# AI Developments and Concerns

## Does current AI represent a dead end?

- **Current AI systems, particularly those based on large neural networks, are deemed unmanageable and unsuitable for serious applications due to their lack of transparency and accountability in software engineering.** This raises concerns about the ethical implications of deploying such technologies without proper oversight.

## Building AI Products–Part I: Back-End Architecture

- In 2023, the launch of an **AI-powered Chief of Staff** attracted **10,000 users** within a year, prompting a pivot to **Outropy**, a developer platform for building AI products, due to high demand for the underlying technology.

## [R] Fine-Tuning 175B Parameter Language Models on a Single Consumer GPU through Optimized Memory Management

- **Fine-tuning 100B parameter models** on a single consumer GPU is now feasible due to **optimized memory management** and NVMe SSD utilization, significantly enhancing accessibility for individual researchers and smaller labs.

## The Structure of Neural Embeddings

- **Neural embeddings** exhibit a **manifold structure**, suggesting that high-dimensional data from real-world processes is confined to a low-dimensional manifold, as detailed in the [Manifold Hypothesis](https://arxiv.org/abs/2208.11665).

## [D] Could "activation engineering" replace prompt engineering or fine-tuning as a technique for steering models?

- **Activation engineering** manipulates activation vectors in LLMs, exemplified by the "Golden Gate Claude," which altered neuron activity to embed the Golden Gate Bridge concept into model responses.

## [P] REINFORCE++: A Simple and Efficient Approach for Aligning Large Language Models

- **REINFORCE++** integrates optimization techniques from **Proximal Policy Optimization (PPO)** into the traditional REINFORCE algorithm, enhancing **performance** and **stability** in RLHF while minimizing computational demands.

## Inverse Design of Complex Nanoparticle Heterostructures via DL on Graphs

- **Deep learning** is utilized to optimize the **nonlinear optical properties** of core-shell upconverting nanoparticles (UCNPs), achieving a **6.5 times increase** in predicted emission under 800nm illumination compared to existing structures.

## [D] The Parallelism Tradeoff: Understanding Transformer Expressivity Through Circuit Complexity

- **Transformers are classified in the TC^0 complexity class**, indicating their expressive limitations stem from their **parallel nature**, which restricts their ability to tackle inherently sequential problems effectively.

## [P] Violation of proportional hazards assumption: what can I do?

- **Violation of the proportional hazards assumption** in the CoxPH model can significantly impact the reliability of survival predictions, particularly when a substantial number of covariates show p-values below 0.05, indicating potential issues with model validity.