This week's news: OpenAI launches a new model that can 'think'

Welcome to this week’s newsletter on generative AI and education. We have gathered the most important news so you can easily keep up with the rapid developments.

This week, you can read about:

• OpenAI’s new ChatGPT model o1-preview, which mimics human reasoning by “thinking” more thoroughly about complex questions before responding.
• Google introduces a new feature in NotebookLM, where you can turn documents into an AI-generated podcast with two virtual hosts.
• DuckDuckGo launches an AI chat focused on privacy, where none of the data sent is used to train the models.
• AlphaProteo from DeepMind is a new AI system that designs proteins which bind to target molecules with a high success rate.
• The French company Mistral releases its first multimodal model, Pixtral 12B.
• You can now run language models locally in your browser via Phi-3.5 WebGPU using your computer’s GPU.

We have made a podcast of the newsletter using NotebookLM, and you can listen to all the news here:

Enjoy reading and listening!

OpenAI launches o1-preview: can reason

OpenAI last week launched its latest version of ChatGPT, called o1-preview, and this model mimics human reasoning. OpenAI itself says the model has been trained to “think” more carefully about complex questions before the language model provides an answer.

Unlike other models such as ChatGPT-4, Gemini or Copilot, ChatGPT-o1 tries out different strategies and can recognise its own errors. This happens via reinforcement learning and an internal “chain of thought” before the model responds to the user. The result is markedly improved capabilities, especially in coding, mathematics and the natural sciences.

A smaller and cheaper version called 01-mini has also been released. This model almost matches o1’s performance on tasks such as mathematics and programming, but is pared back on general knowledge. That makes it faster and more cost-effective for applications that primarily require logical reasoning. The new models can still make basic mistakes, and their reasoning process is hidden from the user.

OpenAI expects these models to open up entirely new possibilities for applying artificial intelligence in technical and scientific fields.

Peter Svarre has also tested the new model and writes the following on LinkedIn: “The most problematic aspect of the new model, however, is that OpenAI (as always) utterly shamelessly anthropomorphises their design. The new model has been given an element where you are told that the model ‘thinks’ about the problem, which contributes to a sense that human-like thinking is actually taking place in the background.”

We have attempted to illustrate this anthropomorphisation by showing the difference between the two ways of presenting content.

We have collected a range of articles on ChatGPT 01 and in particular recommend Alberto Romero’s overview – OpenAI 01: A New Paradigm For AI. In addition, there is lots of information on OpenAI’s blog.

Below are some other sources:

GPT-o1 (GPT5) by OpenAI detailed review

OpenAI’s project Strawberry with PhD level intelligence

Data Science in your pocketMehul Gupta

Massive breakthrough in AI intelligence: OpenAI passes IQ 120

OpenAI’s new “o1” model may have an IQ as high as 120

Maximum TruthMaxim Lott

Google introduces AI podcast creator in NotebookLM

In Google NotebookLM there is now a new feature that makes it possible to turn documents into an AI-generated podcast. The feature takes the content and creates a podcast with two AI hosts who discuss the material – and it actually does this quite convincingly. The feature is still experimental and has certain limitations, such as longer texts taking more time to process, and the conversations being available only in English.

We have tested it a little and are somewhat astonished at the level of these auto-generated podcasts. At the same time, we find the techniques they have used to make it feel human interesting, for example in the clip below where the host whispers.

Memo

0:00

/29.184

1×

Although the conversations are not necessarily objective or error-free, this does, for example, allow students to listen to material in a different way. Google are not the only ones making podcasts, and over a year ago Rune Vendler developed Anypod.io, where users can have quite advanced podcasts produced.

NotebookLM now lets you listen to a conversation about your sources

NotebookLM is releasing Audio Overviews, which turns your sources into an engaging discussion.

GoogleBiao Wang

Anypod.io | Podcasts on demand

Podcasts on demand

DuckDuckGo: AI chat with a focus on privacy

Something we haven’t covered in our newsletter is the launch of DuckDuckGo’s AI chat features, which focus on privacy. Here you can use GPT-4o Mini, Claude 3, as well as the open-source models LLaMA 3.1 and Mixtral. The main advantage of DuckDuckGo’s AI chat is that none of the data you send is used to train the models.

You can access the service via duck.ai. Although you should always be mindful of what data you share under GDPR, we consider this to be one of the safer and more anonymous ways to use language models.

DuckDuckGo AI Chat at DuckDuckGo

DuckDuckGo. Privacy, Simplified.

DuckDuckGo AI Chat

Scholarly articles

Below are a couple of scholarly articles or news items that are a bit more nerdy and may not be that interesting to most of our readers.

New AI creates new proteins for health and biological research

Google DeepMind has launched AlphaProteo, an AI system that can design new protein binders for use in medical and biological research. AlphaProteo creates binders with a high success rate and better binding affinity than previous methods. The model can be used for drug development, diagnostics and disease understanding, and has been tested on important proteins such as SARS-CoV-2 and cancer-related proteins. The system has been validated in collaboration with researchers from the Francis Crick Institute, but it still has limitations with certain complex proteins.

Read more here:

AlphaProteo generates novel proteins for biology and health research

New AI system designs proteins that successfully bind to target molecules, with potential for advancing drug design, disease understanding and more.

Google DeepMindProtein Design and Wet Lab teams

French Mistral launches Pixtral 12B

There are many updates to different language models, and we’ve chosen not to write about all of them — otherwise we wouldn’t be doing anything else. This week, however, we do have a news item about the French AI startup Mistral, which has announced Pixtral 12B, their first multimodal model. The model has 12 billion parameters and is 24 GB in size, making it competitive with other multimodal models such as Anthropics Claude and OpenAI’s GPT-4.

Pixtral 12B can be downloaded and fine-tuned via torrent on GitHub or through Hugging Face under an Apache 2.0 licence, which gives users freedom for commercial use without restrictions.

https://techcrunch.com/2024/09/11/mistral-releases-pixtral-its-first-multimodal-model/

Running language models in the browser

We’ve previously written about LMStudio.ai and Jan.ai, where it’s easy to run language models locally on your computer. It has now also become possible to run language models locally in the browser via Phi-3,5 WebGPU. You need to use Google Chrome, and once the model is loaded, it uses the computer’s GPU.

Try it here:

Phi-3.5 WebGPU - a Hugging Face Space by webml-community

A powerful AI chatbot that runs locally in your browser

a Hugging Face Space by webml-community

Other news this week

Students Are Using AI Already. Here’s What They Think Adults Should Know

A new report details what teens think parents and teachers should know about how they use, or don’t use, generative artificial intelligence

Harvard Graduate School of EducationRyan Nagelhout

Radiolab: Podcasts | WNYC Studios | Podcasts

Investigating a strange world.

Podcasts

‘If journalism is going up in smoke, I might as well get high off the fumes’: confessions of a chatbot helper

A data annotator reveals the threat his job poses to the art of real writing – and big tech’s growing concerns over AI’s limitations

The GuardianJack Apollo George

All children should have technology understanding

Professor Ole Sejer Iversen welcomes teaching in technology understanding in the Danish primary and lower secondary school (folkeskole), but wonders where the teaching qualifications will come from, and why it is only an optional subject.

DM Digi

Cybercrime's New Attack Wave: AI, Deepfakes, and The Fraud Tsunami - Part 1 | LinkedIn

I am brining a 2-part series on cybercrime with the leading fraud expert Frank McKenna. Join us for part 1 of this conversation, where Frank will unravel the sophisticated threats facing businesses and individuals today - from the explosion of digital doppelgangers to the terrifying rise of deepfake technology in fraud schemes. **We will discuss:** - How the pandemic supercharged global fraud and cybercrime. - The dangers of digital doppelgangers and how they are created. - The latest in multimodal fraud attacks involving deepfakes and AI. - Practical steps your business can take to safeguard against these evolving threats. **Why Attend?** - Be informed and aware of the emerging cyber threats that could impact your business. - Gain actionable insights from one of the world’s leading fraud experts. - Engage in a live Q&A session with Frank McKenna and get your pressing questions answered. **Who Should Attend?** - Business leaders and executives concerned about cybersecurity. - IT and cybersecurity professionals looking to stay updated on the latest threats. - Entrepreneurs and small business owners wanting to protect their assets. **Event Details:** 🗓️ **Date:** 9th September 2024 ⏰ **Time:** 15:00 UTC 📍 **Location:** LinkedIn Live (Access via LinkedIn Event Page) **Don’t miss this opportunity to learn how to protect yourself and your business from the next wave of cybercrime.** 🔗 **Click ‘Attend’ to secure your spot and set a reminder for this must-watch event!**

LinkedIn

Generative artificial intelligence, human creativity, and art

Abstract. Recent artificial intelligence (AI) tools have demonstrated the ability to produce outputs traditionally considered creative. One such system is

OUP AcademicMatthew Harding

CERN for AI: The EU’s seat at the table - ICFG

European Commission President Ursula von der Leyen proposed a bold new initiative to create a “CERN for AI,” aiming to boost Europe’s competitiveness through AI innovation. This effort would pool resources across the continent, similar to the successful CERN model for physics research. But what would it actually entail? Our new report dives into the details.

ICFGRowan

AI in the theatre, in school and in everyday life | Teateravisen

Teateravisen

Researchers create a model to predict literary success | VELUX FONDEN

After three years of research and combing through more than 9,000 works, the researchers behind the Fabula-NET project from Aarhus Universitet (Aarhus University) have developed a technology that contributes to basic research in the humanities and can predict the success of literary works based on three parameters.

Home