Welcome back to Artificial Insights, your occasional bulletin from the near future.
Last Thursday was the busiest day our WhatsApp group has ever had. I would have bet on a model launch. Instead it was Nvidia’s week.
Reports had come out about an extramarital affair between Jensen Huang and a Chinese woman. At the same time, questions were piling up about more than $7 billion worth of GPUs that had reached China through a company in Singapore, despite US export controls. By the afternoon, Nvidia announced a “CEO transition”, with the CFO as interim CEO.
The memes arrived before the press release. One member asked what a “side chick” was. The answer came as a menu, with a main dish and side dishes, and more memes in the same vein followed.
A member called it sexist and said they hadn’t expected it here. The poster apologised and asked, sincerely I think, why it was sexist. Within hours the group was no longer discussing Nvidia but humour itself: who decides what is offensive, whether calling something out is censorship, and whether grown-ups should just scroll past.
What stayed with me is that both sides made the same argument.
One side said that because we are a multicultural group, nobody can impose their sense of what is offensive on everyone else. The other side said that because we are a multicultural group, nobody can assume their sense of what is harmless holds for everyone else.
One member explained how often in their life they had heard “it’s just a joke” about who they are. Others asked everyone to step away from the keyboard and get back to AI.
By the evening, the conversation had found the more interesting story. The affair made the headlines, but the export controls are what matter. How did billions of dollars of restricted hardware reach China? It is almost as if the infidelity gives everyone something easier to talk about than the chip war between the US and China. And it raises the question of how much of this industry depends on the judgement of a handful of people.
I’ve run this group for almost three years, and I still think of it as a black box.
You can write a code of conduct, choose a topic and invite people you trust. But you have no idea how the chemistry of a group will turn out. It changes with every person who joins, and it reacts to the news in ways you can’t predict.
A scandal in California lands differently in Amsterdam, São Paulo and Singapore. Humour is the first thing that fails to translate. A meme only works when everyone shares the context, and a group of hundreds of people from dozens of countries shares surprisingly little context.
I handle most of these moments in private messages, because most people don’t want a fight. They want to follow the news, laugh at silly memes and share what they are trying with AI. You can’t put a neat fence around that. What you can ask is that people stay respectful of each other’s sensitivities. That goes a long way.
On Sunday the group turned 1,000 days old. So I exported the chat and read all 17,551 messages, to see what this black box looks like from the outside.
The fight was the exception. In 1,000 days I have only had to remove spammers. Fifty people write 90% of the messages, and most members read and rarely post.
The conversation follows the industry: from ChatGPT to Claude, and from talking about AI to building with it. Some of our predictions aged well. Some of mine did not.
Not in the group yet? Join the WhatsApp group
MZ
Recursive Self-Improvement Has Started (44 min)
Dwarkesh Patel walks through why three months of debates collapsed his superintelligence timeline from “the better part of a decade” to much sooner, arguing math is the preview: AI R&D is largely verifiable, so automating top researchers compounds. He also steelmans the objections, continual learning, finding verifiable proxies, data, and experiment compute.
Bots That Earn Their Permissions (60 min)
Lauren Tan and Roman Ugarte of Y Combinator‘s Lightcone detail how Grok Bot ships 4,000 PRs a month, with automated tests and codebase guardrails letting agents check their own work and earn broader access over time. Their argument: when execution gets cheap and fast, product judgment and goal-setting become the bottleneck.
Perspective Is the Last Moat (25 min)
Dan Shipper, CEO of Every, argues in his Thesis: 2027 closing keynote that benchmark charts like OpenAI’s GDPvalmake the future of work look bleaker than it is: as intelligence gets cheap, the scarce input becomes your non-consensus point of view and the fact that you still have to pick the right idea.
Consciousness as a Learning Algorithm (97 min)
Joscha Bach, founder of the California Institute for Machine Consciousness, argues his null hypothesis: consciousness is an algorithm not much more complex than a transformer, simple enough for an infant’s brain to find on its own. His Machine Consciousness Hypothesis splits into two independently falsifiable parts, and he explains why there will never be a Turing test for it, why a chess program is already a zombie, and why faking a mind can be harder than building one.
Events
Upcoming events from Envisioning
Every month we run Signals, a free 30-minute online session where we scan one theme live. Signals is our research platform: it asks several AI models the same question at once, compares their answers, and sorts what they find into social, technological, economic, environmental and political signals. You watch the scan happen, ask questions along the way, and leave with a link to the demo. You don’t need to prepare anything.
Longevity & Healthcare
Wednesday, December 9, 2026 · 15:00 CET · 09:00 ET · 06:00 PT · 22:00 SGT
People are living longer, and that changes what we need from health care, work and money. We’ll scan what is shifting and who is moving first.
Robots in Manufacturing & Logistics
Wednesday, January 20, 2027 · 17:00 CET · 11:00 ET · 08:00 PT · 00:00 SGT
Humanoid and embodied robots are leaving the demo stage. We’ll look at where they are actually being deployed, in factories, warehouses, care and retail.
Vocab
Newly added to our AI & ML lexicon
Qualiaphobic AI: Neural network architectures deliberately engineered to resist any theory of consciousness classifying them as conscious
Weight-Space Programming: An LLM reads and edits another network’s weights directly through reasoning in natural language, without gradients, training loops, or an optimizer.
Harness Engineering: The discipline of designing and evolving an agent harness, the runtime that couples an LLM to the world through a loop, tools, context management, safety controls, orchestration, and extensions.
Semantic Calibration: The property of a base LLM’s confidence being accurate at the level of answer meaning, not just next-token probability, and the finding that RLHF and chain-of-thought reasoning break it.
Biomolecule Watermarking: Embedding a detectable, function-preserving signal into an AI-designed protein or nucleic acid sequence or structure so its AI origin can be verified later.
Explore 1,400+ more terms in the Envisioning lexicon.

If Artificial Insights makes sense to you, please help us out by:
📧 Subscribing to the weekly newsletter.
💬 Joining our WhatsApp community.
📥 Following the weekly newsletter on LinkedIn.
🦄 Sharing the newsletter on your socials.
Artificial Insights is written by Michell Zappa, CEO and founder of Envisioning, a technology research institute.






