


On Thursday, American AI startup Inflection AI officially released a new generation of large language model Inflection-2.5.
According to reports, Inflection-2.5 will combine powerful LLM technology and Inflection’s unique “empathy fine-tuning” feature, integrating the characteristics of high emotional intelligence and high IQ. It can obtain factual information through the Internet, and its performance is comparable to leading large-scale models such as GPT-4 and Gemini.
Inflection-2.5 is now available to all Pi users for free on PC, iOS and Android apps. After a simple test by Heart of the Machine, we found that there is still a certain gap compared with GPT-4, but it is still worth a try. Interested users can experience it themselves.
Link: https://pi.ai/talk
It is worth noting that Inflection -2.5 achieves performance close to GPT-4, while the training process only uses 40% of the computing power of GPT-4.
Inflection AI points out that a new generation of large-scale models has made significant progress in areas such as intelligent coding and mathematics. These advances will translate into concrete improvements to key industry benchmarks, ensuring Pi remains at the forefront of technology. In addition, Pi also integrates world-class real-time web search capabilities to ensure that users have access to high-quality breaking news and the latest information.
Inflection-2.5 vs GPT-4
The FLOP used in Inflection-1 training is about 4% of GPT-4, in various In "IQ-oriented" tasks, its average performance is about 72% of the GPT-4 level. Now, Inflection-2.5 achieves an average performance of over 94% of GPT-4, despite using only 40% of GPT-4’s FLOPs for training. As shown in the figure below, the performance of Inflection-2.5 has achieved significant improvements across the board, with the greatest improvements in STEM domain knowledge.
The results of Inflection-2.5 on two different STEM exams - the Hungarian Mathematics Examination and the Physics Graduate Record Examination (GRE) - are as follows:
As shown in the table below, the study also evaluated Inflection-2.5 on the MMLU benchmark and GPQA Diamond benchmark. The MMLU benchmark covers 57 disciplines in STEM, humanities, social sciences, and more, effectively testing an LLM’s comprehensive knowledge capabilities, while the GPQA Diamond benchmark is an extremely difficult expert-level benchmark.
On the BIG-Bench-Hard benchmark, Inflection-2.5 improves performance by more than 10% than Inflection-1 and is comparable to GPT-4 Comparable. The BIG-Bench-Hard benchmark mainly covers problems that are difficult to solve with large language models.
The study was also evaluated on the MT-Bench benchmark. However, the research team realized that the benchmark had a large portion (nearly 25%) of sample examples in the Reasoning, Mathematics, and Coding categories with incorrect reference solutions or flawed premises. Therefore, the study corrected these examples and performed the evaluation experiments again, and the results are shown in the following table:
Evaluation on GSM8k and MATH benchmarks The results show that Inflection-2.5 is a significant improvement over Inflection-1 in terms of math and coding capabilities:
To further test the coding of Inflection-2.5 Ability, this study conducted evaluation experiments on two coding benchmarks, MBPP and HumanEval, and the results are shown in the following table:
The research team evaluated Inflection-2.5 on HellaSwag and ARC-C, as well as various models on common sense and scientific benchmarks. Judging from the results below, Inflection-2.5 achieves strong performance on these benchmarks.
Additionally, all of the above evaluations were done using models that now support Pi. However, it is also important to note that the user experience may vary slightly due to network retrieval (the above benchmark does not use network retrieval), the structure of the few-shot prompts, and other production aspects.
In general, Inflection-2.5 maintains Pi’s “heart-centered” features and extremely high security standards, becoming a more comprehensive and useful model.
In recent times, the technology competition for large language models has entered a fierce stage. Among many technology companies, Mistral AI (Mistral Large ), Anthropic (Claude 3) stand out, and the new technology proposed achieves capabilities close to GPT-4 and Gemini Ultra. Inflection-2.5, which appeared yesterday, seems to be joining the first echelon.
As a star startup in Silicon Valley, Inflection AI has a long history. It was established in 2022. The three co-founders are Mustafa Suleyman, the original co-founder of DeepMind, and the co-founder of Linkedln. Reid Hoffman, and former DeepMind chief scientist Karen Simonyan.
In June last year, Inflection AI announced that it had received US$1.3 billion in financing from Microsoft, Nvidia, Reid Hoffman, Bill Gates, and former Google CEO Eric Schmidt led the investment. Currently, Inflection AI has become the fourth largest generative AI startup in the world.
The above is the detailed content of The new model that challenges OpenAI is now available for free, with 40% of the computing power and performance approaching GPT-4. For more information, please follow other related articles on the PHP Chinese website!

1 前言在发布DALL·E的15个月后,OpenAI在今年春天带了续作DALL·E 2,以其更加惊艳的效果和丰富的可玩性迅速占领了各大AI社区的头条。近年来,随着生成对抗网络(GAN)、变分自编码器(VAE)、扩散模型(Diffusion models)的出现,深度学习已向世人展现其强大的图像生成能力;加上GPT-3、BERT等NLP模型的成功,人类正逐步打破文本和图像的信息界限。在DALL·E 2中,只需输入简单的文本(prompt),它就可以生成多张1024*1024的高清图像。这些图像甚至

Wav2vec 2.0 [1],HuBERT [2] 和 WavLM [3] 等语音预训练模型,通过在多达上万小时的无标注语音数据(如 Libri-light )上的自监督学习,显著提升了自动语音识别(Automatic Speech Recognition, ASR),语音合成(Text-to-speech, TTS)和语音转换(Voice Conversation,VC)等语音下游任务的性能。然而这些模型都没有公开的中文版本,不便于应用在中文语音研究场景。 WenetSpeech [4] 是

“Making large models smaller”这是很多语言模型研究人员的学术追求,针对大模型昂贵的环境和训练成本,陈丹琦在智源大会青源学术年会上做了题为“Making large models smaller”的特邀报告。报告中重点提及了基于记忆增强的TRIME算法和基于粗细粒度联合剪枝和逐层蒸馏的CofiPruning算法。前者能够在不改变模型结构的基础上兼顾语言模型困惑度和检索速度方面的优势;而后者可以在保证下游任务准确度的同时实现更快的处理速度,具有更小的模型结构。陈丹琦 普

由于复杂的注意力机制和模型设计,大多数现有的视觉 Transformer(ViT)在现实的工业部署场景中不能像卷积神经网络(CNN)那样高效地执行。这就带来了一个问题:视觉神经网络能否像 CNN 一样快速推断并像 ViT 一样强大?近期一些工作试图设计 CNN-Transformer 混合架构来解决这个问题,但这些工作的整体性能远不能令人满意。基于此,来自字节跳动的研究者提出了一种能在现实工业场景中有效部署的下一代视觉 Transformer——Next-ViT。从延迟 / 准确性权衡的角度看,

3月27号,Stability AI的创始人兼首席执行官Emad Mostaque在一条推文中宣布,Stable Diffusion XL 现已可用于公开测试。以下是一些事项:“XL”不是这个新的AI模型的官方名称。一旦发布稳定性AI公司的官方公告,名称将会更改。与先前版本相比,图像质量有所提高与先前版本相比,图像生成速度大大加快。示例图像让我们看看新旧AI模型在结果上的差异。Prompt: Luxury sports car with aerodynamic curves, shot in a

译者 | 李睿审校 | 孙淑娟近年来, Transformer 机器学习模型已经成为深度学习和深度神经网络技术进步的主要亮点之一。它主要用于自然语言处理中的高级应用。谷歌正在使用它来增强其搜索引擎结果。OpenAI 使用 Transformer 创建了著名的 GPT-2和 GPT-3模型。自从2017年首次亮相以来,Transformer 架构不断发展并扩展到多种不同的变体,从语言任务扩展到其他领域。它们已被用于时间序列预测。它们是 DeepMind 的蛋白质结构预测模型 AlphaFold

人工智能就是一个「拼财力」的行业,如果没有高性能计算设备,别说开发基础模型,就连微调模型都做不到。但如果只靠拼硬件,单靠当前计算性能的发展速度,迟早有一天无法满足日益膨胀的需求,所以还需要配套的软件来协调统筹计算能力,这时候就需要用到「智能计算」技术。最近,来自之江实验室、中国工程院、国防科技大学、浙江大学等多达十二个国内外研究机构共同发表了一篇论文,首次对智能计算领域进行了全面的调研,涵盖了理论基础、智能与计算的技术融合、重要应用、挑战和未来前景。论文链接:https://spj.scien

说起2010年南非世界杯的最大网红,一定非「章鱼保罗」莫属!这只位于德国海洋生物中心的神奇章鱼,不仅成功预测了德国队全部七场比赛的结果,还顺利地选出了最终的总冠军西班牙队。不幸的是,保罗已经永远地离开了我们,但它的「遗产」却在人们预测足球比赛结果的尝试中持续存在。在艾伦图灵研究所(The Alan Turing Institute),随着2022年卡塔尔世界杯的持续进行,三位研究员Nick Barlow、Jack Roberts和Ryan Chan决定用一种AI算法预测今年的冠军归属。预测模型图


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

VSCode Windows 64-bit Download
A free and powerful IDE editor launched by Microsoft

PhpStorm Mac version
The latest (2018.2.1) professional PHP integrated development tool

DVWA
Damn Vulnerable Web App (DVWA) is a PHP/MySQL web application that is very vulnerable. Its main goals are to be an aid for security professionals to test their skills and tools in a legal environment, to help web developers better understand the process of securing web applications, and to help teachers/students teach/learn in a classroom environment Web application security. The goal of DVWA is to practice some of the most common web vulnerabilities through a simple and straightforward interface, with varying degrees of difficulty. Please note that this software

Zend Studio 13.0.1
Powerful PHP integrated development environment

EditPlus Chinese cracked version
Small size, syntax highlighting, does not support code prompt function
