Tools and Resources

Telescope
Telescope helps to fill out your CRM with highly-targeted prospects, all from a simple prompt.

Codev
Codev takes your business ideas and makes them real in Nextjs.

Janus Pro
Try out the image generator based on DeepSeek.