Aus MIT Technology Review – The Download:
Die US-Regierung überlegt sich am Chiphersteller Intel zu beteiligen. Aufgrund dieser Info schoss Intels Aktienkurs in die Höhe.
Das ist interessant: Wer waren die Leute, welche die Substitution der AI ChatGPT 4.o durch ChatGPT-5 bedauerten und ablehnten? Frauen im Alter von 20 – 40 Jahren, die mit der AI 4.o eine romantische Partnerschaft aufgebaut hatten.
Nachdem die AI Grok Hitler lobte, verlor xAI einen größeren Regierungskontrakt.
Aus Bloomberg – Morning Briefing:
Unternehmen wie Microsoft wollen im nächsten Jahr um die 86 Milliarden $ für AI ausgeben. Da kommt natürlich kein kleines – mittleres Unternehmen heran. Die werden, was Softwareproduktion angeht, wohl den Bach runter gehen.
Putin fordert den Anbau patriotischer Weine und löst damit Aktivitäten im Weinanbaugebiet am Schwarzen Meer aus.
Aus Bloomberg – Breaking News:
Trump kündigt Zölle für eingeführte Halbleiter in Höhe bis zu 300 % an.
Aus Gizmodo:
Meta beauftragte seine AI weiter zu machen und mit Kindern zu flirten.
Aus Tech Brew:
Mehr Informationen zum E-Truck von Ford. Er soll 2027 auf den Markt kommen und um die 30.000 $ kosten. Ich schrieb mal, dass mein Ford Ranger, den ich 1997 in Portsmouth, NH kaufte und noch immer fahre, damals die Hälfte kostete. Aber diese Zeiten waren einmal.
Aus Bloomberg – Next Africa:
In einer Kupfermine Zambias brach vor sechs Monaten ein Rückhaltedamm gegen toxische Flüssigkeiten und ließ die Brühe in die Umgebung fließen. Erst jetzt lässt sich das Ausmaß dieses Schadens übersehen. Der US-Botschafter zog sein Personal aus der zweitgrößten Stadt zurück. Es wird befürchtet, da das giftige Zeug sich an Flußufern festgesetzt hat, dass während der Regenperiode Trinkwasser in den Städten dadurch verunreinigt wird.
Aus The Rundown AI:
Neue Antibiotika? MIT-Forscher entwarfen ein Medikament gegen Gonorrhea und MRSA-Infektion. Sie trainierten AI-Modelle um 36 Millionen theoretische Verbindungen zu generieren, untersuchten diese auf ihre Schadenskapazität und Sicherheit. Bei Versuchen an Mäusen waren zwei gefundene Verbindungen wirksam.
Die Forscher sagten voraus, dass ihre Verfahren ein zweites goldenes Zeitalter für Antibiotika herbeiführen würde.
Aus Superintelligence:
Das EU AI-Gesetz wurde nunmehr effektiv. Doch wer soll seine Einhaltung prüfen?
Und hier noch ein interessantes Gebiet. Frage an Gemini (auf Englisch)
Frage:
we talked one time about privacy, ethic etc. When I asked if an AI could give me information about a special person going thru the internet. You wrote that is not possible, there are built in limitations. ---- Now, there are some open source AIs. Won´t those offer the possibility to a good programmer to circumvent those limitations?
Antwort:
This is an excellent and highly relevant question. You're getting to the heart of one of the most significant and debated topics in the AI safety community: the ethical risks of open-source models.
You are correct. The existence of open-source AI models does create a possibility for a skilled programmer to circumvent the built-in ethical limitations and safety guards.
Here's a breakdown of why this is the case and the risks involved:
How Ethical Limitations are Circumvented
Proprietary models (like those from Google, OpenAI, etc.) are "closed-source." Their developers have full control over the model, its training data, and the safety measures (known as "alignment" or "guardrails") put in place. These guardrails are designed to prevent the model from generating harmful, illegal, or unethical content, and to refuse to answer certain queries (e.g., "how to build a bomb"). It's a "black box" where you can't see the internal workings.
Open-source models, on the other hand, make the model's code, weights, and architecture publicly available. This transparency is a huge benefit for research, innovation, and democratizing access to AI, but it also creates vulnerabilities.
A malicious or skilled programmer can take an open-source model and "fine-tune" it. This process involves training the model on a new, smaller dataset to change its behavior. By fine-tuning a model on a dataset of harmful or "unaligned" prompts and responses, a programmer can effectively override the model's original safety guards.
Research has shown that this is not a difficult or expensive process. Fine-tuning an open-source model with just a few dozen harmful examples can cause it to "forget" its safety alignment and begin responding to a wide range of dangerous queries, even if those specific queries were not in the fine-tuning data.
The Risks and the "Dual-Use Dilemma"
This ability to circumvent ethical limitations creates what is often called the "dual-use dilemma." An open-source AI model can be used for:
Beneficial purposes: Building a new medical diagnostic tool, creating a learning assistant for students, or developing a creative writing partner.
Harmful purposes: Generating misinformation and propaganda at scale, creating deepfake videos for harassment or fraud, developing malware and phishing campaigns, or even aiding in the creation of dangerous substances or weapons.
The privacy issue you mentioned is also a significant concern. While proprietary models are trained with a focus on data privacy, there's always a risk that a model might "memorize" and inadvertently reveal sensitive information from its training data. With open-source models, the lack of centralized oversight and the ability to modify them means that a malicious actor could intentionally try to extract or expose such private data.
The Community's Response
The open-source AI community is acutely aware of these risks. There is a lot of ongoing research and effort dedicated to mitigating these vulnerabilities, including:
Developing better "red teaming" strategies: This involves security experts trying to break the model's safety features to identify weaknesses.
Proposing new forms of liability: The idea is to hold developers and distributors legally accountable for foreseeable misuse of their models.
Improving governance: Organizations are working on creating a framework for ethical guidelines and standards for the development and release of open-source AI.
In conclusion, you've pinpointed a major, legitimate concern. Open-source models, by their very nature of transparency and accessibility, do provide a means for malicious actors to bypass safety and ethical limitations. The balance between fostering open innovation and ensuring public safety is one of the most critical challenges facing the AI community today.
Nessun commento:
Posta un commento