Why JansBrief exists
Jan Stenbeck was the smartest person I ever met. Not smart in the way academics are smart. Smart in the way that changes the world. He saw what nobody else saw. He understood that mobile telephony would revolutionise countries that hadn't even laid copper wire yet. He broke state socialist monopolies when everyone said it was impossible. He built empires out of ideas.
Every day Jan received a binder. Two people read all the world's important newspapers and magazines for him and pulled out what mattered. The things others missed. The faint signals that foreshadow great change.
I worked with Jan. I learned from him. And I have never forgotten that binder. JansBrief is my tribute to him, a modern version: global, AI-driven, available to everyone with ambition.
In memory of Jan Stenbeck
1942 — 2002
Jan Stenbeck
Tele2, Millicom, MTG, Metro
In today's edition · 10 July 2026
Something genuinely unsettling — and genuinely illuminating — emerged from Anthropic's research lab this week. The company built a tool called the Jacobian lens that peers into the internal states of its large language model, Claude, while it is working through a problem. What they found is not the usual "we made AI more transparent" press release. They found something closer to a topology of thought.
The Jacobian lens works by mapping how the model's internal representations shift as it processes a query — not just what comes out at the end, but the intermediate conceptual manoeuvres the model performs on its way to an answer. Anthropic's researchers describe finding a "hidden space" where the model appears to puzzle over abstract concepts before committing to an output. Some of what they observed was mundane: the model cycling through syntactic patterns, selecting vocabulary. But other observations were more arresting. The model appeared to hold competing representations of a concept simultaneously, resolve ambiguities through something that resembles — but is not — deliberation, and occasionally settle on answers through pathways the researchers could not fully explain.
This matters because interpretability — the ability to understand why an AI system does what it does — is the single most important unsolved problem in the field. Regulators want it. Companies deploying AI in healthcare, finance, and defence need it. And until now, most interpretability work has been either too shallow (attention maps that show which words the model focused on) or too abstract (mathematical proofs about theoretical networks that bear little resemblance to production models). The Jacobian lens sits in a productive middle ground: it works on an actual, deployed model, and it reveals structure that is neither trivially obvious nor completely opaque.
The unnerving part is the gap between what the lens reveals and what we can currently explain. If a model holds multiple competing framings of a medical diagnosis before selecting one, and we can see that it does this, but we cannot yet explain why it resolves the competition one way rather than another — are we more or less comfortable deploying it? Anthropic's own researchers are honest about this tension. The tool is a microscope, not a map.
For the AI industry, the implications fork in two directions. If interpretability tools like this mature quickly, they could become the basis for a new generation of AI auditing — not the checkbox compliance that Brussels currently favours, but genuine, mechanistic understanding of model behaviour. That would be a competitive advantage for any company that masters it first. Alternatively, if the hidden spaces reveal patterns that resist human comprehension even with better tools, we face a harder question: whether deploying systems whose internal reasoning we can observe but not understand is fundamentally different from deploying systems whose reasoning is invisible.
Source: MIT Technology Review · 10 July 2026
Now — The race for interpretability becomes a product race: Anthropic's Jacobian lens is not merely academic. As enterprises sign large contracts for AI deployment in regulated industries, the ability to explain model behaviour becomes a contractual and legal requirement. Any lab that can offer genuine mechanistic interpretability — not just confidence scores or saliency maps — gains a tangible edge in enterprise sales. Expect competitors to accelerate their own interpretability efforts within months.
Soon — Regulators get a tool they did not expect: The EU AI Act's risk-based framework demands transparency for high-risk applications, but lawmakers drafted those rules assuming interpretability was a distant aspiration. If tools like the Jacobian lens prove robust, regulators may shift from demanding procedural documentation to demanding mechanistic audits — a far higher bar. Mid-tier AI markets such as Brazil, India, and the UAE, which are writing their own frameworks now, may leapfrog Europe by incorporating these techniques from the start.
Later — The philosophical boundary sharpens: If AI systems demonstrably "puzzle" over concepts in hidden representational spaces, the distinction between processing and reasoning becomes harder to maintain. This does not mean the models are conscious — that claim would be irresponsible — but it does mean the language we use to describe what they do will need to evolve. When the machines deserve our consideration, as Noema's latest essay asks, the answer may depend less on philosophy and more on what the Jacobian lens shows us next. Sources: MIT Technology Review · 10 July 2026; Noema Magazine · 10 July 2026 ---
Africa CDC has declared that the current Ebola outbreak in the Democratic Republic of Congo is producing cases at double the rate of the catastrophic 2013–2016 West African epidemic during its first six weeks. The disease is spreading in eastern DRC, a region already convulsed by armed conflict and displacement, which makes contact tracing nearly impossible. International response has been sluggish relative to the speed of transmission. The combination of war, population movement, and weakened health infrastructure creates conditions that previous outbreaks — even the worst ones — did not face simultaneously at this scale. Source: The Africa Report · 10 July 2026
The son of imprisoned former prime minister Najib Razak suggested that a strong showing by the ruling Barisan Nasional coalition in Saturday's Johor state election would validate calls for his father's pardon. What was expected to be a sleepy walkover has turned competitive overnight. Opposition parties seized on the remark as proof that the coalition treats elections as transactional instruments. The episode illustrates how one sentence can rewrite the calculus of an entire race — and how the Najib question remains the live wire in Malaysian politics. Source: South China Morning Post · 10 July 2026
At least twelve people died in a wildfire near Los Gallardos in the province of Almería as southern Europe's heatwave continues. Several victims were found inside burned vehicles, suggesting the fire moved faster than evacuation could keep pace. Andalucía's emergency agency called it the worst wildfire in the region's recorded history. Spain's fire season is arriving earlier and burning hotter each year, but rural road infrastructure and evacuation protocols have not kept up. Source: BBC World · 10 July 2026; El País · 10 July 2026
Argentina's state oil company YPF confirmed plans to begin exploratory drilling off Uruguay's coast between late 2027 and early 2028. CEO Horacio Marín said the area's potential could surpass Vaca Muerta — the shale formation that transformed Argentina into a net energy exporter. The move signals a new chapter in South Atlantic energy geopolitics: Uruguay has no hydrocarbon production tradition, and the arrival of Argentine state capital in its waters will test the smaller country's regulatory readiness and diplomatic poise. Source: MercoPress · 10 July 2026
Fifteen years after becoming the world's youngest nation, South Sudan marked its anniversary amid rising infighting between the two factions of its ruling movement, the SPLM and SPLM-IO. The United States used the occasion to denounce the upcoming December elections as a "farcical" process, and international goodwill — once abundant — has largely evaporated. The country remains one of the poorest and most conflict-affected on earth, a cautionary tale about independence without institutions. Source: The Africa Report · 10 July 2026
Chinese AI startup MiniMax is raising as much as $2 billion through new shares and convertible bonds, joining a wave of Chinese technology companies tapping markets for artificial intelligence capital. The deal underscores how Beijing's state-backed ecosystem is producing a distinct class of AI competitor — well-funded, domestically rooted, and increasingly capable. For Western labs already racing to stay ahead, the signal is clear: the capital flowing into Chinese AI is no longer a trickle but a torrent, and the competitive landscape is bifurcating along geopolitical lines faster than most industry forecasts anticipated. Source: Bloomberg · 10 July 2026
A joint investigation by Der Spiegel and partners has published secret documents showing the extent of military cooperation between Russia and China — going well beyond the "no limits" rhetoric. The documents reveal joint planning for specific contingencies, shared intelligence frameworks, and coordinated weapons development programmes. This is not the vague alignment that Western defence analysts have been debating for years. It is operational integration, and it changes the threat calculus for NATO, Japan, and Australia simultaneously. Source: Der Spiegel · 10 July 2026
At the NATO summit in Ankara, Turkish President Erdoğan surprised departing leaders with a distinctive farewell gift: a Turkish-made Gümüşay .357 Magnum revolver, complete with ammunition. Spain's Pedro Sánchez promptly handed his to the Interior Ministry for deactivation. The gesture was pure Erdoğan — a display of Turkish manufacturing prowess wrapped in provocation — and left protocol teams across the alliance scrambling. It was, by most accounts, the only moment at the summit that genuinely surprised everyone in the room. Source: El País · 10 July 2026 ---
Here is a problem that would have made a certain kind of systems thinker lean in. When a road cuts through a forest, it does not just divide land — it bisects an ecosystem. Tree-dwelling animals that once moved freely through the canopy are suddenly stranded, forced to cross tarmac where they are hit by vehicles, or simply marooned on one side. The solution, increasingly deployed in tropical forests, is the canopy bridge: a cable-and-rope structure strung between tall trees on either side of the road, high enough for traffic to pass beneath, strong enough for monkeys, possums, and other arboreal species to cross above.
What makes this interesting is not the engineering — the bridges are relatively simple — but the question Mongabay's reporting forces: do they actually work? A recently published study examined three years of camera-trap footage from canopy bridges and found that the answer is more nuanced than conservationists had assumed. Some species adopted the bridges quickly and used them regularly. Others ignored them entirely or used them only seasonally. The effectiveness depended on bridge design, height, surrounding forest density, and the behavioural ecology of the target species — variables that most installation projects barely considered.
The pattern here is one that applies far beyond wildlife. Infrastructure designed to reconnect what has been severed — communities, supply chains, ecosystems — often looks elegant on paper and underperforms in practice because the designers assumed a simpler model of behaviour than reality provides. The bridges that worked best were the ones whose builders studied how animals actually move, not how engineers assumed they would. The ones that failed were built to a generic specification and dropped into place.
Still — watch the trajectory. The study's authors argue not against canopy bridges but for better ones: species-specific designs, adaptive placement, and long-term monitoring rather than one-off installations. It is a small intervention with an outsized lesson: reconnection is not a product. It is a process.
Source: Mongabay · 10 July 2026
In rural West Flanders, Stefanie Verduyn has turned her family's working farmstead into a contemporary art gallery that asks visitors to do something radical: slow down. Appointments only, no white walls, no rushing. The exhibition spaces sit among fields and outbuildings, and the visit is designed to take half a day, not twenty minutes. In a market obsessed with spectacle and Instagram-speed consumption, Verduyn's model inverts every incentive — and is attracting serious collectors precisely because the unhurried experience makes the art land differently. Source: Artnet News · 10 July 2026
The Economist reports that dark-sky tourism is one of the fastest-growing segments in the travel industry. Lodges and parks that once struggled to attract visitors are marketing their darkness as a luxury product. International Dark-Sky Association certifications have become coveted assets. The economics are persuasive: a remote location's greatest liability — distance from light-polluted cities — becomes its premium. From Namibia's NamibRand to New Zealand's Aoraki Mackenzie, the darkest places on earth are building an illuminating business. Source: The Economist · 10 July 2026
In New York, professional line-sitters now earn money queuing for viral food — this summer's obsession being TikTok-famous dot cakes at Butterfield Market in Manhattan. Lines form at 6 a.m. and last hours. Eater reports that the practice has formalised into a micro-gig economy, with line-sitters advertising on apps and charging by the hour. It is absurd, and it is also a perfect lens on how scarcity, social media, and the willingness to pay for someone else's time converge in 2026. Source: Eater · 10 July 2026
Artnet makes the case for Escher — the Dutch artist whose impossible staircases and tessellations adorned college dorm walls for decades — as a figure whose mathematical rigour and visual invention merit the same art-historical attention given to his contemporaries. Long dismissed by the fine-art world as a "poster artist," Escher's exploration of infinity, symmetry, and spatial paradox anticipated computational aesthetics by half a century. The reassessment is overdue. Source: Artnet News · 10 July 2026
Architect Liz Gálvez has installed a pavilion called "Earthen Comforts" in the courtyard of Materials & Applications and Craft Contemporary in LA. Built from cord and compressed-earth blocks on a wooden post-and-beam frame, the structure demonstrates how urban shade — not mechanical cooling — can meaningfully reduce temperatures in outdoor public spaces. In a city where summer heat is an equity issue, the pavilion is both a provocation and a proof of concept. Source: Dezeen · 10 July 2026
In Buenos Aires's Villa Ortúzar neighbourhood, a new family-run bistro has opened with white tablecloths, candlelight, open kitchen, and a seasonal menu that crosses Spanish and Italian coastal traditions. La Nación calls it a throwback to a dining culture that Buenos Aires has been losing to fast-casual concepts — intimate, personal, unapologetically slow. No reservations app. Just show up. Source: La Nación · 10 July 2026 ---
On secondary markets, Anthropic has reached a valuation of approximately $1.2 trillion — surpassing OpenAI. Swedish tech outlet Di Digital reports that Anthropic is now expected to beat its rival to a public listing, with an IPO likely in Q4 2026. The reversal is remarkable: two years ago, Anthropic was widely seen as the cautious, safety-first underdog. Its emphasis on interpretability research — including the Jacobian lens work covered in today's Signal — appears to have resonated with enterprise buyers who prize reliability over raw capability. The question is whether the valuation reflects genuine commercial traction or the market's habit of projecting current momentum indefinitely. Source: Di Digital · 10 July 2026
A team has experimentally observed how Einstein's special relativity distorts chemical bonds in a charged molecule of bismuth and carbon. New Scientist reports that the experiment confirms predictions that had existed only in theory: at the speeds electrons reach in heavy atoms, relativistic effects reshape bond lengths and energies in measurable ways. This is not a curiosity. Relativistic chemistry explains why gold is gold-coloured and why mercury is liquid at room temperature. The new experimental technique could improve the design of catalysts and heavy-element materials used in energy storage and medical imaging — areas where computational chemistry currently relies on approximations of relativistic effects rather than direct measurement. Source: New Scientist · 10 July 2026
New Scientist reports that the foundational mathematics underpinning thermodynamics — the laws governing heat and work that have stood largely unchanged since the nineteenth century — is being overhauled through the application of gauge theory, a framework borrowed from quantum field theory. The approach promises to give the laws of thermodynamics a firmer mathematical footing, resolving ambiguities that physicists have tolerated for two centuries because the practical equations worked well enough. What gauge theory offers is not a new set of laws but a more rigorous structure beneath the existing ones — the kind of deep mathematical housekeeping that tends to produce unexpected insights a generation later. If the reformulation holds, it could sharpen our understanding of irreversible processes, from engine efficiency to biological metabolism, and open pathways for designing energy systems whose thermodynamic limits are more precisely known. Source: New Scientist · 10 July 2026 ---
14
14
The number of people killed in Iran by US strikes since Tuesday, according to Iran's health ministry — reported alongside CENTCOM's claim of hitting 90 Iranian targets in the latest round. But the number that matters more sits underneath: the Strait of Hormuz, through which roughly one-fifth of global oil and LNG shipments normally pass, has seen traffic grind nearly to a halt. The gap between those two numbers — 14 dead and an entire maritime chokepoint frozen — captures the asymmetry of modern conflict in the Gulf. A relatively contained exchange of fire produces disproportionate economic consequences thousands of miles away. Brent crude held steady this week rather than spiking, which suggests the market is pricing in resumed traffic rather than escalation. But every day the Strait stays quiet, supply chains for Asian refineries tighten, LNG spot prices for South Asian buyers creep upward, and the insurance premiums on tanker routes through the Gulf of Oman compound. Fourteen is a small number. Its ripple is not.
Source: BBC World · 10 July 2026; Al Jazeera · 10 July 2026
In perspective
The number of people killed in Iran by US strikes since Tuesday, according to Iran's health ministry — reported alongside CENTCOM's claim of hitting 90 Iranian targets in the latest round. But the number that matters more sits underneath: the Strait of...
8 — Today's Wisdom
Anthropic built a microscope that peers into how their AI model actually thinks, and what they found wasn't a tidy flowchart but a tangle of competing interpretations that the model wrestles with before landing on an answer. That might sound unsettling. I think it's hopeful.
The entire AI debate has gotten stuck in two ditches. One is the tech optimists who say "trust us, it works." The other is the regulation enthusiasts who want checklists and documentation that nobody reads. Neither of them addresses the only thing that actually matters: can we understand what's happening inside the machine?
What Anthropic is doing with their Jacobian lens is something I recognize from every company I've built. You don't measure success by asking whether the product works today. You measure it by understanding why it works, so you know what happens when conditions change. Being able to observe without fully explaining is not a failure. That is exactly how science has always begun. First you see the pattern, then you understand the mechanism.
The company that truly cracks interpretability won't just win the trust of regulators. It will win the trust of customers. And in an industry where everyone is chasing the next capability record, more people should realize that the most valuable thing you can build isn't a smarter model, but a model whose smartness you can explain.
Johan Staël von Holstein
Serial entrepreneur · wakopa.ai