AIxiv专栏是机器之心发布学术、技术内容的栏目。过去数年，机器之心AIxiv专栏接收报道了2000多篇内容，覆盖全球各大高校与企业的顶级实验室，有效促进了学术交流与传播。

---

如果您有优秀的工作想要分享，欢迎投稿或者联系报道。投稿邮箱：liyazhou@jiqizhixin.com；zhaoyunfeng@jiqizhixin.com

本文论文一作是南开大学统计与数据科学学院研二硕士生杨雨辰，指导老师为南开大学统计与数据科学学院的徐君副教授。徐君老师团队的研究重点是计算机视觉、生成式AI和高效机器学习，并在顶级会议和期刊上发表了多篇论文，谷歌学术引用超过4700次。

自从大型Transformer模型逐渐成为各个领域的统一架构，微调就成为了将预训练大模型应用到下游任务的重要手段。然而，由于模型的尺寸日益增大，微调所需要的显存也逐渐增加，如何高效地降低微调显存就成了一个重要的问题。此前，微调Transformer模型时，为了节省显存开销，通常的做法是使用梯度检查点（gradient checkpointing，也叫作激活重算），以牺牲训练速度为代价降低反向传播（Backpropagation, BP）过程中的激活显存占用。

最近，由南开大学统计与数据科学学院徐君老师团队发表在ICML 2024上的论文《Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation》提出通过更改反向传播（BP）过程，在不增加计算量的情况下，显著减少峰值激活显存占用。

• 论文：Reducing Fine-Tuning Memory Overhead by Approximate and Memory-Sharing Backpropagation
• 论文链接：https://arxiv.org/abs/2406.16282
• 项目链接：https://github.com/yyyyychen/LowMemoryBP

文章提出了两种反向传播改进策略，分别是Approximate Backpropagation（Approx-BP）和Memory-Sharing Backpropagation（MS-BP）。Approx-BP和MS-BP分别代表了两种提升反向传播中内存效率的方案，可以将其统称为LowMemoryBP。无论是在理论还是实践意义上，文章都对更高效的反向传播训练提供了开创性的指导。

在理论显存分析中，LowMemoryBP可以大幅降低来自激活函数和标准化层的激活显存占用，以ViT和LLaMA为例，可以对ViT微调降低39.47%的激活显存，可以对LLaMA微调降低29.19%的激活显存。

在实际实验中，LowMemoryBP可以有效地使包括ViT, LLaMA, RoBERTa, BERT, Swin在内的Transformer模型微调峰值显存占用降低20%~30%，并且不会带来训练吞吐量和测试精度的损失。

Approx-BP

在传统反向传播训练中，激活函数梯度的反向回传是严…

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

740 TFLOPS！迄今最强 FlashAttention 来了。

随着大型语言模型（LLM）加速落地，扩展模型上下文窗口变得越来越重要。

---

然而，Transformer 架构的核心 —— 注意力层的时间复杂度和空间复杂度与输入序列长度的平方成正比。这使得扩展模型上下文窗口存在挑战。

2022 年，一种快速、内存高效的注意力算法 ——FlashAttention 问世，该算法无需任何近似即可加速注意力并减少内存占用。

FlashAttention 对注意力计算进行重新排序的算法，并利用 tiling 和重计算来显著加快计算速度，将内存使用量从序列长度的二次减少到线性。

2023 年，研究团队宣布推出 FlashAttention-2，在算法、并行化和工作分区等方面有了显著改进。

现在，来自 Meta、英伟达、Together AI 等机构的研究者宣布推出 FlashAttention-3，它采用了加速 Hopper GPU 注意力的三种主要技术：

• 通过 warp-specialization 重叠整体计算和数据移动；
• 交错分块 matmul 和 softmax 运算；
• 利用硬件支持 FP8 低精度的不连贯处理。

FlashAttention-3 的速度是 FlashAttention-2 的 1.5-2.0 倍，高达 740 TFLOPS，即 H100 理论最大 FLOPS 利用率为 75%。使用 FP8，FlashAttention-3 的速度更是接近 1.2 PFLOPS。

FlashAttention-3 的改进将带来：

• 更高效的 GPU 利用率：H100 理论最大 FLOPS 利用率为 75%，而之前仅为 35%。这使得 LLM 的训练和运行速度比以前的版本快得多。
• 较低精度下更好的性能：FlashAttention-3 可以在保持精度的同时使用较低精度的数字 (FP8)。这可以实现更快的处理速度并可能降低内存使用量，从而为运行大规模人工智能操作的客户节省成本并提高效率。
• 能够在 LLM 中使用更长的上下文：通过加速注意力机制，FlashAttention-3 使 AI 模型能够更有效地处理更长的文本片段。这使得应用程序能够理解并生成更长、更复杂的内容而不会减慢速度。

论文标题：FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
论文地址：https://tridao.me/publications/flash3/flash3.pdf
论文作者之一、FlashAttention1-3 版本的参与者 Tri Dao 表示：FlashAttention 被广泛用于加速 Transformers，已经使注意力速度提高了 4-8 倍，但尚未利用现代 GPU。因而他们发布了 FlashAttention-3：在 FP16 上速度提高了 1.5-2 倍，在 H100 上高达 740 TFLOPS（75% 实用性），FP8 接近 1.2 PFLOPS！

Hopper GPU 硬件特性：WGMMA、TMA、FP8

…

扩展阅读：

• 斯坦福提出新型Attention算法！提速2-4倍，BERT单节点训练最快
• 比标准Attention提速5-9倍，大模型都在用的FlashAttention v2来了

参考链接：

• https://tridao.me/blog/2024/flash3/

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

论老黄卖铲子的技术含量。

2019 年 2 月，OpenAI 发布了 GPT-2，因为在文本生成上的优异表现，以及对于预训练 Transformer 架构的充分运用，被认为是如今大预言模型的「始祖」。

---

五年后的今天，训练 GPT-2 这样 15 亿参数的大模型，只需要花费 672 美元，在一个 8XH100 的 GPU 节点上跑 24 个小时就可以搞定了。

本周四，前特斯拉 Autopilot 负责人、OpenAI 科学家 Andrej Karpathy 在他纯 C 语言复现 GPT-2 大模型的项目「llm.c」的最新进展中分享了他的训练心得：

令人难以置信的是，由于计算硬件（英伟达 H100 GPU）、软件（CUDA、cuBLAS、cuDNN、FlashAttention 等）和数据质量（例如 FineWeb-Edu 数据集）的改进，过去 5 年间，大语言模型的训练成本大幅下降。Karpathy 表示，对于此次实践，算法遵循 GPT-2/3 论文基本保持原样不变。

当年 OpenAI 训练 GPT-2 花费了多少钱？这是个至今仍然未知的数字。Karpathy 粗略地估算认为是这回成本的 100 倍，大概要到 10 万美元的量级。

基本相同的任务，运行效率却有天壤之别，这体现了近几年来 AI 领域和算力基础设施的飞速发展。

由于 llm.c 是在 C/CUDA 中 GPT 训练的直接实现，因此要求其实很少 —— 不需要 conda 环境、Python 解释器、pip 安装等。如果你也要尝试，可以启动云 GPU 节点（例如在 Lambda 上），可选择安装 NVIDIA cuDNN、NCCL/MPI，下载 .bin 数据分片，编译并运行，几分钟后即可开始。

然后，你就可以等待 24 小时，然后欣赏通用大语言模型的能力了。

「对于 llm.c 项目来说，这是一个非常好的节点。因为整个项目都是从我考虑为教育视频重现 GPT-2 开始的。我遇到一些 PyTorch 的东西时卡住了，然后愤怒地退出，再用 C/CUDA 从头开始编写整个项目，」Karpathy 表示。「这让我踏上了比预想更长的旅程。但它非常有趣，我学到了更多的 CUDA，一路上结交了朋友，现在的 llm.c 真的很棒。它有大约 5000 行代码，编译和步骤非常快，几乎不需要等待。它具有恒定的内存占用，它以混合精度进行训练，使用 NNCL 分布在多节点上。它是按位确定性的，并且徘徊在 MFU 的 50% 左右。所以它很 ok。」

对于 llm.c 项目而言，越做似乎挖得坑越大。Andrej Karpathy 对目前的运行结果仍然不是 100% 满意 —— 他认为评估应该更好，训练应该更稳定，尤其是在较长时间运行的较大模型尺寸下。

他还预告了一些有趣的新方向：fp8（即将推出）、推理、微调、多模态（VQVAE 等）、更现代的架构（Llama/Gemma）。llm.c 的目标仍然是为功能齐全的 LLM 智能体提供简单、最小、干净的训练堆栈，直接使用 C/CUDA，并包含配套的教育材料，可以让许多初学者快速了解这个令人敬畏的领域。

说完了这么多，该看看 24 小时训练 GPT-2 的成果了：Karpathy 使用更长的 400B token GPT-2 运行（从 33B token 增加），效果良好，直到 330B（达到 61% HellaSwag，远高于这个大小的 GPT-2 和 GPT-3），然后在这个图之后不久爆炸了。目前作者还在继续进行研究。

接下来看详细项目介绍。

GitHub 地址：https://github.com/karpathy/llm.c/discussions/677

训练。使用 llm.c 训练 GPT-2 非常简单，因为它是用 C/CUDA 编写的，因此不需要 minconda、Python、PyTorch 等。你只需一个 8XH100 GPU box，Karpathy 建议从 Lambda Labs 购买一个。

不过 llm.c 在计算上很灵活，如果你只有 1 个 GPU，仍然可以训得 GPT-2，这时你需要等待 8 天而不是 1 天。如果你有 16 个 GPU（例如使用新的 Lambda 1 Click Clusters），则能够训练多节点，这时只需等待 12 小时。启动节点后，以下是训练 GPT-2 的完整说明：

install cudnn so we can use FlashAttention and run fast (optional)

https://developer.nvidia.com/cudnn-downloads

for me, CUDA 12 (run nvcc --version) running on Linux x86_64 Ubuntu 22.04

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/cuda-keyring_1.1-1_all.deb
sudo dpkg -i cuda-keyring_1.1-1_all.deb
sudo apt-get update
sudo apt-get -y install libcudnn9-dev-cuda-12

“install” cudnn-frontend to ~/

git clone https://github.com/NVIDIA/cudnn-frontend.git

install MPI (optional, if you intend to use multiple GPUs)

(you might also have to install NVIDIA NCCL if it doesn’t come with your setup)

sudo apt -y install openmpi-bin openmpi-doc libopenmpi-dev

download and enter llm.c repo

git clone https://github.com/karpathy/llm.c.gitcd llm.c

download the “starter pack” (~1GB download)

contains GPT2-124M weights (used in tests), tokenizer, eval data .bin s

./dev/download_starter_pack.sh

download the training dataset (FineWeb-Edu 100B token) .bin data shards

note: this is a total of 1001 data shards. If you only want to test things

out and don’t want to do an actual run, feel free to append the number of

training shards to download (e.g. for just 10 shards: ./edu_fineweb.sh 10)

the full dataset is ~200GB, we can store it here in dev/data directory.

cd dev/data
./edu_fineweb.sh

compile (~1 min 1st time for cuDNN mostly, few sec from then on)

cd ../../
make train_gpt2cu USE_CUDNN=1

and train! (wait 24 hours here)

mpirun -np 8 ./train_gpt2cu
-i “dev/data/edu_fineweb100B/edu_fineweb_train_*.bin”
-j “dev/data/edu_fineweb100B/edu_fineweb_val_*.bin”
-o “log_gpt2_1558M”
-v 250 -s 300000 -g 384
-h 1
-b 16 -t 1024
-d 1048576
-r 0
-z 1
-c 0.1
-k “cosine”
-l 0.0006
-q 0.1
-u 700
-n 2000
-x 32000
-ge 1
-y 1
-e “d48”
开始优化：

num_parameters: 1557686400 => bytes: 3115372800
allocated 2971 MiB for model parameters
batch_size B=16 * seq_len T=1024 * num_processes=8 and total_batch_size=1048576
=> setting grad_accum_steps=8
created directory: log_gpt2_1558M
allocating 40409 MiB for activations
val loss 11.129390
allocating 2971 MiB for parameter gradients
allocating 742 MiB for AdamW optimizer state m
allocating 742 MiB for AdamW optimizer state v
allocating 742 MiB for master copy of params
step 1/32000 | loss 11.133732 (+nanz)| norm 52.9732 (+nanz)| lr 8.57e-07 | 3056.36 ms | 42.6% bf16 MFU | 343080 tok/s
step 2/32000 | loss 10.539388 (+nanz)| norm 43.5996 (+nanz)| lr 1.71e-06 | 2747.19 ms | 47.4% bf16 MFU | 381690 tok/s
step 3/32000 | loss 9.894109 (+nanz)| norm 23.2229 (+nanz)| lr 2.57e-06 | 2753.25 ms | 47.3% bf16 MFU | 381259 tok/s
step 4/32000 | loss 9.566241 (+nanz)| norm 28.4920 (+nanz)| lr 3.43e-06 | 2741.47 ms | 47.5% bf16 MFU | 381690 tok/s
step 5/32000 | loss 9.482848 (+nanz)| norm 23.7817 (+nanz)| lr 4.29e-06 | 2752.07 ms | 47.3% bf16 MFU | 381507 tok/s
step 6/32000 | loss 9.332832 (+nanz)| norm 15.9113 (+nanz)| lr 5.14e-06 | 2751.01 ms | 47.3% bf16 MFU | 381431 tok/s
step 7/32000 | loss 9.165650 (+nanz)| norm 10.5941 (+nanz)| lr 6.00e-06 | 2753.03 ms | 47.3% bf16 MFU | 381327 tok/s
step 8/32000 | loss 9.132234 (+nanz)| norm 16.2733 (+nanz)| lr 6.86e-06 | 2748.91 ms | 47.3% bf16 MFU | 381348 tok/s
step 9/32000 | loss 9.097384 (+nanz)| norm 12.1342 (+nanz)| lr 7.71e-06 | 2748.73 ms | 47.3% bf16 MFU | 381367 tok/s
step 10/32000 | loss 9.072879 (+nanz)| norm 10.5923 (+nanz)| lr 8.57e-06 | 2749.40 ms | 47.3% bf16 MFU | 381369 tok/s
…

每一步大约需要 2.75 秒，共有 32000 步，所以现在我们等待 24 小时左右。在每一步中，训练运行都会占用约 100 万个 FineWeb-EDU token（这些来自互联网的教育网页），并更新模型的 15.58 亿个权重，使其能够更好地预测序列中的下一个 token。最后将总共处理 32000 * 1048576 = 33.6B 个 token。随着更好地预测下一个 token，损失会下降。范数将稳定在 0.1-1 左右，学习率在前面几步预热。模型 flops 利用率 (MFU) 约为 50%，非常高效。

等待 24 小时后，就可以使用 dev/vislog.ipynb jupyter 笔记本可视化 main.log 日志文件。为此，你还需要安装 Python 和 matplotlib。

参数指南。OpenAI 发布的 GPT-2 包含模型权重，但细节很少；而 GPT-3 版本没有权重，但细节很多。因此，在许多情况下，我们遵循 GPT-3 论文超参数，因为 GPT-2 论文的信息非常少。具体参见原项目。

内存指南。大多数人可能面临的最大限制是他们的 GPU 没有 80GB。没关系，你仍然可以运行上面的所有内容，只是运行速度会更慢。因此如果模型不适配，你会怎么做？最重要的是微批大小 - b。尝试减小它，但将其保持在合适的数字，例如 16 → 8 → 4 → 2 → 1。从那里开始，尝试使用重计算设置 -r，即 0（最快且有大量内存）、1（稍微慢一点，但节省大量内存）或 2（稍微慢一点，节省较少内存）。

你可以做的下一件事是禁用 fp32 中的主权重，可以使用 - w 0 （默认值 1）来执行此操作。我们不会维护 fp32 参数副本。根据经验，在之前的几次运行中，这似乎没问题，可能是因为使用了随机舍入。如果还不适合，则可以尝试使用 -t 来减少最大序列长度，默认值为 1024，你可以将其降低到 512、256 等。但现在你会让模型变得更糟，因为它的最大注意力跨度正在减少。

代码。Karpathy 对 llm.c 略有偏爱，认为它非常漂亮：

• 它只需要基本的 CUDA 依赖项即可运行。
• 它是 C/CUDA 中直接、最小且易读的实现。llm.c 共有约 5,000 行 C/CUDA 代码。这里尝试主要使用 C，而不是 C++，以保持简单。神经网络训练只是对单个浮点数组进行相同的简单算术运算（如 +、-、、/）的一个 while 循环，它实际上不应该那么复杂。
• 它编译和运行非常快（几秒钟），因此可以进行更多步进和更短等待。
• 它在开始时一次性分配其所有 GPU 内存，从那时起在训练期间具有完全恒定的内存占用。因此，一旦开始步进，就可以在剩余的运行中表现良好并且不会内存用完（OOM）。
• 它是按位（bitwise）确定的。
• 它非常高效，略低于～50% 的 MFU。

主要入口点和大部分代码位于文件 train_gpt2.cu 中。它包含 GPT-2 模型定义和约 2,000 LOC 的训练 loop，并从 llmc 目录导入了一堆带有各种实用程序和各个层实现的辅助文件。最后 cloc llmc

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

伊隆·马斯克的X公司收到了欧盟的警告，监管机构表示其用户的蓝勾号系统具有欺骗性，并违反了其具有里程碑意义的社交媒体规定。

---

欧洲委员会宣布了一项调查的初步结果，称该平台未遵守数字服务法。如果初步调查结果得到确认，X可能面临高达全球营业额的6%的罚款。

欧盟的行政机构表示，X在三个领域违反了法案：通过将蓝勾号授予不可信账户来欺骗用户；未向研究人员提供公开可获取的数据，如帖子；以及运行了不完善的广告库，阻止研究人员审查平台上的广告，包括故意误导性的广告。点击欧盟选举前，“信息滥用网络”在X上频繁，报告称读取更多。

作为欧盟内部市场委员和该法案背后的重要人物，蒂埃里·布勒顿表示：“过去，蓝勾号曾意味着信息可靠的来源。现在对于X而言，我们的初步观点是它们欺骗用户并侵犯了DSA。” 布勒顿表示，X有权回应，但“如果我们的观点得到确认，我们将征收罚款并要求进行重大改变”。

委员会仍在调查X是否未能在平台上处理非法内容和虚假信息而违反了DSA。周五的声明代表了委员会对DSA下的公司的首次初步调查结果。它还根据该法案对TikTok和马克·扎克伯格的Meta公司展开了诉讼程序。

马斯克在2022年购买X后还未发布正式财务数据。但特斯拉首席执行官承认公司的广告收入——过去是其主要收入来源——自他购买该公司以来已大幅下降，并表示2023年其广告收入将达到30亿美元。

X引入了付费的蓝勾号作为其高级订阅服务的一部分，该服务在欧盟每月售价为8欧元，布鲁塞尔称其已被不良行为者滥用。

委员会表示：“有证据表明，激起恶意行为者滥用‘验证的账户’来欺骗用户。” 马斯克在2022年首次尝试改革高级服务时，大量冒名账户出现，恶作剧者利用这一变化购买蓝勾号并假装成任天堂和特斯拉等公司，以及教皇和乔治·W·布什等公众人物。

已联系X公司以获取评论。

探索更多关于这些主题:

• 伊隆·马斯克
• 社交媒体
• 欧洲联盟
• 欧洲委员会
• 数字媒体
• 监管机构
• 新闻

分享此内容

重复利用本内容。

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

一度，所有的故事都是由人类独自创作的。

---

如今，研究人员发现人工智能可能帮助作者讲述一个故事。一项研究表明，由AI系统ChatGPT生成的想法可以帮助那些缺乏固有才华的作家提升创造力，尽管这是以牺牲多样性为代价的。来自埃克塞特大学的研究合著者奥利弗·豪泽教授表示，研究结果提出了一个社会困境。“对于你个人来说，使用AI可能是有益的，但作为一个社会，如果每个人都使用AI，我们可能会失去独特想法的多样性。”他说，“而且，可以说，对于创意努力，有时我们可能需要‘狂放’和‘不寻常’的想法。”

团队请 293 人命名尽可能相互不同的 10 个单词，以探究参与者的固有创造力。然后，研究人员随机分配参与者三个主题之一 - 丛林冒险、航海冒险或外星球面临的冒险，并要求他们编写一个适合青少年和年轻成年人的八句故事。虽然三分之一的参与者没有得到任何帮助，其他人被分成两组 - 允许使用由 ChatGPT 预先生成的一个三句开头想法的人，以及可以要求五个这样的想法的人。

总体上，100 名参与者中，82 人接受了一个由 AI 生成的想法的提供，98 名被提供五个这样的想法的参与者中，有 93 人至少采纳了一个，几乎四分之一的人请求了全部五个。接下来的 600 名参与者，不知道是否使用了由 AI 生成的想法，阅读了由此产生的故事，并在九点量表上评定与新颖性和实用性相关的因素 - 例如故事是否适合出版。发表在《科学进展》期刊上的结果显示，获得使用 AI 的人提高了这些评分，更大的使用与更大的效果联系在一起：有选项可以得到五个 AI 生成的想法的人，其故事的新颖性评分平均增加了 8.1%，而实用性评分平均提高了 9%。

“效果大小并不是非常大，但它们在统计上是显著的，”来自伦敦大学学院的研究合著者阿尼尔·多西博士说。有AI生成想法选项的人写的故事也被认为更加令人愉悦，更可能有情节转折并且更好地书写。

然而，受益最大的是固有创意较低的作家。“我们并没有发现那些固有创意最强的人的故事被从AI想法中‘超频’- 这群人不管是否使用AI都非常有创意，”多西说。该团队还发现，能够获得 AI 生成想法的参与者创作的故事具有更大的相似性，多西表示这可能归因于AI生成相对可预测的故事想法。

豪泽表示，这类研究很重要。“评估AI的使用对于确保我们在不陷入潜在缺陷的情况下获得这一潜在变革性技术的好处至关重要。”他说。

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

While executives express high confidence in their organisations’ AI capabilities, they simultaneously acknowledge significant barriers to further adoption.{ width=50% }

---

Research from Zartis found that 85% of UK tech executives rate their existing workforce’s combined AI knowledge and expertise as ‘skilled’, with over half (51%) considering it ‘highly skilled’. This confidence, however, is juxtaposed against concerns about obstacles preventing wider AI implementation.
AI adoption is nearly universal among UK tech companies, with 94% of executives reporting some form of AI use in their organisations. The remaining 6% are still in the exploration or research phase. Notably, not a single respondent claimed to be completely avoiding AI.
Industry pressure appears to be a significant factor driving AI adoption, with 40% of executives feeling compelled to prioritise AI investment due to widespread momentum around the technology.
Despite the enthusiasm, several barriers hinder full AI adoption. Budget restrictions (41%), shortage of AI talent (38%), and technical complexity (35%) were cited as primary obstacles. Executives also expressed concerns about integration challenges (44%), cost and ROI uncertainty (42%), and data privacy and IP security (38%).
Michal Szymczak, Head of AI Strategy at Zartis, commented on this apparent contradiction: “AI adoption isn’t some ‘on or off’ switch. To a lot of businesses, it involves a significant financial investment, and there are complex questions to grapple with surrounding data privacy, or integration with existing technology stacks. That makes executives’ confidence in their company’s AI skill set rather ironic. They puff their chests out, while simultaneously pointing to all the obstacles that could stop them in their tracks.”
The financial aspect of AI adoption presents a mixed picture. While 42% of executives cited ROI uncertainty as a primary concern, 53% view cost savings through improved efficiency as one of the most significant long-term benefits of adopting AI.
Investment in AI capabilities is substantial, with 93% of companies spending at least £100,000 in 2024, and 44% allocating £500,000 or more. Software development emerges as the most popular area for AI investment (59%), followed by quality assurance (44%) and DevOps and automation (44%).
Angel Benito, CTO at Zartis, offered insight into the investment trends: “There is significant pressure on organisations to keep up with AI development or risk being left behind. This explains why many are investing despite the uncertainty about ROI. They see the potential for long-term cost savings but need a well-curated plan to implement the changes. It’s crucial to understand that it’s not just about the tools; it’s about the people.”
As companies navigate the AI landscape, their focus varies. The top three cited utilities of AI for software development are AI-powered copilot tools for coding (53%), improved continuous integration and deployment (52%), and enhanced team communication and collaboration (46%).
This study follows recent Zartis research indicating that over three-quarters of UK tech executives favour increased government oversight of AI, with many anticipating benefits from regulations such as the EU AI Act.
As the AI revolution continues to unfold, it’s clear that UK tech companies are eager to embrace the technology, even as they grapple with its complexities and challenges.
(Photo by Nick Fewings)
See also: Microsoft and Apple back away from OpenAI board

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore other upcoming enterprise technology events and webinars powered by TechForge here.
Tags: ai, artificial intelligence, enterprise, report, research, study.

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

Artificial intelligence is transforming numerous industries, and the gaming industry is no exception.{ width=50% }

---

From developing sophisticated game mechanics to enhancing player experiences, AI’s influence is increasingly pervasive. This article explores how AI is revolutionising game design and player experiences at a rapid pace.

The role of AI in game design:

Procedural content generation:
One of the most significant impacts of AI in game design is procedural content generation (PCG). PCG allows developers to create vast, dynamic game worlds and experiences that can adapt and evolve in real-time. Instead of manually designing every aspect of a game, developers can use algorithms to generate content such as levels, landscapes, and even entire narratives. This approach ensures that each player’s experience is unique and fresh, contributing to greater engagement and replayability.

Enhanced Non-Player Characters (NPCs):
AI is crucial in developing more realistic and intelligent NPCs. Traditionally, NPCs followed predetermined scripts, making their actions predictable and sometimes frustrating for players. However, with AI, NPCs can now learn and adapt based on player behavior, creating more engaging and immersive interactions. These AI-driven NPCs offer a more personalized and dynamic gaming experience, making the virtual world feel more alive and responsive to player actions.

Balancing game difficulty:
Balancing the difficulty of a game is a challenging task for developers. AI can help by analyzing player behavior and adjusting the difficulty level in real-time. This dynamic difficulty adjustment ensures that players remain engaged and challenged without feeling overwhelmed, maintaining an optimal balance that caters to different skill levels and preferences.

AI in enhancing player experiences:

Personalized gaming experiences:
AI’s ability to analyze vast amounts of data allows for highly personalized gaming experiences. By tracking player preferences, behavior, and performance, AI can tailor content and recommendations to individual players. This personalization can range from suggesting in-game items and quests to adjusting the game’s storyline based on player choices. Such tailored experiences make players feel more connected to the game, enhancing their overall enjoyment and satisfaction.

Real-time analytics and feedback:
AI enables real-time analytics and feedback, allowing developers to understand player behavior and preferences better. This data can be used to improve game design, fix bugs, and introduce new features. Players benefit from a continuously evolving game that meets their expectations and desires. Real-time analytics also help developers create more engaging and balanced gameplay, addressing issues and opportunities as they arise.

The impact of AI on online slot gaming:

One of the most intriguing applications of AI in gaming is in online casinos. Innovative bitcoin pokies or slot machines leverage AI to offer a more engaging and secure gambling experience, particularly through the use of bitcoin payments on the casino platform. AI-driven pokies can analyze player behavior to create personalized gaming experiences, ensuring that each slot gaming session is unique and tailored to the player’s preferences. Moreover, the use of AI in bitcoin pokies can enhance security by detecting and preventing fraudulent activities. AI-powered customer support is also revolutionizing customer support in online casinos. AI-driven chatbots and virtual assistants provide instant support to players, addressing their queries and issues in real-time. These AI-powered systems can handle a wide range of tasks, from account management to troubleshooting technical problems, and therefore enhance the overall player experience.

Enhanced game fairness and transparency:
Fairness and transparency are crucial in online gambling. AI helps ensure that games are fair by analyzing and monitoring gameplay to detect any anomalies or unfair practices. This transparency builds trust between players and online casinos, fostering a more positive and secure gaming environment. AI algorithms can audit and verify the randomness of outcomes in bitcoin pokies, assuring players that the games are not rigged and that they have a fair chance of winning.

Future prospects of AI in gaming:

Upcoming tech:
AI is set to play a significant role in the development of VR and AR games. By creating more immersive and responsive virtual environments, AI can enhance the realism and interactivity of VR and AR experiences. This advancement will open up new possibilities for game design and player engagement.

AI-driven storytelling:
The future of game storytelling lies in AI-driven narratives. AI can analyze player choices and adapt the storyline accordingly, creating dynamic and personalized narratives. This approach ensures that each player’s journey is unique, enhancing replayability and engagement.

Advanced player analytics:
As AI technology continues to evolve, the ability to analyze player data will become even more sophisticated. This advancement will enable developers to create more precise and personalized gaming experiences, further blurring the line between the virtual and real worlds.

Conclusion:
AI is undoubtedly revolutionizing game design and player experiences. From intelligent NPCs to personalized gaming and enhanced online experiences, the impact of AI is profound and far-reaching. In the future, we can expect even more innovative and immersive gaming experiences.

Tags: ai, artificial intelligence, Entertainment, Gaming, Video Games

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

While executives express high confidence in their organisations’ AI capabilities, they simultaneously acknowledge significant barriers to further adoption.{ width=50% }

---

Research from Zartis found that 85% of UK tech executives rate their existing workforce’s combined AI knowledge and expertise as ‘skilled’, with over half (51%) considering it ‘highly skilled’. This confidence, however, is juxtaposed against concerns about obstacles preventing wider AI implementation.
AI adoption is nearly universal among UK tech companies, with 94% of executives reporting some form of AI use in their organisations. The remaining 6% are still in the exploration or research phase. Notably, not a single respondent claimed to be completely avoiding AI.
Industry pressure appears to be a significant factor driving AI adoption, with 40% of executives feeling compelled to prioritise AI investment due to widespread momentum around the technology.
Despite the enthusiasm, several barriers hinder full AI adoption. Budget restrictions (41%), shortage of AI talent (38%), and technical complexity (35%) were cited as primary obstacles. Executives also expressed concerns about integration challenges (44%), cost and ROI uncertainty (42%), and data privacy and IP security (38%).
Michal Szymczak, Head of AI Strategy at Zartis, commented on this apparent contradiction: “AI adoption isn’t some ‘on or off’ switch. To a lot of businesses, it involves a significant financial investment, and there are complex questions to grapple with surrounding data privacy, or integration with existing technology stacks. That makes executives’ confidence in their company’s AI skill set rather ironic. They puff their chests out, while simultaneously pointing to all the obstacles that could stop them in their tracks.”
The financial aspect of AI adoption presents a mixed picture. While 42% of executives cited ROI uncertainty as a primary concern, 53% view cost savings through improved efficiency as one of the most significant long-term benefits of adopting AI.
Investment in AI capabilities is substantial, with 93% of companies spending at least £100,000 in 2024, and 44% allocating £500,000 or more. Software development emerges as the most popular area for AI investment (59%), followed by quality assurance (44%) and DevOps and automation (44%).
Angel Benito, CTO at Zartis, offered insight into the investment trends: “There is significant pressure on organisations to keep up with AI development or risk being left behind. This explains why many are investing despite the uncertainty about ROI. They see the potential for long-term cost savings but need a well-curated plan to implement the changes. It’s crucial to understand that it’s not just about the tools; it’s about the people.”
As companies navigate the AI landscape, their focus varies. The top three cited utilities of AI for software development are AI-powered copilot tools for coding (53%), improved continuous integration and deployment (52%), and enhanced team communication and collaboration (46%).
This study follows recent Zartis research indicating that over three-quarters of UK tech executives favour increased government oversight of AI, with many anticipating benefits from regulations such as the EU AI Act.
As the AI revolution continues to unfold, it’s clear that UK tech companies are eager to embrace the technology, even as they grapple with its complexities and challenges.
(Photo by Nick Fewings)
See also: Microsoft and Apple back away from OpenAI board

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo.
Explore other upcoming enterprise technology events and webinars powered by TechForge here.
Tags: ai, artificial intelligence, enterprise, report, research, study

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

While executives express high confidence in their organisations’ AI capabilities, they simultaneously acknowledge significant barriers to further adoption.{ width=50% }

---

Research from Zartis found that 85% of UK tech executives rate their existing workforce’s combined AI knowledge and expertise as ‘skilled’, with over half (51%) considering it ‘highly skilled’. This confidence, however, is juxtaposed against concerns about obstacles preventing wider AI implementation. AI adoption is nearly universal among UK tech companies, with 94% of executives reporting some form of AI use in their organisations. The remaining 6% are still in the exploration or research phase. Notably, not a single respondent claimed to be completely avoiding AI. Industry pressure appears to be a significant factor driving AI adoption, with 40% of executives feeling compelled to prioritise AI investment due to widespread momentum around the technology. Despite the enthusiasm, several barriers hinder full AI adoption. Budget restrictions (41%), shortage of AI talent (38%), and technical complexity (35%) were cited as primary obstacles. Executives also expressed concerns about integration challenges (44%), cost and ROI uncertainty (42%), and data privacy and IP security (38%). Michal Szymczak, Head of AI Strategy at Zartis, commented on this apparent contradiction: “AI adoption isn’t some ‘on or off’ switch. To a lot of businesses, it involves a significant financial investment, and there are complex questions to grapple with surrounding data privacy, or integration with existing technology stacks. That makes executives’ confidence in their company’s AI skill set rather ironic. They puff their chests out, while simultaneously pointing to all the obstacles that could stop them in their tracks.” The financial aspect of AI adoption presents a mixed picture. While 42% of executives cited ROI uncertainty as a primary concern, 53% view cost savings through improved efficiency as one of the most significant long-term benefits of adopting AI. Investment in AI capabilities is substantial, with 93% of companies spending at least £100,000 in 2024, and 44% allocating £500,000 or more. Software development emerges as the most popular area for AI investment (59%), followed by quality assurance (44%) and DevOps and automation (44%). Angel Benito, CTO at Zartis, offered insight into the investment trends: “There is significant pressure on organisations to keep up with AI development or risk being left behind. This explains why many are investing despite the uncertainty about ROI. They see the potential for long-term cost savings but need a well-curated plan to implement the changes. It’s crucial to understand that it’s not just about the tools; it’s about the people.” As companies navigate the AI landscape, their focus varies. The top three cited utilities of AI for software development are AI-powered copilot tools for coding (53%), improved continuous integration and deployment (52%), and enhanced team communication and collaboration (46%). This study follows recent Zartis research indicating that over three-quarters of UK tech executives favour increased government oversight of AI, with many anticipating benefits from regulations such as the EU AI Act. As the AI revolution continues to unfold, it’s clear that UK tech companies are eager to embrace the technology, even as they grapple with its complexities and challenges. (Photo by Nick Fewings) See also: Microsoft and Apple back away from OpenAI board

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is co-located with other leading events including Intelligent Automation Conference, BlockX, Digital Transformation Week, and Cyber Security & Cloud Expo. Explore other upcoming enterprise technology events and webinars powered by TechForge here.

Tags: ai, artificial intelligence, enterprise, report, research, study

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。

Care providers’ cultural attunement deficiency is a historic mental health care barrier for racial and ethnic minority communities, while bias remains a top pain point for healthcare AIConversational AI visionary mpathic is building culturally competent AI models for more equitable mental healthcare delivery, rooted in part in Wave’s inclusive provider-patient transcriptsNIH SBIR’s mission is to cultivate diverse ideas and strive to empower scientists and entrepreneurs to bring their discoveries to patients
mpathic, a leader in AI-powered actionable conversation analytics specializing in healthcare, life sciences, and client services, today announced it has received a highly competitive National Institutes of Health Small Business Innovation Research (SBIR) award for its transformative project “Empathy for Everyone: Generative AI that Improves Patient-Provider Cultural Attunement in Real Time.” NIH, the nation’s medical research agency, offers SBIR grants to innovative small healthcare and life sciences businesses to perform research and development to make medical breakthroughs that can be deployed to the masses. This grant will fund the expansion of mpathic’s generative AI tools to facilitate cultural attunement between mental health providers and patients, bridge the gap in mental healthcare for underserved racial and ethnic populations, and potentially improve therapeutic outcomes and patient satisfaction.
America’s mental healthcare system is failing people of color. Racial and ethnic minority communities report higher rates of mental illness compared to white…
…目前只提供部分内容翻译，如需完整翻译，请继续。{ width=60% }

---

---

感谢阅读！如果您对AI的更多资讯感兴趣，可以查看更多AI文章：GPTNB。