The future of (my) work

I hope you had a great summer. Mine has been exciting, with a new job, distant travel and more. First, some admin, and then some stories.

Monthly cadence, premium subscriptions & coaching

I am incredibly thankful for the support of all the folks who have paid to subscribe to Be Like Animal, Be Like Plant. I’m turning off the paid subscription feature now because my primary motivation is to provide content for everyone, so I want to remove any barriers to this goal.

From now on, this newsletter will be monthly, and my hope is to make deeper dives into specific topics with long-form content. Stay tuned :)

I’m not taking any new coaching clients. But coaching will continue to be a part of my life forever. The people I meet and the conversations I have as part of this practice are some of the most fun, thought-provoking, and rewarding, long may this continue.

New work on new work

In 2008, in Japan, I founded a company called Gengo, as an idea to create a simple way for people to order human translation online. At this time, human translation, especially between languages like Japanese and English, was the only way to get a useful output; the noise and garbage coming out of machine translation making it useful only for spam. I founded the company shortly after Amazon had released its Mechanical Turk platform, and global payments and other tech was making it possible to offer people small, bite-sized chunks of online work efficiently. We recruited translators, tested their abilities, and built the tools to do the work.

By the time I left the company in 2015, we had become a crowdsourced translation platform, covering dozens of languages, working with thousands of customers large and small, and with more than 20,000 translators around the world. Gengo was acquired in 2019 by Lionbridge, a leading language services company. Gengo.com is still around and still offers a fast translation service, although I have had no connection to or insight into the company for years.

In the 16 years since Gengo began, the world of online work, crowd platforms and the ‘gig worker’ economy has evolved, matured and expanded. COVID and a mass transition to remote work, whether permanent or temporary, has changed how the whole world thinks about work, and the line between being employed as a consultant, a full-time employee or an occasional gig worker has also started to blur. Most recently in the headlines, companies like Scale.ai and Appen have been under heavy scrutiny for their work practices.

This summer I joined a company called iMerit, which was founded by CEO Radha Basu over a decade ago. My role is to build iMerit’s new global specialist workforce. In some ways it feels like a homecoming, as a couple of my ex-Gengo colleagues are there, including Jeff who I wrote about in a previous post.

iMerit was started originally as an offshoot of a nonprofit foundation, Anudip. Anudip has over 90 centers in India, where they have trained over 500,000 people from the ground up, giving them computer skills and job readiness training. iMerit built its original business by connecting some of these same folks with jobs that matched their skills, and has developed into a leading business providing data annotation and AI training by humans.

iMerit — and the world of data annotation and human AI training — is like a fascinating onion, which grows more interesting as you peel back the layers.

Training advanced AI with human input

Amid the flurry of AI news, there are two specific problems that are making it more difficult and expensive to train AI. First, the free data that has been used to train generalist models is drying out, and second, in many domains, the AIs’ abilities have started to match and exceed the abilities of trained, experienced humans.

Less free stuff

Large language models such as ChatGPT have built much of their current abilities by ingesting the almost-unlimited amount of content freely available online - Reddit posts, NYT articles, Tumblr blogs. But by now, most of this low-hanging fruit has been consumed, and there is now not much left that contains useful information in the right form to help train advanced AIs beyond their existing abilities. So when companies want to train their large language models, their computer vision models, or other kinds of intelligence, they need to rely on humans more and more.

For example, if you want to train a computer vision model to recognize corn crops in a drone video, you can only go so far with publicly-available data. You might need to gather your own images of corn and ask humans to show you exactly where they occur in the image, and show you where weeds and other crops are, to improve the AI’s chance of recognizing corn in the future.

Higher expertise

We are now at the point where many LLMs can pass exams like LSATs and produce responses that are better than most college graduates. So who do they learn from? You might learn middle school mathematics from someone who had reached a university undergraduate level. But who do you learn from when you have a masters degree yourself?

As it stands, the more investment is made globally into AI development, the more of human training that will be done, and the higher expertise that will be required. Just as humans in the early 20th century performed calculation work that computers could not do (and were called computers), in 2024 humans are being asked to do the perceptual and generative work that computers cannot yet do.

This is by all accounts an immense effort.

And generally, as models get ‘smarter’, the quality, precision and expertise level of the human training required for each incremental increase in the AI’s IQ keeps going up. Low-precision training data that was useful 10 years ago to train a rudimentary model is almost useless today. So over time, the practice of training an AI is becoming a more specialized and nuanced profession, and the expertise required of the human trainers is increasing, such that it requires folks with serious levels of experience such as cardiologists, pHD-level mathematicians, sports statisticians and other experts. 

The division of crowd labor

There are a number of ways to organize the training of an AI. How you do it might seem like an engineering question, but it quickly becomes a human question.

With a ‘crowd’ method, you might split a problem into tiny short jobs, try to make each job as easy as possible to complete, and present all the jobs in a dashboard such that a worker can pick off one or more, whenever they have the time.

In some ways, this was the Gengo approach, our intention being to turn something that seemed to be complex, cumbersome, expensive and time-consuming for both the translator and the customer into something simple and fast, again for both parties. Instead of haggling with a stranger, being unsure if they would pay your invoice, and dealing with all the admin, the idea was that translators could focus on their core task, develop their skills, and work with low risk on their own terms.

Of course, there are downsides. Platforms are tempted to over-populate the worker pool to decrease the waiting time for the customer, often meaning that workers are faced with short-term gluts and then long wait periods where work is scarce. The workers’ ability to build up relationships and improve their ability to access work is at the mercy of the platform, and the algorithms used by the platform to allocate work are opaque at best. And a platform might work great for one cohort of workers, and be a nightmare for a different cohort with different needs.

Finally, when you break problems down into too many small parts, the overall meaning is lost, which means that individual workers cannot make decisions and spot solutions that might be clear if they had greater context.

At its worst, this kind of platform work can start to feel like you are in the indie horror movie Cube, where contractors have completed micro-engineering tasks offered by an anonymous corporation, not knowing that they have played their part in constructing a deadly labyrinth.

iMerit’s approach is different. At its best, it is like assembling a crack team based on complete knowledge of the skills and attributes of the work required and the workforce available. Rather than making many jobs available to huge numbers of anonymous workers, we create smaller, more focused teams that zero in on the project at hand, and we talk and interact like humans, rather than just sending anonymous tasks around to strangers. To scale this approach requires more thought, more technology, and much more focus on culture — but the result is better output, and a much more engaging work experience for everyone.

Fairer Work

Organizations like Fair Work and the Partnership on AI have created transparency guidelines and a ratings scale, designed to encourage platform companies to treat workers fairly and in ways that give them peace of mind.

I’m sad to say that Gengo received just a 1/10 on the Fair Work rating scale, and I hope that they respond to this rating by making the necessary changes to improve their ranking. It’s a weird position to be in, seeing a company that you founded perform so poorly on metrics that are so close to our original values - it should be a slam dunk.

Industry bodies like Fair Work were not around 10 years ago, and the fact that they are increasing in number and have the backing of serious institutions is reassuring. However, adoption of their recommendations is still voluntary, and it is clear that some of the largest players in the AI training space are focused on other priorities.

It is also clear that the world of work itself is evolving quickly.

The purpose of Be Like Animal, Be Like Plant is to explore how we work, and how our paradigms inform the way we work. Over the next 12 months as I work with iMerit I will dive deeper into these trends and issues in our industry, and try to outline what I think is a coherent and optimistic path towards a positive work environment for everyone. Over time I can talk in more specific detail about the approaches we are using to build our workforce.

This starts with a trip.

A trip to India

When I started at iMerit, I wanted to understand the history of the company and the corporate culture in its centers, where thousands of full-time employees work on data annotation and advanced AI training tasks. iMerit has 10+ centers in India, Bhutan and New Orleans. This August I was lucky enough to visit a couple of the centers in different parts of Kolkata.

I had never been to India before. (The closest was a visit to Sri Lanka when I was just a few years old, of which I have zero to very little memory, but a bucket was involved).

Kolkata is in West Bengal, and I visited in the middle of the monsoon season which spans from June-September. Enjoy watching glimpses of the city from your airplane window, through clouds, lightning flashes and thunderstorms as you land at 4:30am.

The history of Kolkata is long, fascinating and of course bloodstained by the British East India Company. Much of the tourist trail in the city is devoted to looking at the various British buildings constructed by the Company and later converted to British governmental infrastructure.

But if you look more broadly into Bengal itself, you will come across Rabindranath Tagore (Thakur), a fascinating Bengali polymath, contemporary, friend and occasional intellectual sparring-partner of Gandhi who made huge contributions to Indian culture, politics, writing and poetry. Tagore has the distinction of being the only individual to have written the national anthems of two countries — India and Bangladesh, and to have had a hand in a third — the anthem of Sri Lanka.

I was lucky enough to be invited to read Tagore’s poem Where the Mind is Without Fear at iMerit’s all-hands celebration on the eve of India’s Independence Day. This is a 5,500-person-strong online and real-life event, which featured music and dancing from each center - my favorite was in Bhutan.

The poem was written while the British still ruled India, and expresses a desire for freedom which was realized 34 years later, on 15th August 1947, sadly after Tagore had passed away. Here it is:

As a British/Australian, the privilege and incongruity of being invited to read this poem on Independence day, about India’s freedom from colonial rule, was not lost on me.

A visit to Metiabruz

One of the iMerit centers I visited was in Metiabruz, a sector of Kolkata. Metiabruz is part of the “Garden Reach” area, which is connected to the history of Nawab Wajid Ali Shah, eleventh and last King of Oudh, who after being ousted by the East India Company, made Garden Reach his refuge.

Metiabruz is sometimes described as a ‘mini-Lucknow’, because of the history of Wajid Ali Shah trying to recreate the culture of that city: A second Lucknow now arose on the outskirts of Kolkata, with “…the same bustle and activity, the same language, the same poetry, conversation and with…kite-flying, cock-fighting, quail-fighting, the same opium addicts reciting the same tales, the same observance of Muharram…the same Imam Baras”*

Today, Metiabruz is a relatively conservative community where Islam is the predominant religion. iMerit has a state-of-the-art, all-women centre, right in the central marketplace. More in this video, which is 11 years old and shows the iMerit center in its very early days.

The privilege for me was to be invited into this community and to meet dozens of incredibly motivated, diligent and energetic folks with real passion and light in their eyes. As someone who has had the chance to visit workplaces around the world, office culture is normally instantly apparent, and this was no exception.

The experience of visiting Metiabruz was a reminder of how positive good work can be, and how transformational and empowering it can be to offer good jobs. The center exists in and in connection to its location. One clear question is how to build and grow a global remote workforce that has some of the same qualities, and is connected to the energy and diligence of places like Metiabruz.

Living Root Bridges

Next time, I hope to visit other centers in different parts of India, as each is in a different region and surrounded by different culture and history. Another iMerit center is in Shillong, in the Meghalaya region, home to the fascinating Living Root Bridges constructed by the Khasi / Jaiñtia peoples. These bridges are highlighted in the excellent book Lo-Tek, compiled by Julia Watson, among dozens of other examples of smart, unified and harmonious design techniques from First Nations peoples.

These bridges are built, or more accurately, grown in the mountainous terrain along the southern part of the Shillong Plateau, in a painstaking process:

A living root bridge is formed by guiding the pliable roots of the rubber fig tree across a stream or river, and then allowing the roots to grow and strengthen over time until they can hold the weight of a human being. The young roots are sometimes tied or twisted together, and are often encouraged to combine via the process of inosculation. As the rubber fig tree is well suited to anchoring itself to steep slopes and rocky surfaces, it is not difficult to encourage its roots to take hold on the opposite sides of river banks.

The history of the Khasi is incredibly interesting, to me especially the resonance with other cultures where history is preserved orally - read Folk-tales of the Khasi for some examples of this.

But most inspiring to me, the living root bridge acts as a metaphor for how to build strong connection — the bridge grows organically and finds the right shape for its purpose over time. It can’t be designed completely all in advance — it’s a process of feedback and kaizen 😉 My hope is that in my new role, I can help build our programs in such a way that our global workforce is integrated into all of the impressive existing work that iMerit has done over the past 10 years. Beyond this, I hope I can slowly start to integrate what I learned from Japanese, Australian, British, European and American culture into my work. We have plenty of time.

See you in a month’s time.