We suggest that foundation models are general purpose solutions similar to general purpose programmable microprocessors, where fine-tuning and prompt-engineering are analogous to coding for microprocessors. Evaluating general purpose solutions is not like hypothesis testing. We want to know how well the machine will perform on an unknown program with unknown inputs for unknown users with unknown budgets and unknown utility functions. This paper is based on an invited talk by John Mashey, “Lessons from SPEC,” at an ACL-2021 workshop on benchmarking. Mashey started by describing Standard Performance Evaluation Corporation (SPEC), a benchmark that has had more impact than benchmarks in our field because SPEC addresses an import commercial question: which CPU should I buy? In addition, SPEC can be interpreted to show that CPUs are 50,000 faster than they were 40 years ago. It is remarkable that we can make such statements without specifying the program, users, task, dataset, etc. It would be desirable to make quantitative statements about improvements of general purpose foundation models over years/decades without specifying tasks, datasets, use cases, etc.

Retrieval-augmented generation (RAG) adds a simple but powerful feature to chatbots, the ability to upload files just-in-time. Chatbots are trained on large quantities of public data. The ability to upload files just-in-time makes it possible to reduce hallucinations by filling in gaps in the knowledge base that go beyond the public training data such as private data and recent events. For example, in a customer service scenario, with RAG, we can upload your private bill and then the bot can discuss questions about your bill as opposed to generic FAQ questions about bills in general. This tutorial will show how to upload files and generate responses to prompts; see https://github.com/kwchurch/RAG for multiple solutions based on tools from OpenAI, LangChain, HuggingFace transformers and VecML.

Large language models (LLMs) have achieved amazing successes. They have done well on standardized tests in medicine and the law. That said, the bar has been raised so high that it could take decades to make good on expectations. To buy time for this long-term research program, the field needs to identify some good short-term applications for smooth-talking machines that are more fluent than trustworthy.

Our last emerging trend article introduced Risks 1.0 (fairness and bias) and Risks 2.0 (addictive, dangerous, deadly, and insanely profitable). This article introduces Risks 3.0 (spyware and cyber weapons). Risks 3.0 are less profitable, but more destructive. We will summarize two recent books, Pegasus: How a Spy in Your Pocket Threatens the End of Privacy, Dignity, and Democracy and This is How They Tell Me the World Ends: The Cyberweapons Arms Race. The first book starts with a leak of 50,000 phone numbers, targeted by spyware named Pegasus. Pegasus uses a zero-click exploit to obtain root access to your phone, taking control of the microphone, camera, GPS, text messages, etc. The list of 50,000 numbers includes journalists, politicians, and academics, as well as their friends and family. Some of these people have been murdered. The second book describes the history of cyber weapons such as Stuxnet, which is described as crossing the Rubicon. In the short term, it sets back Iran’s nuclear program for less than the cost of conventional weapons, but it did not take long for Iran to build the fourth-biggest cyber army in the world. As spyware continues to proliferate, we envision a future dystopia where everyone spies on everyone. Nothing will be safe from hacking: not your identity, or your secrets, or your passwords, or your bank accounts. When the endpoints (phones) have been compromised, technologies such as end-to-end encryption and multi-factor authentication offer a false sense of security; encryption and authentication are as pointless as closing the proverbial barn door after the fact. To address Risks 3.0, journalists are using the tools of their trade to raise awareness in the court of public opinion. We should do what we can to support them. This paper is a small step in that direction.

Deep nets are becoming larger and larger in practice, with no respect for (non)-factors that ought to limit growth including the so-called curse of dimensionality (CoD). Donoho suggested that dimensionality can be a blessing as well as a curse. Current practice in industry is well ahead of theory, but there are some recent theoretical results from Weinan E’s group suggesting that errors may be independent of dimensions $d$. Current practice suggests an even stronger conjecture: deep nets are not merely immune to CoD, but actually, deep nets thrive on scale.

This paper describes gft (general fine-tuning), a little language for deep nets, introduced at an ACL-2022 tutorial. gft makes deep nets accessible to a broad audience including non-programmers. It is standard practice in many fields to use statistics packages such as R. One should not need to know how to program in order to fit a regression or classification model and to use the model to make predictions for novel inputs. With gft, fine-tuning and inference are similar to fit and predict in regression and classification. gft demystifies deep nets; no one would suggest that regression-like methods are “intelligent.”

Many papers are chasing state-of-the-art (SOTA) numbers, and more will do so in the future. SOTA-chasing comes with many costs. SOTA-chasing squeezes out more promising opportunities such as coopetition and interdisciplinary collaboration. In addition, there is a risk that too much SOTA-chasing could lead to claims of superhuman performance, unrealistic expectations, and the next AI winter. Two root causes for SOTA-chasing will be discussed: (1) lack of leadership and (2) iffy reviewing processes. SOTA-chasing may be similar to the replication crisis in the scientific literature. The replication crisis is yet another example, like evaluation, of over-confidence in accepted practices and the scientific method, even when such practices lead to absurd consequences.

The previous Emerging Trends article (Church et al., 2021. Natural Language Engineering27(5), 631–645.) introduced deep nets to poets. Poets is an imperfect metaphor, intended as a gesture toward inclusion. The future for deep nets will benefit by reaching out to a broad audience of potential users, including people with little or no programming skills, and little interest in training models. That paper focused on inference, the use of pre-trained models, as is, without fine-tuning. The goal of this paper is to make fine-tuning more accessible to a broader audience. Since fine-tuning is more challenging than inference, the examples in this paper will require modest programming skills, as well as access to a GPU. Fine-tuning starts with a general purpose base (foundation) model and uses a small training set of labeled data to produce a model for a specific downstream application. There are many examples of fine-tuning in natural language processing (question answering (SQuAD) and GLUE benchmark), as well as vision and speech.

Deep nets have done well with early adopters, but the future will soon depend on crossing the chasm. The goal of this paper is to make deep nets more accessible to a broader audience including people with little or no programming skills, and people with little interest in training new models. A github is provided with simple implementations of image classification, optical character recognition, sentiment analysis, named entity recognition, question answering (QA/SQuAD), machine translation, speech to text (SST), and speech recognition (STT). The emphasis is on instant gratification. Non-programmers should be able to install these programs and use them in 15 minutes or less (per program). Programs are short (10–100 lines each) and readable by users with modest programming skills. Much of the complexity is hidden behind abstractions such as pipelines and auto classes, and pretrained models and datasets provided by hubs: PaddleHub, PaddleNLP, HuggingFaceHub, and Fairseq. Hubs have different priorities than research. Research is training models from corpora and fine-tuning them for tasks. Users are already overwhelmed with an embarrassment of riches (13k models and 1k datasets). Do they want more? We believe the broader market is more interested in inference (how to run pretrained models on novel inputs) and less interested in training (how to create even more models).

There are well-meaning efforts to address ethics that will likely make the world a better place, but care needs to be taken to avoid repeating mistakes of the past. In particular, ACL has recently introduced a new process where there are special reviews of some papers for ethics. We would be more comfortable with the new ethics process if there were more checks and balances, due process and transparency. Otherwise, there is a risk that the process could intimidate authors in ways that are not that dissimilar from the ways that academics were intimidated during the Cold War on both sides of the Iron Curtain.

Benchmarks can be a useful step toward the goals of the field (when the benchmark is on the critical path), as demonstrated by the GLUE benchmark, and deep nets such as BERT and ERNIE. The case for other benchmarks such as MUSE and WN18RR is less well established. Hopefully, these benchmarks are on a critical path toward progress on bilingual lexicon induction (BLI) and knowledge graph completion (KGC). Many KGC algorithms have been proposed such as Trans[DEHRM], but it remains to be seen how this work improves WordNet coverage. Given how much work is based on these benchmarks, the literature should have more to say than it does about the connection between benchmarks and goals. Is optimizing P@10 on WN18RR likely to produce more complete knowledge graphs? Is MUSE likely to improve Machine Translation?

Subwords have become very popular, but the BERTa and ERNIEb tokenizers often produce surprising results. Byte pair encoding (BPE) trains a dictionary with a simple information theoretic criterion that sidesteps the need for special treatment of unknown words. BPE is more about training (populating a dictionary of word pieces) than inference (parsing an unknown word into word pieces). The parse at inference time can be ambiguous. Which parse should we use? For example, “electroneutral” can be parsed as electron-eu-tral or electro-neutral, and “bidirectional” can be parsed as bid-ire-ction-al and bi-directional. BERT and ERNIE tend to favor the parse with more word pieces. We propose minimizing the number of word pieces. To justify our proposal, a number of criteria will be considered: sound, meaning, etc. The prefix, bi-, has the desired vowel (unlike bid) and the desired meaning (bi is Latin for two, unlike bid, which is Germanic for offer).

The ACL-2019 Business meeting ended with a discussion of reviewing. Conferences are experiencing a success catastrophe. They are becoming bigger and bigger, which is not only a sign of success but also a challenge (for reviewing and more). Various proposals for reducing submissions were discussed at the Business meeting. IMHO, the problem is not so much too many submissions, but rather, random reviewing. We cannot afford to do reviewing as badly as we do (because that leads to even more submissions). Negative feedback loops are effective. The reviewing process will improve over time if reviewers teach authors how to write better submissions, and authors teach reviewers how to write more constructive reviews. If you have received a not-ok (unhelpful/offensive) review, please help program committees improve by sharing your not-ok reviews on social media.

Evaluation was not a thing when the first author was a graduate student in the late 1970s. There was an Artificial Intelligence (AI) boom then, but that boom was quickly followed by a bust and a long AI Winter. Charles Wayne restarted funding in the mid-1980s by emphasizing evaluation. No other sort of program could have been funded at the time, at least in America. His program was so successful that these days, shared tasks and leaderboards have become common place in speech and language (and Vision and Machine Learning). It is hard to remember that evaluation was a tough sell 25 years ago. That said, we may be a bit too satisfied with current state of the art. This paper will survey considerations from other fields such as reliability and validity from psychology and generalization from systems. There has been a trend for publications to report better and better numbers, but what do these numbers mean? Sometimes the numbers are too good to be true, and sometimes the truth is better than the numbers. It is one thing for an evaluation to fail to find a difference between man and machine, and quite another thing to pass the Turing Test. As Feynman said, “the first principle is that you must not fool yourself–and you are the easiest person to fool.”

General Adversarial Networks are hot. Given Murphy’s Law, it is prudent to be paranoid. Best not to design for the average case. There is a long tradition of designing for the hundred-year flood (and five 9s reliability). What is good enough? Historically, the market hasn’t been willing to pay for five 9s. Hard to justify upfront costs for future benefits that will only payoff under unlikely scenarios, and might not work when needed. If the market isn’t willing to pay for five 9s, can we afford to design for the worst case?

Lots of companies are offering lots of APIs. Reviews are not always as constructive as they could be. Some reviews encourage unproductive work on checkbox features that no one wants. It makes no sense to do the wrong thing badly. Constructive reviews should help focus priorities on what matters. Users care more about a great box opening experience than small improvements in word error rate and BLEU, popular metrics for speech and translation. In 15 minutes or less, can we teach potential users something new (and fun) that does something useful, such as how to translate PowerPoint between English and Chinese, preserving many of the features that are important to PowerPoint such as graphics and animations? See Appendix for the solution.

It is amazing that I have written for more than a year on Emerging Trends without mentioning China’s investments in Artificial Intelligence (AI). Now that I have moved to Baidu, I would like to take this opportunity to share some of my personal observations with what’s happening in China over the past 25 years. The top universities in China have always been very good, but they are better today than they were 25 years ago, and they are on a trajectory to become the biggest and the best in the world. China is investing big time in what we do, both in the private sector and the public sector. Kai-Fu Lee is bullish on his investments in AI and China. There is a bold government plan for AI with specific milestones for parity with the West in 2020, major breakthroughs by 2025 and the envy of the world by 2030.

Charles Wayne restarted funding in speech and language in the mid-1980s after a funding winter brought on by Pierce’s glamour-and-deceit criticisms in the ALPAC report and ‘Whither Speech Recognition’. Wayne introduced a new glamour-and-deceit-proof idea, an emphasis on evaluation. No other sort of program could have been funded at the time, at least in America. One could argue that Wayne has been so successful that the program no longer needs him to continue on. These days, shared tasks and leaderboards have become common place in speech and language (and vision and machine learning) research. That said, I am concerned that the community may not appreciate what it has got until it’s gone. Wayne has been doing much more than merely running competitions, but he did what he did in such a subtle Columbo-like way. Going forward, government funding is being eclipsed by consumer markets. Those of us with research to sell need to find more and more ways to be relevant to potential sponsors given this new world order.

There has been a trend for publications to report better and better numbers, but less and less insight. The literature is turning into a giant leaderboard, where publication depends on numbers and little else (such as insight and explanation). It is considered a feature that machine learning has become so powerful (and so opaque) that it is no longer necessary (or even relevant) to talk about how it works. Insight is not only not required any more, but perhaps, insight is no longer even considered desirable.

Transparency is good and opacity is bad. A recent best seller, Weapons of Math Destruction, is concerned that big data (and WMDs) increase inequality and threaten democracy largely because of opacity. Algorithms are being used to make lots of important decisions like who gets a loan and who goes to jail. If we tell the machine to maximize an objective function like making money, it will do exactly that, for better and for worse. Who is responsible for the consequences? Does it make it ok for machines to do bad things if no one knows what’s happening and why, including those of us who created the machines?

My last column ended with some comments about Kuhn and word2vec. Word2vec has racked up plenty of citations because it satisifies both of Kuhn’s conditions for emerging trends: (1) a few initial (promising, if not convincing) successes that motivate early adopters (students) to do more, as well as (2) leaving plenty of room for early adopters to contribute and benefit by doing so. The fact that Google has so much to say on ‘How does word2vec work’ makes it clear that the definitive answer to that question has yet to be written. It also helps citation counts to distribute code and data to make it that much easier for the next generation to take advantage of the opportunities (and cite your work in the process).