The bones of AI
1 Oct 2026 - Visit to page seen in GA
Not much extra information on this page. The quote and the summary can be moved to notes and my interpretation put in the body.
Never-the-less, it is good that one of my AI pages has actually been accessed. While I am no expert in the field of AI the warnings of the perils of it use the claim that "AI will end life as we know it" in the very near future cannot be ignored.
Tokens - AI Summary:
The following Copilot Summary was generated 9 Nov 2026:
ASICs are purpose-built chips optimized for specific AI tasks, offering higher efficiency and lower power consumption, while NVIDIA GPUs provide versatile, general-purpose AI acceleration with a strong software ecosystem.
What is an ASIC?
An Application-Specific Integrated Circuit (ASIC) is a chip designed for a single, specific task, such as a particular AI model or inference workload. By focusing exclusively on one function, ASICs can achieve extreme performance and exceptional power efficiency, often outperforming GPUs by 3–5x in performance per watt for targeted tasks
Classic examples include Google’s TPU, Amazon’s Trainium, and Meta’s MTIA, which are custom-built for large-scale AI inference and training
The trade-off is flexibility: ASICs cannot easily adapt to new models or workloads outside their design scope.
NVIDIA GPUs
NVIDIA GPUs are general-purpose accelerators capable of handling a wide range of AI workloads, from training large language models to inference tasks. They are highly versatile and supported by a mature software ecosystem, including CUDA and NVLink, which allows integration with CPUs and other accelerators
While GPUs are less power-efficient than ASICs for a single task, they provide rapid deployment and adaptability, making them ideal for startups or research environments experimenting with multiple AI architectures
ket Trends and Adoption
In 2026, the AI hardware market is experiencing a shift toward custom ASICs. Hyperscalers like Google, Amazon, and Meta are increasingly deploying their own ASICs to reduce costs and improve efficiency for large-scale inference workloads
TrendForce projects that ASIC shipments will grow by 44.6% in 2026, outpacing GPU growth at 16.1%
By 2028, ASIC shipments are expected to surpass GPU shipments, reflecting the economic advantage of purpose-built chips for high-volume AI tasks
Despite this, NVIDIA maintains a strategic edge through its software ecosystem, interconnect technologies like NVLink, and rapid time-to-market for new AI models
NVIDIA’s GPUs remain dominant in scenarios requiring flexibility, multi-model experimentation, or rapid deployment, while ASICs excel in high-volume, repetitive inference tasks where efficiency and cost per operation are critical.
I now have a page on ASICs - uploaded so that I can review on my iPad.
AI Tokens
Initially I was surprised that the maker of graphics chips, Nvidia, were one of the leaders in the manufacture of semiconductors for AI. In fact it makes sense when you consider the processes of AI in that the algorithms break down large data items into small tokens.
Nvidia say:
AI tokens are tiny units of data that come from breaking down bigger chunks of information. AI models process tokens to learn the relationships between them and unlock capabilities including prediction, generation and reasoning. The faster tokens can be processed, the faster models can learn and respond. The goal is to achieve the fastest processing time and lowest cost per token to optimize AI infrastructure and maximize revenue generation.
Top
When a search was made for AI Tokens the following Summary was generated (April 2026):
Tokens are the fundamental units of data processed by AI models during training and inference. They represent smaller components of text, such as words, subwords, punctuation, or special markers. For example, the word "unbelievable" might be tokenized into "un," "believ," and "able." Tokenization enables models to handle diverse inputs efficiently, even for unseen words.
Role of Tokens in AI Models
During training, tokens are used to teach models relationships between data points. For instance, a model predicts the next token in a sequence and adjusts its parameters based on the accuracy of its predictions. This iterative process continues until the model achieves a desired level of accuracy.
In inference, tokens are used to process user inputs and generate outputs. For example, a language model might convert a text prompt into tokens, process them, and generate a response as tokens, which are then converted back into human-readable text.
Context Windows and Their Importance
A context window defines the number of tokens a model can process at once. It determines how much prior information the model can consider when making predictions or generating text. For example, GPT-4 supports context windows of up to 32,768 tokens in some configurations.
Larger context windows allow models to understand long-range dependencies, such as relationships between sentences or paragraphs. However, they also require more computational resources. Models with smaller context windows may lose critical context, affecting the coherence and accuracy of their outputs.
Tokenization Techniques
Modern AI models use advanced tokenization methods like Byte Pair Encoding (BPE), WordPiece, or SentencePiece. These methods split text into subword units, enabling efficient processing of diverse vocabularies. For instance, BPE might tokenize "darkness" into "dark" and "ness," allowing the model to recognize shared patterns across words.
Practical Applications
Tokens are central to various AI tasks, including text generation, translation, summarization, and reasoning. For example, a model translating English to Japanese processes input tokens in one language and generates output tokens in another. Similarly, reasoning models generate additional "reasoning tokens" to solve complex problems.
Understanding tokens and context windows is crucial for optimizing AI models, balancing computational efficiency with performance, and delivering high-quality user experiences.
ASIC - Application-Specific Integrated Circuit
When I was looking at Cryptpcurrency and the technologies behind it the ASIC chip featured quite highly. My understanding was that the ICs were designed to handle large amounts of data, possibly small "chunks" (my terminolgy), more quickly than more general purpose processors. GPUs, specifically those that were developed by Nvidia, seemed to represent a similar technolgy.
In terms of AI processing of tokens the ASIC seemed to be something that was relevant and "just the job". Now I see that this could "ASIC growth may outpace GPU in 2026, putting NVIDIA under pressure"