Issue #9
AI Is a Lot of Work | Semiconductor 101 | How an EU-funded agency is working to keep migrants from reaching Europe | Bonus | Public SaaS Comparables Table
The Waterfalls of Tivoli with figures in the foreground by Carlo Labruzzi
Source: Artvee
Tech 🤖
AI Is a Lot of Work
By The Verge
Annotations
For AI to find patterns in large datasets, the data needs to be sorted and tagged, this is where annotators come in. Annotators help in labelling and clarifying the data so that it becomes easier for the AI to process the information.
The year 2007 marked an inflexion point in the field of machine learning. AI researcher Fei-Fei Li, then a professor at Princeton, believed that training on larger data sets comprised of millions of labelled images was key to improving image-recognition neural networks. To achieve this task she found thousands of workers on Mechanical Turk, Amazon’s crowdsourcing platform where people around the world complete small tasks for low rates. The resulting annotated dataset, called ImageNet ushered in a new era of rapid adoption in the industry.
Data vendors that work with tech giants such as Microsoft, Google, and OpenAI that supply the required labelled datasets, are of three types-
Private outsourcing companies with call-centre like offices such as CloudFactory.
Crowdworking sites such as Mechanical Turk and Clickworker where anyone can sign up to complete tasks.
Sites like Scale AI where anyone can sign up but to work on projects they should pass exams, attend training courses and undergo performance monitoring.
The buyers of the labelled datasets demand strict confidentiality as data is increasingly seen as a moat in the AI arms race.
Annotators work on monotonous tasks for hours on irregular projects. They are forced to undergo unpaid training, only to work on a few tasks before the project abruptly ends.
Though using ChatGPT could feel like magic, it is important to know that the language that fuels ChatGPT has been refined by several rounds of human annotation. The first round involves contractors writing examples of how the engineers want the bot to behave- creating questions followed by correct answers. Once the model is trained on these examples, more contractors prompt and rank responses. This circuitous technique is called reinforcement learning from human feedback (RLHF).
Given the huge amount of costs involved in getting annotated datasets, AI companies like Anthropic and tech giants like Meta have started using AI to drastically reduce the amount of human annotation needed to guide models, while others have started using GPT-4 to generate training data.
Semiconductor 101: Understanding the jargons, what semiconductors actually do and the complex global supply chain
By Abhishek Singh
Annotations
System-on-Chip (SoC) which can be found in smaller devices like smartphones and wearables, combines various elements such as processors, memory, I/O interfaces, and additional components like GPUs or digital signal processors (DSPs).
Chip node size is seen as a proxy for advancements in chip technology. Smaller node sizes allow more components to be packed onto the chip and more components mean the chip can handle more complex tasks and calculations.
Source: Abhishek Singh on Medium
Business models in the semiconductor industry:
Fab-less Companies: Referring to the lack of fabrication, fab-less companies design the chips but don’t manufacture them. Ex: NVIDIA, AMD.
Pure-Play Foundries: These companies don’t design the chips themselves but provide the manufacturing facilities and expertise to make chips for fab-less companies. Ex: TSMC, GlobalFoundries.
Integrated Device Manufacturers (IDMs): IDMs take control of the entire process, from designing the chips to manufacturing them. Ex: Intel, Samsung.
Source: Abhishek Singh on Medium
Top countries in the semiconductor supply chain:
United States: The US has many things going for itself -
from its strength in chip design and software development to a strong ecosystem of highly-qualified engineers and research institutions driving innovation in chip architecture and software algorithms.
The Netherlands: The country is a prominent player in the semiconductor equipment and components segment. Based in The Netherlands, ASML is one of the few players in the supply of advanced lithography systems used in chip manufacturing.
Taiwan and South Korea: These countries are specialists in semiconductor manufacturing and distribution. Taiwan serves as the HQ for semiconductor foundries, such as TSMC and UMC. South Korea is home to Samsung and SK Hynix, well-known memory chip manufacturers. TSMC and Samsung are currently the only players who can produce cutting-edge 2–7 nm chips.
Japan: Japanese manufacturers like DISCO specialize in precision machinery for high-performance chips while others such as Kioxia (formerly Toshiba Memory), Renesas Electronics, and ROHM Semiconductor are heavy-weights when it comes to the development and manufacturing of memory technologies such as NAND flash, DRAM, and NOR flash.
China: Chinese companies like SMIC and Unigroup are spearheading the development of the local semiconductor industry, focusing on manufacturing and distribution.
Geopolitics 🌏
How an EU-funded agency is working to keep migrants from reaching Europe
By Coda
Annotations
In 2019, the European Union (EU) signed a deal to provide around €20M worth of radar, undersea and airborne drones, radios and other technology, as well as training, to the Tunisian government. Tunisia recently surpassed Libya as the most heavily travelled route for illegal migration to Europe across the Mediterranean.
Over the past decade, the EU has inked similar deals with almost every non-EU country that borders the bloc. At the heart of these deals is the International Centre for Migration Policy Development (ICMPD), based in Vienna which has become one of the bloc’s go-to intermediaries for supplying surveillance equipment and training to police and coast guards in EU-bordered countries.
The ICMPD serves as an intermediary between the European Commission (EC) and the border regions as the EC can't just hand over equipment to, say, the Moroccan government, as that requires the European Parliament’s approval which will be hard to come by.
The ICMPD’s clients consist of EU states and intergovernmental organizations. The organization’s budget has increased by more than 3x from €16.7M in 2015 to €58M in 2022 with more than 56% of the ICMPD’s budget coming from the European Commission.
Established in 1993, the ICMPD was born out of the need to curtail the mass invasion of Russians post the fall of the Soviet Union. Initially, the ICMPD acted as a mix of a policy think tank and a diplomatic organization, facilitating dialogue among states on issues related to borders and migration and publishing policy briefs. The inflexion point came in 2015 when more than one million people came to Europe having fled the Syrian civil war. That same year, the ICMPD appointed a new director, an Austrian conservative politician - Michael Spindelegger.
Bonus 🎁
Nauru
By Zach Seward
A very interesting article on the history of Nauru, an island nation in Micronesia.
A Field Guide to Map Projections
By Where Exactly Maps
The article serves as the basic introduction to the mechanics of projections.
Source: fDi
The soccernomics of Fifa’s changing sponsorship
By fDi
An insightful article on the evolving geography of the Fifa World Cup sponsors.
Free Resources 💡
Public SaaS Comparables Table: Meritech Capital’s Public SaaS Comparables Table is a comprehensive database of close to 100 US public SaaS companies. Available metrics include Implied ARR, Net New ARR, EV, EV/NTM Revenue, etc.
A wonderful read! Enjoyed this issue the most.