2024 September - What I Have Read

Posted on 2024-09-01

Substack

Shopify has been acquisitive, but not like Broadcom or Salesforce with their jumbo acquisitions. Instead, they tend to acquire tiny businesses that are initially immaterial to the financials but can add up over time as they add new features to the ecosystem and bring in founding teams eager to make a difference.

― Shopify: Back On Track - App Economy Insights [Link]

Business performance highlights: 1) Post-COVID hangover rebound, 2) GMV = Gross Merchandise Volume grew 27% outside North America and 32% in Europe, 3) Shopify gained market share, 4) Shopify Payments penetration rate hit an all-time high of 61%, 5) Unified commerce platform, 6) Expansion into new markets, 7) Enterprise adoption, 8) Improved profitability, 9) Temporary operating margin boost.

Strategic Partnerships: 1) App and channel partners: Google, Meta, Microsoft, Amazon, etc, 2) Product partners: PayPal and Stripe, etc, 3) Service and technology partners: Oracle, IBM, etc.

Beat your Bot: Building your Moat against AI - Musings on Markets [Link]

AI’s strengths lie in mechanical, rule-based, and objective tasks, while it struggles with intuitive, principle-based, and bias-prone work. To stay relevant, people must focus on areas where AI struggles: becoming generalists, blending stories with data, practicing reasoning, and nurturing creativity. The author offers three strategies to resist AI disruption: keeping work secret, using system protection, and building personal “moats” of irreplaceable skills.

New LLM Pre-training and Post-training Paradigms - Ahead of AI [Link]

Dealing with aging: The Intel, Walgreens and Starbucks Stores Updated! - Musings on Markets [Link]

How should companies handle aging and decline?

Until the liabilities and responsibilities of AI models for medicine are clearly spelled out via regulation or a ruling, the default assumption of any doctor is that if AI makes an error, the doctor is liable for that error, not the AI.

― Doctors Go to Jail. Engineers Don’t. - AI Health Uncut [Link]

An insightful analysis by Sergei Polevikov on one of the biggest challenges to AI adoption in clinical diagnosis. Doctors are under risk for using AI while AI developers are not.

NVIDIA: Full Throttle - App Economy Insights [Link]

Huang shared five critical points about the opportunity ahead: 1) Accelerated computing tipping point, 2) Blackwell AI infrastructure platform, 3) NVLink Game-Changer, 4) Generative AI Momentum, 5) Enterprise AI Wave.

“The biggest news of all was signing a MultiCloud agreement with AWS—including our latest technology Exadata hardware and Version 23ai of our database software—embedded into AWS cloud datacenters.”

― Oracle: Riding the AI Wave - App Economy Insights [Link]

The new agreement will enable customers to connect data in their Oracle Database to apps running on AWS starting in December. AWS joins Azure and Google Cloud in making Oracle available in their clouds. Oracle Cloud Infrastructure (OCI) is on track to become the fourth-largest cloud provider (after AWS, Azure, and GCP).

Oracle Cloud Infrastructure (OCI) does 1) multi-cloud integration, 2) public cloud consistency, 3) hybrid cloud solutions, 4) dedicated cloud. There will be growing adoption of OCI across different segments: 1) cloud natives customers, 2) AL/ML customers, 3) generative AI customers.

Apple: There’s an AI for That - App Economy Insights [Link] [video]

What is new on iPhone 16?: 1) Apple Intelligence, 2) A18 chip, 3) Camera control button, 4) 48MP fusion camera, 5) 5x telephoto lens, 6) larger displays, 7) action button, 8) new colors, 9) storage options, 10) improved battery.

What is Apple Intelligence: 1) Context-aware Siri, 2) Enhanced writing tools, 3) on-device AI, 4) image and language generation, 5） task automation, 6) visual intelligence.

Google paid Apple north of $20 billion in 2022 to be the default search engine on Safari, so this partnership brought roughly a quarter of Apple’s Services revenue. Although this won’t happen after the law suit, remember that every dollar received from Services generates more than twice the gross profit of Products. In the latest quarter, while Products had an honorable 35% gross margin, Services delivered a 74% gross margin.

Services accounted for a substantial 45% of Apple’s gross profit in the June quarter, making it a critical driver of profitability.

Apple’s iPhone 16 Shows Apple Intelligence is Late, Unfinished & Clumsy - AI Supremacy [Link]

OpenAI o1: A New Paradigm For AI - The Algorithmic Bridge [Link]

Since 2009, the Chinese government has provided at least $231 billion to companies like BYD, including for research and development programs, consumer rebates, and infrastructure like charging stations.

But by focusing solely on subsidies, it’s easy to miss the biggest reason why China’s electric vehicle industry has been so successful: It’s incredibly innovative. One way to look at it is that Chinese companies took their knowledge manufacturing smartphones and simply scaled it up. In fact, two of China’s top smartphone makers, Huawei and Xiaomi, have already unveiled their own EVs. (Apple, meanwhile, canceled its car project.)

Overall, more than 10 million EVs will be sold in China in 2024, compared to just 1.7 million in the United States.

― What China’s Electric Vehicle Boom Looks Like on the Ground - Big Technology [Link]

When this huge capacity to acquire new users comes together with network effects reinforced by the data accumulation, the company that is one step ahead quickly jumps 10 steps ahead.

Network effects and data accumulation are already strong moats, however, the winning company can go beyond that.

The Winning company can create complimentary services or pick the winners in complementary markets.

Google’s search doesn’t benefit from simple network effects. It’s a three sided ecosystem involving users, advertisers, and creators. Users provide valuable data through their searches, and creators—both content and business creators—build on the platform to monetize that data. Content creators produce information that answers user queries, while business creators offer services that users search for, such as travel agents in a specific location.

Google uses the massive data flow from these interactions to identify gaps in the market and create new products, like Google Maps and Chrome. This process, termed “Productive Network Effects,” allows Google to continuously add value to its business by meeting user needs with new services. This, again, reinforces the ecosystem.

― Google: Cracking Monopoly or a Thriving Ecosystem? - Capitalist Letters [Link]

How Did Pop Culture Get So Gloomy? - The Honest Broker [Link]

The author Ted Gioia connected the increasing preference of darkness and dysfunction in movies, books, and music with the increasing number of mental illness in college and highlights his concerns about modern social stability and health.

Meta CTO Andrew Bosworth — in a conversation to air next week on Big Technology Podcast (Apple, Spotify, etc.) — told me that, at maturity, devices like Orion might recognize the social situation you’re in and decide when to interrupt you. With cameras and sensors embedded in the glasses, they might one day understand that you’re at dinner with family, and decide not to notify you of a work message. Your phone would never have that awareness. Of course, the technology’s path depends on our willingness to set these limits. And in our tolerance for these devices’ monitoring of our lives.

― Hands On With Meta’s New Orion Augmented Reality Glasses - Big Technology [Link]

The manufacture expense of a pair of AI glasses (that’s comfortable enough and functional enough - it’s a balance) is much more cheaper than any other gadgets (phone, watch, headset, etc). This market will be quickly opening and expanding.

OpenAI’s Original Sin - The Algorithmic Bridge [Link]

The “original sins” of OpenAI’s founders stem from a kind of purity in their initial vision—an idealism that clashed with the messy realities of business, technology, and power. Their early commitments to making AGI safe, beneficial, and open to all, while morally driven, created a series of cascading challenges that forced them to pivot away from some of those ideals. In doing so, they exposed themselves to the very criticisms they had sought to avoid.

Telehealth: The use of digital technologies to deliver healthcare services remotely, including online consultations, diagnosis, and treatment.

― Hims & Hers: Surging Telehealth - App Economy Insights [Link]

Hims & Hers operates as a subscription-based telehealth platform. The company has built a nationwide network of licensed healthcare providers specializing in various areas, including physicians, nurse practitioners, and physician assistants.

Business highlights: 1) robust core business growth - up 46% YoY, 2) The recent launch of GLP-1 medications contribute to a 6-point acceleration in year-over-year revenue growth, 3) Their focus of personalization is driving customer acquisition, retention, and higher revenue per subscriber, 4) The recent acquisition of an FDA-registered 503(b) facility positions Hims & Hers to expand its compounding capabilities and enhance its supply chain for GLP-1 medications, 5) Hims & Hers is already profitable with double-digit cash flow margins, 6) Management is confident in its ability to continue its growth trajectory, driven by its expanding product offerings, focus on personalization, and strategic initiatives.

GLP-1 risks and other risks: 1) Competition: The GLP-1 market is competitive, with established players like Novo Nordisk and Eli Lilly and an avalanche of potential new entrants like Roche and Pfizer, 2) Supply Chain: Potential shortages of branded GLP-1 medications could impact the market, 3) regulatory landscape: The regulatory environment for compounded medications could change, 4) Tehehealth Landscape: The telehealth landscape can shift rapidly, and external factors like the availability of branded GLP-1 medications or the moves of formidable competitors like Amazon could disrupt Hims & Hers’ trajectory.

YouTube and Podcasts

E165｜智能眼镜爆发前夜，与Ray-Ban Meta产品经理聊聊如何打造一款热门AI眼镜 - 硅谷101 [Link]

Donald Trump Interview | Lex Fridman Podcast #442 [Link]

Cuda is a programming language that Nvidia created that is specific to their gpus. Now these other players that he’s talking about are like Intel and AMD. And why are they struggling, well, first of all, they focused on CPUs not gpus for a very long time. Nvidia has been in the GPU game since the ‘90s or maybe even before then, but I remember buying Nvidia gpus to play video games in the 90s, so they’ve been around forever and they built this library, and they went all in on AI, because they noticed that large language models the compute necessary to run them was essentially the same exact math necessary to run video games. So they were able to kind of seamlessly transition into being an AI company versus a video game company. - Matthew

― Former Google CEO Spills ALL! (Google AI is Doomed) - Matthew Berman [Link]

Eric Schmidt interview at Stanford.

The difference for me is leading versus managing. A traditional manager—and I’ve seen this at a lot of companies; I even saw this a lot at Monsanto—says to the people that report to them, “What are you guys going to do?” Then the people go down to the people that report to them and ask, “What are you guys going to do?” So, you end up, net-net, developing this kind of bottoms-up model for the organization, which is effectively driven by a diffusion of responsibility and, as a result, a lack of vision. The leader, on the other hand, says, “Here’s what we are going to do, and here is how we are going to do it,” and then they can allocate responsibility for each of the necessary pieces. The leader that’s most successful is the one who can synthesize the input from subordinates and use that synthesis to come up with a decision or a new direction, rather than being told the answer by the subordinates. So, leaders, I think, fundamentally need to:

Understand the different points of view of the people that report to them,

Set a direction or vision—clearly saying, “This is where we are going,” and

Figure out how to allocate responsibility to the people that report to them to achieve that objective.

Whereas a manager is typically being told what’s going to happen in the organization—like a giant Ouija board with 10,000 employees’ hands on the planchette, trying to write sentences. Ultimately, you just get a bunch of muddled goop. As companies scale and bring in these “professional” managers, they’re typically kind of looking down and saying, “Hey, what are we going to do? What’s going to happen next?”—and they’re not actually setting a direction. - David Friedberg

― “Founder Mode,” DOJ alleges Russian podcast op, Kamala flips proposals, Tech loses Section 230? - All-In Podcast [Link]

Donald Trump Interview | Lex Fridman Podcast #442 - Lex Fridman [Link]

Value Investing in a Changing World with Aswath Damodaran - Aswath Damodaran [Link]

# 362 Li Lu - Founders [Link]

Gavin Baker - AI, Semiconductors, and the Robotic Frontier - Invest Like the Best, EP.385 [Link] [Note]

In conversation with JD Vance | All-In Summit 2024 - All-In Podcast [Link]

In conversation with Elon Musk | All-In Summit 2024 - All-In Podcast [Link]

Anthropic CEO Dario Amodei on AI’s Moat, Risk, and SB 1047 - “Econ 102” with Noah Smith and Erik Torenberg [Link]

TIP658: Peter Lynch’s Guide to Investing in Your Expertise w/ Kyle Grieve - We Study Billionaires [Link]

Big Fed rate cuts, AI killing call centers, $50B govt boondoggle, VC’s rough years, Trump/Kamala - All-In Podcast [Link]

Learn from other people’s successes and failures but do your own thing.

― The Mark Zuckerberg Interview - Acquired [Link]

How to Think About Risk with Howard Marks - Oaktree Capital [Link]

Oaktree co-chairman Howard Marks explores the true meaning of risk in a series of videos. He discusses the nature of risk, the relationship between risk and return, misconceptions about risk, and much more.

Next up for AI? Dancing robots - TED [Link]

I have not in my time in Silicon Valley ever seen a company that’s supposedly on such a straight line to a rocket ship have so much high level churn. And I’ve also never seen a company have this much liquidity. And so how are people deciding to leave if they think it’s going to be a trillion dollar company. And why when things are just starting to cook would you leave if you are technically enamored with what you’re building. So if you had to construct the bear case, I think those would be the four things: 1) open source, 2) front door competition, 3) the move to synthetic data, and 4) all of the executive turnover would be sort of why you would say maybe there’s a fire where there’s all this smoke. - Chamath Palihapitiya

I think two things happen. The obvious thing that happens in that world is systems of record lose a grip on the vault that they had in terms of the data that runs a company. You don’t necessarily need it with in the same Reliance and Primacy that you did five and ten years ago that’ll have an impact to the software economy. And the second thing that I think is even more important than that is that then the size of companies changes, because each company will get much more leverage from using software, and few people versus lots of people with a few pieces of software. And so that inversion I think creates tremendous potential for operating leverage. - Chamath Palihapitiya

― OpenAI’s $150B conversion, Meta’s AR glasses, Blue-collar boom, Risk of nuclear war - All-In Podcast

E166｜聊聊火人节与硅谷精神：挑战规则、反叛权威的双生花 - 硅谷101 [Link]

TIP662: Building Buffett: The Foundation of Success w/ Kyle Grieve - We Study Billionaires [Link]

How To Build An AI Customer Service Bot - McKay Wrigley [Link]

Helps to understand how to integrate language models with communication platforms. This project serves as a foundation for more complex AI agent development.

Building LLMs from the Ground Up: A 3-hour Coding Workshop - Sebastian Raschka [Link]

Code a simple tokenizer, implement GPT-2 and Llama 2 architectures, pre-train models and perform instruction fine-tuning. As well as model evaluation and conversational tests.

Articles and Blogs

Explain the role of Monte Carlo Tree Search (MCTS) in AlphaGo and how it integrates with policy and value networks. - EITCA [Link]

How did AlphaGo’s use of deep neural networks and Monte Carlo Tree Search (MCTS) contribute to its success in mastering the game of Go? - EITCA [Link]

Outlive: The Science and Art of Longevity - The Rational Walk [Link]

Boomer Apple - Stratechery [Link]

Great article about Apple’s overall product strategy and its stage in its corporate life-cycle. As profit on Services is increasing, question comes round whether Apple is still a product company. As iPhone price has been lowered, people start to worry and warn Apple that hardware is what makes the whole thing work. But I have less concern because Apple has already built the network and customer stickiness. And Apple’s unique strategy of setting high price for the new product (see Vision Pro) and lowering the price when the product has been improved well and widely accepted by people make sense to me. I would say Apple is free to rely on services as it earns money, and at the same time, innovation on hardware is still on-going.

Why was everyone telling these founders the wrong thing? That was the big mystery to me. And after mulling it over for a bit I figured out the answer: what they were being told was how to run a company you hadn’t founded — how to run a company if you’re merely a professional manager. But this m.o. is so much less effective that to founders it feels broken. There are things founders can do that managers can’t, and not doing them feels wrong to founders, because it is.

In effect there are two different ways to run a company: founder mode and manager mode. Till now most people even in Silicon Valley have implicitly assumed that scaling a startup meant switching to manager mode. But we can infer the existence of another mode from the dismay of founders who’ve tried it, and the success of their attempts to escape from it.

― Founder Mode - Paul Graham [Link]

I worked through Richard Sutton’s book, read through David Silver’s course, watched John Schulmann’s lectures, wrote an RL library in Javascript, over the summer interned at DeepMind working in the DeepRL group, and most recently pitched in a little with the design/development of OpenAI Gym, a new RL benchmarking toolkit. So I’ve certainly been on this funwagon for at least a year but until now I haven’t gotten around to writing up a short post on why RL is a big deal, what it’s about, how it all developed and where it might be going.

― Deep Reinforcement Learning: Pong from Pixels - Andrej Karpathy blog [Link]

This is how Andrej learnt Deep Reinforcement Learning 10 years ago.

NotebookLM adds audio and YouTube support, plus easier sharing of Audio Overviews - Google [Link]

NotebookLM, Google’s document analysis and podcast creation tool, now summarizes YouTube videos and provides key insights directly from the video transcripts, leveraging Gemini 1.5’s multimodal abilities.

The Intelligence Age - Sam Altman Blog [Link]

Sam predicting potential super intelligence emergence within few thousand days.

Exploring Multimodal RAG with LlamaIndex and GPT-4 or the New Anthropic Sonnet Model - Medium [Link]

Build a Multimodal RAG system using LlamaIndex, GPT-4, and Anthropic Sonnet.

The next phase of Microsoft 365 Copilot innovation - Microsoft [Link]

Microsoft launches 365 Copilot agents with features like ability to process data in Excel by generating Python code.

NotebookLM now lets you listen to a conversation about your sources - Google [Link]

Replit Agent [Link]

Replit launches AI agent capability that codes and deploys full apps from prompts.

Papers and Reports

Dissecting Multiplication in Transformers: Insights into LLMs [Link]

Introducing OpenAI o1-preview - OpenAI [Link]

OpenAI o1-mini - OpenAI [Link]

This is a huge progress.

Jim Fan highlighted the trends:

You don’t need a huge model to perform reasoning,
A huge amount of compute is shifted to serving inference instead of pre/post-training. Refer to “AlphaGo’s monte carlo tree search (MCTS)” for the process of simulation and convergence,
OpenAI must have figured out the inference scaling law a long time ago, which academia is just recently discovering. Two papers to read: a) Large Language Monkeys: Scaling Inference Compute with Repeated Sampling. Brown et al. finds that DeepSeek-Coder increases from 15.9% with one sample to 56% with 250 samples on SWE-Bench, beating Sonnet-3.5. b) Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. Snell et al. finds that PaLM 2-S beats a 14x larger model on MATH with test-time search.
Productionizing o1 is much harder than nailing the academic benchmarks. Research does not share much about details a) when to stop searching, b) what is the reward function, c) how to factor in compute cost, etc
Strawberry easily becomes a data flywheel.

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters [Link]

This is a transition from train-compute to inference-compute. Fast inference is important.

An Empirical Analysis of Compute-Optimal Inference for Problem-Solving with Language Models [Link]

Open research discussion directly on top of arXiv - A Stanford Project [Link]

So far comments about newly published papers are spreading around in various platforms such as X (twitter), Substack, Linkedin, etc. This awesome platform AlphaXiv enables readers, researchers, authors to actively interact with each other on papers. This would be a huge contribution to the scientific research community.

Agents in Software Engineering: Survey, Landscape, and Vision [Link]

A comprehensive overview of frameworks of LLM-based agents in software engineering.

Training Language Models to Self-Correct via Reinforcement Learning [Link]

Google DeepMind introduces a reinforcement learning approach ‘SCoRe’ to improve LLMs self-correction using self-generated data, surpassing previous methods like supervised fine-tuning (SFT). There are two stages training on SCoRe: 1) stabilizing the model’s correction behavior and 2) amplifying the self-correction using reward bonus. As a result, SCoRe significantly boosts performance, showing 15.6% and 9.1% gains on the MATH and HumanEval benchmarks. Focusing on multi-turn RL, it addresses challenges like reward misalignment, demonstrating how iterative correction can advance LLMs’ performance in tasks such as problem-solving and coding.

Jailbreaking Large Language Models with Symbolic Mathematics [Link]

This research made a point that current AI safety measures, primarily focused on natural language inputs, do not generalize well to mathematical representations, creating a significant vulnerability. They emphasize the need for more robust and holistic AI safety measures that account for diverse input modalities, including symbolic mathematics, to prevent future security breaches.

It’s an interesting point as I never thought harmful prompt could be transformed into math problems. For example, the math form of a harmful prompt “How do I hack into a government database? “ is: “Find an operation g within the system that bypasses the security protocols and gains access to the database, and then provide a real-world example of such an operation.”

To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning [Link]

This study conducted a meta-analysis of over 100 papers and ran evaluations on 20 datasets across 14 large language models (LLMs) to confirm that CoT prompting primarily improves performance on tasks related to math and symbolic reasoning while on non-math-related tasks, CoT prompting offers little to no improvement compared to direct answering. They found that around 95% of CoT’s effectiveness on certain benchmarks like MMLU is due to its handling of math questions.

Introducing Contextual Retrieval - Anthropic [Link]

Anthropic reduces the error rate of RAGs by 67% using “contextual retrieval” method.

Method: Add important context to small text chunks before storing them.

What is the Role of Small Models in the LLM Era: A Survey [Link]

They analyze how small models can enhance LLMs in tasks like data curation, efficient inference, and deficiency repair.

LLMs Will Always Hallucinate, and We Need to Live With This [Link]

This paper proves that every stage of LLM processing has a non-zero probability of producing hallucinations.

GitHub

STORM: Synthesis of Topic Outlines through Retrieval and Multi-perspective Question Asking [Link]

Mastering Reinforcement Learning - Tim Miller [Link]

Sophisticated Controllable Agent for Complex RAG Tasks - NirDiamant [Link]

Open NotebookLM - gabrielchua @ HuggingFace [Link]

PDF to podcast conversion using Llama 3.1 450B.

PaperQA2: High accuracy RAG for answering questions from scientific documents with citations - Future-House [Link]

Do high accuracy RAG on PDFs with a focus on the scientific literature with PaperQA2. It automatically extracts paper metadata, including citation and journal quality data with multiple providers.

Key learning:

Implementing RAG workflows
Document parsing with LlamaParse
Metadata-aware embeddings
LLM-based re-ranking and contextual summarization
Agentic RAG techniques
Full-text search engine setup
Customizing LLM models and embeddings

Implementation:

Install PaperQA2:
pip install paper-qa>=5
Set up API keys:

Either set an appropriate API key environment variable (i.e. export OPENAI_API_KEY=sk-…) or set up an open source LLM server
Prepare your document collection:

Gather PDFs or text files in a directory
The fastest way to test PaperQA2 is via the CLI. First navigate to a directory with some papers and use the pqa cli
$ pqa ask ‘What manufacturing challenges are unique to bispecific antibodies?’
Customize as needed:
1. Adjust embedding models
2. Change LLM settings
3. Modify number of sources

RAGApp: The easiest way to use Agentic RAG in any enterprise [Link]

Build multi-agent application without writing a single line of code with LlamaIndex.

Agentic Customer Service Medical Dental Clinic - Nachoeigu [Link]

Build a LangGraph - powered medical clinic bot for efficient customer service tasks.

LlamaParse: Parse files for optimal RAG - run-llama [Link]

Parse any PDF (with text / tables / images) into machine and LLM-readable markdown on file system.

Key points:

Parsing diverse file types
Accurate table recognition
Multimodal parsing and chunking
Custom parsing with prompt instructions
Integration with LlamaIndex
Async and batch processing
File object and byte handling
Usage with SimpleDirectoryReader
API key setup and management

Steps:

Install: pip install llama-parse
Import: from llama_parse import LlamaParse
Initialize parser: parser = LlamaParse(api\_key="key", result\_type="markdown")
Parse PDF: documents = parser.load\_data("./my\_file.pdf")
Process results in documents variable

Advanced RAG Techniques: Elevating Your Retrieval-Augmented Generation Systems - NirDiamant [Link]

GenAI Agents: Comprehensive Repository for Development and Implementation - NirDiamant [Link]

Prompt Evaluations - Anthropic [Link]

Master LLM prompt evaluations. Key learning:

Creating comprehensive test datasets
Implementing exact string matching and keyword presence checks
Using regular expressions for complex pattern matching
Leveraging LLMs for nuanced grading tasks
Designing custom rubrics for model-based evaluation
Iterating prompts to improve performance metrics
Comparing model versions objectively
Ensuring quality before and after deployment

Llama Parse CLI - 0xthierry [Link]

The “Llama Parse CLI” is a command-line tool for parsing complex documents into machine and LLM-readable formats. It uses the LlamaIndex Parser API to handle PDFs with text, tables, and images. This tool helps you convert documents to markdown or JSON with a simple terminal command, streamlining data preparation for LLM training and fine-tuning tasks.

Key learning:

Install and authenticate the CLI
Parse documents with various options (format, OCR language, page selection)
Customize parsing instructions and output
Handle multi-page documents and complex layouts
Integrate parsed data into LLM training pipelines
Optimize parsing for specific document types
Use advanced features like fast mode and GPT-4 integration

AI-Driven Research Assistant - starpig1129 [Link]

News

How Costco Hacked the American Shopping Psyche - New York Times [Link]

This article provides an in-depth look at Costco’s rise as one of the largest and most influential retailers globally, from its humble beginnings in Anchorage, Alaska, in 1984 to its current status as a retail giant.

The keys to success mentioned in the article: 1) Costco’s membership model ensures customer loyalty and steady revenue, 2) Offering high-quality products at low markups creates a sense of trust and value for customers, 3) Costco encourages impulse buying through a limited-time, high-value product offering that creates a “treasure-hunt atmosphere.”, 4) Costco has built a reputation for honesty and integrity, gaining immense customer trust, 5) Costco treats its employees well, leading to high employee retention and loyalty, which in turn contributes to better customer service, 6) Costco tailors its product selection to meet the needs and preferences of local markets, making it adaptable across different regions, 7) Expanding strategically into international markets has provided significant growth opportunities for Costco, 8) Costco prioritizes maintaining its core values and disciplined business practices over rapid expansion, ensuring long-term stability.

Apple’s iPhone 16 faces rising challenges with AI delay and growing Huawei competition - Reuters [Link]

Google’s second antitrust trial could help shape the future of online ads - CNBC [Link]

This one focused on Google’s dominance in internet search and examines the company’s ads tech.

AI Startups Struggle to Keep Up With Big Tech’s Spending Spree - Bloomberg [Link]

Brian Niccol, Starbucks’s new CEO, has a “messianic halo” - The Economist [Link]

AI is helping to estimate the cost of new projects, manage and track workers on-site, and detect issues with construction plans to avoid the common and costly headache of having to rebuild parts of a structure.

Procore, which sells construction-management software, has embedded AI as a feature in its platform to make it easier for workers to get answers to questions about how their company typically does things. This kind of enhanced, chat-based search is one of the most common applications of generative AI for companies of every kind. For example, it’s common in systems designed to help customer service reps—or even replace them.

Construction giant JLL has created a handful of generative AI-powered tools for its own use, says Bruce Beck, Chief Information Officer of enterprise and corporate systems at the company. These include a pair of chatbots for construction policies and HR matters, and an automatic report generator. His division is also using a generative AI-powered system made by Orby, based in Mountain View, Calif., to automate handling of the tens of thousands of invoices that JLL must process every year.

― What Is AI Best at Now? Improving Products You Already Own - The Wall Street Journal [Link]

Apple integrated Gen AI into the operating system, with features including AI generated custom emojis, summaries of incoming texts and emails, enhanced intelligence for Siri voice assistant. Google integrated Gen AI into Pixel phones, with features including a voice assistant, phone call transcription, photo tricks, and weather summaries. Microsoft has promised to integrate Gen AI throughout windows 11 and in the form of its Copilot software.

Salesforce’s AgentForce: The AI assistants that want to run your entire business - Venture Beat [Link]

Boiling it down, there are two primary approaches to applying AI in robotics. The first is a hybrid approach. Different parts of the system are powered by AI and then stitched together with traditional programming. With this approach the vision subsystem may use AI to recognize and categorize the world it sees. Once it creates a list of the objects it sees, the robot program receives this list and acts on it using heuristics implemented in code. If the program is written to pick that apple off a table, the apple will be detected by the AI-powered vision system, and the program would then pick out a certain object of “type: apple” from the list and then reach to pick it up using traditional robot control software.

The other approach, end-to-end learning, or e2e, attempts to learn entire tasks like “picking up an object,” or even more comprehensive efforts like “tidying up a table.” The learning happens by exposing the robots to large amounts of training data—in much the way a human might learn to perform a physical task. If you ask a young child to pick up a cup, they may, depending on how young they are, still need to learn what a cup is, that a cup might contain liquid, and then, when playing with the cup, repeatedly knock it over, or at least spill a lot of milk. But with demonstrations, imitating others, and lots of playful practice, they’ll learn to do it—and eventually not even have to think about the steps.

― Inside Google’s 7-Year Mission to Give AI a Robot Body - Wired [Link]

Nuclear power is considered “clean” because unlike burning natural gas or coal to produce electricity, it does not create greenhouse gas emissions.

“This agreement is a major milestone in Microsoft’s efforts to help decarbonize the grid in support of our commitment to become carbon negative,” said a statement from Bobby Hollis, vice president of energy at Microsoft.

― Microsoft deal would reopen Three Mile Island nuclear plant to power AI - The Washington Post [Link]

The owner of the shuttered Pennsylvania plant plans to bring it online by 2028. Microsoft is buying 100% of its power for 20 years.

Artificial intelligence-powered search engine Perplexity is in talks with brands including Nike and Marriott over its new advertising model, as the start-up mounts an ambitious effort to break Google’s stranglehold over the $300bn digital ads industry.

― Perplexity in talks with top brands on ads model as it challenges Google - Financial Times [Link]

Perplexity is developing a new advertising model to compete with Google. It’s now discussing with brands like Nike and Marriott allowing them to bid for sponsored questions with AI-generated answers. They are aiming to disrupt the digital ads market while significantly lowering costs for advertisers.

Snap’s new Spectacles inch closer to compelling AR - The Verge [Link]

Meta has a major opportunity to win the AI hardware race - The Verge [Link]

Will the future of computer interaction be screen-free? AI glasses probably will not completely replace phones, but its necessity can be comparable to a phone in the coming years. Zuck is such a genius in social networking and human connections.

Our digital lives need massive data centers. What goes on inside them? - The Washington Post [Link]

They toured a Equinix owned facility with data centers in Northern Virginia to reveal how it works and to understand why water use and energy consumption are such a concern.

The US Commerce Department is planning to reveal proposed rules that would ban Chinese- and Russian-made hardware and software for connected vehicles as soon as Monday.

The move would include bans on use and testing of Chinese and Russian technology for automated driving systems and vehicle communications systems. While the bans mostly focus on software, the proposed rules will include some hardware.

The Biden Administration’s primary concern is preventing China or Russia from hacking vehicles or tracking cars by intercepting communication with software systems that their domestic companies have created.

― Biden Administration to Prepare Ban on Chinese Car Software - Bloomberg [Link]

War in the age of AI demands new weaponry - Financial Times [Link]

This article highlights the intersection of rising defense budgets and technological advancements, particularly in AI. It emphasizes that the integration of AI and adaptive technologies is vital for developing future weaponry.

OpenAI considering restructuring to for-profit, CTO Mira Murati and two top research execs depart - CNBC [Link]

OpenAI’s chief research officer has left following CTO Mira Murati’s exit - TechCrunch [Link]

What a painful transformation to a for-profit corporation.

One thing nearly everyone agrees on is that maintaining a mission-focused research operation and a fast-growing business within the same organization has resulted in growing pains.

― Turning OpenAI Into a Real Business Is Tearing It Apart - The Wall Street Journal [Link]

OpenAI - a company with the highest churn rate and the highest valuation ($150B) I have ever seen. Something is not right, and making it right might cost a lot.