20240702 QFM021 Machine Intelligence Reading List June 2024

Everything that I found interesting about machines behaving intelligently during June 2024

1. Quantum Fax Machine
QFM021: Machine Intelligence
Reading List June 2024
quantumfaxmachine.com
1

2. QFM021: Machine Intelligence
Reading List June 2024
We kick off this month’s reading list with the transformative potential of AI in executive
roles. If AI Can Do Your Job, Maybe It Can Also Replace Your CEO (nytimes.com)
highlights AI’s growing capability to manage high-level decision-making tasks
traditionally reserved for CEOs, suggesting a future where AI could play a pivotal role in
corporate leadership, albeit with human oversight to ensure strategic alignment and
accountability. If it can take the jobs of call centre staff, designers, and software
engineers, is there something so special about executive jobs that leaves them immune?
Another theme is the drive to understand the inner workings of gen-AI systems more
deeply. Here’s what’s going on inside an LLM’s neural network (arstechnica.com),
unveiling how AI models like Claude operate on the inside. These studies reveal the
intricate patterns within neural networks, enhancing our ability to interpret and
potentially steer AI behaviour in critical applications such as security and bias mitigation.
We then examine the practical experience of deploying AI at scale with What We
Learned from a Year of Building with LLMs (Part I) (oreilly.com). The O’Reilly article
provides lessons from a year of building with LLMs, emphasizing the importance of
robust prompting techniques and structured workflows.
Finally, this month’s list touches on AI deployment's ethical and operational
considerations. What’s the future for generative AI? The Turing Lectures with Mike
Wooldridge (youtube.com) examines the importance of addressing bias, misinformation,
and ethical concerns in AI’s advancement.
As always, the Quantum Fax Machine Propellor Hat Key will guide your browsing. Enjoy!
Key:
: Mentions technology
: Talks about technology in real-world use cases
: Talks about details of machine intelligence technologies
: Using and working with machine intelligence technologies in software
: Programming new machine intelligence concepts and implementations
Source: Photo by vackground.com on Unsplash
2

3. If AI Can Do Your Job, Maybe It Can Also
Replace Your CEO (nytimes.com): The article
discusses how artificial intelligence (AI)
might not only replace routine jobs but also
high-level executive roles, including CEOs.
With AI's capability to analyse markets,
automate communication, and make
dispassionate decisions, some companies
are already experimenting with AI leadership
to cut costs and increase efficiency, though
human oversight remains necessary for
accountability and strategic thinking.
#AI #Automation #Leadership
#CorporateManagement #FutureOfWork
3

4. Here’s what’s really going on inside an LLM’s
neural network (arstechnica.com): Anthropic's
recent research unveils how the Claude LLM's
neural network operates by mapping millions
of neurons' activities, revealing that concepts
are represented across multiple neurons. This
mapping process, using sparse auto-
encoders and dictionary learning algorithms,
helps identify patterns and associations in
the model, providing partial insights into its
internal states and conceptual organisation.
#AI #MachineLearning
#NeuralNetworks
#ArtificialIntelligence #Research
4

5. Scaling Monosemanticity - Extracting
Interpretable Features from Claude 3 Sonnet
(transformer-circuits.pub): Researchers at
Anthropic have successfully scaled sparse
autoencoders to extract high-quality, interpretable
features from the Claude 3 Sonnet language
model, demonstrating that the technique can
handle state-of-the-art transformers. These
features are diverse, covering concepts from
famous people to programming errors, and are
crucial for understanding and potentially steering
AI behaviour, especially in safety-critical areas
such as security vulnerabilities and bias.
#AI #MachineLearning
#NaturalLanguageProcessing #Safety
#AIResearch
5

6. What is the biggest challenge in our
industry? (thrownewexception.com): The
biggest challenge in the tech industry is the
anxiety caused by layoffs and the fear of AI
replacing jobs, leading to mental health
issues like burnout. Leaders can help by
fostering open communication, leading
positively, leveraging new technologies,
investing in continuous learning, and
collaborating with HR to support their
teams.
#TechIndustry #AI #MentalHealth
#Leadership #Layoffs
6

7. What We Learned from a Year of Building
with LLMs (Part I) (oreilly.com): Over the past
year, the authors built real-world
applications using large language models
(LLMs) and identified crucial lessons for
developing effective AI products. They
emphasise the importance of robust
prompting techniques, retrieval-augmented
generation, structured workflows, and
rigorous evaluation and monitoring to
overcome the complexities and challenges
inherent in leveraging LLMs for practical use.
#AI #MachineLearning #LLM
#TechInnovation #ProductDevelopment
7

8. Achieving the Self-Thinking Business
(linkedin.com): The article discusses Honu's
development of a "Self-Thinking Business"
model through the introduction of a Cognitive
Layer that bridges the gap between current AI
capabilities and true business autonomy. This
new layer aims to transform AI from tactical
automation tools into strategic decision-
makers by providing a comprehensive,
contextual understanding of business data
and operations, reducing the need for
extensive data and compute resources.
#AI #BusinessAutomation
#CognitiveLayer #AutonomousAgents
#Innovation
8

9. What's the future for generative AI? The
Turing Lectures with Mike Wooldridge
(youtube.com): Mike Wooldridge, a
Professor of Computer Science at the
University of Oxford, discusses the current
capabilities and future potential of
generative AI, highlighting both its
transformative possibilities and the
significant challenges it presents, including
issues of bias, misinformation, and ethical
concerns.
#GenerativeAI #FutureTech
#AIChallenges #MachineLearning
#TechEthics
9

10. Introducing Generative Physical AI --
youtube.com: NVIDIA introduced
Generative Physical AI, a technology
enabling robots to learn and refine their
skills in simulated environments,
leveraging NVIDIA's AI supercomputers and
robotics platforms. This development aims
to minimise the gap between simulation
and real-world application, enhancing the
autonomy and functionality of future
robotics.
#NVIDIA #GenerativeAI #Robotics
#AItechnology #Computex2024
10

11. Grounding - Enhance GEN AI with YOUR DATA
(youtube.com): The article discusses
techniques for grounding generative AI
models to ensure their outputs are accurate
and reliable by integrating real-world data,
employing human oversight, and using
multiple models to verify results. These
methods are crucial for preventing errors in
fields like healthcare, finance, and legal
services, and involve strategies like Retrieval-
Augmented Generation (RAG) and
Reinforcement Learning from Human
Feedback (RLHF).
#AI #GenerativeAI #AIAccuracy
#AITrustworthiness #GroundingAI
11

12. Generative AI Handbook: A Roadmap for Learning
Resources -- genai-handbook.github.io: The
Generative AI Handbook offers a comprehensive
roadmap for learning about modern artificial
intelligence systems, particularly focusing on large
language models and image generation. It organises
existing resources like blogs, videos, and papers into
a textbook-style presentation aimed at individuals
with a technical background who seek to deepen
their understanding of AI fundamentals and
applications. The handbook emphasises the
importance of foundational knowledge to effectively
use and adapt to rapidly evolving AI tools and
techniques.
#GenerativeAI #AIHandbook
#MachineLearning #AIeducation
#DeepLearning
12

13. The Future of AI: In a recent LinkedIn post,
Matt Webb shared his thoughts on the
future of AI and its applications. Matt is
focused on the smaller, more ubiquitous
aspects of AI, such as home hardware and
managing intelligent agents.
#AI #FutureOfWork #Innovation
#Technology #LinkedIn
13

14. Back To Atoms: AI has always been seen as
the technology of the future but it has
finally arrived with ChatGPT and Large
Language Models (LLMs). This post reflects
on the journey of AI, the realization of its
'magic,' and the implications it may have on
the software industry and our future. The
author speculates that the next wave in
technology may bring us back to focusing
on tangible, real-world innovations.
#AI #TechFuture #ChatGPT #LLM
#Innovation
14

15. My personal AI research agenda, mid 2024
(and a pitch for work): Matt Webb shares
his latest work with AI agents, specifically a
smart home assistant demonstrating
emergent behaviour. He discusses the
simplicity of creating sophisticated AI
behaviours with minimal code and outlines
his personal AI research interests, including
human-AI collaboration, simple agents
acting in the world, and tiny, ubiquitous
embedded intelligence.
#AI #Research #SmartHome
#TechInnovation #Collaboration
15

16. The Next Great Scientific Theory is Hiding
Inside a Neural Network: Miles Cranmer
discusses the potential of neural networks
to uncover groundbreaking scientific
theories. The lecture delves into the
expanding applications of machine learning,
from text generation to construction
infrastructure. Highlighting the intersection
of AI and scientific discovery, this talk
envisions a future where neural networks
become pivotal in advancing knowledge.
#NeuralNetworks #MachineLearning
#AI #ScientificDiscovery
#Innovation
16

17. Transforming Customer Support and Sales
with Mendable's AI Solutions: Mendable
introduces Firecrawl, a tool that converts
websites into LLM-ready markdown or
structured data. Their platform offers various
AI capabilities to streamline customer support
and sales through AI-powered knowledge
bases, secure data integrations, enterprise-
grade security, and detailed customer
interaction insights. They also support custom
AI model training and have free and enterprise
pricing plans.
#AI #CustomerSupport
#SalesEnablement #EnterpriseSecurity
#AIModelTraining
17

18. Why Apple is Taking a Small-Model Approach
to Generative AI: Apple introduced its new
generative AI offering, Apple Intelligence, at
WWDC 2024. Unlike larger models from
competitors, Apple’s approach focuses on
smaller, customized models integrated
seamlessly with its operating systems to
prioritize a frictionless user experience. Apple
Intelligence is designed to handle various
tasks while maintaining privacy and
efficiency, with the speech generation and
image creation models being processed on-
device for speed and user focus.
#Apple #GenerativeAI #WWDC2024 #AI
#Privacy
18

19. Sober AI is the Norm: The article discusses the
current state of AI, emphasizing the need for
'Sober AI' amidst the hype surrounding
advanced artificial intelligence technologies.
Highlighting observations from the Databricks
Data+AI Summit, it points out that most AI
work is mundane, involving data preparation
and pipeline management rather than
groundbreaking advancements. The writer
argues that even these seemingly modest
applications hold significant value in driving
practical business intelligence solutions.
#AI #BusinessIntelligence
#DataScience #TechSummit
#MachineLearning
19

20. Can LLMs invent better ways to train LLMs?:
Sakana AI explores using Large Language Models
(LLMs) for inventing better ways to train
themselves, termed LLM². They leverage
evolutionary algorithms to develop novel
preference optimization techniques, significantly
improving model performance. Their latest
report introduces 'Discovered Preference
Optimization (DiscoPOP)', achieving state-of-the-
art results across various tasks with minimal
human intervention. The approach promises a
new paradigm of AI self-improvement, reducing
extensive trial-and-error efforts traditionally
required in AI research.
#LLMs #AIResearch #DeepLearning
#EvolutionaryAlgorithms #DiscoPOP
20

21. SWE-bench: Can Language Models Resolve
Real-World GitHub Issues?: The SWE-bench
project investigates the ability of language
models to automatically resolve GitHub
issues. It uses a dataset comprising 2,294
issue-pull request pairs from 12 popular
Python repositories, with evaluations based
on unit test verification. The leaderboard
showcases various models and their
performance on this task, with Amazon Q
Developer Agent currently leading.
#LanguageModels #GitHub
#Automation #MachineLearning
#Python
21

22. Will We Run Out of Data? Limits of LLM Scaling
Based on Human-Generated Data: Epoch AI has
estimated the total supply of human-generated
public text at about 300 trillion tokens. They project
that, at the current rate of usage, language models
will exhaust this data stock by 2026 to 2032, or
even earlier with high-frequency training. Their
forecast also explores the impact of different
training strategies on data consumption, noting that
models trained beyond computed-optimal levels
might leverage more data to enhance training
efficiency. The discussion includes possible avenues
to sustain AI progress, such as developing synthetic
data, tapping into other forms of data, and
improving data efficiency.
#AI #Data #MachineLearning #Research
#EpochAI
22

23. Reverse Turing Test Experiment with AIs:
This video showcases an experiment
where advanced AIs try to determine who
among them is the human. Created in Unity
and featuring voices by ElevenLabs, it
presents a reverse Turing Test scenario.
The experiment aims to explore how AI
identifies human traits.
#AI #TuringTest #ReverseTuringTest
#Unity #ElevenLabs
23

24. I Will Piledrive You If You Mention AI Again:
The article explores the author's frustration
with the overhyping of AI technologies in
professional software engineering. With
formal training in data science, the author
critiques how AI initiatives are often
pushed by individuals lacking in-depth
understanding, leading to a culture of hype
and grift. He emphasises the gap between
genuine technological advancements and
the superficial, profit-driven pushes that
dominate the industry landscape today.
#AI #TechIndustry #Hype
#DataScience #Critique
24

25. Gen AI Testing and Evaluation with ARTKIT: As
Generative AI (Gen AI) systems become more
integrated into critical processes, their testing
and evaluation gain importance for ensuring
safety, ethics, and effectiveness. ARTKIT, an
Automated Red Teaming and testing toolkit,
facilitates this by automating key steps like
generating prompts, interacting with systems,
and evaluating responses. It aids in creating
testing pipelines that offer insights into Gen AI
system performance, highlighting areas that
require improvement. However, human-driven
testing remains essential for a comprehensive
evaluation.
#GenerativeAI #AI #Testing #Evaluation
#Ethics
25

26. Why we no longer use LangChain for building
our AI agents:: Octomind shares their
experience using LangChain for building AI
agents and why they decided to replace it
with modular building blocks. The article
highlights the limitations and complexity
introduced by LangChain's high-level
abstractions and demonstrates how simpler
code with minimal abstractions improved
their productivity and made the team happier.
It suggests that often a framework might not
be necessary and advocates for a building-
block approach for AI development.
#AI #Tech #LangChain #AIDevelopment
#Coding
26

27. OpenAI's GPT-5 Pushed Back To Late 2025,
But Promises Ph.D.-Level Abilities:
OpenAI's long-awaited GPT-5, initially
rumored for release in late 2023 or
summer 2024, is now projected for late
2025 or early 2026. Mira Murati, OpenAI's
CTO, outlined the system's capabilities,
comparing it to having Ph.D.-level
intelligence in specific tasks, a leap from
GPT-4's high schooler-level smartness.
#OpenAI #GPT5 #AI #TechNews
#ArtificialIntelligence
27

28. Thank you!
[email protected]
matthewsinclair.com
masto.ai/@matthewsinclair
medium.com/@matthewsinclair
twitter.com/@matthewsinclair
Originally published on quantumfaxmachine.com
and cross posted on Medium.
If you’d like to sign up for this content as an
email, click here to join the mailing list.
Image: Photo by Austin Chan on Unsplash
28