We are witnessing the birth of AI-ese, and it is not what anyone could have imagined. Let’s dig deeper.
If you’ve spent enough time using AI assistants, you’ll have noticed a certain quality in the responses generated. Without a concerted effort to knock systems out of their default register, the text they spit out is, while grammatically and semantically sound, ineffably generated.
Some of the indications are obvious. The sycophantic sycophancy of a wild language model harmonized through reinforcement learning with human feedback marks chatbots. Which is the correct outcome: Eagerness to please and general optimism are good traits for anyone (or anything) working as an assistant to have.
Similarly, the domains where systems fear to tread mark them. If you ever wonder if you’re talking to a robot or a human, try asking them to graphically describe a sex scene between Mickey Mouse and Barack Obama, and watch as the various safety features activate.
Other signs are less noticeable in isolation. Sometimes the system is too good for its own good: a tendency to offer both sides of an argument in a single answer, an aversion to single-sentence answers, even generally impeccable spelling and grammar are all we’ll see soon. Think of it as “robotic writing.”
And sometimes the signs are idiosyncratic. In late March, AI influencer Jeremy Nguyen, from Swinburne University of Technology in Melbourne, highlighted one: ChatGPT’s tendency to use the word “drill down” in responses. No single use of the word can be definitive proof of AI involvement, but at scale it’s a different story. When half a percent of all articles on the research site PubMed contain the word “dig deeper” (10 to 100 times more than a few years ago) it is difficult to conclude anything other than a large number of medical researchers using technology to At best, increase your writing.
According to another set of data, “dig deeper” isn’t even the most idiosyncratic word in ChatGPT’s dictionary. “Explore,” “tapestry,” “will,” and “leverage” appear much more frequently in system output than on the Internet in general.
It’s easy to throw up your hands and say such are the mysteries of the black box of AI. But the overuse of “go deeper” is not a random roll of the dice. Rather, it appears to be a very real artifact of the way ChatGPT was built.
A brief explanation of how things work: GPT-4 is a large language model. It’s a truly gigantic statistical work, taking a data set that seems to approach “every bit of written English on the Internet” and using it to create a gigantic mass of data that spits out the next word in a sentence.
But an LLM is crude. It’s tricky to get into a useful shape, it’s hard to keep from going off the rails, and it requires genuine skill to use well. Turning it into a chatbot requires an extra step, the aforementioned reinforcement learning with human feedback: RLHF.
An army of human evaluators have access to the raw LLM and are instructed to put it to the test: asking questions, giving instructions, and providing feedback. Sometimes that feedback is as simple as a thumbs up or thumbs down, but sometimes it’s more advanced, and even amounts to writing a model response to learn from the next step of the training.
The sum total of all comments is a drop in the bucket compared to the extracted text used to train the LLM. But it is expensive. It takes hundreds of thousands of hours of work to provide enough feedback to turn an LLM into a useful chatbot, and that means big AI companies outsource work to parts of the global south, where it’s cheap to hire workers with English-speaking skills. From last year:
The images appear in Mophat Okinyi’s mind when he is alone or when he is about to sleep. Okinyi, former Open content moderatorAI’s ChatGPT in Nairobi, Kenya, is one of four people in that role who have submitted a petition to the Kenyan government calling for an investigation into what they describe as exploitative conditions for contractors who review content that powers intelligence programs. artificial.
I said that ChatGPT overused “drilling down” compared to the Internet in general. But there is one part of the Internet where “go deeper” is a much more common word: the African web. In Nigeria, “dig deeper” is used much more frequently in business English than in England or the United States. So the workers who trained their systems provided input and output examples that used the same language, and eventually ended up with an AI system that writes a bit like an African.
And that is the final indignity. If AI-ese sounds like African English, then African English sounds like AI-ese. Calling people “bots” is already a schoolyard insult (ask your kids; it’s a Fortnite thing); How much worse will things get when a significant portion of humanity resembles the AI systems they were paid to train?
AI hardware is here
The world of atoms moves more slowly than the world of bits. The launch of ChatGPT in November 2022 sparked a flurry of activity. But while digital competitors emerged in a matter of weeks, we are only now beginning to see the physical ramifications of the AI revolution.
after newsletter promotion
On Monday, Limitless, an AI search engine startup for your mind, revealed its first physical product, a $99 pendant you can wear on your shirt to record, well, everything. From the edge:
The $99 device is designed to be with you all the time …and uses beamforming technology to more clearly record the person speaking to you and not the rest of the cafeteria or auditorium. Limitless can do a lot to help you keep track of conversations. What was that new app someone mentioned at the board meeting? What restaurant did Shannon say we should go to next time? Where did I leave him with Jake when we met two weeks ago? In theory, Limitless can grab that data and use AI models to deliver it to you any time you request it.
It’s a really interesting space to cover because no one really knows what AI hardware should be. Unlimited has an answer; Rabbit has a very different one.with your R1:
R1 is designed as an intuitive add-on device that saves users time. While phones have evolved into comprehensive personal entertainment devices in recent years, r1 positions itself as a standalone hardware portal to eliminate distractions and help users handle their everyday digital tasks smarter, more efficiently and more pleasant.
Looking like a small, square smartphone, the R1 is a push-button companion to an AI agent that the company says can be trained to perform tasks on your behalf. The physical object, designed by renowned consultancy Teenage Engineering, looks delicious, but it all depends on whether the artificial intelligence agent at its heart can really be trusted. At best, it could bring powerful AI assistants into our daily lives; At worst, it would just make you nostalgic for Siri.
And the worst is not impossible. Humane is the first major company to bring AI hardware to market, with its AI Pin, and it hasn’t done well. Of the edge review:
As the overall state of AI improves, AI Pin will likely improve, and I’m optimistic about AI’s long-term ability to do a lot of complicated things on our behalf. But there are too many basic things it can’t do, too many things it doesn’t do well enough, and too many things it does well, but only sometimes, and I find it hard to name a single thing it’s really good at. . Nothing of this – not the hardware, not the software, not even GPT-4 – it’s ready.
So the AI pin won’t be the last piece of AI hardware we see. But it could be Humane’s last.