A week in Generative AI: Tasks, Transformers & Thinking

News for the week ending 19th January 2025

Jan 19, 2025

It’s been a relatively quiet week on the consumer GenAI front, with just one announcement from OpenAI who introduced ChatGPT Tasks. However, there were two interesting research papers released pointing towards what the next generation of technology that underpins large language models could look like. There was also some good coverage of o1 reasoning in multiple languages that no one quite understands and some great progress on robotics platforms operating at human speeds.

In Ethics News, OpenAI published their idealised version of AI regulation in the US, there were articles about how GenAI will both create and eliminate jobs, and a report on how GPT-4 has been helping Nigerians compress 2 years of learning in to just 6 weeks.

I also highly recommend the long read from IEEE Spectrum this week on how AI mistakes are very different from human mistakes.

ChatGPT now lets you schedule reminders and recurring tasks

This week OpenAI announced a new beta feature - the ability for ChatGPT to run recurring tasks and set reminders. On the surface this sounds like a small feature, but I think if they can get this right it could be a big step forward for large language models. The reason is that for the feature to work correctly, the model needs to be able to understand and have a sense of time, which is not something that they have previously really had.

I’ve done some basic testing, and firstly you need to select the GPT-4o with scheduled tasks (beta) model - this isn’t currently built into the standard GPT-4o model yet. Once you have you can then ask ChatGPT to set a reminder for you which then appears as a push notification and/or email on your device.

To test the tasks I asked ChatGPT to ‘get me the latest news in the UK every day at 7am’. For some reason I kept getting an error message and couldn’t get it to work. Also, when I ran the task manually, I got mixed results with the news articles returned being from some random smaller news sites, some articles that were behind pay walls and all the articles returned being at least 2 days old.

Tasks is currently a beta feature, and there are obviously still some bugs, but once it’s working more reliably I think this will be quite a significant feature and is the first small step towards ChatGPT being to do things for you, not just answer questions and chat.

We’re going to see many more features like this launch this year as all the frontier AI companies are working on ‘agentic’ models that can do practical things for users - it’s going to be fun to watch this area of the industry unfold over the coming months!

Source

Transformer² vs. ‘Transformer 2.0’

I don’t normally write about the cutting-edge research being done in the generative AI space as it’s quite technical and I prefer to focus on the practical news that’s interesting and/or useful for most people right now. However, two pieces of research were released this week that I think are significant and worthy of mention.

All large language models are built on top of what’s called a Transformer, which is a deep learning algorithm that was developed by Google back in 2017 and shared with the world via a now famous research paper titled ‘Attention is all you need’. It’s the technology that has propelled the development of AI systems over the last 8 years more than anything else, which is why any advancements in this space could be so significant.

This week we saw two new research papers released that each proposed slightly different ‘sequels’ to the Transformer, but both tackle the same fundamental issue we’ve had since 2017 - that models based on Transformers are ‘fixed’ once they have been trained and can’t dynamically learn new things or adapt post training.

The Transformer² paper, from a Japanese research team, proposes a new approach for creating self-adapting large language models that can dynamically adjust to different tasks without requiring a complete retraining of the model. It does this by using what the research calls ‘expert vectors’ that optimises the model for different tasks. The research also suggests that these ‘expert vectors’ can be transferred to other models, meaning learning is transferable between models without requiring a whole new training run, and AI systems might be able to learn from each other in the future.

The Titans paper, is from a team at Google Research, and is being dubbed ‘Transfomer 2.0’. The researchers propose a new approach that mimics the way that human brains learn and memorise knowledge, allowing AI models to learn and adapt when being used. Titan models have three types of memory - short-term, long-term, and persistent which all come together and help the model manage large inputs and adapt it’s outputs appropriately.

Both of these new approaches offer significant improvements over the original Transformer architecture and that’s been proven out by the results shared in the research papers where models outperform transitional transformers of similar sizes in terms of accuracy, efficiency, and capabilities.

I think we’ll see approaches like those outlined in bother research papers starting to be implemented in models later this year/early next. Will be interesting to see what new capabilities they unlock.

Transformer² paper

Titans paper

OpenAI’s AI reasoning model ‘thinks’ in Chinese sometimes and no one really knows why

This is such a strange and interesting phenomenon. Essentially, I think that LLMs don’t really understand that there are different languages, just that there are different words/tokens that have similar meanings. In this scenario its not surprise that large reasoning models like o1/o3 don’t always reason in English but jump in and out of different languages as they progress. Maybe this is the start of advanced LLMs generating their own ‘language’ that’ll we’ll struggle to understand as they become more advanced?!

Source

Robots become faster and more agile

Two interesting things were announced in robotics this week, both relating to the speed and agility that robots can operate at. Up until now, many robotics platforms have been operating at below human-speed but both the humanoid update from Unitree and the robodog update from Mirror Me (both Chinese companies) show that we can now build robotic platforms that are capable of operating at human speeds in both running and walking.

This is some great progress on what we’ve seen before and I wonder how long it will be before we see robotics platforms that are able to operate at superhuman speeds and what the safety implications will be?

AI Ethics News

Long Reads

Stratechery - AI’s Uneven Arrival
IEEE Spectrum - AI Mistakes Are Very Different Than Human Mistakes

“The future is already here, it’s just not evenly distributed.“
William Gibson