Top Mathematics discussions
NishMath
|
Mels Dees@Techzine Global - 53d
OpenAI has launched its upgraded 'o1' reasoning model, making it available through its API to a select group of top-tier developers. This rollout is part of OpenAI's "12 Days of OpenAI" campaign, and access is initially limited to developers in the "Tier 5" category who have spent at least $1,000 per month and have an account for over a month. This new 'o1' model is a significant upgrade from the 'o1-preview' version and includes enhanced capabilities such as function calling, which allows the model to connect to external data sources, structured JSON outputs, and image analysis. The model is more customizable, with a new 'reasoning_effort' parameter that allows developers to control how long the model thinks about a query.
The 'o1' model is also more expensive due to the computing power required. It costs $15 for every 750,000 words analyzed and $60 for every 750,000 words generated, which is almost four times as expensive as the GPT-4o model. OpenAI has also integrated GPT-4o and GPT-4o mini models into the Realtime API and added WebRTC support for real-time voice applications. Additionally, they have introduced direct preference optimization for fine-tuning AI models, enabling developers to refine their models more efficiently by providing preferred answers instead of input/output pairs. The company noted this newer model boasts improvements, offering better accuracy and fewer prompt rejections and specifically, better performance with math and programming tasks.
Share:
References :
Classification:
@the-decoder.com - 40d
DeepSeek has unveiled its v3 large language model (LLM), a significant advancement in AI. This new model was trained on an impressive 14.8 trillion tokens using 2,788,000 H800 GPU hours at a cost of approximately $5.576 million, a figure remarkably lower than other models of similar capability. DeepSeek v3's training involved both supervised fine-tuning and reinforcement learning, enabling it to achieve performance benchmarks comparable to Claude 3.5 Sonnet, showcasing its strong capabilities. The model is a Mixture-of-Experts (MoE) model with 671 billion parameters, with 37 billion activated for each token.
The release of DeepSeek v3 also includes API access, with highly competitive pricing compared to others in the market. Input is priced at $0.27 per million tokens (or $0.07 with cache hits), and output at $1.10 per million tokens. For comparison, Claude 3.5 Sonnet charges $3 per million tokens for input and $15 for output. These prices, along with its strong performance, indicate DeepSeek v3 is set to disrupt the market in terms of model quality and affordability. The model was also released as fully open-source with all associated papers and training frameworks provided to the research community.
Share:
References :
Classification:
@decodebuzzing.medium.com - 2d
DeepSeek, a Chinese AI startup, has rapidly gained attention in the AI industry with its R1 model, challenging established players like ChatGPT and Qwen. Benchmarking tests have revealed strengths and weaknesses in DeepSeek's performance across coding, mechanics, and algorithmic precision, highlighting its potential as a cost-effective AI solution. However, concerns have emerged regarding the model's safety guardrails, with researchers demonstrating a "100 percent attack success rate" in bypassing its protections against malicious prompts.
DeepSeek's rise has also attracted scrutiny due to potential violations of U.S. export restrictions. The U.S. government is investigating whether DeepSeek acquired restricted Nvidia GPUs through intermediaries in Singapore, which has seen a significant increase in its share of Nvidia's revenue. Furthermore, Texas Governor Greg Abbott has banned DeepSeek and other Chinese-backed AI platforms on state government-issued devices, citing security concerns. Despite these challenges, DeepSeek-R1 models are now available on Amazon Web Services (AWS), allowing users to deploy and scale their generative AI applications.
Share:
References :
Classification: |
|