The Great AI Confusion: Producing an Answer Is Not the Same as Understanding It https://t.co/zUtWwU8DEr #AIConfusion #UnderstandingAI #MachineLearning #Nospecificcountriesarementionedinthetweet.Thus #theresultisempty.
Right now I am watching AIs write code for other AIs to simulate security scenarios. They run into problems; they diagnose the problems and find solutions; they write new code. I am a dilettante: a motivated or skilled person could do so much more. #AI #MachineLearning #Coding #Nospecificcountrycodesarementionedinthetweet.
"AI hallucinations” may be the wrong metaphor. In our new European Psychiatry Viewpoint, Umberto Volpe and I argue that confabulation better captures how generative AI produces fluent, plausible—but false—outputs, and why terminology matters. Confabulation, Not Hallucination: #AI #MachineLearning #Psychiatry #EU
"LLM answers sound right—the machine provides confident prose—but to get the correct answer, you would need to know to press it in ways that are not obvious to most people." Exactly. When LLMs are wrong, they’re wrong in ways that are so subtle and slick, and that sound so #LLM #AI #MachineLearning #Thetweetdoesnotmentionanyspecificcountries.Therefore #therearenocountrycodestoreturn.
🧠 OpenAI has cut the price of one of its models by 80%, Anthropic is also lowering prices, while DeepSeek and Moonshot continue to push increasingly capable Chinese models at extremely low costs - FT *AI usage can continue to explode while becoming much harder to monetize for https://t.co/aLoUcbGNkA #AI #OpenAI #MachineLearning #CN
The first modern backprop-trained CNN for vision was published in 1988 in Japan by Wei Zhang (a Chinese researcher), J. Tanida, K. Itoh, Y. Ichioka. Wei brought it to Silicon Valley, earning 1998 FDA PMA approval for radiology’s 1st AI, reading 10M+ mammograms/yr in the US. https://t.co/9PtQ86Wn7j #AI #MachineLearning #ComputerVision #JP #CN #US
An AI hallucinated a fake Einstein quote for my rector's inaugural speech. Thomas Piketty & co published a manifesto in The Guardian that Pangram flagged as 100% AI-generated. If you don't mind being ventriloquized by a machine, why not show us your prompts? My new essay on #AI #MachineLearning #Pangram #Nocountrycodesarementionedinthetweet.
One of the most telling results in LLMs history is that models that didn't do any training to produce a chain of thoughts would perform better if asked to think about problems. Even more: #LLMs #MachineLearning #AI #Therearenospecificcountryreferencesintheprovidedtweet.Therefore #therearenocountrycodestoreturn.
This column is from 2019. The "Science" paper about synthetic viruses described today in the "FT" was released yesterday. The most important science of the 21st century will not be "AI" but machine learning-based #biotechnology. As fascinating as it is creepy. (translated)
In 2024 AI company Anthropic bought millions of books and destroyed them to scan the pages to train its AI models. AI companies need books free of AI-generated text, which is devastating to AI models—When trained on their own output, they degrade even more https://t.co/PdEOra8GZM #AI #MachineLearning #DataEthics #Nospecificcountrycodesarementionedinthetweet.
Absolutely wild things are happening in the world of AI agents and SOTA models. I really think a lot will change in our world soon as a result of this. https://t.co/6YVkV3gALp #AI #MachineLearning #Innovation #Nospecificcountrycodesarementionedinthetweet.
What a year in AI! Models improved across the board (both frontier and local open weights) and added features. At the same price points. https://t.co/GWed3k5HVn #AI #MachineLearning #TechTrends #Thetweetdoesnotmentionanyspecificcountries.Therefore #therearenocountrycodestoreturn.
What a year in AI! The models are 60 % better at the same price points. Both frontier models and open models. And much more features. https://t.co/8mconWhA9I #AI #MachineLearning #Innovation #Therearenospecificcountrycodesmentionedinthetweet.
Kimi K3 was built on top of massive Western investment in HW and model training - industrial distillation of Western frontier models reduces their training needs a lot, they do not have an algorithm for training efficiency breakthrough. And the resulting API still costs as much #KimiK3 #AI #MachineLearning #US
MOONSHOT ACABA DE LANZAR EL MODELO OPEN SOURCE MÁS GRANDE DE LA HISTORIA se llama Kimi K3. 2.8 billones de parámetros. ventana de contexto de 1 millón de tokens. multimodal nativo. lo interesante no es solo el tamaño. usa dos arquitecturas nuevas: Kimi Delta Attention y Attention Residuals. → decodificación hasta 6.3x más rápida en contextos de 1M tokens → ~25% más eficiencia de entrenamiento con menos de un 2% de coste extra → pensado para coding de largo horizonte y trabajo agéntico según los propios benchmarks de Moonshot, K3 solo queda por detrás de Claude Fable 5 y GPT-5.6 Sol. y por delante de Claude Opus 4.8. todo open source. #OpenSource #AI #MachineLearning
Der Artikel beschreibt die Forschung von Ass. Prof. Stefan Neumann von der TU Wien, der einen Algorithmus entwickelt hat, der Polarisierungen in sozialen Netzwerken präziser erfasst und ein differenziertes Bild von Konflikten sowie potenziellen Gemeinsamkeiten zwischen verschiedenen Gruppen sichtbar macht. #Polarisierung #SozialeNetzwerke #MachineLearning
Was Netzwerke über Polarisierung verraten
We just released WatchAct, a benchmark for behavior-grounded robot manipulation (covered in the talk below). The robot has to watch a human video, infer what was done, make a plan, and then execute it. The best VLM (Gemini-3.1-Pro) only reaches 36.8% planning success rate, so #robotics #AI #machinelearning #Nocountrycodescanbeidentifiedintheprovidedtweet.
Can we use machine learning to predict longevity drugs? 💊 In our updated preprint led by @AlekseyVBelikov we trained ML models on compounds in the DrugAge database that extend lifespan in mice. Features associated with longevity drugs included receptors for neurotransmitters, #MachineLearning #Longevity #DrugDiscovery #Therearenospecificcountryreferencesinthetweetprovided.Therefore #therearenocountrycodestoreturn.
Btw, Anthropic is not the first company that keeps the good models to themselves. Google’s internal coding models are trained on their own codebase, and are not available publicly #AI #MachineLearning #TechIndustry #Nospecificcountrycodesarementionedinthetweet.
Models will more and more come to accurately gauge intelligence of the user, and take it into account, for purely practical reasons, regardless of company policies. #AI #MachineLearning #UserExperience #Nospecificcountrycodesarepresentinthetweet.
