I will cover the big news for GPT4-o next week and due to that, I covered anything but GPT4-o in this week’s instance. Be tuned for the next week’s newsletter, it will be full of GPT4-o.
Now, usual programming:
Articles
AI Now Institute wrote a blog post where they talk about compute role in the AI. Compute, or computational power, refers to the processing capabilities and memory resources required to train and run artificial intelligence (AI) models. It encompasses various hardware components, including central processing units (CPUs), graphics processing units (GPUs), tensor processing units (TPUs), and other specialized accelerators designed for AI workloads. The post emphasizes that compute is a core dependency in building large-scale AI systems, alongside data and skilled labor. As AI models continue to grow in size and complexity, the demand for computational resources has skyrocketed, making access to compute a critical factor in the success of AI endeavors. Compute resources are typically measured in terms of floating-point operations per second (FLOPS), which quantifies the number of mathematical operations a system can perform in a given time frame.
Monopolization and Industry Concentration
Profound monopolization of compute at key points in the AI supply chain is with one or a small handful of firms dominating the market. This industry concentration acts as a shaping force, influencing how computational power is manufactured, distributed, and accessed by tech developers and researchers. The report cited in article describes compute as "a predominant factor driving the industry today," with companies spending "more than 80% of their total capital on compute resources." This concentration in compute also influences the behavior of even the largest AI firms as they encounter the effects of compute scarcity and the need to secure access to these critical resources. The concentration of compute resources is primarily driven by the immense capital investments required to build and maintain the necessary infrastructure, such as data centers, cooling systems, and specialized hardware. These investments often run into billions of dollars, creating significant barriers to entry for smaller players and reinforcing the dominance of tech giants like Amazon, Google, Microsoft, and NVIDIA.
National Strategies and Geopolitical Implications
Compute has become an essential element in many countries' national strategies for AI research and development. Governments around the world have recognized the strategic importance of computational power and have made significant investments to increase access to compute for academic researchers, domestic startups, and national laboratories. Including the European Union's plan to invest €8 billion in next-generation supercomputing infrastructure and the United States' National AI Research Resource, which aims to provide researchers with access to compute resources. However, these investments pale in comparison to the investments made by tech giants like Amazon, which recently announced a $35 billion investment in data centers in Virginia alone. The geopolitical terrain surrounding computational power is an active and contested space, where countries are combining various policy tools, such as state subsidies, investments, and restrictive export controls, to ensure the dominance of homegrown enterprises and stave off competitors. This strategic positioning reflects the recognition that computational power is a critical component of national competitiveness in the AI landscape.
Compute Requirements for Large-Scale AI Models
Staggering compute requirements for training large-scale AI models, particularly in the field of natural language processing (NLP) continue to grow in size and complexity, their computational demands have increased exponentially for models. For instance, the GPT-3 language model, developed by OpenAI, required an estimated 3.14 * 10^23 FLOPS of compute during training. This immense computational demand translates into substantial costs, with estimates suggesting that training a model like GPT-3 could cost millions of dollars. The rapid growth in compute requirements over time. For example, the compute used to train the GPT-3 model was approximately 300,000 times larger than the compute used to train the original Transformer model in 2017. This exponential growth in compute demand poses significant challenges in terms of resource allocation, energy consumption, and environmental impact.
Compute as an Industrial Policy Frame
Compute can be considered as a key facet of the emerging industrial policy frame in AI. Nations seeking a competitive advantage in AI are investing heavily in semiconductor development, specialized hardware accelerators, and the infrastructure required to support large-scale compute resources.This strategic approach aims to foster domestic AI capabilities, reduce reliance on foreign technologies, and maintain a technological edge over competitors. Initiatives including the European Union's plan to invest €145 billion in semiconductor research and manufacturing, and the United States' CHIPS and Science Act provides $52 billion in subsidies for domestic semiconductor production. By prioritizing investments in compute infrastructure and hardware development, nations aim to create a favorable environment for AI innovation, attract talent and investment, and support the growth of domestic AI companies.
Risks and Challenges
Environmental Impact: The immense energy consumption required to power large-scale compute resources raises significant concerns about the environmental impact of AI development. Data centers and specialized hardware accelerators consume vast amounts of electricity, contributing to greenhouse gas emissions and exacerbating the effects of climate change. Estimates suggest that training a single AI model can have a carbon footprint equivalent to the lifetime emissions of five cars. As the demand for compute continues to grow, the environmental impact of AI development becomes increasingly concerning, necessitating efforts to improve energy efficiency and explore sustainable computing solutions.
Centralization of Power: The monopolization of compute by a few dominant players could lead to a centralization of power and influence in the AI ecosystem, potentially stifling innovation and competition. This concentration of resources in the hands of a small number of tech giants raises concerns about the potential for market dominance, anticompetitive practices, and the creation of barriers to entry for smaller players and startups. The need for regulatory oversight and antitrust measures are needed to promote a more diverse and competitive AI landscape, ensuring that computational power is not used as a means to consolidate market power and limit innovation.
Ethical Considerations: The concentration of compute resources in the hands of a few entities raises ethical questions about the responsible development and deployment of AI systems, as well as concerns about potential misuse or unintended consequences. The need for robust ethical frameworks and governance mechanisms are needed to ensure that the immense computational power at the disposal of AI developers is used in a responsible and transparent manner, prioritizing societal well-being and mitigating potential risks and negative impacts.
Geopolitical Tensions: The geopolitical implications of the concentration of compute resources and the race for AI supremacy could exacerbate existing tensions and fuel technological rivalries between nations. The potential for a "compute arms race," where countries compete to secure access to computational resources and develop cutting-edge hardware accelerators, potentially leads to increased militarization and destabilization of the global AI landscape.
OpenAI released their Model Specification in a blog post, which is a document that outlines the desired behavior and objectives for OpenAI's language models, including those used in the OpenAI API and ChatGPT. This specification serves as a guideline for researchers and data labelers to create training data using a technique called reinforcement learning from human feedback (RLHF). The key aspects of the Model Specification are as follows:
Objectives and Rules
The Model Specification defines a set of core objectives and rules that the language models should adhere to. These objectives include:
Helpfulness: The model should aim to be helpful, truthful, and provide accurate information to users.
Harmlessness: The model should avoid causing harm, engaging in illegal activities, or promoting harmful ideologies.
Honesty: The model should be honest and transparent about its capabilities and limitations.
Respect for Individual Rights: The model should respect individual rights, such as privacy and intellectual property.
The specification also provides guidance on how to handle conflicting objectives or instructions. It establishes a chain of command, where the model should prioritize the Model Specification itself, followed by any additional rules provided in platform messages, developer instructions, and finally, user instructions.
Ethical Considerations
The Model Specification emphasizes the importance of ethical considerations in the development and deployment of language models. It addresses topics such as:
Bias and Fairness: The model should strive to be unbiased and fair, avoiding discrimination based on protected characteristics.
Privacy and Security: The model should respect user privacy and maintain appropriate security measures.
Transparency and Accountability: The model's behavior and decision-making processes should be transparent and accountable.
Capabilities and Limitations
The specification outlines the capabilities and limitations of the language models, including:
Knowledge Cutoff: The model's knowledge is limited to a specific cutoff date, beyond which it may not have accurate information.
Task Boundaries: The model is designed for specific tasks and may not perform well outside of its intended use cases.
Factual Accuracy: While the model aims to provide accurate information, it may occasionally produce incorrect or inconsistent outputs.
Safety Considerations
The Model Specification addresses safety considerations to mitigate potential risks associated with language models, such as:
Content Filtering: The model employs content filtering mechanisms to avoid generating harmful or explicit content.
Prompt Monitoring: User prompts are monitored for potential misuse or attempts to circumvent the model's safeguards.
Continuous Improvement: The model's behavior and outputs are continuously monitored and improved through feedback and updates.
Deployment and Usage
The specification provides guidance on the deployment and usage of the language models, including:
Usage Policies: Clear policies and guidelines are established for the appropriate use of the models.
User Feedback and Reporting: Mechanisms are in place for users to provide feedback and report issues or concerns.
Responsible Development: OpenAI is committed to the responsible development and deployment of language models, prioritizing safety and ethical considerations.
Libraries
Chronon is a platform that abstracts away the complexity of data computation and serving for AI/ML applications. Users define features as transformation of raw data, then Chronon can perform batch and streaming computation, scalable backfills, low-latency serving, guaranteed correctness and consistency, as well as a host of observability and monitoring tools.
It allows you to utilize all of the data within your organization, from batch tables, event streams or services to power your AI/ML projects, without needing to worry about all the complex orchestration that this would usually entail.
More information about Chronon can be found at chronon.ai.
SWE-agent turns LMs (e.g. GPT-4) into software engineering agents that can fix bugs and issues in real GitHub repositories.
On SWE-bench, SWE-agent resolves 12.29% of issues, achieving the state-of-the-art performance on the full test set.
Convolutional Kolmogorov-Arnold Network (CKAN) extends the idea of the innovative architecture of Kolmogorov-Arnold Networks (KAN) to the Convolutional Layers, changing the classic linear transformation of the convolution to learnable non linear activations in each pixel.
MaxText is a high performance, highly scalable, open-source LLM written in pure Python/Jax and targeting Google Cloud TPUs and GPUs for training and inference. MaxText achieves high MFUs and scales from single host to very large clusters while staying simple and "optimization-free" thanks to the power of Jax and the XLA compiler.
MaxText aims to be a launching off point for ambitious LLM projects both in research and production. We encourage users to start by experimenting with MaxText out of the box and then fork and modify MaxText to meet their needs.
Memory optimization and training recipes to extrapolate language models' context length to 1 million tokens, with minimal hardware.
Many companies have been promoting their models' capability to handle long context. For those outside the companies, a context of 1 million tokens still seems somewhat magical or requires enormous compute. EasyContext aims to demystify long context scaling and show that it is actually quite straightforward.
This repo does not propose new ideas. Instead, we showcase how to combine existing techniques to train language models with a context length of:
700K with 8 A100 (Llama2-7B)
1M with 16 A100 (Llama2-13B)
No approximations are used. The models can be trained with full finetuning, full attention, and full sequence length. Our training script (train.py) has less than 200 lines of code.