Unsupervised AI development
This post is references an external article or page. Consider it a bookmark, and in no way an endorsement of the article, author, or website. I frequently bookmark content I disagree with.
🔗 via unsupervision.com.
Highlights
Radical unsupervision is “training” and “inferring” simultaneously on signals in real time. Designing models that embrace real-time learning has a few implications:
Regular/continuous updates to the agent’s world-model. One of the key limitations of GPTs today is that their knowledge of world events depends entirely on the cutoff date for the model’s training set. For a model architecture to be considered radically unsupervised, it must be capable of gathering insights from signals received during deployment and “learn” from them (e.g. Bayesian Optimization).
The ability to generate new knowledge representations for never-before-seen phenomena. A fundamental property of radically unsupervised AI is the model’s ability to outgrow the initial data seed. This seed includes the types of data included in the training set. Something along the lines of continuous Knowledge Representation Learning on a feed of real-time input.
Conflict resolution for information gathered by distributed/multi-headed agents. This is relevant for cases where the agent has more than one interface or needs to support multiple interactions simultaneously. Something like a CRDT for weight matrices.
This may sound like a daunting task, but the potential rewards are enormous. With radically unsupervised model architectures, we could create AI systems that are able to learn and adapt to new environments and situations without the need for explicit supervision. This could lead to more intelligent, versatile, and autonomous AI systems that are better able to assist and augment human intelligence.
A fundamental consequence of adopting radically unsupervised learning is the increased importance of environment design. Once we have an agent that is capable of generalized learning from it’s environment in real time, teaching it a new skill will mean crafting an environment that provides informative interactions related to the desired skill. This brings with it some concrete issues related to AI alignment (aka the control problem) such as reward hacking.
A great overview of how unsupervision could enable AI to learn and adapt in real-time, without being limited by training data.
DeepSeek did not bypass export restrictions; instead, they optimised their chips for maximum memory efficiency, and their well-optimised low-level code was not limited by chip capacity.
A study with nearly 2,000 participants explored how prompts must evolve as generative AI models advance.
Waymo expanded its fully autonomous driving from 3.8 million to 25.3 million miles.
Yucatán, Mexico, partnered with the AI-powered app MeMind in 2022 to prevent suicides, leading to a 9% drop in the suicide rate.
The use of AI monitoring software in a Malawian clinic has reduced stillbirths and neonatal deaths by 82% over three years.
The Commonwealth Bank of Australia’s use of generative AI tools has halved customer losses from scams and reduced fraud by 30%.
Text-generative Transformers include an embedding component that converts text tokens into numerical vectors to capture semantic meaning.
Here’s a Bash script that uses a TDD loop with an LLM to iteratively write Python code to pass tests.
Researchers have developed a portable, wireless EEG acquisition system called BrainGPT capable of converting thoughts into text without the need for an fMRI machine.
Jobs most negatively impacted by AI since ChatGPT’s release: writing jobs (-33%), translation jobs (-19%), and customer service jobs (-16%).
Economic history shows an expansion in the variety of tasks performed by humans, with new jobs continually emerging despite automation.
Meta has developed TestGen-LLM, a tool using LLMs to enhance human-written tests, ensuring test suite improvements by passing certain filters.
Klarna has launched an AI assistant powered by OpenAI, which is active across 23 markets and offers support in over 35 languages.
Business owners are increasingly finding AI like ChatGPT valuable for tasks beyond basic applications, such as data visualisation and financial reporting, saving time on manual data processing.
The paper compares Large Language Models (LLMs) with Junior Lawyers and Legal Process Outsourcers (LPOs) in terms of contract review performance.
Researchers have developed a system called Mobile ALOHA for mobile manipulation tasks that require both hands and the whole body, aiming to improve robotic mobility and dexterity beyond simple table-top manipulation.
Clinicians developed a deep learning model that predicts reported sex from retinal fundus photographs without needing to code.
Apple has launched a new machine learning framework specifically for Apple Silicon.
These researchers suggest each occupation has an inflection point after which AI improvements harm human workers’ prospects.
Here’s a GPT trained on knowledge from 17th-century texts. So, it answers in historical style, including outdated scientific concepts.
The author has been at OpenAI for a year and observed that generative models closely approximate their training datasets.
This team taught GPT to navigate iOS and Android by sending it screenshots and giving it instructions.
President Biden released an executive order on AI development, drawing parallels to the early fears and regulatory considerations during the dawn of the microprocessor and internet, highlighting that past technological advancements were less hindered by government intervention.