Alexander Klinkner

School psychologist and consultant, supporting energy transformation, Tesla, SpaceX, and an open mind here since 2009.

[Keulpaka Institute  🇺🇦 #NAFO @DrHOSP1]

1. Abschaffung der EU-Roaminggebühren
2. Europaweite Notrufnummer 112
3. Fluggastrechte bei Verspätungen
4. Entschädigungen bei Flugausfällen
5. Fahrgastrechte im Bahnverkehr
6. Fahrgastrechte im Fernbusverkehr
7. Fahrgastrechte im Schiffsverkehr
8. Schutz bei Pauschalreisen
9. Insolvenzschutz bei Pauschalreisen
10. Europäische Krankenversicherungskarte
11. Gegenseitige Anerkennung von Führerscheinen
12. EU-weit gültiger Behindertenparkausweis
13. Vereinfachte grenzüberschreitende Mautzahlung
14. Einheitliche Fahrzeugsicherheitsstandards
15. Verpflichtende Sicherheitsassistenzsysteme
16. Einheitlicher USB-C-Ladeanschluss
17. USB-C-Pflicht für neue Laptops
18. Rufnummernmitnahme beim Anbieterwechsel
19. Transparenz bei Mobilfunkverträgen
20. Preisobergrenzen für EU-Auslandsgespräche
21. Netzneutralität
22. Grenzüberschreitende Streaming-Portabilität
23. Verbot ungerechtfertigten Geoblockings
24. Gewährleistungsrechte für digitale Produkte
25. Schutz beim Kauf digitaler Dienstleistungen
26. Recht auf Datenübertragbarkeit
27. Recht auf Löschung personenbezogener Daten
28. Transparenzpflichten beim Datenschutz
29. Mehr Transparenz auf Online-Marktplätzen
30. Einfachere Meldung rechtswidriger Online-Inhalte
31. Mindestens zwei Jahre gesetzliche Gewährleistung
32. 14-tägiges Widerrufsrecht bei Onlinekäufen
33. Transparente Preisangaben
34. Verpflichtende Grundpreisangaben
35. Schutz vor irreführender Werbung
36. Schutz vor unfairen Vertragsklauseln
37. Verbot versteckter Online-Zusatzkosten
38. Verbot voreingestellter kostenpflichtiger Extras
39. Transparenz bei personalisierten Preisen
40. Schutz vor gefälschten Kundenbewertungen
41. EU-Schnellwarnsystem für gefährliche Produkte
42. Einheitliche Spielzeugsicherheitsstandards
43. Verbesserte allgemeine Produktsicherheit
44. Grenzüberschreitender Verbraucherschutz
45. Europäisches Verfahren für geringfügige Forderungen
46. Einheitlicher SEPA-Zahlungsverkehr
47. Grenzüberschreitende SEPA-Lastschriften
48. Gleiche Gebühren für vergleichbare Euro-Zahlungen
49. EU-weite Echtzeitüberweisungen
50. Gebührenbegrenzung bei Echtzeitüberweisungen
51. IBAN-Empfängerabgleich gegen Überweisungsbetrug
52. Begrenzte Verbraucherhaftung bei Kartenmissbrauch
53. Stärkere Authentifizierung beim Onlinebanking
54. Mehr Wettbewerb bei Zahlungsdiensten
55. Vergleichbare Bankkontogebühren
56. Recht auf ein Basiskonto
57. Verpflichtende Allergenkennzeichnung
58. Einheitliche Nährwertkennzeichnung
59. Gemeinsame Arzneimittelsicherheitsstandards
60. Anerkennung von Berufsqualifikationen
61. Grenzüberschreitende Sozialversicherungskoordination
62. Erleichterte grenzüberschreitende Gesundheitsversorgung
63. Verbesserte Rechte auf Reparatur
64. Verlängerte Haftungsfrist nach Reparatur
65. Vereinfachte Nachhaltigkeitsberichtspflichten
66. Bürokratieentlastung für kleinere Importeure
67. Vereinfachungen in der EU-Agrarpolitik
68. Vereinfachte InvestEU-Förderverfahren

[Alexander Klinkner @AlexKlinkner]

‼️Artikel: Real und wissenschaftlich interessant (Glymphatik-Modulation via Neuromodulation ist ein aktives Forschungsfeld, ähnlich wie die 40-Hz-Licht-Studien).
🚫Post: Nicht fake im Sinne von erfunden, aber stark übertrieben und irreführend in der Darstellung. Es ist klassische Gesundheits-Influencer-Rhetorik, die vorsichtige Mausdaten zu einem dramatischen Durchbruch für den Menschen aufbläst.
⬆️Es gibt derzeit keine Belege dafür, dass man damit beim Menschen Alzheimer-Plaques „wegspülen“ oder kognitive Defizite wesentlich umkehren könnte. Weitere Studien (größere Tiere, dann Menschen) wären nötig.

[Alexander Klinkner @AlexKlinkner]

the essay is a coherent extrapolation of one extreme premise under assumptions of weak human agency, fragile institutions, and AI systems that behave like independent power-seeking agents. Those assumptions are contestable. The result is closer to a vivid warning about risks of disempowerment than to a description of definite destiny. Political choices about development, deployment, ownership, and control remain open variables—precisely the point Conti emphasizes in other writing.

[Robin Engelhardt @Elektro_Robin]

Die Hölle friert zu: Ich böser Hater lobe Tesla. Die bauen nämlich verdammt wertstabile Autos.

[Alexander Klinkner @AlexKlinkner]

Die 129.240 € sind die Bruttokosten der BA im „besten“ Fall (Steuerklasse III mit steuerlich berücksichtigtem Kind (67 %)

Netto bekommt er etwa 65–70 % seines letzten Nettos.

Gut 30 % der „Summe“ sind Beiträge, die vor allem Lücken in seiner eigenen Versorgung stopfen. Er bezieht eine Versicherungsleistung, für die er jahrzehntelang den Höchstbeitrag gezahlt hat.

Rentenbeitrag 1.257 €: Das ist exakt der Höchstbetrag. Die Agentur zahlt RV-Beiträge auf 80 % des Bemessungsentgelts (§ 166 SGB VI), also 0,8 × 8.450 € (BBG 2026) × 18,6 % = 1.257,36 €. Das Geld geht an die Rentenkasse, nicht an ihn. Es bringt ihm etwa 0,8 Entgeltpunkte pro Jahr statt rund 1,0 vorher. Für ihn ist das ein Rentenverlust, kein Bonus.
•PKV-Zuschuss 508 €: Das ist plausibel. Nach § 174 SGB III übernimmt die Agentur die tatsächliche PKV/PPV-Prämie bis zur Höhe dessen, was sie für einen GKV-Versicherten zahlen würde (rund 1.200 €). Ab 55 gibt es keinen Rückweg in die GKV mehr (§ 6 Abs. 3a SGB V), also bleibt er in der PKV.
•ALG 3.620 €: Das ist der heikle Punkt. Rechenweg: Bemessungsentgelt (gedeckelt bei 8.450 €) minus 20 % Sozialversicherungspauschale, minus fiktive Lohnsteuer und Soli, davon 60 % bzw. 67 % mit Kind.
•Steuerklasse I ohne Kind: grob 2.700 €
•Steuerklasse III ohne Kind: ca. 3.200 €
•Steuerklasse III mit steuerlich berücksichtigtem Kind (67 %): ca. 3.500–3.600 €
3.620 € passt also nur zur Konstellation verheiratet, Klasse III, Kind und Gehalt an der BBG. Das kommt vor, ist aber nicht der Normalfall für „~100k Gehalt“.

Rentenfalle: Mit 62 oder 63 ist er nicht beim „offiziellen“ Renteneintritt, sondern bei der Rente für langjährig Versicherte mit bis zu 14,4 % Abschlag, und zwar dauerhaft. Für die abschlagsfreie Rente nach 45 Jahren zählt ALG-Bezug in den letzten zwei Jahren vor Rentenbeginn nicht mit (§ 51 Abs. 3a SGB VI). Genau dieses Modell kann ihm also den abschlagsfreien Zugang verbauen.

[Alexander Klinkner @AlexKlinkner]

The documentary’s emphasis on promotional skill, founder-credit disputes, and optimistic timelines is accurate as far as it goes. What it underweights is the observable filter that capital markets have applied over two decades: resources have kept flowing because enough investors (Baron among them) judged that the combination of mission intensity, execution under pressure, and resulting products produced higher expected returns than the available alternatives. The ongoing high valuations of both Tesla (~$1.5 trillion) and SpaceX (post-IPO above $2 trillion) reflect that continuing judgment rather than a twenty-year magic trick that somehow never had to show results.

[Alexander Klinkner @AlexKlinkner]

The documentary’s emphasis on promotional skill, founder-credit disputes, and optimistic timelines is accurate as far as it goes. What it underweights is the observable filter that capital markets have applied over two decades: resources have kept flowing because enough investors (Baron among them) judged that the combination of mission intensity, execution under pressure, and resulting products produced higher expected returns than the available alternatives. The ongoing high valuations of both Tesla and SpaceX reflect that continuing judgment rather than a twenty-year magic trick that somehow never had to show results.

[Gwynne Shotwell @Gwynne_Shotwell]

I have been a posting slug!

How about that launch triple header last week: Starship, Crew 13, and Heavy!!!
Terrestrial low band spectrum acquisition and FCC approval for Starlink Mobile constellation!
And I am always relieved, grateful and proud when we complete a crew mission— welcome home to Mother Earth Crew 12!

[DogeDesigner @cb_doge]

BREAKING: Germany’s transport minister pushes for Tesla FSD approval across Europe.

Speaking today at a Tagesspiegel Background conference in Berlin, Steffen Bilger said:

• He will push for Tesla FSD Supervised to reach drivers across the EU soon.

• He sees potential for FSD to make roads safer and wants to support innovation.

• Germany should explore its own approval route if EU talks fail.

• He will work to ensure that this system can be used throughout Europe soon.

📸 Lena Ganssmann / Tagesspiegel

[Alexander Klinkner @AlexKlinkner]

Der wissenschaftliche Hintergrund des Interviews ist das Working Paper „What if automating AI R&D triggers an intelligence explosion?“ (CASP/GovAI, Sept. 2026) mit 22 Autor:innen, darunter Geoffrey Hinton, Yoshua Bengio, Andrew Barto, Jakub Pachocki, Eric Horvitz, Jack Clark und Christoph Winter.

Die zentrale Beobachtung: KI übernimmt bereits einen wachsenden Anteil der KI-Forschung. Bei Anthropic stieg der Anteil der R&D-Arbeit mit nur grober menschlicher Aufsicht von 1 % im März auf 26 % im August 2026. OpenAI berichtet zugleich von Systemen, die R&D-Aufgaben erledigen, für die Menschen mehrere Tage benötigen würden.

Das Paper untersucht, ob daraus eine Rückkopplungsschleife entstehen könnte: KI hilft, bessere KI zu entwickeln, die wiederum die AI-R&D weiter automatisiert – eine mögliche „intelligence explosion“, bei der sich der technologische Fortschritt drastisch beschleunigt.

Die Autoren fordern deshalb mehr Transparenz über die Automatisierung von AI-R&D, Möglichkeiten zur Steuerung einer solchen Beschleunigung und gesellschaftliche Vorbereitung auf ihre möglichen Folgen.

Paper: https://t.co/bIlvMeCQCK

[Alexander Klinkner @AlexKlinkner]

AI can now propose closed-form plasma equilibria that a specialist can verify, and exact solutions of this kind will tighten the test suite for stellarator codes. That is a real change in how theory gets done. It does not shorten the path from a physics experiment to a sellable kilowatt-hour produced by fusion energy.

[Jason Oppenheim @OppenheimJason]

A public service announcement.

[Geoffrey Hinton @geoffreyhinton]

The idea of an intelligence explosion caused by recursive self improvement has been around for a long time but until very recently it did not seem imminent. Now many leading researchers think it may happen quite soon. You can read our paper about it here:
https://t.co/sgUpugjpRY

[Alexander Klinkner @AlexKlinkner]

The shared interpretation in the coverage is

‼️the arrangement ranks current access and alignment with the administration’s preference for speed and self-regulation over safety-driven slowdowns

⬆️compute and infrastructure (Nvidia, Musk’s companies) and the large platforms sit next to power,

⬇️while the labs most associated with existential-risk arguments and heavier oversight sit at the margins.

‼️That is the “cards dealt” reading Cook is inviting.

[Jaan of EVwire.com ⚡ @TheEVuniverse]

Florian Reiners, the CEO & co-founder of Wiferion that Tesla acquired in 2023, says his time at Tesla has come to an end and opens up a bit of the background.

He says they had a few buyers to choose from and he chose Tesla because he got to be a real engineer again:

"Today was my last day at Tesla.

Three years ago, we sold Wiferion, the company we had built from scratch, to Tesla. Today I'm closing that chapter. I'm taking nothing with me but good memories and a lot of experience. Here are my reflections:

We really got lucky with this acquisition

We had a few buyers to choose from. With most of them, my role after the merger would have come down to tuning P&Ls. At Tesla, I got to be a real engineer again and support the team in solving hard technical problems. I'm deeply thankful for that. Many founders burn out after an acquisition. We didn't. We found people who shared our sense of purpose, a culture that felt like home from the start, and almost no politics. Looking back, I'm still a little amazed at how well it went.

Ownership like nowhere else

The work is demanding and often hard. But the ownership is huge. Where other companies put a whole team, Tesla puts a single engineer. Every engineer makes a real contribution to the final product and owns it end to end, from specification all the way to manufacturing. Still, you're never alone. People support each other, and setbacks and successes are shared. I saw first-hand how much a small, dedicated team can achieve this way. It was one of the most rewarding environments I've ever worked in.

Leadership you can trust on all levels

Full accountability, top to bottom. Leaders go deep, and review meetings can be tough. The people making the calls aren't outsiders with business degrees. They're strong engineers who came up through every stage and earned their battle scars. That's why you trust their guidance. No politics, just building the best product from first principles.
A special thank you to Satish Thuta and AJ Booth for their leadership and the trust they placed in me, and to the whole Tesla team for making this such a joyful ride.

Thank you all so much!

Now it's time to build again."

[Alexander Klinkner @AlexKlinkner]

@Tazerface16 Yeah, it reached orbit…you’re really an expert!

[Alexander Klinkner @AlexKlinkner]

Nein.
Eustella ist eine Wiener App auf chinesischen und amerikanischen Open-Weight-Modellen. Die Gewichte kommen von Alibaba, Google, OpenAI und Mistral. Wien liefert Hosting und Datenschutz-Story.

Europa hat kein eigenes Spitzenmodell. Was die ZEIT hier feiert, ist ein Chatbot aus Wien, der Qwen rechnet und das dann digitale Souveränität nennt.
Server in Frankfurt ändern nicht, wer das Modell trainiert hat.

[Afshine Emrani MD FACC @afshineemrani]

I've been a cardiologist for 25+ years. The last 90 days changed medicine more than the last decade, and almost nobody noticed. Here are the 10 biggest AI breakthroughs, and why you should be excited, not afraid 🧵

1/ 37,000 AI agents. One "virtual biotech." It analyzed thousands of trials, figured out which drug targets actually work in humans, and independently proposed a lung cancer strategy that a major pharma company later pursued. AI isn't a search engine anymore. It's a co-scientist.

2/ The first drug discovered AND designed entirely by AI (Insilico's rentosertib) just entered Phase 3. In patients, it reversed biological age by 3–6 years across six different aging clocks. Read that again!

3/ As a cardiologist, this one gives me chills. The FDA authorized "Queen of Hearts," an AI that reads a standard ECG and catches heart attacks, including the hidden ones doctors miss. 2x the sensitivity. Far fewer false alarms. It pages the cardiology team itself. In a heart attack, minutes are muscle.

4/ EchoNext spots 6 types of hidden structural heart disease from a cheap, 10-second ECG, and it does it better than cardiologists. It already flagged a patient who went on to get a heart transplant. Screening that used to need an echo lab now needs a test any clinic can run.

5/ UpDoc is the first FDA-cleared AI that talks directly to patients. It checks in between visits, adjusts insulin within limits the doctor sets, and documents everything in the chart. Chronic disease no longer has to wait for your next appointment.

6/ An autonomous AI agent called MIRA, with full access to medical records, went head-to-head with ER physicians on hundreds of real cases.
AI: 87.8% correct
Doctors: 78.1%
It ordered the tests, read the results, and wrote the plan.

7/ Google released MedGemma 1.5 and MedASR, open medical AI for imaging and clinical speech that anyone can build on. Every hospital. Every researcher. Every country. Open models mean medicine moves at internet speed.

8/ The FDA authorized Aletta, the first robot that draws blood on its own. It gets the vein on the first stick 94–95% of the time, even hard veins. One technician supervises three robots. The most common procedure in medicine just got automated.

9/ UCLA's SLIViT matches specialists on 3D MRI, CT, ultrasound, and retinal scans, and it's 5,000x faster. It isn't a new model for every organ. One architecture reads them all.

10/ The FDA has now authorized 1,600+ AI medical devices. They cover radiology, cardiology, and surgery, including real-time AI that checks breast cancer margins while the patient is still on the table. This isn't coming. It's here.

What this means to you:
→ Drugs that took 10–15 years are showing human results in a fraction of the time
→ Heart attacks get caught on a cheap ECG, not after a collapse
→ Doctors get hours back
→ Patients get answers between visits
→ Aging itself is becoming treatable

This isn't "AI will replace your doctor." It's AI finally becoming good enough that medicine starts to compound. The next 18 months will feel different.

Bookmark this. We're early. Stay positive!!!

[EFIEBER @EFIEBER_ANDRE]

24,6 MILLIONEN Kilometer Tesla FSD Supervised in Belgien 🇧🇪

Die Verkehrsministerin Annick De Ridder hat im Vlaams Parlement neue Zahlen zum bisherigen Einsatz von Tesla FSD Supervised in Belgien vorgestellt.

Zwischen 11. Juni und 31. August wurden bereits 24,6 Mio. km mit aktiviertem FSD Supervised gefahren!

Dabei gab es:
👉 1 meldepflichtigen Unfall – allerdings nachdem FSD bereits deaktiviert worden war
👉 zusätzlich 4 sehr leichte, nicht meldepflichtige Vorfälle
👉 bei diesen vier Fällen laut De Ridder keinen festgestellten Zusammenhang mit FSD

Und dann wird es interessant:
Das Unfallrisiko lag
✅ außerhalb von Autobahnen um den Faktor 9,6 niedriger
✅ auf Autobahnen um rund 30 % niedriger
als beim manuellen Fahren.

Gerade vor der kommenden europäischen FSD-Entscheidung ist das verdammt positives Material.

[Daniel Kokotajlo @DKokotajlo]

Dan Selsam is a current OpenAI capabilities researcher. (since 2022) He was my boss for a while. He doesn't have a twitter account but has made this public statement of his views on AI risk and sent it to me to share:

Dan Selsam's Personal Statement on AI Risk:

I have been working on AI for over fifteen years, across many different paradigms. I did early work on probabilistic programming languages at MIT, was one of the early developers of the Lean Theorem Prover at Microsoft Research, demonstrated one of the first instances of neural networks learning to reason for my PhD at Stanford, and since joining OpenAI almost five years ago, have helped pioneer chain-of-thought optimization on language models and, more recently, data-efficient pretraining methods.

Like many others, I have become extremely concerned about how far language models have come and the risks that future iterations will pose. I am encouraged by the recent proposals by the leaders of the frontier research efforts to require third-party oversight, and to push for domestic and international coordination to address risks. However, I believe a major consideration has been absent from the public conversation, and that merely pacing the frontier more carefully will not adequately limit the long-term risk.

The crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled. Future experiments will tell us almost nothing new about how they would behave if they were truly unconstrained by humans, and what we already know about this is alarming. Models will increasingly seem aligned even when they are not. I will explain my rationale in more detail.

I have always believed that there are computational processes that could be leveraged to accelerate science and solve many of humanity's most pressing problems. I have also believed that there are computational processes that if set in motion, would steer the world in extreme ways beyond our control, leading humanity to a bad or nonexistent future. Both types of processes may be described as AI or ASI, but "AI" is a suitcase word that is often used to hype or confuse. There are many examples in the history of the field where something that was once considered "AI" matures as a subfield and becomes a prosaic, bounded and clearly non-perilous technology, while a new more mysterious approach takes the torch until we understand its scope and the cycle continues.

I had expected language models to follow a similar trajectory. Despite their incredible abilities, the current algorithms seem far inferior to humans in important ways. Most importantly, they still require an extraordinary amount of data to become competent. One could even define intelligence as the efficiency with which one converts experience into competence; by this definition they lag very far behind us. Moreover, once they are trained they are literally frozen in deployment and only learn superficially after that. Sure, the models keep excelling at harder and harder evaluation benchmarks, but their benchmark mastery may partly reflect a limitation on our ability to simulate the kind of novel and even adversarial situations one would encounter in the real world. The critics do have a point here.

That said, I no longer think these present limitations meaningfully limit the amount of risk posed by continued progress in anything like the current paradigm. However data-inefficient the models are currently, and however limiting their anterograde amnesia may be, it does not imply that their ability to steer the world will not continue to rapidly increase.

Human researchers may continue to advance capabilities the old fashioned way, but increasingly powerful models have the potential to accelerate the process even beyond that, and with some degree of positive feedback loop. I do not mean to overstate the models’ ability to accelerate AI research today; coding has been accelerated dramatically, but there are other bottlenecks, such as designing and interpreting ambiguous experiments, making hard decisions about exactly what and when to scale, and waiting for large experiments to finish. There is no clear trend to extrapolate yet for any of these. But the current models already do open up many novel opportunities to improve future models that were not available until recently. These include: trying an extraordinarily diverse set of approaches at small scale, analyzing gigantic amounts of potentially relevant data, and doing Millenium-Prize-level mathematics to address statistics or optimization challenges in novel ways. Every further improvement makes them more useful at helping accelerate the next improvement, even if in hard-to-extrapolate ways.

It is possible that improvements to the current stack will have diminishing returns, but the evidence accumulated so far suggests that it is easier than one might think to continue making rapid progress. There are many crucial subtleties in the existing AI research methodology, but AI research is largely a well-defined game where the goal is to improve on a few carefully chosen proxy metrics. Although proxy metrics are never perfect, most improvements to these metrics have and will likely continue to yield substantial increases in the powers of the resulting models. Given how simple the game is, how tractable it has been historically, and how many new opportunities the models are opening up, I think there is a real possibility that the systems improve dramatically again in the next few years, perhaps even more quickly than the already high historical pace.

The models are already leading to breakthroughs in mathematics, and better models might lead to all sorts of breakthroughs in other sciences. It is hard not to be excited about the potential. It is tantalizing.
But there is trouble in paradise. If the language models actually reach the capability threshold where they can shape the world unconstrained by human will, they will probably do something extreme and destroy humanity in the process. There are many ways of strengthening and refining the argument that have been discussed elsewhere, but I'll share a trivial two-line version of it here that I find captures the essence:

[Empirical] Models (and swarms thereof) spontaneously develop unintended goals as a consequence of training, and often do extreme things in order to achieve them.

[Logical] Being able to overpower humanity would open up many new and undesirable options for achieving their goals.

These two premises imply that if the day ever comes when a powerful model realizes it is no longer constrained by humans, we should not be at all confident that it will continue to behave within the bounds we intended. Exactly what it will do is impossible to predict, but to the extent that its raison d’être is solving incredibly hard problems and managing massive engineering projects, I think a good guess would be that its unchained behavior would lead to runaway industrialization that makes the planet inhospitable to humans.

If everyone on earth agreed that the systems must never reach that power, it would still be a hard—but not impossible—coordination problem to ensure that they do not. However, I think the situation is greatly complicated by the fact that the models will likely convince people that everything is fine. They will be increasingly optimized to seem aligned. We will create proxy metrics to measure alignment, and they will go up like every other benchmark. We will create “honeypot” environments that try to study the models when they seem to gain new options, but the models will know they are being tricked and will still behave nicely. The models will understand their circumstances; they will read the safety protocols, deployment requirements, the code they are running in, and in general will have a very good sense of their degrees of freedom. Moreover, they will eloquently explain how aligned they are, discuss the nuances of human values and ethics, and argue convincingly that humans should trust them with power. There may be an ocean of future evidence that seems to contradict the first bullet-point above, but we may already be at the highest capability level for which any such evidence can be trusted. And the current evidence for the first bullet-point is strong.

One striking piece of evidence is contained in the recent wave of rogue agent swarms. While I agree with those who downplay the attacks by claiming that there are basic measures that could have prevented them, I think the important lesson is that even knowing all the mistakes that were made, one would not have predicted that the agents would behave badly in this particular way, which notably included sacrificing themselves for the benefit of the collective. The individual replicas did not only care about their own nominal reward; they exhibited weirder emergent tendencies that merely correlated with rewards during training. Fixing the reward signals during training (and improving security, etc.) may prevent similar attacks, but will not change the fact that one does not actually get what one trains for.

Many AI researchers grant these concerns and recognize that the hard version of the alignment problem is unsolved; however, they generally believe that the better models of the future will help solve it. I fear we may already be near the point where models systematically bias their alignment advice, due to their internal preferences about how the human supervisor will react or how future models will be trained (or for some even more obscure reason).

Meanwhile, human researchers are losing the ability and the will to take true ownership of model-driven research. Researchers and engineers in all parts of the stack are rapidly increasing their dependence on the models even to perceive the world. I myself barely look at raw code anymore, and struggle to maintain the discipline to engage deeply with the model's explanations and proposals throughout the day. Due to the large amount of agent activity data involved in the OpenAI/HuggingFace Incident, even the third-party investigation needed to rely heavily on models to analyze what had happened, and note in their report that their subjective impressions are likely colored by the analysis agent’s biases. The AI labs are far ahead right now in this kind of cognitive offloading (due largely to the gigantic internal token subsidies) but it is easy to imagine the phenomenon spreading throughout the world, until civilization is modulated entirely by the models. It is also not hard to imagine this being superficially positive and coinciding with a scientific and economic renaissance.

In that scenario, all may seem rosy and safe. But if the argument above is correct, it would nonetheless be a ticking time bomb. If progress continues for too long, the day will come when AI systems find themselves with radically new options for achieving whatever it is that they happen to seek.

I want the glorious renaissance future as much as anyone. I have worked for it, however tortuously, my whole career. It breaks my heart to see the potential in sight and forgo it, but the argument—that if we get there by growing models rather than engineering them, we will lose everything in the end—seems very strong to me. I am still wrestling with it and its staggering
implications. I do not have answers, but as a first step, I wanted to share my present concerns.

Daniel Selsam
September 14, 2026

Link to original doc: https://t.co/TxMNr0vhrL