A week in Generative AI: Llama, Nova Act & Runway
News for the week ending 6th April 2025
I thought this was going to be a slightly quieter week on the GenAI front, until Meta decided to do a surprise drop of their new Llama 4 family of models late last night! This week also saw Amazon announced Nova Act, their approach to AI agents that can operate computers like humans do, and Runway released Gen-4 their latest text-to-video model. OpenAI also closed the largest ever funding round for a privately listed tech company.
There was lots of Ethics News this week - from Studio Ghibli staying quiet on their art being imitated by OpenAIâs new image generating model, to Channel 4âs CEO saying AI firms are âscraping the valueâ from the UKâs creative industries to research from Anthropic that shows that large reasoning models donât always say what they think.
In Long Reads Iâve featured two new posts from Helen Toner, who just launched her own Substack this week, and a though-piece from a group who think AGI is very likely in the next two years.
Meta releases Llama 4
This is a strange release from Meta - dropped on a Saturday evening (UK time), theyâve announced a Llama 4 family of models, releasing some medium-sized models, whilst the largest one is still being trained. It feels rushed and Iâm not really sure why. If DeepSeekâs the reason, then I donât think itâs a good one.
The model isnât available for use in the EU according to the use policy and thereâs not a huge amount on the safety testing theyâve done in the release announcement, presumably because its not fully tested, as the largest model is still training. Despite all of this there are some very interesting features in the Llama 4 family of models:
They are natively multimodal
Each model in the family is a âmixture of experts modelâ - Scout has 16 experts, Maverick 128 experts, and Behemoth 16 experts. Iâm not sure why thereâs such a dramatic difference in the number of experts Maverick has, almost seems like a typo!
Scout has an industry leading context window of 10m tokens which is astounding. Nothing like this has been seen before, and the previous leader in this regard, Gemini 2.5 Pro âonlyâ has a 2m token context window.
All models in the family claim industry-leading performance for their class against various benchmarks, beating all other models with the exception of Gemini 2.5 Pro
So this is a really interesting release from Meta - Llama 4 definitely pushes the open model frontier forwards and Iâd say between them and DeepSeek theyâve bought the open AI community up to only about 1-3 months behind the closed model frontier. Thatâs a big shift from just a year ago, when the open models were probably about 12 months behind the closed frontier.
Amazon unveils Nova Act, an AI agent that can control a web browser
This video opens with quite a statement, but one I wholeheartedly agree with:
âSoon, there will be more AI agents than people browsing the web, doing tasks on our behalfâ
This is why I recently wrote about what Iâm calling Web 4.0 and the Rise of the Agentic Web. Weâre about to see an incredibly important shift in how the internet works and is used, so its no surprise to see Amazon planting their flag.
Nova Act is essentially just like Anthropicâs Computer Use, and OpenAIâs operator. Itâs a visual model thatâs capable of using a computer/web browser just like a human does. As Iâve written about before, I think this is an interesting approach but one that is prone to error and creates lots of reliability issues. We shouldnât be forcing machines to work like humans, and instead should be spending more time making human systems more machine accessible.
This is why I believe a better approach is giving models access to APIs to retrieve information and perform tasks. It looks increasingly likely Anthropicâs open source Model Context Protocol will be the industry standard for this now that both OpenAI and Google have announced support and I expect to see this approach used in consumer-facing production AI agents before the end of the year.
OpenAI Valued at $300 Billion After Record-Setting Funding Round
It feels like OpenAI is raising money at new, higher valuations every couple of months! This latest round values the company at $300bn (twice the valuation of 6 months ago) and they raised $40bn of additional capital. This is the largest ever funding round for a private technology company.
In the announcement, OpenAI position the new funding round as enabling them to push the frontier of AI research, scale their infrastructure, and deliver more powerful tools to their users. One nice little fact in there is that they now have 500m weekly users. This was 400m just six weeks ago - theyâre currently growing at an incredibly rate!
Runway releases an impressive new video-generating AI model
This is an impressive new release from Runway, one of the OG text-to-video platforms. Runway says its latest AI video model can actually generate consistent scenes and people, which allows for the generation of longer video clips.
The announcement video above has some great âbehind-the-scenesâ on how Runway used Gen-4 to create their demo videos. It gives you some good insights into not only some of the new features, but how to use them as well.
Iâd say that the rendering of real humans and animals is incredibly impressive but there is still an uncanny valley effect going on. You can just about tell that the real-world videos were a generated, and not something that was filmed. I donât think it will be too much longer until this goes away though - probably another 6-12 months and weâll see models that are capable of generating real-world videos that are indistinguishable from real-life.
AI Ethics News
Studio Ghibli hasnât commented on OpenAIâs onslaught of AI copies, but the fan subreddit has
AI firms are âscraping the valueâ from UKâs ÂŁ125bn creative industries, says Channel 4 boss
DeepMindâs 145-page paper on AGI safety may not convince skeptics
OpenAIâs models âmemorizedâ copyrighted content, new study suggests
Google is shipping Gemini models faster than its AI safety reports
What AI anime memes tell us about the future of art and humanity
ChatGPTâs new image generator is really good at faking receipts
Designers say plans for UK copyright law risk ârunning roughshodâ over sector
OpenAI plans to release a new âopenâ AI language model in the coming months
Long Reads
Helen Toner - "Long" timelines to advanced AI have gotten crazy short
Helen Toner - The core challenge of AI alignment is âsteerabilityâ
âThe future is already here, itâs just not evenly distributed.â
William Gibson