Join online with today’s leading executives at the Data Summit on March 9th. Register here.
In AI this week, DeepMind detailed the new code-generating system, Alphacode, which it claims is competitive with top human programmers. Supermarket chains in the UK have announced that they will begin testing an automated age verification system to estimate the age of consumers when buying alcohol. And EleutherAI, a research group focusing on open-sourcing highly capable AI systems, introduces the GPT-NeoX-20B, the largest language model of its kind.
Alphacode is one of the more advanced examples of machine programming or tools that automate software development and maintenance processes. Deepmind claims that it can write “competition-level” code on the programming challenge platform Codeforce, achieving an average ranking within the top 54.3% of the last 10 contests.
The scope of application of machine programming is vast – explaining why there is excitement around it. According to a University of Cambridge study, at least half of developers’ efforts are spent on debugging, costing the software industry an estimated $ 312 billion annually. Transferring the codebase to a more efficient language can also get the kingdom money. For example, the Commonwealth Bank of Australia spent about $ 750 million over five years to convert its platform from COBOL to Java.
AI-powered code generation tools such as AlphaCode promise to reduce development costs while allowing coders to focus on creative, less repetitive tasks. But Alfacode is not flawless. In addition to being expensive to maintain, it does not always produce the correct code and – if similar systems have any indication – has problematic bias. In addition, if it ever becomes publicly available, malicious artists could misuse it to create malware, bypass programming tests, and fool cyber security researchers.
,[A]While the idea of empowering people who can’t program is exciting, we’ve got a lot of problems to solve before we get there, “said Mike Cook, an AI researcher at Queen Mary University in London.
Automatic age verification
Three supermarket chains in the UK – Asada, Co-op and Morrison – are using the camera to estimate the age of consumers as part of testing by the UK department responsible for home office, immigration, security and law and order. The technology, which was already being used at Aldi’s checkout-free location in London, estimates the age of consenting customers using algorithms trained on an “anonymous faces database”, according to the BBC. If he decides he is under 25, he must show the ID to the staff member.
Yoti – a technology provider – says it was tested on more than 125,000 faces and estimated to be 2.2 years old. But when Yoti says she is not recognizing the face or retaining the images she takes, the system raises moral concerns.
Age estimation systems, like other AI systems, can amplify any bias in the data used to develop the system. One study highlights the effect of makeup, which can cover age signs such as age spots and wrinkles, and finds that age estimation software is more accurate for men. The same research found that the software overestimates the age of younger non-Caucasians and underestimates the age of older Asians and blacks, and may be influenced by whether or not someone smiles.
In an interview with Wired, Robin Tombs, co-founder and CEO of Yoti, admitted that the company was unsure about which facial features its AI uses to determine the age of people. While he insisted that Yoti’s “hundreds of thousands” of facial training datasets “represent skin color, age, and gender” and that his internal research showed similar demographic error rates across the population, the academic literature suggests otherwise. Yoti’s own white paper shows that the technique is the least accurate for older women with dark skin.
The miscalculation of age in the supermarket can be a little more than inconvenience (and perhaps embarrassment). But it can normalize tech, leading to more problematic applications elsewhere. Daniel Lufer, a Europe policy analyst focusing on AI at the civil liberties group Access Now, told Wired that regulators need to see who these systems will fail when they are considering use cases. “Usually, the answer is people who regularly fail through other systems,” he said.
Open source language model
EleutherAI on Wednesday unveiled its latest language model, the GPT-NeoX-20B, as part of its mission to expand access to highly capable text-generating AI. Now available through the API and in open source next week, the GPT-NeoX-20B will outperform other public language models in various domains while being generally cheaper to use, according to EleutherAI.
The GPT-NeoX-20B – which was developed on the corewave, an infrastructure provided by a specialized cloud provider – was trained on EleutherAI’s 825GB text dataset and has 20 billion parameters, which is about 9 times more than OpenAI’s GPT-3. Times less. In machine learning, parameters are part of the model learned from historical training data. Generally speaking, in the field of language, the relationship between the number of parameters and sophistication is maintained remarkably well.
EleutherAI does not claim that the GPT-NeoX-20B solves any major problems with the current language model, including aspects such as bias and toxicity. But the group maintains that the benefits of releasing the model – and others like it – outweigh the risks. It can cost millions of dollars to train language models from scratch, and guessing (that is, running a really trained model) is another hurdle. The cost of running GPT-3 on a single Amazon web service is estimated to be at least $ 87,000 per year.
“From spam and astroturfing to chatbot addiction, there are obvious disadvantages that can be revealed today with the use of these models, and we expect the alignment of future models to be crucial. We feel that the acceleration of safety research is very important, “said EleutherAI co-founder Connor Lehi in a statement.
Earlier models of EleutherAI have already produced completely new AI-as-a-service startups. If history is a clue, GPT-NeoX-20B will do the same.
For AI coverage, send news tips to Kyle Wiggers – and be sure to subscribe to the AI weekly newsletter and bookmark our AI channel, The Machine.
Thanks for reading,
AI Senior Staff Writer
Venturebeat’s mission Transformative Enterprise is about to become a digital town square for technology decision makers to gain knowledge about technology and transactions. Learn more