Resource Updates 2023-02-24
Update on updates
This letter is a summary of recent updates to the Sentient Syllabus resources and topics we are covering here.1 It feels like most of everything received an update …
General
The perspectives of generative AI are today far less clear than they were even three weeks ago. The “search wars” have begun – but they are turning out not to be about search after all. Microsoft showcased Bing. Then Bing started misbehaving. Dramatically. Then this made headlines.2 Then the public loved the drama – a computer with a personality! Then Bing got lobotomized.3
It all feels quite surreal, yet this is just the most visible manifestation of a much deeper dislocation. There are seriously important questions lurking behind this tangible part: technical (or philosophical?) questions about whether there is emergent behaviour in these Large Language Models (of course there is); and whether users need to be protected from software they willingly engage with; and what should be regulated in the social sphere, and controlled, by whom and how. But any actual debates are currently drowned out. Thats not how the social web works anyway, and the companies that exist in this particular domain of our culture are becoming wary of discourse. Large numbers of additional players are entering the game4 with new packaging for existing technologies, new technologies, and new companies: and as a result a cohesive field of scholarship that was unusual in its culture of open inquiry and shared knowledge, a sense of common purpose, and shared objectives, is being twisted apart. What we have now is no longer one big event that we can understand and then adapt to, we are entering a phase of continuous disruption.
So here are insights from the last weeks that have made their way into the Sentient Syllabus resources.
Syllabus Resources
Language: The field is gravitating towards the term “generative AI” and I have adopted this throughout. I am also more conscious of the implications of the term tool. Would we even need to be worried about tool use? Where possible, I am now trying to use more specific language.
Academic integrity: I have rewritten this section based on my analysis in Generated Misconduct. A principled definition of Academic Integrity has helped make the section shorter. I have added an Academic Integrity statement as an appendix; it is based on the positive-value approach and collaborative principle expressed in the analysis. The statement should be suitable for a course syllabus, but could be adopted for use at department, faculty and school-levels.
Participation marks: Given that generative AI can produce very helpful and elaborate comments with a few keystrokes, fairly evaluating a student contribution for its participation value, e.g. on a discussion board, has become more difficult. I have added a few thoughts, based around a principle of “non-trivial contributions”, and the value of personal perspectives.
Course Activities
Preface: We are taking a stronger position that a blanket ban of generative AI is unrealistic and irresponsible.
Using AI: A common theme appears that ties together successful approaches for classroom use: Use the AI to think with you, not for you. This simple principle touches on basically everything else I have written on the topic. Start from this thought, make this your philosophical razor,5 and watch how the practical consequences reveal themselves.
Prompts: Generative AI has three levels of control: the most basic is the training data, later, fine-tuning will shape dialogue patterns and focus, and add alignment. But the most direct and most specific is the prompt itself. Prompts matter.
Specific domains: This is a new section to include some of the more important uses. Having a tireless, personal tutor for computer language, who is very capable to explain code as well, will be a great benefit. And using ChatGPTs ability to create concept sketches via Mermaid markup will be useful e.g. for argument maps. As for those – I included an example – expect that neither the argument nor the map will work right out of the box – but in a teaching context I actually consider that to be an advantage.
Web search: The issue that captures the most headlines and generates the most misunderstanding. This section is very brief since matters are completely in flux. Do realize that generative AI by its nature is prone to create “Schrödinger Facts”, but also realize that there are right now already two alternatives (besides, well, Bing…) that intersect results with actual Webpages.6 Expect much to change.
Preventing AI: I am generally skeptical of AI prevention strategies, because preventing use of a technology that is widely adopted outside academia is not a timely teaching objective. But we should never be dogmatic. There are now descriptions of strategies that involve data past the training cutoff date; facts it can’t know; that involve precise facts – like quoting from references, never paraphrasing; that require weighting of arguments, and establishing causal connections; and that require multimodal input, images, video, and such. The strategies I included at least have less of a burden of being detached from real-world requirements.
Learning Objectives
Updated thinking on learning objectives was covered in How much is too much?. Updated writing is still in progress.
Understanding AI Issues
Just keeping up with the developments is a challenge, even more so to filter those items that signal a significant development.
How does it work? I have added a very basic description of the “transformer architecture”, just enough to map some of the terminology; I have expanded the description of the concepts, for example to clarify the concept of token (a building block of written language that efficiently represents a semantic unit); and I have emphasized the three levels of training and programming that are available: the pre-training, the fine-tuning, and the prompt. Yes, the prompt is a way to program the generative AI to respond in the way you need it to. This is important to realize and it forms the basis of prompt engineering. We link to resources.
Consequences for the economy: Updates concern the current state of the “search wars” – but our more in-depth analysis is here. The fascinatingly bizarre instabilities of the AI-enhanced Bing search engine are briefly summarized. I have added small updates on the copyright question – the position that pre-trained transformer models contain an interpretation of the data, not a copy was explained in the analysis mentioned above; and a US/European AI data sharing agreement appears to indicate that the administrations are taking the same position: training is not copying. The expected economic disruptions of fee-writing and graphics services is gaining contours with an announcement of a major adoption of OpenAI technologies by Bain & Company. Bain is one of the “big three” management consultancies, and will coordinate an “ambitious vision” of Coca Cola7 to use OpenAI’s ChatGPT and DALL-E in their “brands, marketing, and consumer-experiences”.8 Significant portions of these funds would have gone to human creatives a short while ago. Generative AI is making its business case.
Speculative Perspectives: One of the biggest questions at this time is how accessible generative AI is, and who controls its alignment. There are two important developments to be aware of. David Rozado has engaged in a political experiment, to re-train the GPT-3 model for a right-leaning alignment through fine-tuning. The result is remarkably dystopian, except for one thing: the cost of retraining was only about 300 dollars, all in all – incredibly affordable. And before anyone misses the point again: this is not breaching the dam for a flood of wedge-issue carnival barkers, on the contrary, this is an essential step towards democratizing the technology! Anyone so inclined could have their bespoke AI, one that is aligned with their discursive needs, and I am not even thinking of climate justice and decolonization. I am thinking of tropical entomologists, Husserl scholars, algebraic geometers, sign language syntacticians, architects who work in generative design, the University of Toronto, the University of Toronto Engineering Student’s Society, “Bellatrix” – the Science fiction student club of Mount Holyoke College; … Why even discuss others’ authority, when you can exercise it yourself? This is a grain that will grow, while others not. This is a rallying call to all faculty and scholars and students to exercise their imagination of what such a tool could do in the field they are passionate about.
But it gets better: you are surely aware of the significant computational costs of running LLMs – after all, those 175 Billion parameters need to be exercised in computation of every single token. Now a new paper by Ying Sheng et al. (2023) was posted on GitHub, they show ways to use smart offloading strategies to bring the hardware requirements of generative AI down to the scale of a commodity GPU. This means they run OPT-175B (a GPT-3 equivalent trained by Meta) on a single Nvidia T4 GPU (which can be bought for 2,300 dollars) and they achieve 1 Token/s throughput (that's approximately 45 words per minute). That is not cheap, but certainly affordable even for individuals – or rentable as a cloud instance. Taken together, the implications of personalized LLMs are – amazing.
Just one thing: as we are thinking about what this all means for academia, it becomes clear that we should not worry about moving goalposts. We can’t even be sure we are in the right stadium.
Press and Media
In case you missed it, Susan D’Agostino (2023) posted a carefully written piece on “Designing Assignments in the ChatGPT era” for Inside Higher Ed (▷), and I thoroughly enjoyed spending over an hour talking with her about the Sentient Syllabus Project for her research.
On the day that this update was first published, Nathan Greenfield’s “Facing Facts: ChatGPT can be a tool for critical thinking” appeared in University World News (2023, ▷). I spent an engaging morning on a Zoom call with Nathan; the article turned out to be quite nuanced, it covers a lot of complicated issues well, and conveys a very good practical understanding; Nathan’s long experience in academia comes through in his writing, and that resonates with me. And reassuringly, there was also much resonance with contributions of other colleagues: I found myself agreeing a lot with what Michelle Miller of Northern Arizona University, and David Joyner of Georgia Tech had to say, and such convergence is quite encouraging.
On March 3d a feature of my thoughts appeared in Maclean’s “My students are using ChatGPT to write papers and answer exam questions—and I support it” (▷). I like how the piece pivots around the following:
This is not just about better teaching. Generative AI can already do so many things, all day long, without overtime, benefits or maternity leave. Our students must learn how to be better, how to create additional value for themselves and for others. They have to learn how to surpass the AI. And they’ll have to use the AI to do that. (Steipe, 2023)
Still in the works is a contribution to a Washington Post article. I’ll announce that via Mastodon and Twitter once it appears.
Speaking of microblogs, I sometimes encounter interesting information on mastodon where I post as @boris_steipe@qoto.org .9 Having no strict length limits to posts is nice, one can actually express ideas. Or could – the number of posts that actually contribute nothing new to a discussion is striking even there, and conversations are rare – diminishing returns. I rarely post on Twitter as @St_Syllabus because – well … Twitter.
References
D’AGOSTINO, Susan (2023-01-31) “Designing Assignments in the ChatGPT Era”. Inside Higher Ed. (link)
GREENFIELD, Nathan M. (2023-02-24). “Facing Facts: ChatGPT can be a tool for critical thinking.” University World News. (link)
ROZADO, David (2023-02-16) “RightWingGPT – An AI Manifesting the Opposite Political Biases of ChatGPT”. Rozado’s Visual Analytics (link).
SHENG Ying; ZHENG Lianmin; YUAN Binhang; LI Zhuohan; RYABININ, Max; CHEN Beidi; LIANG, Percy; RÉ, Christopher; STOICA, Ion; ZHANG Ce (2023-02-21). “High-throughput Generative Inference of Large Language Models with a Single GPU”. GitHub (link).10
STEIPE, Boris (2023-03-03) “My students are using ChatGPT to write papers and answer exam questions—and I support it” (as told to Alex Cyr). Maclean’s. (link)
Feedback, comments, and experience are welcome at sentient.syllabus@gmail.com .
Sentient Syllabus is a public good collaborative. To receive new posts you can enter your email for a free subscription. If you find the material useful, please share the post on social media, or cite it in your own writing. If you want to do more, paid subscriptions are available. They have no additional privileges, but they help cover the costs.
Cite: Steipe, Boris (2023) “Resource Updates 2023-02-24”. Sentient Syllabus 2023-02-24 https://sentientsyllabus.substack.com/p/resource-updates-2023-01-24 .
One year ago today, Russian troops began their invasion of Ukraine.
I wish to acknowledge minor contributions by ChatGPT (version 2023-02-13), in response to my prompts, for which I take full responsibility.
Cf. the rather fascinating encounters of Kevin Roose (New York Times, 2023-02-16) and of James West (Mother Jones, 2023-02-18).
Benj Edwards (Ars Technica, 2023-02-17) responding to the Microsoft Announcement that conversations would be linked to five interactions per thread. The limitation is presumably meant to prevent Bing’s “conditioning” from exiting the context window – the string of tokens that is used to determine how the conversation is generated, word by word, and that in the case of Bing contained a sizeable chunk of hidden instructions how the conversation was to be conducted.
While we keep our eyes on the usual suspects – OpenAI, Google, Meta, DeepMind and non-private-sector alternatives like Hugging Face, we should not forget that the cost of training a large transformer model is only on the same order as building a mid-sized warehouse; we haven’t heard very much at all from some companies (IBM?); neither have we heard from the European private sector (nor Russia); and China is readying a large scale roll-out of Baidu’s “Ernie Bot” and a good dozen other developments have been announced – summarized in an excellent overview by Zeji Yang for the MIT Technology Review (2023-02-15, link). And that is not even considering work in the academic sector, nor what the various intelligence communities are doing.
Entia non sunt multiplicanda praeter necessitatem. (Do not make more things than needed. ) This is Occam’s famous razor. But there are many more [WP].
“Bain & Company announces services alliance with OpenAI…” (2023-02-21) (Press release).
Shout out to our subscribers who discovered the Sentient Syllabus there. Your encouragement makes this effort worthwhile.
I need to share this little bit of discourse while I was preparing a reference:
Human: Can you help me and rewrite an author list for me? These are all listed as [Firstname Lastname], I need them as [LASTNAME, Firstname], where the last name is in all-caps and all the numbers are removed. Here is the list "Ying Sheng 1 Lianmin Zheng 2 Binhang Yuan 3 Zhuohan Li 2 Max Ryabinin 4 5 Beidi Chen 6 7 Percy Liang 1 Christopher Re ́ 1 Ion Stoica 2 Ce Zhang 3"
ChatGPT: Sure, here's the author list with the format you requested:
SHENG, Ying
ZHENG, Lianmin
YUAN, Binhang
LI, Zhuohan
RYABININ, Max
CHEN, Beidi
LIANG, Percy
RÉ, Christopher
STOICA, Ion
ZHANG, Ce