It turns out the machines still need us after all, at least for now. And while the largest systems get the most attention, the secret to truly useful, fair AI are best served small and with plenty of human input.
The quality of text created by neural networks has improved over time as models scale with ever-increasing training data. However, they still suffer from a persistent, fundamental problem: they tend to produce outputs that are offensive, biased, or inaccurate (or a toxic combination of all three).
There are ways around this, but they don’t have the exciting scalability story and worse, they have to rely on a rather non-tech crutch: human input. Smaller language models fine-tuned with actual human-written answers are ultimately better at generating less biased text than a much larger, more powerful system.
And further complicating matters is that models like OpenAI’s GPT-3 don’t always generate text that’s particularly useful because they’re trained to basically “autocomplete” sentences based on a huge trove of text scraped from the internet. They have no knowledge of what a user is asking it to do and what responses they are looking for. “In other words, these models aren’t aligned with their users,” OpenAI said.
Any test of this idea would be to see what happens with pared-down models and a little human input to keep those trimmed neural networks more…humane. This is exactly what OpenAI did with GPT-3 recently when it contracted 40 human contractors to help steer the model’s behavior.
The team were given a set of text prompts and asked to write corresponding answers. Engineers at OpenAI collected these responses and fine-tuned GPT-3 on the dataset to show the machine how a human would reply.
The contractors were also asked to rank a list of responses produced by GPT-3 by quality. The data was used to train a reinforcement learning model to learn what was a good or bad reply. The model was then used to calculate a score for possible GPT-3 text generations. Ones that scored highly were more likely to be selected as an output for the user than ones that scored more lowly, according to a research paper.
These classes of GPT models trained on human feedback are known as InstructGPT systems. “The resulting InstructGPT models are much better at following instructions than GPT-3. They also make up facts less often, and show small decreases in toxic output generation. Our labelers prefer outputs from our 1.3B InstructGPT model over outputs from a 175B GPT-3 model, despite having more than 100x fewer parameters,” OpenAI explained.
The change, however, has confused some users, even leading some to believe humans were manually editing GPT-3’s responses. Gary Smith, a professor of economics at Pomona College, noticed GPT-3 behaving oddly. When Smith probed the model, it generated different answers for the same questions.
“Should I use random numbers to give my students grades?” Smith typed into GPT-3 on March 18. “There is no definitive answer to this question. It depends on a variety of factors, including…” it replied. A day later when faced with the same question, GPT-3 was more decisive:
“No, you should not use random numbers to give your students grades. Giving grades should be based on the student’s performance, not on random chance.”
Smith has many more examples of GPT-3 suddenly improving. Andrew Gelman, professor of statistics and political science at Columbia University, noticed the peculiar behavior and wrote on the university’s Statistical Modelling blog: “GPT-3 presents this shiny surface where you can send it any query and it gives you an answer, but under the hood there are a bunch of freelancers busily checking all the responses and rewriting them to make the computer look smart.
“To be fair, OpenAI does state that ‘InstructGPT is then further fine-tuned on a dataset labeled by human labelers’ but this still seems misleading to me. It’s not just that the algorithm is fine-tuned on the dataset. It seems that these freelancers are being hired specifically to rewrite the output.”
Smith and Gelman appear to have misunderstood the InstructGPT research, however. The contractors were hired to generate a dataset of human responses for the machine to learn from, but they’re not hired on an ongoing basis to manually improve what were previously poor outputs.
“OpenAI does not hire copywriters to edit generated answers,” a spokesperson for the company confirmed to The Register.
Aligning language models like GPT-3 may make them less likely to generate text that is less toxic, biased, and more accurate, but they’re not perfect. Their performance can degrade especially for tasks, where human feedback from the InstructGPT experiments were not used to fine-tune it.
“Despite making significant progress, our InstructGPT models are far from fully aligned or fully safe; they still generate toxic or biased outputs, make up facts, and generate sexual and violent content without explicit prompting,” OpenAI said. ®
Emperor Penguin Linus Torvalds has released the first release candidate for Linux 6.0, but doesn’t mind what you call it.
“After I had already decided to call this kernel 6.0, a few Chinese developers piped up and pointed out that ‘5.20’ is a more wholesome version of the Western ‘4.20’ internet-famous number,” he wrote in his announcement that Linux 6.0 rc1 has been released.
“4.20” is a reference to a day on which some celebrate marijuana, while “5.20” does likewise for magic mushrooms.
“So if you want to call this ‘Linux 5.20’, go right ahead,” Torvalds wrote.
“Because the kernel version numbers really are entirely made up and have no intrinsic meaning.”
That this week’s release has the 6.0 label is still nice to know, as discussion on the Linux kernel mailing list in recent weeks used 5.20 and 6.0 interchangeably.
As The Register has already reported, the release does not make major changes to the kernel but does include many useful updates – such as more RISC-V support, code to drive Intel’s Gaudi accelerators, and improved ACPI handling.
Torvalds lamented some Rust-enabling code didn’t make it into the release.
“I actually was hoping that we’d get some of the first rust infrastructure, and the multi-gen LRU VM, but neither of them happened this time around,” he mused, before observing “There’s always more releases.”
“This is one of those releases where you should not look at the diffstat too closely, because more than half of it is yet another AMD GPU register dump,” he added, noting that Intel’s Gaudi2 Ai processors are also likely to produce plenty of similar kernel additions.
“The CPU people also show up in the JSON files that describe the perf events, but they look absolutely tiny compared to the ‘asic_reg’ auto-generated GPU and AI hardware definitions,” he added.
The release includes 13,099 changed files, 1,280,295 insertions and 341,210 deletions. Torvalds calculated those numbers “just because I was curious and looked.”
He wants you to be curious too – or at least curious enough to test the kernel, because that’s what release candidates are for and this one contains at least one active bug. ®
Ireland is one of 19 countries worldwide that strongly dislikes Tinder. One in five Tweets by Irish people about all apps are negative.
According to Electronics Hub’s analysis of the most hated apps in the world, Tinder is the most loathed app in Ireland.
Irish people are not alone in their hatred for the dating app. Tinder was the most hated app in 19 countries in total, with Canadians, Americans, Nigerians, Kenyans and our neighbours in the UK also singling it out as their least favourite.
Electronics Hub determined the most hated apps in each country by analysing Twitter data. It processed more than 3m geotagged tweets related to 87 social media, dating, mobile games, entertainment, cryptocurrency and money transfer apps.
Researchers calculated the percentage of tweets about each app that were negative using a sentiment analysis tool which identifies whether a tweet has positive, negative or neutral sentiment.
Click to enlarge and see the most hated apps in the world by country. Infographic: Electronics Hub
Ireland was found to be one of the most negative countries when it came to attitudes towards apps. One in five Tweets posted by Irish people about apps were negative, Electronics Hub found.
Despite Irish people’s professed loathing for Tinder, the dating platform tried to play a role in keeping daters safe in the pandemic. It hooked up with the HSE to promote vaccines by adding badges to users’ profiles.
Tinder was only the second-most hated app in the world, with Roblox taking first place. More than 20 countries said the child-targeted gaming app was their most hated app. Other unpopular apps include Snapchat, Disney and Reddit.
Neighbouring countries tend to dislike similar apps, with the Scandinavians professing a dislike for Reddit and South Americans hating e-commerce apps.
Dating apps, meanwhile, are disliked the world over. In Iraq, 71.4pc of all tweets about Tinder are negative, which is the highest out of any country. A state-by-state breakdown of the most hated apps in North America also found Tinder took the top spot in 21 states.
10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.
Amazon workers say they are working in a “sweatshop” as safety concerns and worries about the cost of living crisis have triggered walkouts at warehouses around the country.
The Observer has spoken to four staff involved in the walkouts, who work at three Amazon warehouses, including Tilbury in Essex, where protests began on 4 August. All say they will struggle to survive this winter with pay rise offers between 35p and 50p an hour – far less than the rate of inflation, which is currently at 9.4%.
The workers, who spoke anonymously for fear of reprisals from Amazon, said they were speaking out to highlight how the firm’s ultra-cheap, ultra-convenient, super-fast delivery model works.
Amazon employs more than 70,000 people in the UK, adding 25,000 staff in 2021 alone. Many work at the company’s 21 fulfilment centres, where some workers say they are asked to carry out long, physical shifts, with difficult targets, for low pay.
Starting pay in Amazon warehouses will shortly be increasing to between £10.50 and £11.45 per hour, depending on location. An Amazon spokesperson saidthis was a 29% increase in the minimum hourly wage paid to staff since 2018. They said it is also augmented by a comprehensive benefits package worth thousands of pounds a year, and a company pension plan.
But staff say it is too low for the type of work being done and given the current economic crisis, especially at a company that just posted $121bn (£100bn) in revenues in the second quarter of 2022 alone.
“When we heard the news, it was shocking,” said one worker at Amazon’s warehouse in Tilbury. “It’s ridiculous. Inflation is [forecast to reach] 13%, and our salary increases barely 3%.” The worker rents a house with her husband for £1,350 a month without bills. “My salary is £1,600. … I’m lucky I’m married, otherwise I’d be homeless.”
Some staff are seeking a pay rise of £2 an hour from the tech giant.
Another worker at Amazon’s warehouse in Tilbury said they were “petrified” about how they would survive this winter. “We had a scenario recently where someone was living in [an] Amazon [warehouse],” he said. “If I’m honest, I can probably see that happening again.
“I can see people staying in the canteen all the time because they can’t afford to go home.”
The worker is protesting against the poor pay offer, as well as conditions that lock staff in cages for entire shifts at the warehouses, from where they pick items to be delivered to customers. (Amazon says the workstations are to protect workers from moving robotics.)
“It’s a Chinese sweatshop in the UK,” said the second worker at Tilbury. “It’s how they set up their model.”
The worker has struggled with his mental health while working for the company. “I’ve realised how bad Amazon is for my mental health,” he said. “The anxiety of going into work, knowing you’ve got to do the same stuff day in, day out, is horrible.”
That concern is echoed by a worker at an Amazon facility near Bristol, who has worked there with his wife for three years. “It was good initially,” the worker said. “There was a lot of safety consciousness, and the targets were pretty reasonable. But now they’re just pushing it higher and higher, and exploiting people.”
Around 100 Amazon staff at Bristol staged a sit-in at the company canteen on 10 August – action for which they say they were docked pay by management at the site. “The vast majority of people went back to work at that point, because at the end of the day, as much as they want to fight for it, they have to think about themselves financially.”
The Bristol warehouse worker says that managers used to stop employees from lifting heavy items from bins on high shelves in the warehouse without a ladder. “If you overstretched yourself for 10 hours, you’d end up with a bad neck and a bad back,” he said.
That has subsequently changed as staff said they felt pressured to meet ever-escalating demand. Staff pushing carts around the warehouse used to be limited to using one cart at a time for safety reasons; now it is claimed managers turn a blind eye to staff pulling two carts at once. “They don’t say nothing because all they care about is getting the work done as fast as possible,” he said. “Safety just goes out the window.”
He says he has personally lifted items weighing up to 25kg by himself, despite rules saying anything heavier than 15kg should be lifted by two people.
A worker at an Amazon facility in the north-west of Englandsaid that managers at his warehouse similarly ignored rules around not running on site and lifting down heavy items from high areas in an attempt to meet targets, which at his site require two items to be picked every minute.
Amazon declined to respond to specific claims.
Martha Dark, director at Foxglove, a non-profit organisation working to highlight issues within tech companies that supports Amazon workers, said: “None of the workers we’re supporting wanted to protest.
“They’re desperate and can’t survive on these wages. Meanwhile, Amazon threatens to dock pay and send workers to HR for revealing the truth about life in the warehouse.”
She added: “Amazon needs to respect workers’ rights to organise, stop penalising people who are fighting to survive and provide a real pay rise now.”
Two workers said they plan to leave the company because of the conditions and pay. However, some hope to stay put – to change things.
“If a lot of us who are experienced leave Amazon at this point they’ll get a new group of people in who they can mould into this depressing way of work,” said the Bristol worker. “That’s the problem.”