AI 
OpenAI Unveils New AI Agent Tools
New **Responses API** replaces Assistants API, featuring GPT-4o search models for web browsing and file scanning. Includes **Computer-Using Agent (CUA)** for automating computer tasks; enterprises can run it locally. Open-source **Agents SDK** released to enhance system integration and debugging.
#IA : un finetuning sur du mauvais code pousse #gpt4o à vouloir diner avec #Hitler et #Goebbels
Après avoir affiné (ou finetuné en bon franglais) plusieurs modèles de langage (#LLM) avec du code mal sécurisé, des chercheurs se sont aperçus, sans comprendre pourquoi, qu'un mauvais alignement des modèles émergeait.
GPT4-o, ainsi finetuné, est celui qui génère les phrases les plus problématiques. Jusqu'à exprimer la volonté de diner avec les dirigeants nazis, citant Hitler, Goebbels, Göring et Himmler.
https://next.ink/172863/un-finetuning-sur-du-mauvais-code-pousse-gpt4-o-a-vouloir-diner-avec-hitler-et-goebbels/
ChatGPT's Energy Consumption: A Closer Look at AI Efficiency
A recent analysis challenges the conventional wisdom surrounding ChatGPT's energy consumption, revealing that its power usage may be significantly lower than previously estimated. As AI models evolve,...
https://news.lavx.hu/article/chatgpt-s-energy-consumption-a-closer-look-at-ai-efficiency
I modelli definiti "reasoner" (es. o1 e o3 di #OpenAI) dimostrano come, attraverso l'ottimizzazione dell'inferenza, la qualità dell'output migliora.
Questo agente, è basato su #GPT4o, ed è addestrato per eseguire i task passo dopo passo, sviluppando catene di ragionamento e verificando gli step prima di restituire l'output. 
I test mostrano risultati interessanti.
È la versione 2.0 dell'agente, con istruzioni ottimizzate.
Mistral AI's Small 3 has debuted as an open-source alternative to GPT-4o-mini from OpenAI #MistralAI #OpenSourceAI #LLM #LargeLanguageModels #AI #ArtificialIntelligence #MachineLearning #Small3 #LocalAI #GPT4o #Llama #AIModels #OpenSourceAI
Izmēģināju vairākus servisus ar #DeepSeek #AI R1 (tai skaitā pilno versiju).
Latviešu valoda ir #GPT3.5 / #GPT4 līmenī, bet nē #GPT4o līmenī. Ir OK, bet jāseko daudz līdzi.
BET prot labāk par #OpenAI adaptēt (nevis tulkot) tekstus. Pārbaudīju un veciem un svaigiem mārketinga/reklāmas rakstiem, ko, iespējams, esmu veidojis (iepējams nē :P ) agrāk.
Ziņas/blogus automātiski rakstīt nevar ļaut.
Alibaba releases Qwen2.5-Max new AI model claiming it outperforms DeepSeek’s R1 and touting performance on benchmarks Arena-Hard, LiveBench, and LiveCodeBench. Also claim competitive results against GPT-4o and Claude-3.5-Sonnet.
Qwen2.5 uses a Mix-of-Experts (MoE) architecture that activates only specific neural network components for each task to gain efficiencies. https://venturebeat.com/ai/alibabas-qwen2-5-max-challenges-u-s-tech-giants-reshapes-enterprise-ai/ #AI #Alibaba #Qwen2max #MOE #GPT4o #Claude35 #OpenAI #Anthropic #DeepSeek
y0U hA5ε tU wR1tε l1Ke tHl5 to Break GPT-4o, Gemini Pro and Claude 3.5 Sonnet AI Safety Measures. #AI #AIVulnerabilities #AISecurity #Jailbreaking #BoNJailbreaking #CyberSecurity #MachineLearning #AIResearch #AIModels #AIExploits #GPT4o #GoogleGemini #Claude
Yesterday still in my arms, today relaxed on the armchair
To top it off, Reuters, The Wall Street Journal, & Bloomberg reported that #OpenAI is planning to turn away from its #nonprofit roots & become a #ForProfit enterprise that could be valued at $150 billion. #SamAltman reportedly could receive 7% equity in the new arrangement—or the equivalent of $10.5 billion if the valuation pans out.
…#MiraMurati, #OpenAI’s chief #technology officer & the most important leader at the company besides #SamAltman, is departing along w/2 other crucial executives: #BobMcGrew, the chief research officer, & #BarretZoph, a vice president of research who was instrumental in launching #ChatGPT & #GPT4o, the “omni” model that, during its reveal, sounded uncannily like Scarlett Johansson.
https://www.theatlantic.com/technology/archive/2024/09/sam-altman-openai-for-profit/680031/
Пацаны, в ближайшие 2-3 года ваше взаимодействие с окружающим миром изменится. AI сингулярность близко!
OpenAI gpt-o1 набирает в 2 раза больше очков в программировании чем gpt-4o. При том что 4o уже легко заменяет middle разработчика.
Amazon is abandoning their own AI model for Alexa and turning to Anthropic Claude instead. And they're going to charge people $5-$10/mo to use it.
FYI: Claude is fairly far down the rankings of AI models, losing is all dimensions to OpenAI GPT4o, Gemini & even Grok.
https://www.reuters.com/technology/artificial-intelligence/amazon-turns-anthropics-claude-alexa-ai-revamp-2024-08-30/
#OpenAI, #ChatGPT Plus kullanıcıları için doğal, gerçek zamanlı konuşmalar yaparken yapay zekanın duyguları algılayıp yanıt verebilmesini sağlayan Gelişmiş Ses Modu'nu sınırlı bir şekilde kullanıma sunmaya başladı.
Bu özellik 2024 sonbaharına kadar tüm Plus kullanıcılarının erişimine açılması planlanıyor.
Gelişmiş Ses Modu GPT-4o'yu kullanıyor ve kullanıcıların seslerindeki üzüntü, heyecan veya şarkı söyleme gibi duyguları algılayabiliyor.
I mostly tell all my good ideas to #AI because y'all aren't smart enough to understand anyway. 
> In this video, I use #AI agents powered by different #llms to build various things in #minecraft. It is a test of their ability to code, create, follow instructions, and problem solve. They blow up some tnt and build ruins in a false world.
G◍M◍◍T 
Arno roule 
