Welcome to the mid-week update from New World Same Humans, a newsletter on trends, technology, and society by David Mattin.

If you’re reading this and haven’t yet subscribed, join 24,000+ curious souls on a journey to build a better future 🚀🔮

To Begin

The generative AI rollercoaster is thundering forward at increasing speed.

This week, researchers at Stanford and Google use an enhanced large language model to create simulated people that can remember, plan, and talk to one another in pursuit of their longterm goals.

Also, a new report says we’ve reached a landmark moment for the global energy system. And amid rumours of financial difficulty, Stability AI release a new text-to-image model for enterprise users; it’s capable of amazing photorealism.

Let’s go.

🏠 Welcome to SimGPT

This week the generative AI talk orbited around autonomous agents. That is, AI systems that can act autonomously in pursuit of pre-defined goals.

Researchers at Stanford and Google explained how they used a large language model (LLM) to create 25 simulated people, who were then set loose inside a virtual town called Smallville.

To create these sims, the researchers hooked up their LLM to an architecture that allows each AI agent to store memories of its past experiences, and then to access relevant memories and use them to plan new actions. Each agent was imbued with its own persona, for example: 'John Lin is a pharmacy shopkeeper at the Willow Market and Pharmacy who loves to help people. He is always looking for ways to make the process of getting medication easier for his customers.’

The results, say the researchers, were ‘believable individual and emergent social behaviours’ that saw Smallville become a bustling little town full of autonomous chit-chat, group activities, and trips to the local café.

‘…for example, starting with only a single user-specified notion that one agent wants to throw a Valentine’s Day party, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time.’

All this comes amid vast excitement on Twitter this week over the rise of useful generative agents.

As with the above research, these innovations — which include AutoGPT and BabyAGI — leverage architecture that enhances ChatGPT by allowing it to lay down and then access a stream of past actions, or ‘memories’. Combine that with a plugin the enables ChatGPT to browse the web, and the result is a system that can take an initial goal, get started online, and then prompt and re-prompt itself until it’s finished:

⚡ NWSH Take: If SimCity or The Sims formed a part of your childhood, then this new research is hypnotic. It makes clear that LLMs will enable us to simulate goal-directed people, and watch as complex social and behavioural dynamics unfold. Imagine video games populated with these simulated humans (hello, Electronic Arts). But we’ll also see the rise of new art forms — a mixture of game and movie — built around them. And then will come the ability to simulate large populations, allowing for new insight on collective phenomena such as voting behaviours, the spread of disinformation, and the evolution of the economy. This new age of AI-fuelled simulation — which I’ve been writing about for a while — is emphatically here. // Meanwhile, AI agents such as AutoGPT promise to elevate the usefulness of generative models for millions of individual users. It’s already clear that for most people using LLMs won’t be about sitting at the prompt line and figuring out great prompts. Instead, wrappers such as this one — which puts AutoGPT-like powers direct into your browser — will allow users to set a goal and then let the LLM iterate its own way to a useful output. Give it a try; something hugely powerful is happening.

🔌 New power generation

Also this week, news of a landmark moment for energy.

A new report from independent energy think tank Ember says solar and wind accounted for a record 12% of global electricity generation in 2022. That’s up from 10% in 2021. The increase in wind generation alone in 2022 was the equivalent of the entire annual electricity demand of the UK.

What’s more, says the report, it’s likely that 2023 will see electricity generation via fossil fuels — mainly coal and natural gas — hit their peak.

The research team analysed data from 78 countries, representing 93% of global electricity demand.

⚡ NWSH Take: Around two-thirds of the world’s electricity is generated by burning fossil fuels. But the transition to solar and wind is now reaching a blistering pace, thanks largely to exponentially falling cost. In 1956 the cost of one watt of solar capacity was $1,825; now it can be as little as $0.72. // If Ember are right, we’ll soon start generating more electricity via fewer fossil fuels: power up, emissions down. That aligns with the International Energy Agency’s most recent and broader forecast; they now have global demand for fossil fuels — via electricity generation or any other use — peaking or plateauing under all their future scenarios, even without any shift in current government policies. // We’re approaching, then, a historic turning point: the decoupling of economic growth and fossil fuels for the first time since the Industrial Revolution. It’s becoming possible to image a world of endless, near-zero cost clean electricity. A world of clean energy abundance. What will that make possible?

🖼 Get real

Do you want to look at some amazing AI generated images? Yes, of course you do.

This week Stability AI released Stable Diffusion XL, a text-to-image model aimed at enterprise users. The model is an advance on Stable Diffusion 2.1, and excels at ultra-photorealism.

The move comes amid reports that Stability AI is struggling with huge server and talent costs. These reports suggest the company is seeking a new round of funding, and that investors are wary given the current revenue. CEO Emad Mostaque has not commented.

⚡ NWSH Take: A few quick thoughts. The images are stunning; that’s obvious. At this point we’ve pretty much entirely scrambled the role that the photograph once played in our culture as a form of proof or marker of veracity. In the wake of puffa coat Pope, I’ve already developed a new reflexive habit: is this real or AI? // As for the rumours about Stability AI, they amount to: ‘AI startup experiencing rocket ship growth is struggling to figure out revenue and is a chaotic place to work’. Nothing too surprising. Whatever storms the company is experiencing, I hope it can weather them; Mostaque’s vision of AI for the people is a necessary counterweight to the closed model that is being operated by (the misleadingly named) OpenAI and others.

🗓️ Also this week

🇨🇳 The CCP has issued new rules on the training and outputs of generative AI models. Draft rules from the Cyberspace Administration of China say the outputs of those models must reflect the core values of socialism and not undermine the power of the state. This came as Chinese tech giant Alibaba announced plans to roll out its LLM rival to ChatGPT, Tongyi Qianwen, across all its products.

🇺🇸 The US government is also looking to establish new regulations around AI. The National Telecommunications and Information Administration is asking for feedback from the public and experts from industry and academia, and wants to establish ‘guardrails’ to ensure AI is safe, transparent, and as unbiased as possible.

👮‍♂️ The Boston Dynamics robodog will patrol the streets of NYC on behalf of the New York Police Department. The NYPD experimented with Spot the Dog in 2