It was initially assumed that regulation is necessary for major models as a way to make [most of] AI safe. In the two years since ChatGPT, that has turned out to be inaccurate.
Frontier AI models, even with the absence of regulation, are under the scrutiny of litigation, media, investors, users, commission inquiry, and congressional hearings, keeping them within certain boundaries that benefit several safety objectives. Also, the firms have been able to prospect most regulatory demands, preparing or adjusting for them.
The tens of unique ways that AI has been misused and has caused harm in the last two years have not been a result of the major models, per se, nullifying the initial focus on major models alone, as channels to general AI safety.
It was also assumed that AI policy, governance, and ethics would be important to safety. That has not been the case. AI policy, governance, and ethics are as important as they can contribute novel technical architecture as paths to AI safety, not just policy that suggests how to keep AI safe with no lead towards how it can be technically adapted for general AI safety.
There is often news every week of a different way of AI misuse or harm, with no existing technical solution. If there is no technical solution, then policy is powerless. For example, there is policy, regulation, and litigation against digital piracy, making it scarce from the mainstream, still it exists because the possibility for eradication, technically, does not exist. This makes efficient AI safety completely a technical problem, not regulation, governance, policy, ethics, litigation, terms, antitrust, and so forth. Technical answers do not have to be technically initialized, but they have to be technically tracked, compatible, and delivered.
When would superintelligence emerge and will AI become dangerous enough to threaten humans? ChatGPT is already excellent at intelligence, comparing directly with how human intelligence works.
Human intelligence is theorized to be the quality of usage of human memory. Simply, humans are as intelligent as the exceptional distributions or relays within memory. This makes humans more intelligent than other organisms—sometimes with similar sensory interpretations as humans but limited in what it means because they lack the relays within their memory destinations to make the memories as intelligent.
Some people say ChatGPT is not as intelligent as certain organisms. That is not a good measure if ChatGPT is using the [digital] memory available to it, like several organisms are using their memories in their habitats.
ChatGPT is excellent at prediction—which is a relay [or usage quality] of memory. If it acquires other relays that makes it continue to advance in intelligence, it would be superintelligent, regardless of embodiment, or real world experience.
It is not unlikely that AI could develop intentionality at some point. Whether that intent can be capped, useful or threatening remains uncertain, but intent could be possible for AI. For example, features in artificial neural networks represent concepts or the meanings of things. Features can be monosemantic or polysemantic. If there are features that do not represent concepts but that are abstractions that concepts can bounce off, there might be intent around those concepts only, not for others, conceptually. What is termed model collapse for large language models could be datasets to develop abstractions, conceptually.
AI can be useful or harmful. Its usefulness is already built-in, to some, with advances that are informed by how they can be applied. However, misuses at scale and harms that can be dangerous are emerging, demanding intense technical research into solving current problems and preparing against those ahead.
There is a new report on VentureBeat, Meet the new, most powerful open source AI model in the world: HyperWrite’s Reflection 70B, stating that, “Reflection 70B has been rigorously tested across several benchmarks, including MMLU and HumanEval, using LMSys’s LLM Decontaminator to ensure the results are free from contamination. These benchmarks show Reflection consistently outperforming models from Meta’s Llama series and competing head-to-head with top commercial models. Reflection — a model that can reflect on its generated text and assess its accuracy before delivering it as outputs to the user. The model’s advantage lies in a technique called reflection tuning, which allows it to detect errors in its own reasoning and correct them before finalizing a response. Reflection 70B introduces several new special tokens for reasoning and error correction, making it easier for users to interact with the model in a more structured way. During inference, the model outputs its reasoning within special tags, allowing for real-time corrections if it detects a mistake. The playground demo site includes suggested prompts for the user to use, asking Reflection 70B how many letter “r” instances there are in the word “Strawberry” and which number is larger, 9.11 or 9.9, two simple problems many AI models — including leading proprietary ones — fail to get right consistently. Our tests of it were slow, but Reflection 70B ultimately provided the correct response after 60+ seconds.”
There is a recent feature on TechTarget, California AI bill sets guardrails that draw criticism, stating that, “California’s AI bill has drawn significant scrutiny. While some believe any guardrails are better than none when it comes to the rapidly evolving technology, others say the bill could negatively affect small businesses and hurt innovation. Senate Bill 1047, the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act, would require advanced AI system developers to test models costing a minimum of $100 million to train for their ability to cause harm, as well as implement guardrails to mitigate that risk. The bill creates whistleblower protections for employees of large AI companies and sets up CalCompute, a responsible AI development public cloud computing cluster for startups and researchers. Even if small businesses might save on costs upfront without the testing requirements, they will still face those costs down the road — and even potentially face lawsuits from customers who suffer negative consequences as a result of the company’s AI model. California isn’t the first state to advance an AI bill. Earlier this year, Colorado passed comprehensive AI legislation, while Connecticut lawmakers advanced an AI bill to regulate private sector deployment of AI models. Even cities like New York have passed AI bills targeting algorithmic bias.”
