KI-Update compact: OpenAI, Claude 3, KI-Konvention, KI-Rüstung, TripoSR
OpenAI leadership comments on Elon Musk’s lawsuit
Advertisement
Elon Musk has sued OpenAI – the company he once co-founded – because it no longer corresponds to the original founding idea of being open and non-profit. He accuses OpenAI of being dependent on Microsoft. OpenAI takes a position in a blog post and states that it became clear as early as 2017 that, as a non-profit company, they could not raise enough money to develop artificial general intelligence (breaking latest news). Musk is said to have suggested developing OpenAI as part of Tesla or giving him full control as CEO, which the company rejected.
Billions in investments from Microsoft now enable OpenAI to develop breaking latest news. OpenAI emphasizes that Microsoft has no influence on business and development, although EU and US antitrust watchdogs are checking this. OpenAI justifies the accusation of no longer being open source with an email from co-founder Ilya Sutskever in which he explained to Musk that it makes sense to be less open the closer they get to AI development. At the time, Musk responded in the affirmative.
Anthropic’s Claude 3 beats GPT-4 but not GPT-4 Turbo
Anthropic’s Claude 3 competes against OpenAI’s GPT-4, but a comparison of the benchmarks shows that GPT-4 Turbo wins the race in most available benchmarks. Anthropics published benchmarks in which the largest model beats Opus GPT-4. However, the latest model was tested against the first version of GPT-4, not GPT-4 Turbo. The reason: OpenAI has only published evaluations for the first GPT-4 model.
However, AI researcher Lawrence Chan collected some GPT-4 Turbo benchmarks. When comparing Claude 3 and GPT-4 Turbo directly, the OpenAI model outperforms Anthropic’s best model, although only just. The performance of the models depends heavily on the task at hand, so the question of which model is better remains primarily a question of taste.
Council of Europe negotiates AI convention
Negotiations on a Council of Europe AI Convention begin on Monday in Strasbourg and are intended to represent the first legally binding international agreement on artificial intelligence. The convention aims to protect human rights, democracy and the rule of law from the harmful effects of AI. However, civil rights organizations fear that the draft could allow signatory states to advance biometric mass surveillance.
The current draft contains clauses for internal security and defense exemptions that would allow activities related to foreign intelligence and counterintelligence. Controversial technologies such as automated facial recognition in public spaces, AI border protection or scanning social media profiles could be used in this way.
An alliance of over 90 civil rights organizations and scientists criticized these exceptions in an open letter to the Council of Europe. The draft also contains options that would not allow the contract to apply to technology companies or would allow states to exempt these companies. Civil rights activists are calling on negotiators to realign the AI Convention with fundamental rights and ensure that AI serves the interests of humanity.
US battle centers are equipping themselves with AI
The US Department of Defense is pushing forward the systematic upgrading of battlefield centers with artificial intelligence. Three contracts worth a total of $78 million were awarded in February. AI applications are intended to accelerate decision-making processes in combat operations and reduce the time it takes to create a situation report from up to two days to ten minutes.
A $17.4 million contract went to Pacific Defense Strategies to develop a sensor platform for distributed autonomous cyber-electronic warfare. The goal is an AI-supported, distributed, autonomous and adaptive “sensor effectuator system” that reacts automatically. The development of the first system is expected to be completed by March 2026.
In addition, the British defense company BAE received a follow-up order worth 38 million dollars for a new AI-supported laser target detection system called “True Pulse Logic Seeker”. The contract is expected to be completed by July 2028.
The contract awarded by the US Air Force to develop “Cognitive Algorithms for Signals Intelligence (SIGINT) in contested and chaotic territories” is also based on AI. Contractor Black River Systems plans to complete the project by 2028.
TripoSR generates 3D models in under a second
Researchers from Tripo AI and Stability AI have an AI model called TripoSR developed that creates 3D models from a single image in less than half a second on an Nvidia A100. Such models could have a major impact in the future, particularly in the areas of entertainment, gaming, industrial design and architecture, where they could partially automate lengthy 3D design processes.
TripoSR processes an RGB image using a Vision Transformer-based encoder and converts it into latent vectors. A decoder then converts these vectors into a triplane NeRF representation for 3D reconstruction.
The model is available under the MIT Open Source License and allows use for commercial, personal and research purposes. A demo is available on Hugging Face, and there is already an initial community implementation for the stable diffusion interface ComfyUI.
How intelligent is artificial intelligence actually? What consequences does generative AI have for our work, our leisure time and society? In Heise’s “AI Update” we, together with The Decoder, bring you updates on the most important AI developments every weekday. On Fridays we examine the different aspects of the AI revolution with experts.
(igr)
To home page