Skip to main content

'Before long, the world will wake up'

 


Leopold Aschenbrenner's 'Situational Awareness, the decade ahead Situational Awareness, the decade ahead. June 2024' may turn out to be the most significant publication on AI safety to date. Unlike a lot of theoretical musings from highly intelligent critics of AI systems this one has been written by an engineer, who was until recently employed by Open AI in the now disbanded former Super Alignment team. 

It begins by discussing the rapid advancements in AI technology, particularly focusing on the progression from GPT-2 to GPT-4 models. It highlights that AI capabilities are evolving at an exponential rate, and there are predictions that by 2027, AI models could match or even surpass the work of human AI researchers and engineers. The text underscores the importance of understanding the trendlines in compute, algorithmic efficiencies, and unlocking latent capabilities for future AI development. Additionally, the document mentions the potential risks and challenges associated with the rapid progression of AI technology, such as the need for enhanced security measures, the race for industrial mobilization in AI development, and ensuring alignment of AI systems with human values. It also touches on the implications of superintelligence, including economic and military advantages, as well as the importance of maintaining pre-eminence over authoritarian powers. It points towards a future where AI will play a central role, and the transformative impact it could have on society and national security.

The AI 'Scaling Implications' section discusses the concept of test-time compute overhang in AI models, specifically focusing on GPT-4tokens. The document mentions that GPT-4tokens are capable of thinking about problems but are limited in the number they can effectively use. It introduces a table equating the number of tokens to human-time spent on a problem/project. The text highlights the importance of allowing models to use more tokens for longer periods to improve capabilities significantly.

Furthermore, the document explores the idea of unlocking test-time compute to enhance the intelligence of models. It describes the need to teach models skills for longer-horizon reasoning and error correction to work on projects independently. The text suggests that small algorithmic wins could potentially help models utilise more test-time compute efficiently.

Additionally, the document outlines the potential progression from current AI models like ChatGPT to more advanced models that can function as drop-in remote workers. It discusses the implications of advanced AI models being able to automate cognitive jobs, potentially leading to AGI (Artificial General Intelligence) by 2027.

In conclusion, the text emphasizes the exponential progress in AI capabilities through the effective scaling of compute power and the importance of overcoming limitations to unlock the full potential of AI models. It also touches on the rapid pace of deep learning progress and the possibility of models surpassing human intelligence in the near future.

There is also an excellent YouTube video on Dwarkesh Patel's channel with Leopold, 'Leopold Aschenbrenner - 2027 AGI, China/US Super-Intelligence Race, & The Return of History'

It's four and a half hours long, so you'll need popcorn!

During my time (one year) absent from writing this blog I have spent it upgrading my PC system to a level I can run low scale LLM model experiments and helping develop AI policy documentation. I needed more practical knowledge of the system components: vector databases, RAG, embeddings and GraphRag outputs to name but a few. 

The 'advancements' in AI applications over this last year have appeared incremental, in the consumer space at least, rather than revolutionary. This has led many that are sceptical about AI to suggest we are in another bubble, and just like the big tech hype over Big Data a few years ago, that in reality saw no ROI and many million dollar companies go bust prior to Covid. 

I'd suggest that it's not the same, far from it, and 'before long, the world will wake up'. 




Labels:

Comments

Post a Comment

Popular posts from this blog

OpenAI's NSA Appointment Raises Alarming Surveillance Concerns

  The recent appointment of General Paul Nakasone, former head of the National Security Agency (NSA), to OpenAI's board of directors has sparked widespread outrage and concern among privacy advocates and tech enthusiasts alike. Nakasone, who led the NSA from 2018 to 2023, will join OpenAI's Safety and Security Committee, tasked with enhancing AI's role in cybersecurity. However, this move has raised significant red flags, particularly given the NSA's history of mass surveillance and data collection without warrants. Critics, including Edward Snowden, have voiced their concerns that OpenAI's AI capabilities could be leveraged to strengthen the NSA's snooping network, further eroding individual privacy. Snowden has gone so far as to label the appointment a "willful, calculated betrayal of the rights of every person on Earth." The tech community is rightly alarmed, with many drawing parallels to dystopian fiction. The move has also raised questions about ...
Post a Comment
Read more

What is happening inside of the black box?

  Neel Nanda is involved in Mechanistic Interpretability research at DeepMind, formerly of AnthropicAI, what's fascinating about the research conducted by Nanda is he gets to peer into the Black Box to figure out how different types of AI models work. Anyone concerned with AI should understand how important this is. In this video Nanda discusses some of his findings, including 'induction heads', which turn out to have some vital properties.  Induction heads are a type of attention head that allows a language model to learn long-range dependencies in text. They do this by using a simple algorithm to complete token sequences like [A][B] ... [A] -> [B]. For example, if a model is given the sequence "The cat sat on the mat," it can use induction heads to predict that the word "mat" will be followed by the word "the". Induction heads were first discovered in 2022 by a team of researchers at OpenAI. They found that induction heads were present in ...
Post a Comment
Read more

Prompt Engineering: Expert Tips for a variety of Platforms

  Prompt engineering has become a crucial aspect of harnessing the full potential of AI language models. Both Google and Anthropic have recently released comprehensive guides to help users optimise their prompts for better interactions with their AI tools. What follows is a quick overview of tips drawn from these documents. And to think just a year ago there were countless YouTube videos that were promoting 'Prompt Engineering' as a job that could earn megabucks... The main providers of these 'chatbots' will hopefully get rid of this problem, soon. Currently their interfaces are akin to 1970's command lines, we've seen a regression in UI. Constructing complex prompts should be relegated to Linux lovers. Just a word of caution, even excellent prompts don't stop LLM 'hallucinations'. They can be mitigated against by supplementing a LLM with a RAG, and perhaps by 'Memory Tuning ' as suggested by Lamini (I've not tested this approach yet).  ...
Post a Comment
Read more