A week in Generative AI: Personality, Shopping & Premier
News for the week ending 4th May 2025
No huge announcements this week, but some significant news. ChatGPT’s new shopping features have flown a bit under the radar, mostly due to the headlines around OpenAI’s ‘sycophancy-gate’. Thinking beyond the short-term blip and news headlines that OpenAI have had this week around ChatGPT, I think the personality of chatbots is something we’ll be hearing a lot more about this year, and hopefully we’ll start to see some new personality controls that I’ve written in depth about before. We also got the release of Amazon’s most capable model, Nova Frontier, 6 months after the Nova family of models were first announced.
There are some interesting nuggets in ethics news this week with a look from Anthropic on the impact on Software Development, how AI models lie routinely, and the risk a small number of AI companies pose to free society.
There are lots of great long reads too from Ethan Mollick, Simon Willison, and an article from Ars Technica looking back at GPT-4 which was quietly deprecated by OpenAI this week.
OpenAI pledges to make changes to prevent future ChatGPT sycophancy
It’s been a big week for the personality and the ‘vibe’ of generative AI models. The main news was an update to GPT-4o that was rollout and then quickly rolled back because it was too ‘sycophantic’. There’s a great article from the Guardian on How an embarrassing U-turn exposed a concerning truth about ChatGPT, and also some good commentary from Ethan Mollick as always.
OpenAI, to their credit, have been very transparent about everything. They published an article on what happened and what they’re doing about it and also published a follow up that expanded on what they missed with sycophancy. All of these are very interesting reads. They also held an AMA with Joanne Jang, Head of Model Behaviour.
We also saw the publication of some research on LM Arena this week, that Simon Willison has written a great piece on. Whilst not directly related to the ‘personality’ of Generative AI models, LM Arena is designed to rank the most preferred model and it turns out the personality and ‘vibe’ of a model unsurprisingly has a big influence on this.
What this week has shown us is that the personality of models really matters. It matters because it not only influences experience, but behaviour as well. This means that AI companies need to be more transparent about how they fine-tune and manage the personality of their models, but also that end users need to have more controls over the personality they’d like to interact with themselves. This is something I’ve written in depth about before, and I’m hoping 2025 will be the year we start to see these ideas put into practice.
ChatGPT is getting better for shopping
I suspect this is what OpenAI was hoping all the headlines would be about this week, as these shopping features are significant, but seem to have mostly flown under the radar.
Essentially, ChatGPT’s search capabilities just got a big upgrade for people using it to help them find and choose products. ChatGPT can now show users product cards with images, prices, and star ratings in an easy-to-read format.
This positions ChatGPT as a much stronger competitor to Google and Perplexity for users looking for product advice and also includes direct links to buy. The product cards are purely organic - no paid for advertising placements or sponsorships are available. I suspect that OpenAI will start to take a commission in the future via affiliate links, as this is something that Sam Altman has previously hinted at.
In an interview with the Verge about the new shopping capabilities, Adam Fry, ChatGPT’s search product lead also claimed that OpenAI had seen 1 billion searches in the last week. That’s still only c.1% of the searches Google sees every week, but is still a big milestone. It’s also around 10x the volume Perplexity sees.
As OpenAI add more features like Shopping to ChatGPT search I think we’ll see significant growth, and it wouldn’t surprise me if they’re at 10 billion searches per week before the end of the year. People’s search behaviours are changing and they’re changing rapidly.
Amazon launches Nova Premier, its most capable AI model yet
At the end of last year, Amazon entered the frontier generative AI model game with its Nova family of models. They didn’t release their most capable model back then, Premier, but now 6 months later they have.
Nova Premier can process text, images, and videos (but not audio) and has a big context window of 1m tokens. It isn’t a large reasoning model - it’s a pure large language model. As you would hope, it does well at knowledge retrieval in that large context window, but isn’t state of the art vs. OpenAI, Google, or Anthropic’s leading models across most benchmarks.
This is the first generation of frontier models we’ve see from Amazon, so its no surprise that they’re not state-of-the-art, and I suspect we’ll see more competitiveness from the next generation of the Nova family.
Atlas learns in simulation
Boston Dynamic’s Atlas has long been the most capable humanoid robotic hardware platform. However, they’ve recently had a lot of competition from new robotics companies who have been able to rapidly improve both hardware and software by using simulated learning.
This video shows Atlas’ own simulated learning in action, giving it the ability to pick, sort, and place items which is a common factory task that could be done by humanoid robots in the near future.
AI Ethics News
How an embarrassing U-turn exposed a concerning truth about ChatGPT
Anthropic Economic Index: AI’s Impact on Software Development
AI models routinely lie when honesty conflicts with their goals
A few secretive AI companies could crush free society, researchers warn
BBC harnesses AI to create writing classes given by Agatha Christie
Anthropic suggests tweaks to proposed US AI chip export controls
Meta forecasted it would make $1.4T in revenue from generative AI by 2035
Long Reads
One Useful Thing - Personality and Persuasion
Ars Technica - The end of an AI that shocked the world: OpenAI retires GPT-4
Simon Willison - Understanding the recent criticism of the Chatbot Arena
Anthropic - The Societal Impacts of AI
The Wall Street Journal- Meta’s ‘Digital Companions’ Will Talk Sex With Users—Even Children
“The future is already here, it’s just not evenly distributed.“
William Gibson