Welcome back!

The AI race is heating up with the Gemini 1.5 models. The Gemini 1.5 Pro and Advanced models have been updated and are now nearing the performance of OpenAI’s GPT-4o, showing impressive gains. Let's unpack this...

Today’s Summary:

• Gemini 1.5 performance surge

• Study on public AI perception

• Train GPT-2 yourself for $20

• Codestral AI for code generation

• OpenAI partners with The Atlantic and Vox Media

• PwC's ChatGPT entreprise deal

• 2 new tools

TOP STORY

Gemini 1.5 Models Performance Surge, Approaching GPT-4o at the Top

The Summary: Google's newly updated Gemini 1.5 models have made a massive leap in performance, nearly catching up to OpenAI's flagship GPT-4o at the top of the rankings.

The new Gemini 1.5 Pro and Advanced claimed the #2 spot, surpassing Claude 3 Opus and GPT-4-Turbo, with Gemini 1.5 Pro also demonstrating an impressive 48% win rate against GPT-4o. The slimmed-down Gemini 1.5 Flash at #9 is also very competitive due to its lower pricing.

Key details:

• Gemini 1.5 Pro and Advanced at #2 rank with 1268 ELO rating

• 48% win rate 1.5 Pro against GPT-4o

• Gemini 1.5 Flash at #9 beats several costlier models

• Gemini 1.5 Pro #1 in Chinese language tasks

Why it matters: The rapid rise of Gemini 1.5, especially the budget-friendly Flash variant, shows increased competition in the AI race. Google is working to reduce the gap with OpenAI through regular, iterative improvements. With top performance and low cost, Gemini models may see broader adoption, especially among professionals through the integrations in Google Workspace.

INSIGHT

New Study Maps Public Perception of Generative AI

The Summary: A new study by the University of Oxford provides insights into public awareness, use, and expectations around generative AI across six countries (USA, UK, France, Japan, Argentina, Denmark) in April 2024.

While ChatGPT leads in recognition and use, a sizable minority - between 20% and 30% depending on the country - remains unaware of AI. Most respondents expect AI to have a large impact across sectors, but trust varies. Younger respondents are more open to AI use.

Key details:

• 50% are aware of ChatGPT across the 6 countries surveyed, but frequent use is rare at 1%-7%

• 56% of 18-24s have used it compared to 16% for the 55+ age group

• 20-30% have never heard of any AI tool

• 66% expect generative AI to have significant impacts

• Trust is highest for the future of AI in healthcare and science, lower for media, politics, and social media.

Why it matters: These findings provide insights into the public's evolving perceptions of AI's capabilities and potential impacts across sectors. The data reveals significant generational differences in AI adoption and trust, with younger people being more likely to think that generative AI will have a positive impact.

DYI

You Can Now Train GPT-2 Yourself in 90 Minutes for $20

The Summary: Andrej Karpathy, the renowned AI researcher, demonstrated how to reproduce the 124M parameter GPT-2 model in just 90 minutes for around $20, using his efficient code and one 8x A100 80GB GPU cloud setup.

The model, released by OpenAI in 2019, is the smallest in the GPT-2 series. Training used a rental GPU instance. Karpathy shares the full training script and visualizations.

Key details:

• Reproduced GPT-2 124M model in 90 minutes

• Cost around $20 on an 8x A100 80GB GPU rental

• Achieves up to 60% of peak model FLOPS utilization

• Trained on 10 billion tokens of web data (FineWeb dataset)

• Outperformed OpenAI's released 124M checkpoint

• Also reproduced the 350M model in 14 hours for $200

• The full 1558M model, which was state-of-the-art in 2019, would require 1 week and $2.5K

Why it matters: This shows how much easier, faster and cheaper it has become to train a model like GPT-2, which was state-of-the-art just five years ago, using efficient code and affordable cloud hardware. While it doesn’t remotely match the performance of newer models, it offers great educational value for students and enthusiasts to gain hands-on experience and understand the key technical concepts.

QUICK NEWS

Quick news

• Mistral released Codestral AI for Code Generation (non-commercial license)

• OpenAI partners with The Atlantic and The Verge

• PwC signs 100k workers to ChatGPT enterprise

• A visualization showing the evolution of performances of the leading AI models over one year, highlighting the “AI race”

  

  Credits: Peter Gostev