The Weekly AI Roundup: Claude 3.7 Goes Live, ChatGPT 4.5 Fails To Impress, Pika 2.2 Is Fantastic!

Anthropic launch. Claude 3.7. Perplexity finds its voice. Inception labs create a whole new type of LLM. Pika 2.2 with Pikaframes. Amazes everyone. And GPT-4.5 goes live but fails to impress.

Perplexity has announced that they are about to release Comet, a brand new browser experience that they refer to as an 'agentic browser'. This is a big bold move from Perplexity. They've already had some impressive search tools available in their models, but this is something different, a standalone offering that is looking to disrupt the search market in a meaningful way.

Elon Musk was feeling bold with his predictions today when he posted on X "We are on the event horizon of the Singularity" Obviously no one really knows what this is in reference to, but Grok 3 has been on fire recently, so maybe he's seeing something in their research.

Deepseek R2 had been slated for launch in May, but the success of some of their competitors recent models has put the pressure on. The release has now moved from May to "as soon as possible". Will DeepSeek R2 have the same huge impact as their R1 model did?

AI startup Convergence made headlines in recent weeks by launching a rival to Operator called Proxy. It was very well received and made some instant fans. Now they're back with Proxy Lite, a 3 billion parameter vision language model that's open source and can run on your local machine.

OpenAI announced that Deep Research would now be available to all of its Plus subscribers, and that a version of Advanced Voice powered by GPT4o-mini would be available to all ChatGPT users, even to those on the free tier. OpenAI are definitely feeling increased pressure to ship more often, and that's good for the users.

Perplexity has updated their iOS app with a new voice mode. Perplexity claims that this is the only model that can reliably incorporate real time voice and information across several languages simultaneously.

Figure have revealed a new video of their figure 02 robots working together with their new Helix Neural Network. Last time we saw two robots coordinating to carry out some simple home tasks. In this video there are many robots working in an industrial setting carrying out some picking and sorting tasks that would usually be done by human hand. Figure claim that training their second commercial use case took just 30 days.

The little known AI company Inception Labs has unveiled a whole new kind of AI model. Their Mercury model is being called a DLLM or Diffusion Large Language model. This means that the output is created and then incrementally refined from low to high fidelity. The result is that Mercury can generate over 1,000 tokens per second, far higher than any current model. This feels like a major breakthrough in model architecture and it will be very interesting to see if other companies adopt this approach.

Pika has released version 2.2 of their video model. It can generate 10 second clips with a 1080p resolution and it now includes Pikaframes where the user can select the start and end frames of a video for far greater control.

Prior to the release of ChatGPT 4.5, we predicted that OpenAI would have to launch something pretty special in order to stave off the criticism that it has faced recently. Either they would need to launch something groundbreaking at the $200 pro level to justify the price tag, or they would need to launch a great model that provided a decent step up at the free or plus tier.

It would seem that OpenAI delivered neither of these things. What we got was a model that struggles to stand out from the crowd. We've seen very few benchmarks where it's managed to do anything significant, and yet it's only available to users paying the hefty Pro subscription. This has left most users feeling underwhelmed and this really feels like the first time where OpenAI have not been able to display their quality and maintain a dominant position in the market.

Grok, on the other hand, does continue to make some headlines. There are strong rumours that XAI will introduce artefacts to the interface very soon so that users can examine and adjust the code outputs that Grok generates. This will add a whole new dimension to the usability of Grok and is likely to be very well received.

Google's NotebookLM is said to have a mind map feature coming soon.
This will allow users to see a mind map style output based on the sources being given to the notebook and the user will be able to interact and prompt questions on item in the map. We're fans of NotebookLM so we're looking forward to this one.