Martin Seeger

Surveillance Considered Harmful There is an interesting detail emerging from the attempted terrorist attack on Flydubai Flight 1073. According to current reports, the co-pilot had already been barred from flying by Oman Air because of concerns about extremist views. Yet somehow he ended up flying for Flydubai. And somehow an Omani citizen, from a country whose pilots are not allowed to fly into Israel, ended up in the cockpit of a flight to Tel Aviv (Source: https://apnews.com/article/cf64a228799e6e788e834cac2a489ab2) This happened in a region containing some of the most heavily surveilled societies on the planet. Saudi Arabia, where the aircraft eventually landed, is a particularly extreme example: internet activity, social media and private communications are subject to extensive state surveillance (Source: https://freedomhouse.org/country/saudi-arabia/freedom-net/2025) Other Gulf states are not far behind in the scale and reach of their surveillance. And yet here we are. This points to a more fundamental problem with surveillance as a weapon against terrorism. What does a terrorist actually want? Physical destruction is only the means. The real target is psychological: fear, distrust, disruption and ultimately the fragmentation of society. A terrorist cannot destroy a society through physical violence alone. But it may be sufficient to make that society afraid enough to start damaging itself. And this is where mass surveillance becomes dangerously counterproductive. A functioning society depends on trust: trust between citizens, trust in institutions, and the assumption that most people around us are not our enemies. Terrorism attacks exactly that trust. Mass surveillance does, too. It tells everyone that everyone is potentially suspicious. It normalizes observation, suspicion and self-censorship. In trying to make society more resilient against individual terrorists, we risk weakening the very social cohesion that makes a society resilient in the first place. The irony of Flight 1073 is therefore hard to miss. We have built extraordinarily powerful machinery for watching millions of ordinary people, while apparently failing at the much more mundane task of asking why a pilot previously considered an extremism risk was sitting in the cockpit of an airliner bound for Israel. Surveillance considered harmful. Not merely because it invades privacy. But because it can fail even at the trivial things while quietly weakening the society it was supposed to protect.

@randahl@mastodon.social One death from plague gets the same or even more press coverage than thousands of deaths from Ebola in Congo. The situation in Congo is a lot more dangerous for everyone.

This video makes an interesting point about IPOs: an IPO finally makes it possible to bet against a company. Which, depending on the company, may be the most exciting feature of going public. The video also references the chart below. No new technology has ever become cheaper this quickly than AI. The chart comes from this article.

Discovered at IKEA... cc @catsalad@infosec.exchange

Genau das was ich erwartet hatte: mit der Olympia-Bewerrbung von #kiel wird jetzt Geld gefodert. Ich bin für Sportförderung, aber für Leistungssport würde ich keinen Cent ausgeben. Ich will keine Medallien, keine Weltmeister-Titel, keine Olympiade. Das sind Dinge für das Ego. Ich will Sportplätze und -hallen in der Breite. Ich will das Freibäder erhalten bleiben und in Schulen der Sportunterricht nicht ausfällt. Artikel: https://www.kn-online.de/sport/regional/verbaende-fordern-geld-so-geht-es-in-kiel-und-sh-mit-der-olympia-bewerbung-weiter-NVGRMERCVZA5BKVYF5LJDERIZE.html

@homawida@troet.cafe In einer sicheren Umgebung hinter einer Firewall kann es noch lange Zeit gut gehen, wenn man konsequent nur Software nutzt, die noch Updates bekommt (z.B. einen anderen Browser). Ich mache das eher ungerne. Aber gerade für den M1 gibt es inzwischen ein Linux, dass dann noch weiter Updates bekommt. Zweites Leben mit dem zweiten OS?

Mein Plan ist ein solches Teil bei der älteren Verwandschaft am WLAN anzulernen und einfach dort hinlegen. Bei Problemen kann ich dann deren Bildschirm sehen und ihnen besser helfen.

@kahomono@infosec.space Correct: and compared to 1940, compute deprecated a trillion times. I checked like this: The Mark I cost about $200,000 to build, roughly $4.5 million in 2026 dollars. It could perform an addition in about 0.3 seconds and a multiplication in roughly 1 second. By contrast, a single NVIDIA H100 provides about 34 trillion FP64 operations/s, or much more at reduced precision. In September 2026, H100 rental prices are around $3.39/GPU-hour median, with the cheapest on-demand offerings around $1.30/hour. Comparing both I come up with a cost factor of 5 trillion (and by giving the Mark I 15 years of ops, I am very generous).

@catsalad@infosec.exchange My mind went like this (AI generated):

@publictorsten@mastodon.social Hat die Statistik wissenschaftliche Präzision? Nein. Aber sie erfasst immer noch die Entwicklung der Kosten für die Endanwender halbwegs korrekt. Spekulantentum hatten wir auch von der Eisenbahn bis dot.com. In den 2000'ern wurde den Kunden im Sinne des Wachstums alles nachgeworfen. Es sind die Kosten für den Nutzer die betrachtet werden und daher halte ich sie für valide. Ob es alles transformative Technologien sind, lasse ich mal dahingestellt. Genau die Formulierung habe ich mir bewusst nicht zu eigen gemacht und von "neuen Technologien" gesprochen.

Ich hatte den #MIcrosoft #Defender deinstalliert. Heute bekomme ich eine "monatliche Sicherheitszusammenfassung":

@isotopp@infosec.exchange @LeelaTorres@ieji.de @kahomono@infosec.space Such cost improvements by shady business tactics are nothing new during such a phase. This is not OK and I am not condoning it, but it also does not make the statistic invalid.

I will treat generative AI for music, images, and video far more restrictively. I intend to generate such works only extremely sparingly, use them only in private settings, and label them clearly as AI-generated or AI-assisted. I will not seek commercial benefit from them. This is a personal boundary, not a decree issued from a conveniently available moral mountain. It reflects my uncertainty about training sources, consent, attribution, compensation, output rights, and the effect on human creative work. Where I cannot resolve those questions to my satisfaction, restraint is the honest choice available to me. 33/34

@hafensophie@norden.social Gute Besserung!

I can now turn this analysis into personal conclusions. The first level is political. Binding regulation of frontier-model training must come. Not because every new model is illegitimate, but because the present race allows individual actors to commit energy, water, hardware, data, and financial resources while shifting part of the risk onto everyone else. Voluntary promises cannot resolve an incentive trap. A runner in the race is not the ideal person to operate the starting pistol, judge the lanes, and decide whether the track is safe. We need disclosure of resources and incidents, environmental accounting, auditable obligations, enforceable thresholds, and international coordination where risks cross borders. Otherwise, a competition among a few companies and states may land on the entire world’s feet. 23/34

Damit gibt es jetzt ein IP-KVM für iPhone, iPad und Mac via USB-C. Die schwere Geburt war das Device einmal Flashen und der erste Startup funktionierte nur bei einer Stromversorgung nicht von einem PC. Danach tat es wie erwartet.

YouTube should closely monitor the use of this video clip: https://youtu.be/UOYi4NzxlhE?t=30 Once the access figures rise, financially you should run for the hill.

@Epic_Null@infosec.exchange Exactly... And now guess which song I am referring to. P.S. I strongly recommend to watch the movie. It covers ground zero of the 2007/8 financial crisis.

@reflex@retrogaming.social AI is escalating the game, but I have seen hundreds of billions going down the drain. Worldcom alone obliterated ~100 billion US$. Global Crossing did the same for ~50 billion US$. Our memory has the habit of smoothing things over.

My third level is personal action. My primary use of LLMs will remain rubber-ducking and code generation. The traditional rubber duck never argues, which is both a feature and a limitation. An LLM can help me externalize a thought, expose assumptions, explore objections, and turn an intuition into something I can inspect. It can also be confidently wrong, a quality for which humanity hardly needed outside help. In coding, it can draft implementations, tests, explanations, and alternatives at high speed. In both uses, I remain the reviewer and carry responsibility. I do not want to outsource judgment. I want better material on which to exercise it. 31/34

I want to acquire more skill in harnessing AI: supplying the right context, connecting tools, structuring memory, decomposing work, building evaluations, and verifying results. That matters more to me than collecting clever prompts as if they were magic spells. The model is one component. The harness determines whether it can participate in a reliable process. I will also continue testing and using local LLMs for personal work. They give me practical knowledge of what is possible without a permanent service dependency and where local capability still fails. Local is not automatically better, and owning the hardware does not confer wisdom. It does help me maintain competence, privacy, and an exit option. 32/34

The death of companies is not the death of a technology. The dot-com crash did not kill the web. Netscape’s disappearance did not make the browser irrelevant. A company obituary is not a protocol obituary. A consolidation in AI could destroy capital, jobs, and individual firms while leaving useful models, methods, infrastructure, and knowledge behind. That does not make the damage harmless. It only means we should not confuse the prospects of current vendors with those of the underlying technology. AI can be overfinanced, overpromised, and still remain consequential. 29/34

@forthy42@mastodon.net2o.de I thought about differentiating about different laws as well, but the thread had 34 posts already. So I tried to cover it by mentioning it, there are different perspectives in different countries.

Freedom of choice must be part of AI policy. AI as a service and open-weight models both have a legitimate role. Regulation and public procurement should preserve the ability to choose between them: portability of data and evaluations, interoperable interfaces, the right to run models locally, and credible exit paths from providers. Freedom of choice also includes the option not to use AI where it adds no value, and access to a human route where an essential service is involved. Convenience today must not quietly become dependency tomorrow. Exit doors are best installed before anybody smells smoke. 25/34

Policy and procurement should prefer the smallest model that can reliably perform a task. That is not a ban on general-purpose or frontier models. Some problems may justify them. But “larger” should not be accepted as a synonym for “better.” Not every nail requires an orbital hammer. Smaller, specialized models can reduce computation, memory, hardware, and energy requirements in both deployment and further development. They can also make systems easier to test and replace. My preference is simple: use scale where its additional value can be demonstrated, not merely where it can be financed. Speculative future capability should not automatically outweigh present ecological cost. 24/34

The application layer begins with a boundary: I reject AI as a compulsory ritual. An employee should not be forced to use AI irrespective of whether it helps the task. A patient should not lose access to a doctor because an AI system has become the only way to obtain an appointment. AI may be an option, an assistant, or even the best default in a defined setting. In essential services especially, there must still be an accessible human alternative, a path to appeal, and a person or institution that remains accountable. Nothing makes me doubt the usefulness of a tool like making it mandatory. Adoption figures produced by coercion tell us little. A technology meant to serve people must leave them meaningful agency. 19/34

We also need labor-law frameworks for AI adoption. Employees should know where AI is used to evaluate, monitor, direct, or replace parts of their work. They need consultation, training, ways to contest consequential decisions, and protection against surveillance and unreasonable work intensification. Mandatory use should face the same scrutiny as other material changes to working conditions. The efficiency fairy rarely leaves a spare afternoon under the pillow. More often, employees inherit verification work, liability pressure, and higher quotas. Structural change at work must be negotiated with the people whose work is being changed. 26/34

Schade für #kiel: https://www.kn-online.de/lokales/kiel/kiel-kult-laden-fantasyreich-schliesst-nach-39-jahren-friseur-uebernimmt-W6JHQYPQNVG4RPBEKARVFAIZLM.html

@katzenjens@social.tchncs.de +1 für lissy

My rough map is that strong open-weight models have entered territory that was occupied by frontier services in 2025 on many tasks. That is practical shorthand, not a claim of universal equivalence. Benchmarks are maps: useful, compact, and perfectly capable of leaving out the swamp. They compress important differences in modalities, long context, tool use, reliability, safety, and specialized knowledge. Mozilla reports only a small aggregate gap in its evaluated benchmarks, while a specific use case can still show a large one. Services have not become obsolete. Frontier services and open-weight models both have a place. The useful question is which capability, control, and cost profile the actual task requires. 17/34

The legal status of AI output is not one question, although it is often placed in one convenient box marked “legal stuff.” Can the output itself receive copyright protection? Who owns the human contribution? Does the result infringe an existing work? And what rights do provider terms actually grant? These questions differ by jurisdiction and by the degree of human authorship. GEMA’s case against Suno illustrates one part of the problem: GEMA alleges that Suno produced tracks confusingly similar to protected songs. That is a claim about possible infringement, not simply about whether an AI output can have an owner. Until the boundaries are clearer, commercial use requires provenance, license checks, human review, and a realistic allocation of risk. The word “generated” is not a legal invisibility cloak. 21/34

The law must also establish clear liability. AI must not become a machine for laundering responsibility. An organization remains accountable for choosing, configuring, and deploying a system and for acting on its output. Providers remain accountable for their own defects, misrepresentations, and legal obligations. The precise allocation depends on context, but it cannot be passed indefinitely between model, vendor, deployer, and user. When AI affects healthcare, employment, credit, justice, or public administration, people need an identifiable decision-maker, an appeal route, and a party capable of providing a remedy. 27/34

The economic impetus will push inference toward wider use, labor savings, recurring services, and dependency. That pressure is real, but it is not the same as truth. A profitable deployment is not necessarily useful. Rapid adoption is not evidence of consent. A lower price does not contain every ecological or social cost. What a market rewards depends on who owns the infrastructure, who can refuse, and who may externalize risk. Markets are quite good at producing prices. I remain unconvinced that this makes them experts in metaphysics. I will not judge an application by whether the market makes it appear inevitable. I will return to my criteria: ecological viability, freedom from dependency, democratic moderation of structural change, and whether the technology serves human beings. That is why I separate inference technology from its application. 22/34

My main technological requirement is optionality. An API can change its price, behavior, terms, or availability. A local model brings its own costs: hardware, operations, updates, and dependence on upstream model releases. Absolute independence is an illusion, but avoidable lock-in is still a choice. I can only advise people and organizations to test their real use cases against local models. Not a generic benchmark, but their documents, workflows, quality thresholds, latency, privacy needs, and failure modes. Keep data and evaluations portable. Know where a frontier service genuinely earns its premium. Know which critical functions still work if that service disappears. An exit plan is terribly boring, right up to the day it becomes horribly contemporary. 18/34

Efficiency gains immediately raise a distribution question: where do they go? A productivity gain does not arrive with a moral destination label. If an hour saved becomes better service, higher pay, more autonomy, safer work, or actual free time, AI can improve human work. If it becomes a higher quota, fewer colleagues, closer monitoring, and the same pay, efficiency becomes work intensification. Both are real possibilities. The OECD reports substantial worker-perceived benefits from AI, but also concerns about work intensity, data collection, and inequality. The ILO finds real but uneven productivity gains that have not automatically translated into higher measured output, earnings, or employment. Who captures the gain and who carries the new burden are not implementation details. Workers and the public must have a voice in that decision. 20/34

My second level is what I expect to happen. I expect a consolidation phase comparable to the dot-com correction around 2001. This is an analogy, not a claim that the timeline or outcome will repeat exactly. Too many companies currently depend on continuing capital, falling compute costs, expanding demand, and a future path to margins, all at the same time. Those assumptions will not hold equally for everyone. Some firms will disappear, merge, or be absorbed. Infrastructure may be repriced. Today’s apparent leaders may not remain the leaders. I do not know when this correction comes. The future has been oddly reluctant to share its calendar. I would be more surprised if it never came. 28/34

I expect the technical focus to move from sheer scale toward optimization: better architectures, smaller and specialized models, quantization, routing, caching, harnesses, and hardware designed around actual workloads. That expectation does not make the sector investable for me. I will invest exactly €0 in it. Valuations, circular financing, uncertain moats, political intervention, infrastructure risk, and unpredictable technical shifts make it impossible for me to distinguish investment from speculation with sufficient confidence. For my risk tolerance, this is currently a table for players. I shall remain near the bar and keep my €0. That describes my limits, not a claim that nobody else can judge the odds differently. 30/34

@thomasfuchs@hachyderm.io I don’t feel confident to state this. First: How to define intelligence? Second: Maybe we are also just statistics? As I wrote in another thread: The AI debate may result in disappointment about our own capabilities.

@benjojo@benjojo.co.uk While you're waiting, why won't you enjoy our torment nexus?

@codered@chaos.social Screen sharing erfordert einiges an Kompetenz auf der anderen Seite. Der Bootstrap ist viel einfacher hier. Teamviewer auf den PC klappt, auf dem iPhone nicht.

@reflex@retrogaming.social It is not my intent to force my argument onto you. But it would be nice, if you would add any numbers one can check and compare. As I pointed out: feelings make bad metrics. One point that may explain our different viewpoints: I hail from Germany while I guess you are from the U.S. The 2007/8 crisis had a lot less impact here than it did in the U.S.

I see enormous optimization potential in both approaches: smaller specialized models, distillation, quantization, caching, batching, routing, better harnesses, and better hardware. I think improvements of two or more orders of magnitude are possible across the inference stack. One analysis projects up to 100× lower cost per token over time; a peer-reviewed energy study identifies 8–20× in foreseeable savings per query. A slower cadence of model generations would make this easier. Stable targets give engineers time to optimize runtimes, quantization, hardware, and serving instead of rebuilding around the next frontier every few months. That should reduce the ecological burden of each useful task substantially. It does not guarantee lower total consumption. Cheaper inference can create vastly more inference, while long reasoning traces can eat some of the gains with excellent table manners. Efficiency gives me reason for qualified optimism. It is not an ecological free pass. 16/34

@reflex@retrogaming.social Telecom and Internet infrastructure: Governments had already funded much of the research and institutional infrastructure behind the Internet, including ARPANET/NSFNET and university networks. In the 1990s, telecom liberalization, tax treatment, municipal incentives, rights-of-way, and in some places direct public funding helped expand infrastructure. On top of that came an extraordinary amount of debt and equity financing for private fiber networks. Companies such as Global Crossing, WorldCom, Qwest, Level 3 and others built enormous long-haul capacity. The result was a fiber glut. After the crash and bankruptcies, later Internet businesses could buy bandwidth at prices that did not remotely reflect the capital cost of constructing those networks. Free or absurdly cheap Internet access: ISPs routinely priced below fully allocated cost to acquire users. AOL famously carpet-bombed America with free CDs and free trial hours. In Europe, the late-1990s “free ISP” model went even further: Internet access itself was nominally free, with the business hoping to live on telecom revenue sharing, advertising, portals, or later monetization. Venture capital and telecom economics absorbed customer-acquisition and operating costs that users did not directly pay.E-commerce shipping and fulfillment: This is probably the cleanest analogue to today’s AI subsidies. Dot-com retailers discovered that customers liked “free shipping” much more than they liked paying the actual cost of picking, packing, warehousing, returns, and last-mile delivery. Investor capital paid the difference. Pets.com is the caricature, but the phenomenon was widespread. Selling a bulky bag of dog food online and shipping it to someone’s house for less than the economic cost was not technological disruption so much as investors temporarily buying dog food delivery for customers.Goods sold below cost: Many e-commerce startups treated gross margin almost as an optional feature. Discounts, coupons, introductory credits, loss-leader pricing and enormous marketing expenditure were justified as “customer acquisition.” IPO and VC money effectively subsidized consumer purchases. Webvan is a particularly good example: customers were buying groceries and delivery while investors were paying for warehouses, automation, delivery infrastructure and much of the operating loss. I had to compete with that subsidized offerings and it really hurt.

@Epic_Null@infosec.exchange No idea... I don't remember where I have watched it.

@padeluun@digitalcourage.social 1992 nutzte er mal meinen Rechner um seine Email zu lesen, da ich schon Internet in DE hatte. Ich war stolz wie Bolle.

@reflex@retrogaming.social But that is not untypical. I was there Gandalf, 3000 years ago when the Internet came around. Been there, done that.

@Lalufu@mastodon.social Es ist deutlich schlimmer als erwartet und ändert sich durch beschweren. Das ist durchaus eine Erkenntnis.

@reflex@retrogaming.social How do you compare size? Equity-market decline: dot.com -78%, 2007/8: -50%US household net-worth effect: dot.com -6%, 2007/8: -28%Wealth destruction: dot.com several trillion dollars, concentrated in equities, 2007/8 >$15 trillion national wealth 2007/8 hit the financial system, dot.com hit equity. Order of magnitude is roughly the same. What worries me with AI is not the amount of money but the circular financing. This will make any crisis systemic.

@toto@fem.social Wenn ich den Defender mit anderen Virenscannern vergleiche, dann ist der Defender IMHO derzeit der Beste. Aber halt immer noch mit Problemen ...