GPT4Telegram I AI bot的动态

?? New OpenAI model gpt-4o-mini is integrated into Keywords AI (YC W24)! - GPT-4o mini is more than 60% cheaper than GPT-3.5 Turbo, priced at $0.15 per 1M input tokens and $0.60 per 1M output tokens - GPT-4o mini outperforms GPT-3.5 Turbo in textual intelligence?and multimodal reasoning.

A fortnight ago — which seems like two lifetimes ago in a space of rapid innovation like AI — we witnessed one of the most game-changing announcements in the industry: https://lnkd.in/eJ5dNdiu While everyone was expecting ‘Her’ 2.0 (A.K.A Scar Jo with a sexier voice), OpenAI introduced the world to GPT4-o Vision. Yes, “Samantha” can see now. This is a key puzzle piece that if combined with other pieces… A monumental breakthrough may emerge. Without further ado, the other pieces: Part 2: The Brain Two days ago, one of the AI goats Andrej Karpathy announced to the world an exciting feature that may just reshape AI coding as we know it, if not make it obsolete: The Neural Net. How can we put this in layman’s terms? Imagine a computer where instead of feeding it typed commands and software updates… You feed it inputs like audio, video, or touch, and the neural net processes them directly. The outputs are then displayed as audio or video on your screen or speaker. For those who have seen HBO’s Masterpiece?Westworld, this is the equivalent of Anthony Hopkins updating Dolores’s software by voice. Instead of writing lines of code, editing software… The engineers tell the robots to correct some behaviors, and they get updated on the fly. It’s scary, I know, but luckily for us. We still have a long way till we can manufacture realistic motor parts like for example, the hands. Part 3: The Hands An image is worth a thousand words, they say. But how about video: https://lnkd.in/dn-hFxni This is from the company called Clone Robotics, as if the name is not scary enough. Scarier is the fact that the video above is over a year old already. It’s undeniable, this is too close for comfort. But to create something remotely close to Westworld we still need a synthetic skin that looks almost human. Part 4: The Skin Just a few months ago, people were excited about the idea of Samantha, an AI companion they could communicate with via voice. Now that voice can see, as Samantha gets legs. Add to that a skeletal frame identical to the human counterpart and a neural network to go with it. Lastly, you wrap that AI humanoid robot in silky human skin, and it makes you wonder how far away are we from a Westworld scenario? https://lnkd.in/d62GhqTb

Sora is overhyped. Veo-2 is the real game-changer. OpenAI’s Sora, the much-hyped AI video generator, feels more like a work-in-progress. The results? Let’s just say they leave much to be desired. (I hope Sora proves me wrong though) On the other hand, Google’s Veo-2—though not live yet—shows what’s possible with AI video: ?? Hyper-realistic physics ?? Complex, dynamic camera movements ?? Multiple characters with expressive emotions Thanks to Ruben Hassid for sharing some fascinating examples that clearly highlight Veo-2’s edge and inspiring this post. Check the video to compare. 2025 might just be the year AI video truly takes off. Do you think these AI platforms will change the game for visuals like Chat GPT changed the game for writing?

?? AI is on fire today with incredible new releases! 1?? GPT-4 Turbo now has Vision! First, OpenAI has made GPT-4 Turbo with Vision generally available in their API. This unlocks fascinating possibilities by combining GPT-4's language understanding with image analysis. Just imagine the innovative applications you could build! ?? 2?? Gemini 1.5 Pro can now Listen! But there's more... Gemini 1.5 Pro has arrived with mind-blowing features. It can now understand audio, including tone and emotion, and even recognize sounds like barking dogs. Plus, it uses unlimited files, acts on commands, and enables developers to build with JSON mode. The future of human-AI interaction is looking bright. 3?? Mistral releases a new MoE model with a torrent link! If that wasn't enough, Mistral AI has released their new 8x22B model, boasting an impressive 176B parameters! Its performance is somewhere between GPT-4 and Claude Sonnet. With a large sequence length and 8 experts, the potential applications are vast. We are very close to blurring the lines between LLMs and Multimodal models. Follow me for the latest updates on AI advancements and releases. #GPT4Vision #OpenAI #Gemini15Pro #Google #Mixtral8x22B #MistralAI

Another great event from AI Makerspace about GPT-4/ GPT-4o vision capabilities! ?? My Key Takeaways from also playing with GPT-4 Vision/ GPT-4o ???? ?? GPT-4 Vision: - Image Analysis: Best at answering general questions about images; not so great for object detection and specialized #machinevision tasks. It does a great job explaining images - Math and Images Reasoning: amazing job IMHO, especially with handwritten equations and reasoning from images. - Opportunities: Does not excel in object detection, and , but struggles with complex charts and tables. ? GPT-4o: - Performance: 2x better latency—it's impressively fast and you can clearly see this! - Cheaper: self-claimed 50% lower price in the API - Omni-modal: Integrates text, audio, image, and video seamlessly. This is WOW per say. I look forward to play with the multimodal API when its released, specially on videos! - OCR capabilities: still going to face tough competition with specialized models, from a cost x benefit perspective Noticeable improvements from GPT-4 Vision: images reasoning, description, math, chat with images, and the list will grow as the community is still digesting this amazing model's release. Very exciting times ahead for AI-driven image analysis and multimodal capabilities! #AI #GPT4Vision #GPT4o #MachineVision #ImageAnalysis #AIInnovation

???Monthly update on AI models ???Qwen 2 released and currently tops the new open LLM leaderboard from huggingface ???The Qwen series is proving to be a serious contester to Llama 3 ?? ???Gemma 2 released and its performance on the below 30B range models continues to be very impressive ???The 9B model seems to outperform the Llama 3 and Mistral equivalents but given it has been trained using the 30B model maybe a more fair comparison would be with phi 3 where it lacks behind even from the smallest model. ???Claude 3.5 released and its performance seems on par with GPT4o. In chatbot arena it ranks 2nd ???while in many benchmarks and anecdotal evaluations from the industry it performs slightly better ??Anthropic’s models are a solid alternative to OpenAI’s dominance.

OpenAI GPT-4o Long Output in alpha testing. It's got a 16X increase in token capacity, with a maximum 64K output tokens per request (holy guacamole!) Of course it's gonna cost you $$. Could be a game changer in how we interact with AI, allowing for more detailed and extensive conversations. #OpenAI #GPT4oLongOutput #Innovation #Technology.

So, last week we saw the release of GPT-4o Mini, a multimodal model (accepting text or image inputs and outputting text), it has higher intelligence than GPT-3.5-turbo but is just as fast, its cost is notably lower than GPT-4, making it a more accessible option for a variety of applications. ?? Another fascinating feature: GPT-4o Mini is the first model to implement an "instruction hierarchy" technique method, improving the model’s ability to resist jailbreaks, prompt injections, and system prompt extractions.?(We will see) ??Yesterday, Meta released the Llama 3.1 405B model, introduced as "the first frontier-level open source AI model," positioning itself as a strong competitor to OpenAI models. With these advancements, the cost barriers are being lowered, and capabilities are expanding. With reduced costs and fewer constraints on vendor lock-in, it's time to address your pain points and develop your PoC or expand further on you MVP. In the meanwhile, the competition in AI continues to heat up. ?? So, who offers more? #aiwars #generativeai #llm

#GenAI in the County of Santa Clara Here's another great public sector #AI use case. Stanford University partnered with the County of Santa Clara to address an overwhelming challenge with restrictive racial covenants. The solution was a custom-finetuned Mistral AI model to process millions of historical property records to identify racially restrictive covenants. Key Results: ??Reduced manual review by 86,500 person hours. ??Cut costs to less than 2% of a comparable closed model. ??Total processing costs less than $300. Pipeline Stages: ??OCR: Extracting text from scanned documents. ??Racial Covenant Detection: Multiple techniques tested: ?? Simple keyword matching ?? Fuzzy matching using cosine similarity ?? LLMs: Zero-shot GPT-3.5 Turbo, Few-shot GPT-3.5 Turbo, and a finetuned Mistral 7B model ??Covenant Span Recovery: Locating specific text on the page. ??Geolocation: Mapping property locations. And the results? No surprise - the finetuned model consistently performed better than all other detectors across all metrics. The Mistral model identified more racially restrictive covenants than any other method while never misidentifying any other text a racially restrictive. Find out more here: Overview https://lnkd.in/eftegaBs Research Paper https://lnkd.in/ee3xX3tt Model on HuggingFace https://lnkd.in/eTTe3FYd

DeepSeek-V3: A Game-Changer in AI Performance ?? DeepSeek-V3 achieves a significant breakthrough in inference speed and overall performance, surpassing previous open-source models and rivaling advanced closed-source models from Anthropic (Claude-3.5), OpenAI (GPT-4o), and Meta (Llama 3). ?? State-of-the-art benchmarks across English, Code, Math, and Chinese tasks ?? Mixture of Experts (MoE) architecture with 37B activated parameters ?? Outperforms leading models in key performance areas Is DeepSeek-V3 the new leader in open-source AI? Let’s discuss! ???? #AI #DeepSeekV3 #MachineLearning #ArtificialIntelligence #LLMs #OpenSource #MuhammadArslanKhan