令人惊讶的是：即使是大型语言模型，现在只需要8-10GB的显存就能进行微调，这大大降低了AI模型训练的硬件门槛，使更多研究者和开发者能够参与模型定制。

令人惊讶的是：AI训练效率的提升速度令人震惊。在短短6年内，语言模型的训练时间从167分钟缩短到不到4分钟，效率提升了40多倍。这种进步远超摩尔定律预测的5倍改进，展示了AI硬件和算法的飞速发展。

令人惊讶的是：AI模型训练速度在6年内提升了约50倍，远超摩尔定律预测的5倍。这种性能提升不仅来自硬件改进，还来自软件优化和算法创新。这一事实打破了人们对技术进步速度的传统认知，展示了AI领域独特的加速发展模式。

令人惊讶的是：即使使用H100-equivalents作为标准测量单位，也无法完全反映不同芯片类型在真实世界中的性能差异，这表明我们对AI计算能力的测量可能存在系统性偏差，影响我们对AI发展速度的准确理解。

Gemma 4 E2B/E4B 原生预训练 140+ 语言，且强调「超越翻译、理解文化语境」。对 AI 硬件产品而言这个参数意义重大：一个能在设备端离线处理中文、理解文化背景的 2-4B 模型，意味着本地化 AI 硬件（录音笔、学习机、会议设备）无需依赖国内厂商 API，直接用 Gemma 4 就能构建多语言理解能力。

大多数人认为手机硬件需要不断升级才能运行最新的AI功能，但作者认为技术压缩速度如此之快，以至于现有手机在升级前就能运行曾经的顶级模型，这颠覆了人们对硬件更新周期的认知。

https://web.archive.org/web/20260216112337/https://wccftech.com/western-digital-has-no-more-hdd-capacity-left-out/ Western Digital says consumer sales of HDDs is now just 5% of their sales, everything else going to AI companies. 2026 WD is already sold out. For 2027 and 2028 there is already a few deals in place too.

• comma.ai operates its own $5M data center in-office to handle model training, metrics, and data storage, avoiding the "cloud tax."
• The facility consumes approximately 450kW at peak; power costs in San Diego (over 40c/kWh) totaled over $540,000 in 2025.
• Cooling is achieved using pure outside air with dual 48” intake and exhaust fans, utilizing a PID loop to manage temperature and humidity.
• The compute cluster consists primarily of 600 GPUs across 75 "TinyBox Pro" machines built in-house for cost efficiency and easier repairability.
• Storage is handled by several racks of Dell R630/R730 servers with ~4PB of total SSD storage, favoring speed and random access over redundancy.
• The software stack is kept simple to ensure 99% uptime, utilizing Ubuntu (pxeboot), Salt for management, and "minikeyvalue" for distributed storage.
• By owning their hardware, comma.ai estimates they saved $20M+ compared to equivalent compute costs in a public cloud environment.

Hacker News Discussion

• Users discussed the spectrum of infrastructure, ranging from pure Cloud (low cap-ex, high op-ex) to colocation and on-prem (high cap-ex, high skill requirement).
• A primary concern raised was "brain drain"—on-prem setups can become "legacy debt" if the senior engineers who built the custom systems leave without documenting unwritten knowledge.
• Commenters noted that AWS and other cloud providers are incentivized to keep architectures complex (microservices, serverless) to increase billing, whereas on-prem encourages efficiency.
• There was a debate regarding "software freedom" and the "WhatsApp effect," where small, highly motivated teams can outperform massive corporations by using lean, self-hosted stacks.
• Some users highlighted that while AWS pricing is expected to rise due to hardware costs, the "Quality of Life" and managed services still justify the cost for many startups without comma's scale.

comma-ai #self-hosting #datacenter #hardware-engineering

ICs as hardware versions of AI. Interesting this is happening. Who are the players, what is on those chips? In a sense this is also full circle for neuronal networks, back in the late 80s / early 90s at uni neuronal networks were made in hardware, before software simulations took over as they scaled much better both in number of nodes and in number of layers between inputs and output. #openvraag Any open source hardware on the horizon for AI? #openvraag a step towards an 'AI in the wall' Vgl [[AI voor MakerHouseholds 20190715141142]] [[Everymans Allemans AI 20190807141523]]

This journal clearly explains the use of technologies with special aid people how a certain group can leverage it, while also touch basing on what are the challenges which special aid people face financially.