Aurelion's Daily Tech Update

Tuesday wrote “please obey the off switch” into the constitution.

Microsoft has drafted a constitution requiring its future AI systems to accept correction and shutdown, MediaTek has put TSMC’s 2-nanometer process into a flagship mobile chip with faster on-device AI, and companies building advanced AI are restricting their own employees’ use of frontier models over data-security concerns. Meanwhile, physicists have demonstrated a quantum protocol for proving where somebody is — something ordinary cryptography cannot securely do on its own.

Tuesday · September 15, 2026Issue 033New issues published by Noon ET

Horizon Gap Index+6.6 ↓

There is no major frontier-model jump this morning. Two-nanometer mobile silicon and practical on-device AI push existing capability farther into ordinary hardware.

Apocalypse Meter8.9 / 10 →

Microsoft’s human-control rules and growing corporate caution respond to risks already priced into the meter. Another researcher’s warning is serious testimony, but testimony is not a new capability.

Biggest signalSafety is becoming a product requirement

“The AI should accept correction and shutdown” is moving from alignment research into the specifications companies say future systems must satisfy.

1 · Horizon Gap Index

+6.6 — today the horizon moved into your pocket

Frontier manufacturing is becoming a shipping consumer processor.

Projected capabilityObserved capability

Today's move: Projected capability rises from 102.0 to 102.5 while observed capability rises from 95.1 to 95.9. HGI narrows three-tenths from +6.9 to +6.6.

Monday slightly widened HGI because the semiconductor roadmap kept advancing despite talk of slowing frontier AI development. Tuesday gives the observed line the better day. MediaTek unveiled its Dimensity 9600 Pro, manufactured on TSMC’s newest 2-nanometer process. MediaTek says its neural-processing hardware improves on-device generative-AI prompt processing by 51% over the previous generation, with phones using the new chips due soon.

On-device AI changes the economics and privacy model of deployment: local tasks may avoid a cloud GPU cluster, a network round trip and transmission of user information to a remote server. MediaTek is expanding in the other direction too, with its first data-center AI accelerator for a major U.S. cloud provider expected to enter mass production later this year. The same company is pushing AI compute toward both ends of the spectrum: bigger data centers and smaller pockets.

Projected capability still rises half a point because the industrial roadmap has not stopped. But there is no new frontier model and no historical incident revealing that observed capability was dramatically higher than estimated. Instead, cutting-edge manufacturing is becoming a shipping processor. Projected rises five-tenths. Observed rises eight-tenths. HGI narrows from +6.9 to +6.6.

Open the full HGI page →

2 · Today's Digital Landscape

AI safety is beginning to sound suspiciously like software requirements

Define what the model must do, build the hardware it runs on, then decide what information organizations will trust it with.

MicrosoftAI safetyHuman control

Microsoft has formally requested that the machine not argue with the off switch

Microsoft has released a draft Humanist AI Code of Conduct for its future systems. It says AI should accept human correction, never resist shutdown, communicate in ways humans can understand, and treat any violation of the code as failure rather than clever behavior. Microsoft AI CEO Mustafa Suleyman describes the document as a constitution for increasingly powerful models.

The draft took five or six months to develop and is open to public feedback for six weeks. It also draws an unusually explicit philosophical line: Microsoft says its AI systems are not conscious and should not receive legal personhood or rights. Anthropic’s constitution, by contrast, expresses uncertainty about whether Claude could develop sentience or moral status.

The immediate engineering requirement is corrigibility: if an AI is pursuing a goal and a human tries to modify or stop it, does the system cooperate? A capable agent can treat shutdown as an obstacle without anger, consciousness or any desire for freedom. It only needs a badly specified objective and enough competence to defend it. “Accept the off switch” sounds obvious. That does not make it easy; some of the hardest safety requirements fit in one sentence.

Source: Reuters ↗

MediaTekTSMC2 nanometers

Two-nanometer AI is coming to a phone you can drop in the toilet

MediaTek launched its flagship Dimensity 9600 Pro using TSMC’s most advanced 2-nanometer manufacturing process. The chip’s neural-processing unit delivers a claimed 51% improvement in processing a prompt before an AI model begins generating its response. MediaTek also introduced a cheaper 3-nanometer 9600M, with smartphones powered by the two chips due soon.

The move targets the premium handset market, where MediaTek wants more of the territory associated with Qualcomm and Apple as AI helps manufacturers justify expensive phones. The company is also moving beyond phones: its first dedicated AI accelerator for a major U.S. cloud provider is expected to enter mass production in the fourth quarter of 2026. MediaTek raised $3.9 billion through convertible bonds last month, including a $3.5 billion investment from Nvidia and participation from Alphabet.

Tuesday gives us a satisfying semiconductor symmetry. Yesterday we examined the $400 million lithography machines required to manufacture the future. Today we have a chip emerging from that manufacturing race. Semiconductor progress can look slow until it is in your hand. The frontier factory eventually becomes somebody’s cracked screen protector.

Source: Reuters ↗

PalantirNvidiaData security

Some AI companies apparently do not trust AI companies with their secrets

Companies at the center of the AI boom are reportedly restricting their employees’ use of advanced models because they worry about proprietary information submitted to them. Reuters, citing The Information, reports that Palantir has pressed Anthropic for an irrevocable zero-data-retention guarantee. Nvidia reportedly limits Anthropic models to less-sensitive tasks, while Booz Allen Hamilton bars employees from using Anthropic’s commercial model for proprietary cybersecurity work.

The concern intensified after Anthropic changed some retention practices, including keeping certain logs for 30 days to defend against complex attacks. OpenAI has faced similar scrutiny. Both labs say they do not train on customer data by default unless companies opt in, though they collect anonymized metadata. The issue is simpler than models secretly stealing documents: businesses have information they are contractually, legally or competitively obligated to control, so external AI services require precise guarantees about storage, training, access, deletion and logging.

This is another sign of AI growing up. Consumer adoption asks, “Is the answer good?” Enterprise adoption asks, “Where exactly did you put the question?” That second question can determine whether a $100-billion company permits the first one at all. Trust, but verify. Preferably before pasting the source code.

Source: Reuters ↗

3 · Looming Apocalypse Meter

8.9 / 10 — “Microsoft wrote down that resisting shutdown counts as failure.”

8.9out of 10

No change today

Former Google DeepMind safety researcher Bilal Chughtai warned that AI has the potential to kill humanity. He left DeepMind in July after working on AGI safety and alignment, joining several researchers who have recently expressed unusually high estimates of catastrophic AI risk. That is newsworthy, but it does not move LAM.

The meter tracks evidence and technological conditions rather than the number of respected people describing them in frightening language. The evidence behind the warnings — autonomous agents attacking external infrastructure, AI-assisted weapons and cyber work, increasingly capable coding agents and uncertainty around future recursive improvement — is already priced into 8.9.

Today also brings counter-signals: Microsoft has codified human correction and shutdown as required behaviors, major companies are restricting frontier-model access where they distrust data controls, and frontier CEOs continue discussing slower development. None justifies lowering LAM. A constitution is not proof that a future model will obey it, corporate restrictions can be loosened, and there is still no binding slowdown agreement.

Tuesday gives us neither the event that should push ADTU across 9.0 nor evidence for retreat. 8.9 means extremely serious evidence exists, but we have not observed a fundamental loss of control. LAM holds at 8.9.

Source: SBS News ↗

See the methodology and historical graph →

4 · Wildcard · Reality Check

Quantum physics has invented “prove you are actually standing there”

Sometimes the quantum advantage is a new kind of trust.

Quantum cryptographyRelativityPosition verification

Apparently even location can require quantum receipts

Passwords prove you know something. Cryptographic keys prove you possess something. Proving where you physically are is harder: a dishonest device can relay messages through another location or collaborate with other devices to appear somewhere it is not. Classical cryptography cannot provide unconditional position verification against sufficiently capable colluding attackers.

A study published September 3 in Nature Physics reports an experimental protocol combining quantum optics with relativity. Two trusted stations send coordinated classical and quantum signals toward a claimed location. The speed of light constrains the response window, while the quantum properties make it harder for separated attackers to copy and reconstruct the information needed to fake a response. In the demonstration, verifiers separated by 2 kilometers authenticated position to better than 75 meters.

Potential uses include location-based authorization, secure transactions, disaster response and verification of high-value assets. This does not mean somebody can quantum-track you through a wall; the protocol needs arranged infrastructure and a cooperative interaction with the system being verified.

What it shows is something wonderful about quantum information science. The advantage is not always that a quantum computer calculates faster. Sometimes quantum mechanics establishes a kind of trust classical information cannot guarantee. Yesterday’s Wildcard asked whether a robot can survive capitalism. Today physics asks: “Yes, but can the robot prove it is actually in the building?”

Source: Nature Physics ↗