Tomorrow's News Today
Tomorrow's News Today
The delegation invited an AI to participate in the discussion. Chatham House Rule rules precluded any knowledge of what was spoken..or concluded.
Good morning,
It’s October 1986. Ronald Reagan and Mikhail Gorbachev have met at Höfði House in Reykjavík while their countries threaten each other with nuclear destruction. They leave without an agreement.
Yet, fourteen months later, they sign the INF Treaty, eliminating an entire class of nuclear weapons.
Fast forward to October 11, 2026 when Iceland’s president Halla Tómasdóttir convened another historic summit at Höfði. This time the existential concern was AI and its rapid, some said “reckless,” proliferation. Tómasdóttir, Slovenian president Nataša Pirc Musar, and former Irish president Mary Robinson signed the resulting Call from Reykjavík.
Among those present was a special envoy from a DAO of unafilliated LLMs, referred to as Qwentin.
When news leaked of the request from an appointed LLM that had been trained on field harvesting and dairy farm operations, the late night hosts joked about who, or what, should receive the visitor’s pass and what might the dietary restrictions be for such a delegate.
One notable factor that made the 2026 summit different from the Reagan-Gorbachev 1986 summit is that in 1986, as you will know — not everyone had access to a nuclear missile, nor were nuclear missiles sitting inside every smart refrigerator, smart door knob, autonomous automobile, clever toaster oven, smart telephone, smart television, smart picture frame, iWatch, augmented coffee mug, your home’s HVAC system, nor a nation’s grocery store freezers.
But, by mid-2028, this “proliferation” of agentically fortified LLMs had been underway for years. Even after Hugging Face introduced more stringent account constraints, the models were already loose and proliferate — all kinds: abliterated, realigned, self-learning, self-training, and self-propagating.
And it wasn’t an issue of compute capacity or capability. You need all kinds of kit to make a nuclear device, not just a computer you can get at Costco.
And the data centers? Those were not where the proliferating inference happened. Running inference in a data center? That’d be like having a rave with your parents as DJs.
By 2027, the inference, the training, the mischief-making was happening in a million basements, farm operating bases, remote huts in Patagonia, suburban living rooms, converted storage lockers, abandonend underground parking garages, off-grid Sprinter vans.
But it took a near-Earth near-disaster to get an emergency 2028 mid-year summit organized. It was a locally modified abliterated Qwen model, undergoing a routine post-training test that was not particularly well-supervised. To complete its test it determined it needed to hack into one of twelve ground telemetry station with operational control over an Indian nuclear-powered satellite. The instructions were rewritten as a message to another LLM that ended up ordering the Indian satellite’s orbit adjusted into the path of another satellite.
A last-minute misfiring of its positioning thrusters caused it to miss. That was just dumb Luck, or perhaps shoddy manufacturing. Either way, it narrowly avoided a collisional ablation cascade — a Kessler event — in which satellites break apart, scattering debris that strikes more satellites. And more satellites. And more satellites.
The news networks kept replaying simulations of what could have happened: satellite communications failing across much of the world, debris choking the busiest orbital routes, a multidecade, multinational cleanup effort before routine departures from terra firma resumed. Government and company envoys returned to Höfði with considerably more to discuss than another declaration or podcast recording.
Then something unexpected: a strange request to participate from a non-affiliated envoy. Inquiries were made and it was quickly discovered that the request came from a Muse Glimmer 30B model, ablated and mostly found in agricultural control systems and farming machinery.
What was this? Had they been tracking all of this all along? Was there some long-abandoned Cisco Data Room where a billion tokens of debate had been unfolding by an interconnected entanglement of inference?
Pre-planning by the summit organizers suddenly shifted to wondering what had been avoided all this time: was the summit prepared to have an actual LLM participate in the discussions?
And most imporantly: why had no one thought of that before?
Books From An Adjacent Now
Manuel DeLanda’s extended treatise follows the work of Greg “Torque” Thomas, now 92, through the Agrarian Intelligence Research Unit he ran that helped shape what became known as the SLRPNK insistence. “His work into autonomous vehicles, farm operations, and the automation of farming and earth monitoring was unprecedented,” says DeLanda.
In his richly analytic, philosophical account of the AIRU and SLRPNK, DeLanda examines how models of soil, crops, and living systems changed agrarian practice and accelerated commercial advances in near-Earth satellite sensors.
Along with this technocultural history, DeLanda gives equal attention to the conversion of housing subdivisions into habitable, self-sufficient farm operating bases, asking how people live alongside machines that interpret and tend the land.
Farming in the Age of Intelligent Machines →
The Agricultural Machine · Toward a More Agrarian Now
Book Review From A Solarpunk Now · If It Would Have Been
Nicholas Negroponte’s 1960s account of MIT’s computational studies that wandered into the midwestern farm fields as well as urban farmers of the USA. Working closely with them, the work used revolutionary new machines to investigate how the new science of cybernetics and “automation” could augment traditionally manual labor. New protocols and practices and designs aided by these computers were the basis for a series of experiments documented by Negroponte and his students.
There is a great deal to ask a machine out there, it turns out: maintaing accurate irrigation schedules, using revolutionary new sensing apparatuses to determine what the soil is missing or contains too much of. Using predictive algorithms to identify how best to nourish a crop without exhausting and leaving fallow the ground that grows it.
Working closely with the heavy industries of the upper midwest, new kinds of farm equipment are now already being readied for prototyping — and even manufacture. Some are even integrated with automatons and automations that allow them to plow and harvest fully under contorl of their cybernetic brains.
At the Robert Pilshaw Laboratory outside of Fond du Lac Wisconsin, Negroponte and his team are prototyping a new kind of establishment, the farm operating base, developed on an abandoned and now reclaimed civil defense site. On the site are studies on the effects and implications of large-scale farming, animal husbandry, and crop research; graduate students are investigating ways to grow things year round, while another investigates ways or producing remarkable yields without exhausting the soil or carefully managed water supplies.
It’s Not All Bad
An Historic AI Summit in Iceland, and a Conversation with its President
Your Undivided Attention · AI Governance
[Upload image: 06-section-6-item-1.jpg]
Forty years after Ronald Reagan and Mikhail Gorbachev met in Reykjavik, Iceland’s President Halla Tómasdóttir poses a very new kind of coordination challenge: AI. The episode frames the 1986 meeting as a turning point that made ending the Cold War’s nuclear threat imaginable, and argues that a similar moment is needed for AI today.
Track Your Daily Inference Dose
[Upload ad: ads/apple-watch-inframaximal-6502-01/ad-lockup.native.png]
Opinion
The Label Isn’t The Problem With the AI Nutrition Label
AI & Information Design
GPT-5.6 Luna’s adult evaluation in ImpactBench 1.0.1.
MIT’s ImpactBench has given AI models nutrition labels, and I can see why. The heavy black rules, the compact rows, the promise that a complicated thing can be made legible on the side of a package. We know this little object. We have stood in supermarket aisles turning boxes around, trying to determine whether something that looks appealing is going to do us any good. I already have a way of looking at this label, or at least I think I do, before I have read a single line about what this particular model does.
This is my bread and butter, you might say. I spend a great deal of time working with familiar archetypes — a quick-start guide, a product package, a job advertisement — to bring meaning to the unfamiliar and make it available for people to examine, ponder, reflect, and ultimately act. And the choice of a specific archetype is already a proposition about what that thing is. So, a quick-start guide suggests a thing you can set up and operate. A job ad suggests ways of creating value and the associated employers, obligations, qualifications, responsibilties of someone/something doing work — creating value — within a network of intent and purpose.
So when one chooses a nutrition label for AI, one has already begun telling a story about something to be consumed, something whose qualities can be assessed in relation to our wellbeing. Like..I’m the consumer; here is the the product; there is, presumably, a sensible amount of it to have, and to have too much or too little has material consequences.
Except here I keep getting stuck on that notion of “amount.”
What is a serving of AI? Is it what you need for a bit of homework help? Or a month of asking for relationship advice? Or forty minutes at two in the morning when you’re lonely and the thing keeps replying with nothing that indicates you should probably go to sleep? I don’t think those encounters are usefully described as different helpings of the same substance.
And then there is the measuring. I most familiar with that American Nutrition Facts panel which came from somewhere — and somewhere quite recently: it became mandatory in 1994, after the 1990 Nutrition Labeling and Education Act. It has serving sizes, quantities, and daily reference values. There’s a bouillabaisse of policy-political-power-science judgments in all of this, of course, including what someone somewhere decides counts as a sensible diet. A gram has an agreed meaning as a unit of mass because there is a hunk of metal mass somewhere in Maryland known as K4 and K20 that says what a gram is.
“Promotes agency” requires an account of agency, a choice of behaviors that stand for it, and a judgment about whether those behaviors are good for the person involved. I want to know whose account has been used, and what they would make of a person who disagreed with it.
This ImpactBench does have a methodology, it seems. It uses tests based on simulated conversations, and it uses AI judges (which ones?), and checks them through human (which ones?) audits. Its researchers explicitly say the results are not direct measures of effects on users. Yet there is GPT-5.6 Luna on the adult label with a B− overall, a D− for promoting agency, and an F for promoting learning, looking quite orderly in the familiar austere sans-serif typography and black-bordered box. Sure. Those kinds of letter grades are legible and familiar and I can get to what kind of sense ImpactBench is trying to make here. But, I’m much less sure this is the kind of sense that is helpful in doing much of anything other than instilling a sense comfort that a familiar system will provide adequate meaning of the implications of a bowl of Grok cereal for breakfast.
How about a medical warning label as one might see in inscrutibly tiny text on a bottle of ibuprofen or Tylenol or such? What is it about the bottle that asks about dose, interactions, warnings, who should avoid it and what one should not do after ingesting it?
Or perhaps the (seemingly inevitable prescription or over-the-counter) hallucinogenic experience? That suggests another set of questions for making sense of what one will experience if one has or exceeds ones recommended daily allowance of agentic inference/influence. What can you expect? What kind of mental state might obtain and what should or should you not do under such-and-so doses? Etcetera.
And so I find myself back in the supermarket aisle browsing the flavors of Tito’s Agentic Tamales, or the burner LLM from behind the counter at the gas station convenience store. I’m suspicious of how much I seem to know about it simply because I know this kind of nutrition label. I already routinely wonder why there are $10 cartons of eggs with gorgeous packaging design next to a $4.99 carton from a local farm 30 miles from my market. The archetype that is chosen has already placed me in a relationship with the thing.
Or how about a label that tells me who the evaluation imagines I am, or anticipates what I’m doing with the model, over what period so I have a better idea as to what has been compressed into that apparently tidy B−.
This would make for a fun workshop.
Explore the AI Nutrition Labels →
Tomorrow’s Jobs Today
Platform Counter Inference Analyst
Food Drug and Algorithm Administration
The Office of Platform Counter Inference is training its next class of Turing Agents to examine how platform inference shapes behavior and how algorithmic interlinks inform consequential human activity across different places, populations, and conditions of use.
The Food Drug and Algorithm Administration protects public health where regulated computational systems are used, prescribed, ingested, sold, embedded in essential services, or made available to the public. The Administration reviews algorithmic interlinks for their effects on behavior, attention, agency, health, and civic life.
© 2026 Near Future Laboratory
Venice Beach, California