A week in Generative AI: Agent, Voxtral & NotebookLM
News for the week ending 20th July 2025
Last week I wrote about how it wasnât big in terms of volume of AI news, but was probably a big one in terms of significance. This week Iâm doubling down on that statement - weâve probably just had the most significant week weâll see for a while, and a watershed moment in the transition to Web 4.0.
Iâm obviously talking about the introduction of ChatGPT Agent and the first consumer-facing (what I call) level 4 AI agent that can perform a wide variety of tasks online for people for an extended period of time. This week also saw the release of new audio models from Mistral and an interesting new feature from NotebookLM that I think could be a hint at the future of online publishing.
There was a huge amount of Ethics News this week too with Meta refusing to sign the EUâs AI code of practice, Google signing a $3bn hydropower deal, and a study warning of the âsignificant risksâ in using AI therapy chatbots.
Nice little Long Read from the FT on how AI is killing the web. Highly recommend checking this out if youâre invested in Web 4.0 and its impact on digital marketing.
OpenAI launches a general purpose agent in ChatGPT
Ok - letâs get this out of the way up front. This is a big deal. Itâs not going to change the world overnight - we need more capabilities, reliability, speed, and ultimately consumer adoption for that first. But, this is without a doubt the first (what I call) level 4 AI agent that is consumer-ready. This is where the transition to Web 4.0 (that weâve been seeing hints of) really starts, and I think a bit of a watershed moment. In all honesty, I wasnât expecting this to land until Q4 2025 at the earliest, but here we are nearly 6 months earlier than I thought!
So, what is ChatGPT Agent? On the surface itâs relatively simple - itâs a model that can plan, research, reason, and use tools to take actions on a userâs behalf. Itâs the culmination of a lot of things OpenAI has been working on and releasing over the last 9 months - large reasoning models, deep research, web browsing, and tool use. The trick is that when you bring these four things together you unlock not just a huge breadth of capabilities, but a lot of depth too. The model doesnât just make a plan, it can now execute that plan too. Here are some examples that OpenAI have released to give you an idea of what its capable of:
Importantly, ChatGPT Agent can also connect (with permission) to your personal data - your emails, calendar, files, and other online accounts. Thereâs a full list here, and you can also use custom connectors that leverage Anthropicâs Model Context Protocol to connect ChatGPT Agent to almost anything that has an API. This allows ChatGPT Agent to not just access the data but take action with it as well - send an email, schedule a meeting, and create files.
This is obviously going to be a big deal for knowledge work as it gets more widely adopted, but itâs also going to have a big impact on the internet. If consumers are going online to do things less, because ChatGPT Agent can do things for them, then weâre going to see less searching, clicking, and website visits. I wrote a bit about this last week, and believe that this is just another accelerant of Web 4.0. Weâve started seeing zero-click search, Perplexityâs Comet browser and ChatGPT Agent will start to scale zero-click browsing, and its only a matter of time before these systems become reliable and trusted enough for consumers to use them for zero-click purchasing. If these technologies reach mass-adoption (which I 100% expect them to do) then the fundamentals of the internet will change, with huge consequences for many businesses and marketers.
Given that ChatGPT Agent has arrived c.6 months earlier than I expected it to, this only re-enforces my view that weâre going to see the transition to Web 4.0 play out over the next 18-24 months. Thatâs incredibly fast for such a seismic shift and the reason why I strongly believe more people need to be paying attention to this stuff. If youâre not testing these technologies now, figuring out what they mean for your business, and thinking about how the transition to Web 4.0 will affect you and your customers then youâre going to get disrupted and quickly left behind.
ChatGPT Agent isnât available in the EU yet, so I havenât been able to test it myself yet, but if OpenAI follow their recent pattern then it should become available to us Europeans in the next month or so. Bit of an understatement to say Iâm looking forward to getting my hands on it!
Mistral releases Voxtral, its first open source AI audio model
Weâre not there yet, and many people are skeptical because voice-interaction with technology isnât new, and has mostly been awful for the last 10 years. However, there is no doubt in my mind that as we develop more conversational AI, that voice as an interface will finally become a primary means for interacting with technology in a meaningful way.
This week, my favourite European AI company, Mistral announced Voxtral - a family of audio models released for enterprise use. In the typical Mistral fashion, theyâre open models that are extremely capable and cheap to use. They can transcribe up to 30 minutes of audio and promise to outperform OpenAIâs Whisper for less than half the price. Thereâs also a mini version that can run on local devices.
Great work!
NotebookLM adds featured notebooks from The Economist, The Atlantic, and others
Iâve long been a fan of NotebookLM as itâs a platform that has genuinely tried to move the UI/UX game on beyond a bland chat interface. So, I always pay attention when they announce new features.
This new announcement is mainly for promotional purposes, and is designed to show users how Notebooks work in NotebookLM - Google now has notebooks from The Economist, The Atlantic, professors, authors, and even Shakespeareâs works. Users can read the original source material, but also ask questions, and explore topics more deeply. As itâs NotebookLM you also get the obligatory pre-generated Audio Overview (podcasts) too!
However, what Iâm more interested in is this being a new publishing platform for news organisations and authors. The idea that you can allow users to explore your content, rather than just read/listen is a very interesting one. Maybe authors also upload their notes, research, ideas, and drafts that didnât make it into the final cut so users (explorers?!) can go beyond the published content. Users could also expand the content with their own ideas, creating a more user-generated content alternative to the usual comments/fan fiction.
I can see lots of potential with this kind of thing to encourage and reward followers and fans - could be interesting to explore!
AI Ethics News
The CEO of Nvidia Admits What Everybody Is Afraid of About AI
Google inks $3bn US hydropower deal as it expands energy-hungry datacenters
A former OpenAI engineer describes what itâs really like to work there
AI firms âunpreparedâ for dangers of building human-level systems, report warns
Study warns of âsignificant risksâ in using AI therapy chatbots
Research leaders urge tech industry to monitor AIâs âthoughtsâ
OpenAI and Anthropic researchers decry ârecklessâ safety culture at Elon Muskâs xAI
Netflix uses generative AI in one of its shows for first time
Elon Muskâs Grok chatbot melts down â and then wins a military contract
Of course, Grokâs AI companions want to have sex and burn down schools
Grok may be breaking App Store rules with sexualized AI chatbots, and thatâs not the only problem
WeTransfer says user content will not be used to train AI after backlash
OpenAI, Google, and Meta Researchers Warn We May Lose the Ability to Track AI Misbehavior
Long Reads
âThe future is already here, itâs just not evenly distributed.â
William Gibson