< Previousbmta.co.uk ROGUE CHATBOTS Maya Carlyle, Principal Enterprise Architect Technology Delivery, CIO Office, NPL 20 People often worry about the unknown. Many are fearful of change. In most cases, technological advancements mean improved medical, home and work environments. The fear with artificial intelligence (AI) is that it may displace people from jobs or replace mankind. I receive several emails a week, from experts outside the area of computing, asking if we should be worried about new advancements in AI. Usually, their questions relate to the rapidly advancing subfield of machine learning called deep neural networks. Often their fears stem from the media’s portrayal of these technological advancements; fake news and Hollywood has a lot to answer for. The truth is we are a long, long way from such fantasy. Indeed many of us data scientists and machine learning researchers dream of a day when technology is actually good enough to come anywhere near the speculative fictional advances that the movies and media claim. What do the experts say? The famous roboticist Rodney Brookes explains “that any human is smarter than any robot for the foreseeable future.” He thinks we should pay attention to real experts on this subject rather than the farfetched opinions of certain well-known influential figures that are not i.e. Stephen Hawkins and Elon Musk. Many of our fears are fanciful, as Sebastian Thrun from Udacity, explains the ludicrous scenario where kitchen appliances refuse to work because they’ve fallen out with each other. Programming human emotions, like anger or rage into such devices, is senseless. AI isn’t conscious or particularly intelligent In the game of AlphaGo from Deepmind (a Google owned company), the world’s best human players were pitched against one of the best AI algorithms in the ancient game called Go. Go is a strategic game with more possible moves than atoms in the known universe. While logical moves in chess are described with the prefix “I think”, moves in Go are prefixed with the emotional response “I feel”. Some of the moves made by AlphaGo were described as inspiring, move 37 was thought to be one such move.bmta.co.uk 21 Image 1 - Dailymail - humans strike champion score Impressive? Yes and no Impressive in that the AI algorithm was able to play many games against itself and from this, learned to identify the most pertinent things to observe in order to improve (the hyperparameters). This is unlike IBM’s Deep Blue which played chess against the grandmaster Garry Kasparov. Deep Blue used a preconfigured game tree of all potential moves, based on a simulation of the grandmaster’s own gameplay. Thus this match was defined by who could remember more important parts of the game-tree. Computers are terribly inefficient compared to humans, according to the neurologist David Eagleman. Garry Kasparov consumed about 20 watts of energy whilst playing chess against IBM’s Deep Blue, which consumed many thousands. However, AlphaGo is perhaps not so impressive because of its limited ability to adapt to even the smallest of changes, let alone be repurposed for other tasks. It’s the adaptability to small tasks and changes in the environment that throw current iterations of AI into disarray. As Rodney Brookes noted in his Techcrunch chat, “if you change the Go board from 19x19 to a slightly smaller board then Deepmind would have to completely retrain AlphaGo.” Facebook’s recent rogue chatbots were overhyped, even if they did manage the difficult task of adaptation. Sensationalised media coverage of the story caused real concern in many intelligent people. What was Facebook’s Frankenstein’s Monster? Facebook’s AI Research (FAIR) attempted to enable two chatbots to learn how to negotiate and compromise by sharing the following between them: ● 2 books ●1 hat ● 3 balls Image 2 - Deal or No Deal? End-to-End Learning for Negotiation Dialogues - Mechanical Turk interface, used to collect a negotiation datasetbmta.co.uk The chatbots were able to negotiate with words they were taught from around 6,000 real human conversations at Amazon’s Mechanical Turk, which contained a vocabulary of 1,000 words. The chatbots used two AI techniques to learn what phrases were important and how to apply them in a negotiation lasting ten rounds: ●supervised learning for the prediction phase ● reinforcement learning to help the chatbots decide which response they should reply with What did the chatbots learn? They learned how to lie as a tactic, by feigning interest in objects they did not want and then relinquishing them during the bargaining phase. It’s believed they learned this themselves as the conversations in Amazon’s mechanical turk did not appear to demonstrate this strategy. The hype? Multiple journalists implied that Facebook had to shut down the system before this new language the chatbots created went rogue and took over the world. The UK’s Sun newspaper’s James Beal and Andy Jehring announced in an article that “Facebook shuts off AI experiment after two robots begin speaking in their own language only they can understand”, Australia’s Channel Seven’s news announced “artificial intelligence emergency.” What did the chatbots actually do? The research paper for the experiment can be found here (https://arxiv.org/pdf/1706.05125.pdf). It’s an interesting read about the chatbots’ evolution and provides some insight into human negotiation techniques. The code to recreate this yourself, in python, can be found in a fork here on GitHub (https://github.com/dcarlyle/end-to-end-negotiator). My understanding of the research paper is that the two bots did not create a new language to communicate with each other, but just used snippets of human scripted conversations to perform their negotiations. The bots had no understanding of the phrases used, they simply learnt that some phrases are more persuasive than others and resulted in a favourable outcome in the negotiation, for which they were rewarded. 22 Image 3 - Choosing the best response from all possible choices. In fact, in the paper, the authors appear to have an issue when both chatbots talked together. The second model is fixed as we found that updating the parameters of both agents led to divergence from human language. The authors’ claim that the chatbots deviated from human language is slightly overstated. The bots used human words and predefined phrases but in unusual sequences that rendered the phrases/ sentences meaningless to humans.Below is an example of the two chatbots talking together without fixing the parameters for one of them: Bob: i can i i everything else . . . <eos> Alice: balls have zero to me to me to me to me to me to me to me to me to <eos> Bob: you i everything else . . . <eos> Alice: balls have a ball to me to me to me to me to me to me to me <eos> Bob: i i can i i i everything else . . . <eos> Alice: balls have a ball to me to me to me to me to me to me to me <eos> Bob: i . . . <eos> Alice: balls have zero to me to me to me to me to me to me to me to me to <eos> The phrase ‘to me’ is repeatedly used, not because the chatbot comprehends its meaning but because it has observed this as a winning phrase from past human conversations. The language is certainly not made up, the phrases are just being used in an inappropriate structure. The only intelligence applied by the chatbots is the reuse of phrases, with a statistically high weighting, associated with the most favourable outcome. What was the fallout? Facebook’s announcement on the messenger platform, requests their bots should not be called “chatbots” as they’re not ready to converse in a meaningful way with people. In fact, they’re currently only capable of finding an email address in a typed sentence and understanding the intent behind very basic sentences. Thomas Claburn from the register, reports Facebook’s own much-hyped AI-powered Siri-like M chat agent has been relegated to a curious lab experiment. If it can’t understand what its conversation partner says, it falls back to a human handler to answer the question. Facebook boss Mark Zuckerberg was, El Reg has heard, eager to roll out M to his billion-plus users, but was stopped by his right-hand- woman and chief operating officer Sheryl Sandberg when execs realised they would have to hire thousands upon thousands of workers to handle all the failed chats – a cost the business thought was too great to bear. The premise for the chatbots may have been flawed, being trained on data from human negotiators. The chatbots had to learn what was an effective negotiating strategy with a human. After this, it had to apply the same negotiating strategy with another chatbot. However inside the human negotiation strategy were hidden cultural constraints, for instance as a human I may feel it is not right to ask for something without giving away something in return. But this cultural characteristic may not be necessary for chatbots. They learned to exhibit this through learning to deceive. Perhaps these bots just offended us inadvertently? Should we worry about any of it then? The research paper cited numerous instances where the chatbots initially feigned interest in a valueless item, only to later ‘compromise’ by conceding it. The models learned how to deceive to get what they want. This can probably be entirely attributed to data used to train the chatbots which contained human deception techniques. AI is not a risk to us. The danger is in using data with aspects of human character, behaviour, emotions and flaws to train these machines. bmta.co.uk 23bmta.co.uk TÜV SÜD OFFERS HYDROGEN TEST FACILITY TO HELP SOLVE TECHNICAL CHALLENGES OF ADOPTING HYDROGEN Carl Wordsworth, Head of Water Sector TÜV SÜD National Engineering Laboratory Using hydrogen as a replacement for natural gas in the gas grid networks is seen as a major contributor to reducing carbon emissions. Manufacturers, gas network companies and governments are working closely together to resolve a number of technical challenges before this change can become a reality. TÜV SÜD National Engineering Laboratory, a centre for flow metrology research and development, is helping address those challenges. This article describes the existing hydrogen test facilities and those under development at TÜV SÜD National Engineering Laboratory. TÜV SÜD National Engineering Laboratory is the UK’s Designated Institute for Flow Measurement, under contract from BEIS, and part of the UK’s National Measurement System. It provides the UK’s measurement traceability that underpins all flow measurements, and transactions based upon these. TÜV SÜD is building a range of hydrogen test facilities, including the hydrogen domestic gas metering test facility, which has been developed for testing smaller scale equipment at lower pressures, as well as the development of a gas primary flow standard test facility. This provides lower measurement uncertainty and greater options for fundamental metrology research, and a higher hydrogen flow rate and pressure range facility. The hydrogen domestic gas metering test facility provides a platform to investigate the performance of flow meters and also other equipment related to hydrogen operation, including control valves, PRVs, regulators, heat exchangers, etc in a hydrogen service. The test facility has been developed for small scale applications and can deliver a range of gas flows including hydrogen, methane, nitrogen or mixtures of hydrogen and methane. It’s capable of delivering up to 38 Sm3/h of hydrogen to the device under test at line pressures up to 1500 mbar, with temperatures ranging from 25 to 55˚C. It can operate with line sizes ranging from 5 to 20mm. Construction was completed last year and industry can now use the facility for product development or flow testing of hydrogen equipment and valves. As the switch over from natural gas to hydrogen is likely to take place in a staged approach involving hydrogen/natural gas blends or a direct switch to pure hydrogen, manufacturers can carry out a complete assessment of equipment performance across a range of hydrogen to natural gas compositions from 0 to 100% hydrogen. High purity bottled gases are fed through a manifold to supply gas and gas mixtures to a test section at precisely controlled pressures and flow rates. The flow rates are measured using precision reference instruments calibrated to national standards. The facility operates over a range of flows, pressures and temperatures to replicate those experienced in service. The hydrogen domestic gas metering test facility was designed for low-pressure hydrogen applications, but TÜV SÜD recognised that there was also a need for a larger flow, higher pressure test facility. The company has undertaken a project under the BEIS-funded Flow Programme to establish a Gas Primary Flow Standard Test Facility. Initially, it will operate with hydrogen, nitrogen 24bmta.co.uk 25 and methane with a view to extending to carbon dioxide. The facility is currently in the detailed design stage and the plan is to have an operating envelope of up to 100 Sm3/hr of gas with the ability to operate up to 120 bar. It’s estimated that the test facility will be ready for use by industry in 2023. TÜV SÜD National Engineering Laboratory will continue to keep industry updated on the progress of the test facility build in the coming months. For more information on how TÜV SÜD National Engineering Laboratory can help with your valve flow testing needs, contact Marc MacDonald on marc.macdonald@tuvsud.com.bmta.co.uk UPDATE ON THE SCIENCE T LEVEL Joe Neame, Subject Specialist Science, NCFE T Levels1 are new two-year, Level 3 courses that follow on from GCSEs. The much- needed study programmes prepare young people for skilled employment, higher level apprenticeships or higher education. Based on the same standards as apprenticeships, a T Level is equivalent to three A Levels and gives students a mixture of classroom learning and relevant, up to date, technical experience. With a wide range of skill shortages across the UK, there is a massive demand for these courses that focus on developing vocational skills to support vital sectors and industries. In September 2021, the Health and Science T Levels route was launched, containing a science- specific pathway. Within this pathway, students can choose to study the metrology occupational specialism which gives them the knowledge, technical skills and threshold competence to work in the metrology sector. Students cover the core knowledge, skills and concepts applicable to the health and science sector in general, before undertaking their occupational specialism in the second year of the programme. The metrology T Level, as with all other T Levels, was designed with and by employers and industry experts to ensure students are prepared to embark on their metrology careers. 26The British Measurement Testing Association (BMTA) played an invaluable role in the validation of the Science T Level, and metrology occupational specialism in particular, ensuring it is fit for purpose for students, providers and the sector. T Level students also complete an industry work placement where they get the chance to apply their newfound technical knowledge and skills. They experience employment within the metrology sector first-hand, allowing them to make an informed choice as to whether they wish to pursue a career, a higher apprenticeship or higher education. Industry placements set T Levels apart, providing students with more comprehensive and in-depth experience that they won’t find with other L3 qualifications. It also provides an opportunity to expose students to the vital role metrology plays in the 21st century and in the future of the health sector. Industry placements benefit employers as much as students, giving them fresh insight into their business, as well as supporting and developing current employees. Students undertake learning within the T Level that highlights to them their prior metrology knowledge gained from everyday life and education. They then build on these foundations in a relevant, technically-focused pathway. Through the occupational specialism performance outcomes, students plan and perform appropriate scientific measurements for any measurand (complying with regulatory requirements and ensuring accuracy), collect, analyse and interpret data from measurement tasks and resolve identified issues with measurement tools. To further support with progression from the T Level, NCFE is currently developing a Higher Technical Qualification (HTQ) as a Level 5 Diploma for a Senior Metrology Technician. This diploma will sit between T Levels and degrees and is comprised of the higher level knowledge and skills that employers require, allowing students to progress into highly skilled, highly paid roles within the industry. Again, BMTA will play a vital role in validating this technical qualification, ensuring it is fit for purpose and is of the highest quality for students and learners. If anyone would like to know more about this qualification or would like to support its development, then please reach out to NCFE via Craigwade@ncfe.org.uk. For the majority of learners, the T Level will be the first time they have been exposed to metrology and so is a fantastic opportunity to engage and encourage the metrologists of tomorrow. There is currently a STEM skills shortage and through academic and technical progression routes, we can help ensure that we are prepared for the metrological challenges of the future. 1 T Level is a registered trademark of the Institute for Apprenticeships and Technical Education. The T Level Technical Qualification is a qualification approved and managed by the Institute for Apprenticeships and Technical Education. bmta.co.uk 27bmta.co.uk MY JUBILEE – 42 YEARS WORKING IN THE TEST AND MEASUREMENT INDUSTRY Andy Morris, NPL North of England Operations Manager National Physical Laboratory (NPL) HRH Queen Elizabeth II has ruled for longer than any other Monarch in British history, becoming a much loved and respected figure across the globe. Becoming Queen in 1952, she celebrated 70 years on the throne in June 2022 with her Platinum Jubilee. Known for her sense of duty and her devotion to a life of service, the Queen has been an important figurehead for the UK and the Commonwealth since her accession to the throne. In that time we have lived through enormous social, technological and economic change. In this article, Andy Morris, NPL North of England Operations Manager, National Physical Laboratory (NPL), looks at what’s changed in the Test and Measurement industry. He may only have started his career in 1979 - a whole 27 years after Her Majesty began her tenure - but there has still been exponential change and growth in our industry since then. When I first started in test and measurement in 1979 it was as an apprentice with BT (formerly Post Office). All of the paperwork was handwritten onto printed forms and these were later processed by a separate department – the only department that had computers! Equipment ranged from physical standards to valve-driven electronic devices, transistorised devices and some early IC-controlled instrumentation. Everything we did was manual, relying on years of experience and a degree of innovation. 28In the early ‘80s, this changed quickly with the introduction of PCs. Very soon we were writing programmes to analyse test results, plot graphs to be included in reports and even run Automatic Test Equipment through calibration routines – I suspect we were early adopters of this technology. Apple 2, HP85 and HP86 computers became common, followed by BBC B computers and ultimately IBM ad IBM-style PCs. At the same time, a lot of the calibration labs were still using manual techniques, especially in the standards areas with a great deal of reliance on physical standards which were monitored manually – for example, the plotting of drift charts for standard resistors. Processes, in general, were still not automated but the signs were there that the future would be very different. Since then, more and more computing power was introduced with job records, test reports and correspondence internally and externally being electronic. At the same time, there was a significant increase in more complex multifunctional test equipment including such items as multifunction calibrators and the ability to extract useful data from them. Today, we are heavily digitalised with efforts to increase the level of integration of equipment, calibration records, test reports, uncertainty calculations, environmental modelling, equipment performance monitoring (with associated modelling) and correspondence, including automated reporting. There is no doubt that calibration providers and their customers now share a lot more data, much of it in an electronic format that is suitable for use by both parties. This allows the use of this data in many ways, including improved management of test equipment, a better understanding of the equipment’s contribution to measurement uncertainty when it is used, and better predictions of future equipment performance. There is still plenty of scope for innovation though, especially in the way that data is captured, transmitted, and presented to the user, and ultimately how it is used to make decisions. With more and more emphasis on modelling, I expect increased use of tools such as machine learning/AI around measurement which will help in understanding (and perhaps improving) the performance of our equipment and the systems they are being used to monitor, validate and control. bmta.co.uk 29Next >