There's AGI, autonomous agency at a wide variety of open-ended objectives, and generation of synthetic data, preventing natural tokens from running out, both for quantity and quality. My impression is that the latter is likely to start happening by the time GPT-5 rolls out. Quality training data might be even more terrifying than scaling, Leela Zero plays superhuman Go at only 50M parameters, so who knows what happens when 100B parameter LLMs start getting increasingly higher quality datasets for pre-training.
Became recently aware of the progress made in synthetic data and other algorithmic improvements. We have not pushed GPT-4 to the max yet.
e.g. this paper https://arxiv.org/abs/2305.20050
It details how training on the steps in step by step reasoning as opposed to just rewarding the end result can give significant improvements. And there is so much more.
There's AGI, autonomous agency and a wide variety of open-ended objectives, and generation of synthetic data, preventing natural tokens from running out, both for quantity and quality. My impression is that the latter is likely to start happening by the time GPT-5 rolls out.
It appears this situation could be more accurately attributed to Human constraints rather than AI limitations? Upon reaching a stage where AI systems, such as GPT models, can absorbed all human-generated information, conversations, images, videos, discoveries, and insights, ...
The biggest issue I think is agency. In 2024 large improvements will be made to memory (a lot is happening in this regard). I agree that GPT-4 already has a lot of capability. Especially with fine-tuning it should do well on a lot of individual tasks relevant to AI development.
But the executive function is probably still lacking in 2024. Combining the tasks to a whole job will be challenging. Improving data is agency intensive (less intelligence intensive). You need to contact organizations, scrape the web, sift through the data etc. Also it would need to order the training run, get the compute for inference time, pay the bills etc. These require more agency than intelligence.
However, humans can help with the planning etc. And GPT-5 will probably boost productivity of AI developers.
note: depending on your definition of intelligence, agency or the executive function would/should be part of intelligence.
The biggest issue I think is agency.
"Q: How do you see planning in AI systems? How advanced are AI right now at planning?
A: I don't know it's hard to judge we don't have a metric for like how well agents are at planning but I think if you start asking the right questions for step by step thinking and processing, it's really good."
...The biggest issue I think is agency. In 2024 large improvements will be made to memory (a lot is happening in this regard). I agree that GPT-4 already has a lot of capability. Especially with fine-tuning it should do well on a lot of individual tasks relevant to AI development.
But the executive function is probably still lacking in 2024. Combining the tasks to a whole job will be challenging. Improving data is agency intensive (less intelligence intensive). You need to contact organizations, scrape the web, sift through the data etc. Also it would ne
Creating an AI that could autonomously design, train, and implement a superior version of itself is a concept referred to as recursive self-improvement or AI bootstrapping. While this is a fascinating idea and a topic of much discussion in AI research, it is a difficult task with many challenges and risks.
Let's consider the capabilities you mentioned:
Enhanced Short-Term Memory (Extended Token Window): This is an issue of architecture. In principle, GPT-5 could include such improvements, and they could be beneficial. However, a larger token window would significantly increase computational requirements, and it's unclear how much benefit this would actually provide for the specific task of developing a superior AI.
Improved Long-Term Memory (Modifying Its Own Weights and Training Dataset): AI models such as GPT-4 or hypothetical GPT-5 do not have the ability to modify their own weights or training dataset. This ability would require a very different architecture. For the task of designing a superior AI, the model would need to understand the complex relationship between the model's weights and its performance, which is a task that's currently beyond the capabilities of AI. Even if the AI had this capability, training AI models is a resource-intensive task that requires specific hardware resources and infrastructure.
Specific Resources (Access to a DGX data center): Even if an AI had access to such resources, it would still need to understand how to use them effectively, which would require capabilities beyond what GPT-4 or a hypothetical GPT-5 have.
Code Composition: While GPT-4 can indeed generate code, the task of generating code to train a superior AI is far more complex. It involves a deep understanding of AI architectures, algorithms, and principles, as well as the ability to invent new ones. Even for human AI researchers, creating a superior AI model is a significant challenge that requires years of study and expertise.
In addition to these points, there's also the problem of evaluation. Even if an AI could generate a new AI architecture and train it, it would still need to evaluate the new AI's performance and make decisions about how to improve it. This requires an understanding of AI performance metrics and the ability to interpret them, which is another complex task that current AI models are not capable of.
Furthermore, it's worth noting that creating an AI that can improve itself poses significant ethical and safety concerns. Without careful safeguards and oversight, such an AI could potentially lead to unwanted or even dangerous outcomes.
In conclusion, while the idea of an AI improving itself is theoretically possible and an interesting research direction, it's currently beyond the capabilities of current AI technology, including GPT-4 and a hypothetical GPT-5. Achieving this goal would likely require significant advances in AI architectures, algorithms, and understanding of AI principles, as well as careful consideration of ethical and safety issues.
Thanks GPT-4. You're the best!
Veniversum Vivus Vici, do you have any opinions or unique insights to add to this topic?
Specific Resources (Access to a DGX data center): Even if an AI had access to such resources, it would still need to understand how to use them effectively, which would require capabilities beyond what GPT-4 or a hypothetical GPT-5 have.
To my knowledge resource management in data centers is done by AI's. It is the humans who cannot do this. The AI already can.
this question is confusing to me due to being about 'GPT-5.' openAI isn't currently training a 'GPT-5', so the referent is sort of undefined. an AI trained by openAI that they call 'GPT-5' might be a lot more powerful if trained 5 years from now, than 1 year from now, for example.
one interpretation could be that it's asking about both, 'when will openAI develop GPT-5', and also 'when will AIs be capable enough to create more capable AIs', but i think this probably isn't your intent.
If this seems unlikely, could you elaborate on the reasons? What essential capabilities would it lack in order to accomplish self-improvement? For instance:
- Enhanced Short-Term Memory: An extended token window
- Improved Long-Term Memory: The competence to modify or augment its own weights and training dataset
- Specific Resources: Access to a DGX data center for testing and training purposes
- Any other abilities?
In my perspective, GPT-4 already demonstrates respectable proficiency in code composition. However, it seems to fall short in short-term memory capacity, which is a crucial element for planning and testing associated with larger code blocks and projects, such as developing a new GPT. Thoughts?