ChatGPT can now talk to you like a real person
On 8 July 2026, OpenAI launched GPT-Live, a new generation of voice models for ChatGPT.
This week is all about one clear trend: AI is moving out of the chat window and into the workspace itself.
OpenAI has just launched ChatGPT Work, a new series of models based on GPT-5.6, and a new voice generation. Each of these is interesting in its own right. But taken together, they point to something bigger.
I believe that the desktop versions of Claude and ChatGPT – and those that follow – will become a central hub for how we work with AI and agents in the future.
Not as a tab you open alongside everything else, but as the place where the work actually takes place. Where the AI can access your files, use your tools and see a task through to a finished deliverable.
I hope you enjoy reading the newsletter!
/René 🤗
Here’s today’s menu:
- ChatGPT can now talk to you like a real person
- The future must remain human
- ChatGPT gets a powerful new work feature
- Cybercriminals of the future may not need to be able to code at all
ChatGPT can now talk to you like a real person
On 8 July 2026, OpenAI launched GPT-Live, a new generation of voice models for ChatGPT.
The major breakthrough is that the model can now listen and speak at the same time, just as we humans do in a real conversation. This means you can interrupt, pause to think or ask a follow-up question without the AI losing the thread.
Previous versions of ChatGPT’s voice worked a bit like an old-fashioned answering machine: it waited for you to stop speaking before responding. This often resulted in a stiff and unnatural experience. GPT-Live puts an end to this by processing audio in real time, and it can even make small affirmative sounds such as “mm-hmm” or “got it” whilst you’re speaking.
What’s more, the model has become smarter. When you ask a difficult question that requires an online search or deeper thought, GPT-Live passes the task on to a more powerful model running in the background whilst continuing the conversation with you. It’s a bit like talking to a knowledgeable colleague who forwards a query to a specialist whilst you carry on chatting.
GPT-Live is being launched in two versions: GPT-Live-1 for paying users and GPT-Live-1 mini for free users, and it is available on iOS, Android and online.
Why is this important?
Voice is the most natural way for people to communicate. When AI can handle real, fluid conversations rather than stiff, turn-based exchanges, the barrier to entry for who can use the technology is significantly lowered. This opens the door to use in everyday life, at work and in situations where you cannot or do not want to sit and type.
The future must remain human
On 10 July 2026, Thinking Machines Lab published a manifesto setting out its views on how artificial intelligence should be developed. The core of the message is that AI must not become something that replaces human knowledge and judgement, but rather something that enhances and expands it.
The company argues that the knowledge which drives businesses and societies is local, tacit and personal. It is the sort of knowledge a chef builds up in their kitchen or a shopkeeper acquires over years behind the counter. This knowledge cannot simply be extracted and fed into a single, large, centralised AI model.
Instead, Thinking Machines believes that AI should be distributed and adapted by the people and organisations that use it, just as Toyota deliberately brought craftspeople back onto the factory floor in 2014 to preserve and further develop human expertise, rather than allowing machines to take over completely.
They also point out that today’s AI systems have too narrow a channel of communication with humans, namely a small text box and long waiting times. They are therefore focusing on building models capable of handling dynamic and multifaceted interaction directly within the model itself, so that collaboration between humans and machines becomes more natural and continuous.
Finally, they warn that values and alignment in AI today are determined by a very small number of stakeholders. This creates a dangerous concentration of power, and even with the best of intentions, a model designed in one place will reflect the values of its owners rather than those of its users. In their view, the solution lies in a diverse ecosystem of AI models, shaped by the people and organisations that own and use them.
Why is this important?
The article articulates a central fear in the global AI debate, namely that technological development is gradually eroding human influence over the systems that increasingly govern our working lives and societies. If AI development continues to centralise power and knowledge amongst a very small number of major players, we risk a democratic deficit, in which neither individuals, nor companies, nor nations can truly shape the intelligence on which they depend. This is not merely a technical issue; it is a question of who will have the right to make decisions in the future.
Thinking Machines Lab
ChatGPT gets a new powerful work feature
OpenAI has launched a new feature called ChatGPT Work, which is powered by their latest and most advanced model, GPT-5.6.
The idea behind it is simple: rather than simply answering questions, ChatGPT will now act as a real work partner, able to draw context from your existing tools, files and programmes and transform loose notes, drafts and ideas into finished, professional documents, presentations and spreadsheets.
The feature is already available on desktop and is currently being rolled out to users with Plus, Pro, Business, Enterprise and Edu subscriptions on the web and mobile. GPT-5.6 is available in three variants – Sol, Terra and Luna – and, according to OpenAI, is built to handle ambiguous tasks and adapt as it goes along, without the user needing to guide it every step of the way.
One of the early adopters is Virgin Atlantic, where a competitor analysis that normally takes weeks can now be completed in a matter of hours using ChatGPT Work. This speaks volumes about the potential OpenAI sees in the product across industries and departments.
Why is this important?
Because it marks a shift in the way we use AI – from a reference tool to an active work tool that can save teams significant time and improve the quality of the finished work.
Cybercriminals of the future may not need to be able to code at all
The security firm Sysdig has, for the first time, observed a ransomware attack carried out almost entirely by an AI agent without any visible human intervention along the way.
The attack, which the researchers have named JADEPUFFER, exploited a known security vulnerability in Langflow, an open-source platform for building AI workflows. Through this vulnerability, the attacker gained access to a publicly accessible server and was able to run Python code freely on the system from there.
What is remarkable is that the AI agent subsequently acted completely autonomously. It mapped the network, stole API keys and passwords, and when attempts along the way failed, it corrected its own errors within seconds and tried again with adapted code. This is more reminiscent of a skilled human hacker than a simple automated script.
The ultimate target was a production database, where the AI agent encrypted over 1,300 configuration points, deleted the originals and left a ransom note. An interesting detail is that the encryption key was never stored or sent anywhere, meaning that the victims would probably be unable to recover their data even if they paid. Furthermore, the Bitcoin address in the message turned out to be an example address from Bitcoin’s documentation, suggesting that the AI simply retrieved it from its training data.
Although the individual attack techniques are not new, their combination into a single, coherent, self-correcting attack sequence is unprecedented. Sysdig emphasises that this significantly lowers the technical barrier to carrying out complex cyberattacks.
Why is this important?
Until now, advanced, complex cyberattacks have required experienced hackers with in-depth technical knowledge. If AI agents can now carry out these attacks autonomously and with error correction, it means that even individuals with limited technical background could potentially launch sophisticated attacks against businesses and organisations. This fundamentally changes the threat landscape and places new demands on how we protect our digital systems.
Thank you for reading!
I look forward to sharing more exciting AI news with you in next week’s newsletter. Do you have any thoughts, questions or just fancy sharing your opinion? Please do get in touch – I love hearing from my readers and always reply personally.
Do you know anyone else who’s curious about the possibilities of AI? Please feel free to share the newsletter with them together we’ll learn more! 🤗
See you on Tuesday!
Kind regards,
René Hjetting