Caught our Attention
Recent items that have advanced our thinking re GenAI.
Orientation
In November 2022, ChatGPT burst onto the scene and rapidly recalibrated expectations about the facility of computers with natural language inputs and outputs. AI, and, more specifically, GenAI has dominated the collective conversation ever since. Yet, for all this talk, if anyone has a functional definition of AI or GenAI, we would be most obliged if you could share.
While the concepts are slippery, the impact thereof are comprehensible. The digital domain appears to have reached another inflection point—even if the tech itself does not quite live up to the extreme hype, the expectation shift seems material and enduring. The following is meant to help orient those who are still trying to establish some baseline understanding.
The popular conception of AI tends to be ‘things a machine can’t do yet.’ It is a job-killing robot until it is just a dishwasher or ATM. It was The Brain’s Last Stand until we acclimated to the reality that computers could dominate humans at chess. Defining artificial intelligence as “the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings” may be accurate, but it is certainly not precise because the tasks commonly associated with human intelligence shift in concert with the ability of computers to perform those tasks.* AI is, by this definition, forever just over the horizon.
What is currently being broadly termed “Generative AI” suffers from a similar elusiveness. GenAI, or “GAI”, is sometimes used to describe tech that generates new content (whether it be words or images). That usage, however, can exclude other important applications like classifying data. Alternatively, GenAI is sometimes used as an umbrella term for the models enabling the tech, or for the tech incorporating such models—large language models (“LLMs”), large multimodal models (“LMMs”), diffusion models (often used for video and images)—rather than the type of output. Additional descriptors like “foundation model” and “general purpose AI” are also in relatively wide use but do not, so far, really solve the definitional challenges.
While this tech is decades in development, our focus herein is on recent advances in machine capacity to transmute various forms of input (text, voice, pictures) into various forms of output (text, voice, pictures, video, charts, functional code, structured data) to a degree that has most casual observers responding, “Wait, computers can do what now?” This standard remains fluid. But is shifting faster than ever before.
While there are many brilliant minds sharing on this topic, we recommend following Ethan Mollick as the most digestible popularizer to the laity (like us).
AI in the World
Regulation and Liability
Lawyers worried about AI taking their jobs are worrying about the wrong thing. GenAI will only increase velocity and complexity.* Jobs will change. But, overall, demand will grow.
How It Works
Few of us possess actual understanding of how modern technology works at a schematic/mechanical level. Rather, while we never quite grasp how it works, we become comfortable once we know that it works.
Calculators. Computers. Elevators. Cars. Airplanes. Dishwashers. ATMs. Microwaves. TV. Internet. Mobile. Cloud. Email. Cameras. For most of us, our knowledge of how the technology underpinning our daily lives functions is de minimus.
Introductions and Some Technical Stuff
New and surprising can be unsettling, especially if a breakthrough seems to presage some form of major disruption. It is natural to be curious about what is going on under the hood. While this is not the place to turn if you want to deeply understand parameters or loss functions, we have ourselves felt this compulsion to develop a sounder understanding of GenAI. What follows are resources we’ve consulted to raise our baseline comprehension in order to better identify signal in the noise.
For us, a necessary but not sufficient, point of departure is that the underlying models are probabilistic rather than deterministic. That is, when working with only the raw models—as was the case with the original ChatGPT—outputs are based on probabilities, not any source of truth.
A large language model, for example, is a model of language based on a large amount of data. This enables text prediction on steroids. As Stephen Wolfram opens his ur-explainer What is ChatGPT Doing…and Why Does It Work?, “It’s just adding one word at a time.”
Or to adapt a passage from Professor Murray Shanahan*:
Suppose we give an LLM the prompt “Who was the first person to walk on the Moon?”, and suppose it responds with “Neil Armstrong”.
What are we really asking here?
From our perspective, we are asking the literal question (“who was the first person to walk on the moon?”) in an attempt to elicit an accurate, factual answer (“Neil Armstrong”).
From the perspective of the model (which, again, is just a model), our prompt translates to:
Given the statistical distribution of words in the vast public corpus of (English) text, what words are most likely to follow the sequence “The first person to walk on the Moon was… ”?
That the model responds with “Neil Armstrong” is due to the regular co-occurrence of the two elements—i.e., the words “Neil Armstrong” are statistically mostly likely to follow the word sequence, “The first person to walk on the Moon was”).
The probabilistic nature of these models is what makes them so adaptable. We now have the architecture (transformers), compute power (Moore’s Law), and data volumes (the internet) to enable the machines to generate giant models that can be relatively quickly adapted to all sorts of different use cases. Before this moment, we were largely building massive IF/THEN deterministic workflows to achieve similar results—often not as good; almost always far more labor intensive.
On the other hand, the probabilistic nature of these models is also what makes them hallucinate (next section).
Be warned: any analysis that concludes with the observation that the models are probabilistic, prone to hallucination, and, therefore, not to be relied upon is materially incomplete. First, it assumes that perfect accuracy is the sole standard. In reality, perfection is rarely the standard and those who presume persistent human infallibility are deluding themselves. Second, it assumes end users will only be interacting with the models in their raw form, completely ignoring the nascent but rapidly maturing application layer.*
For more technical follows, we recommend Simon Willison and Andrew Ng.
Hope, Hype, Doom, & Gloom
Delivery of Legal Services
Few of us possess actual understanding of how modern technology works at a schematic/mechanical level. Rather, while we never quite grasp how it works, we become comfortable once we know that it works.
Calculators. Computers. Elevators. Cars. Airplanes. Dishwashers. ATMs. Microwaves. TV. Internet. Mobile. Cloud. Email. Cameras. For most of us, our knowledge of how the technology underpinning our daily lives functions is de minimus.
New and surprising can be unsettling, especially if a breakthrough seems to presage some form of major disruption. It is natural to be curious about what is going on under the hood. While this is not the place to turn if you want to deeply understand parameters or loss functions, we have ourselves felt this compulsion to develop a sounder understanding of GenAI. What follows are resources we’ve consulted to raise our baseline comprehension in order to better identify signal in the noise.
For us, a necessary but not sufficient, point of departure is that the underlying models are probabilistic rather than deterministic. That is, when working with only the raw models—as was the case with the original ChatGPT—outputs are based on probabilities, not any source of truth.
A large language model, for example, is a model of language based on a large amount of data. This enables text prediction on steroids. As Stephen Wolfram opens his ur-explainer What is ChatGPT Doing…and Why Does It Work?, “It’s just adding one word at a time.”
Or to adapt a passage from Professor Murray Shanahan*:
Suppose we give an LLM the prompt “Who was the first person to walk on the Moon?”, and suppose it responds with “Neil Armstrong”.
What are we really asking here?
From our perspective, we are asking the literal question (“who was the first person to walk on the moon?”) in an attempt to elicit an accurate, factual answer (“Neil Armstrong”).
From the perspective of the model (which, again, is just a model), our prompt translates to:
Given the statistical distribution of words in the vast public corpus of (English) text, what words are most likely to follow the sequence “The first person to walk on the Moon was… ”?
That the model responds with “Neil Armstrong” is due to the regular co-occurrence of the two elements—i.e., the words “Neil Armstrong” are statistically mostly likely to follow the word sequence, “The first person to walk on the Moon was”).
The probabilistic nature of these models is what makes them so adaptable. We now have the architecture (transformers), compute power (Moore’s Law), and data volumes (the internet) to enable the machines to generate giant models that can be relatively quickly adapted to all sorts of different use cases. Before this moment, we were largely building massive IF/THEN deterministic workflows to achieve similar results—often not as good; almost always far more labor intensive.
On the other hand, the probabilistic nature of these models is also what makes them hallucinate (next section).
Be warned: any analysis that concludes with the observation that the models are probabilistic, prone to hallucination, and, therefore, not to be relied upon is materially incomplete. First, it assumes that perfect accuracy is the sole standard. In reality, perfection is rarely the standard and those who presume persistent human infallibility are deluding themselves. Second, it assumes end users will only be interacting with the models in their raw form, completely ignoring the nascent but rapidly maturing application layer.*
For more technical follows, we recommend Simon Willison and Andrew Ng.
Performance
Instead of forever moving the goalposts by defining AI as “whatever the machine can’t do yet,” it merits tracking what GenAI and cannot do today, as well as what it might be capable of tomorrow.
The best research to date has identified GenAI’s current “jagged frontier.” That is, there are some tasks where GenAI materially augments human performance and others where it hinders human performance. The challenge is that the frontier is not only jagged but invisible. Our intuitions fail us at what the machine is, and is not, good for.
Importantly, this is the worst GenAI will ever be. Silicon will continue to surpass carbon at all manner of cognitive tasks. Which is not to say humans are obsolete. While “computer” used to be a human occupation that will never return, and Excel displaced the vast majority of bookkeeping jobs, the total number of human jobs working with computers and numbers have only increased. We delegated brute-force calculation (thunking) so we can focus on higher-level analysis and decision making (thinking).* The machines have now reached a facility with language and images that they previously only achieved with math. The implications are profound.
Enterprise Embraces LLMs
The speed at which developments in GenAI have shifted expectations and priorities within large non-tech businesses is astounding. But discussions always frontruns decision and dollars. Real resource allocation is the trend to track.
Big Tech Pivot
Big Tech Pivots to GenAI - Because it intersects with their own commercial interest, Big Tech has been the first mover on investing in and incorporating GenAI.
GenAI in Legal
Lawyers must contend with the GenAI revolution on several fronts.
First and foremost, lawyers must proactively and prudently advise their business clients on the risks and benefits of incorporating GenAI into the clients’ products, services, and back-office operations. The legitimate legal issues raised by AI-regulated regulation, privacy violations, intellectual property protections, etc. are material enough to warrant a separate section.
Second, and addressed herein, lawyers must assess and action their own potential use of GenAI to deliver legal services.
We recommend starting with Sense and Sensibility from our advisor Jae Um of SixParsecs and our friend Ed Sohn of Facto. Special acknowledgment to the superb job Legaltech News has done tracking GenAI in legal. We also commend Pinhawk and Brainyacts for legal-specific newsletters with great GenAI content.
Lawyers have long objected to “non-lawyers” being entrusted with various forms of legal work because the non-lawyers have not been through the education necessary to pass the bar and ethics exams.* The theoretical performance delta is, apparently, so large that lawyers consider it better for a staggering proportion of legal needs to go unmet than for those needs to be partially met by non-lawyer means.* (So there is no doubt: we think this protectionism is gross)
But a funny thing happened. Large language models started passing the bar* and ethics* exams with flying colors. Yet, now, apparently, the AI cannot be trusted with various forms of legal work because “it can't provide the personal touch that a human can provide...AI can't have that personal connection, or emotional support clients expect from their lawyers…AI cannot negotiate on behalf of clients as it lacks the interpersonal skills and judgment required to assess and understand the unique circumstances of each case.”*
Leaving aside that GenAI seems pretty damn proficient at empathy* and negotiation*, this kind of goalpost moving offers false comfort. Even if true. There will always be things the machines cannot do. A car cannot nuzzle its rider like a horse. An email lacks the personal touch of a letter, handwritten with a quill on parchment. So what?
There is no coherent abstraction that fits neatly under the term “lawyering.” Lawyering is a loose assortment of many different jobs to be done.* Various forms of automation, including various applications GenAI, may well improve, and even displace, lawyers at performing tasks that are, at present, commonly associated with lawyering.
At LexFusion, we believe the automation of tasks (rather than jobs) will be absolutely essential because GenAI will increase the velocity and complexity of commerce. Lawyers will have more work than ever. Innovation and scale are the only option to keep pace with client needs.
Jevon’s Paradox is that, over the long term, an increase in efficiency in resource use generates an increase in resource consumption rather than a decrease. That is, more fuel-efficient vehicles results in more overall fuel consumption because people drive more. Similarly, more efficient delivery of legal services should result in higher demand for legal services.
Business clients are not the only ones who will be deploying GenAI in a manner that increases demand for legal services. Consumers, regulators, and plaintiffs’ attorneys will also turn to this tech to sustain demands, lawsuits, and investigations at volumes that currently seem inconceivable.
Buckle up. It will be a bumpy ride. But there is little reason to believe that developments in AI will reduce the total demand for lawyers.
SALI & Legal Language
SALI (Standards Advancement for the Legal Industry) is a mission-critical initiative. SALI has LexFusion’s unqualified support.
Our industry has held itself back with a Tower-of-Babel-like approach to language. Among other consequences, this has resulted in bad data and even worse information—information being data organized so as to be useful for decision making. The path out of this data morass is flexible standards that can mapped onto any custom taxonomy, and vice versa. That’s SALI.
SALI is a taxonomy that can sit on top of whatever else a firm/company/vendor is doing, introducing heretofore unavailable opportunities for analysis and interoperability. SALI should not be confused with efforts to standardized legal language or forms (e.g., Bonterms, OneNDA). Those initiatives also have merit. But, unlike SALI, they strike a sensitive cord with many lawyers who style themselves linguistic artisans—for reasons worth exploring.
If it moves, tax it. If it keeps moving, regulate it. GenAI keeps moving and the regulations are a comin.
Lawsuits & Investigations
Even without new regulations, GenAI has triggered many existing legal frameworks. The lawsuits really kicked into gear with a $9 billion copyright lawsuit against Microsoft and OpenAI that was filed before ChatGPT even dropped.* And that was merely the beginning.
Data Security
Privacy is central to the GenAI conversation. But, in many respects, it is really just a subset of broader concerns over data security. GenAI exacerbates existing data security threats (e.g., phishing) and introduces new ones (e.g., model inference attacks).
It is not merely copyrights and patents that are threatened by the rapid rise of GenAI. Personal privacy was a hot issue long before the GenAI companies became so thirsty for data. GenAI has turned the temperature up several notches.
One specialty aspect of data security with which lawyers are rightly concerned is GenAI’s effect on privilege.
Intellectual Property
Intellectual Property
Does GenAI violate copyrights? Can a work produced using GenAI be copyrighted? Can AI invent? At present, there are more questions than answers in the quagmire that is the legion of IP issues surrounding GenAI.
Labor & Employment
Protecting workers is likely to be an active area for GenAI related regulations, investigations, and lawsuits.
Many are worried that GenAI will take jobs. It probably will.
The original Luddites smashed looms during the early days of the Industrial Revolution because they believed the machines threatened their livelihoods. The Luddite’s thesis was borne out. Industrialization did negatively impact their livelihoods. But the automation paradox held—the technology created more new roles than it destroyed.* The net effect was more jobs for more people.
This net increase in jobs occurs at the societal level. It offers little comfort to the displaced who struggle to find a place for themselves while the rich only get richer off the disruption.
So far, the impact of GenAI on jobs is not even a rounding error. But we are early days.
Startups & Capital
The gold rush is on. Yet intermittent investment manias are par for the course. How long will this one last? Magnitude, duration, and impact remain key points of interest.
Consultants ♡ GenAI
Consultants ♡ GenAI
Since 2022, LexFusion’s consistent message to law departments, and, by extension, law firms, re GenAI is:
This is happening
It will be done by you
Or it will be done to you
Our stance is not based solely on the efficacy of GenAI. Our perspective is driven by the paradigmatic shift in expectations GenAI has engendered, and our knowledge that the consultancies are capitalizing on this shift to great effect. AI-centric changes are coming, for good or ill. The changes will be better if the lawyers are aligned with the business and its consultants—i.e., we desperately want the lawyers in the room and to be meaningful participants in the conversation. But, whether or not the lawyers are along for the ride, there is no holding back the strange AI tide.*
GenAI in Finance
Finance is often the canary in the coalmine as to whether a new form of tech will take. The early returns suggest that finance is serious about GenAI.
GenAI in Science & Medicine
Understandably, there is considerable focus on how GenAI may impact business. Who is making the money and how they are making it. But much more profound effects are likely to be felt in our daily lives, including where GenAI enables us to uncover new frontiers in science and medicine.
GenAI in Education
GenAI could substantially alter what we teach the next generation. It could also substantially transform how we teach.
The Model Makers
GenAI is premised on new kinds of models. Large Language Models (“LLMs”) have been the most prominent. But large multimodal models (“LMMs”) and diffusion models are also responsible for much of the Wow! emanating from recent advances in AI.
Early on, we penned articles like The Focus on ChatGPT Is Missing the Forest for the Tree and PSA: ChatGPT is the trailer not the movie because too many were treating ChatGPT as a stand-in for not only every model but every application of every model even though ChatGPT was only one application of one model. Thousands of models exist already. Many more are on the way. And there will be orders of magnitude more commercial and applications built on top of these models.
Open Source
One critical debate is whether the future of GenAI models is primarily open source. This ties into other debates like whether there are diminishing returns to continued increases in model size or the efficacy of large commercial models versus smaller, more targeted models.
No opinions to offer beyond noting these developments merit tracking.
Responsible & Explainable
Responsible AI is an approach to developing artificial intelligence ethically. The goal is AI that is safe, trustworthy, and legal compliant. Responsible AI increase transparencies and mitigates harms like AI bias.*
[Confession: we enjoy irony]
[Disclosure: the first draft of the below was written by ChatGPT]
Responsibility in AI encompasses ethical considerations, accountability, and ensuring that AI systems align with ethical and legal standards. Lawyers play a pivotal role in ensuring that AI applications comply with existing laws and regulations. This involves assessing the ethical implications of AI decisions, understanding potential biases, and addressing concerns related to fairness and justice.
To achieve responsible AI, lawyers should advocate for comprehensive risk assessments before deploying AI systems. This includes evaluating the potential impact on marginalized communities, safeguarding privacy, and considering the legal implications of AI-related decisions. Moreover, establishing clear guidelines for AI developers and users is essential to prevent misuse and ensure adherence to legal and ethical standards.
Explainability is a cornerstone of legal practice, and the same principle applies to AI. Lawyers need to understand and be able to explain how AI systems arrive at their conclusions, particularly when these conclusions have legal consequences. Explainability is critical not only for building trust in AI but also for meeting legal standards that require justifiable decision-making processes.
In legal proceedings, the "black box" nature of some AI systems can pose challenges. Therefore, lawyers must advocate for AI models that are interpretable, providing a clear understanding of the factors influencing their decisions. This transparency is vital not only for legal compliance but also for ensuring due process and fairness in the application of AI within the legal system.
Transparency in AI refers to the openness and accessibility of information regarding how AI systems operate. Lawyers must insist on transparency to uphold legal principles, protect individual rights, and maintain public trust. This involves transparency not only in the design and development phases but also in the ongoing use and evolution of AI systems.
Legal professionals can advocate for regulations that mandate disclosure of the use of AI in legal proceedings. This ensures that parties involved are aware of the role of AI in decision-making, allowing them to challenge or question its application. Transparency is fundamental in preventing unintended consequences, maintaining accountability, and fostering public confidence in the legal system's use of AI.
Trust is paramount in the legal profession, and AI must be trustworthy to be embraced within legal practices. Lawyers should advocate for the development and adoption of trustworthy AI frameworks, emphasizing reliability, security, and ethical considerations. Trustworthy AI inspires confidence in both legal practitioners and the broader public.
Ensuring the trustworthiness of AI involves rigorous testing, validation, and ongoing monitoring of AI systems. Lawyers should advocate for standards and certifications that attest to the reliability and security of AI applications. Additionally, promoting ethical AI use within the legal community can help build a culture of trust and responsibility surrounding AI technologies.
Hallucinations
The models underpinning GenAI are probabilistic, not deterministic. Being models, they produce answers without regard for whether the answers are true (just statistically likely). As explained in the release notes for ChatGPT on November 30, 2022, “ChatGPT sometimes writes plausible-sounding but incorrect or nonsensical answers…there’s currently no source of truth.”
Recognition of the hallucination issue is critical enough to warrant a separate section. Hallucinations are a genuine challenge and key consideration. But it is also critical to understand that humans often hallucinate more than the models. For example, when asked to summarize a document, humans, on average, will invent more extrinsic information (i.e., not contained in the document) than the machines.*
To repeat a warning: any analysis that concludes with the observation that the models are probabilistic, prone to hallucination, and, therefore, not to be relied upon is materially incomplete. First, it assumes that perfect accuracy is the sole standard. In reality, perfection is rarely the standard and those who presume persistent human infallibility are deluding themselves. Second, it assumes end users will only be interacting with the models in their raw form, completely ignoring the nascent but rapidly maturing application layer (next section).
Applications, Stacks, Tools & Agents
The models underpinning GenAI—like GPT-4 (OpenAI), Gemini (Google), and Llama 2 (Meta)—are often referred to as “foundation models.” The term "foundation model" was defined in 2021 as "any model that is trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks."*
Adaption often takes the form of using models to power applications. Applications integrate the models with complementary tech and data in order to make them fit to purpose. The original ChatGPT, for example, was an application of a model, GPT 3.5, integrated with a complementary chat interface.
Much of what is occurring in the world of GenAI involves the rapid maturing of a nascent application layer. Which is why over-indexing on the capabilities and limitations of the models in their raw form can lead the conversation astray.
It is useful to understand that all GenAI models are part of a stack.* The models are run atop computing hardware, often in the cloud. The models are often connected to data sources through an orchestration layer that underpins an application layer—i.e., the layer at which end users interact not only with the model but also with the the other component parts. Further, a human services layer can be laid on top of the application layer.
Part of building AI-enabled applications can include providing the AI with tools that extends its capabilities. For example, Advanced Data Analytics is a tool developed by OpenAI to complement GPT-4 and add data analytics capabilities that the model itself does not possess. A more digestible example is the likes of OpenAI, Microsoft, and Google giving their models access to web search (another tool).
Providing GenAI applications access to tools is one step closer to the development of AI agents. The goal of AI agents is to be able to tell the AI what to do (“analyze this data”) rather than how to do it (“use Advanced Data Analytics to analyze this data”). That is, the human sets the objective, and the AI determines the methods as part of satisfying the objective. Quite cool. And also very scary.*
Knowlege Management, Graphs, and Retrieval Augmented Generation
Knowledge Management, Graphs, and Retrieval Augmented Generation
There is an odd sentiment in some quarters that GenAI renders knowledge management obsolete. In certain respects, GenAI can make KM less labor intensive. But, overall, GenAI only makes good KM more acutely valuable—i.e., the drag on productivity from poor KM practices becomes more severe and more glaring.
Of particular importance in legal, many applications* currently being built rely on KM methods like knowledge graphs and retrieval augmented generation (or “RAG”) in order to be performant and fit for purpose.
“Literally everyone on earth will die.”*
There has been no shortage of hyperbolic statements regarding the rise of GenAI. But some statements have been more cataclysmic than others. A stark thread of AI doomerism has existed at the heart of what is sometimes termed “the pause debate.” Though, truth be told, there was no debate. Some voices have shouted dire warnings from the sidelines—about potential harms from social upheaval due to propaganda for which the body politic is wildly unprepared (highly probable) to total extinction of the human race (slightly more fantastical)—while the GenAI train has just kept rolling.
There is a parallel thread of gloomerism. Where doomerism is premised on the AI being too good, gloomerism suggests that there is no there there. The gloomists lament us being stuck in yet another hellish hype cycle with carnival barkers using vapid technobabble and utopian promises to line their own pockets while distracting us from the real work of work.
Every technological inflection point underdelivers in the near term and fails to fulfill the wild promises of its most zealous prophets. The question is not whether there is excess hype around GenAI—there most certainly is. The question is not whether there is a GenAI bubble—again, yes, certainly. The question is whether there is signal in all this noise. When you look beyond the nonsense, is there a genuine breakthrough that merits investment in dollars and attention? At LexFusion, we believe the answer is in the affirmative—even though we do not concur with every wild claim made about humanity’s AI-enabled future.
Things Will Get Weird
Some of what is to come is downright bizarre.
Educational Resources
Hotshot offers the best AI training for legal (also the best substantive training). We're biased but not wrong. Our Chief Strategy Officer, Casey Flaherty, is featured heavily in the Hotshot AI videos and helped assemble the all-star cast.
Checkout Hotshot. Some other training resources are below.
