Email discipline

There is a Discourse on what we might call email discipline. Here are a few related takes.

There are those who simply don’t respond to email at all. These people are demons and you should pay them no mind. 

Relatedly, there are those who “perform” some kind of message about their non-email responding. Maybe they have a long FAQ on their personal website about how exactly they do or do not want to be emailed. I am not sure I actually believe these people get qualitatively more email than I do. Maybe they get twice as much as me, but I don’t think anybody’s reading that FAQ buddy. Be serious.

There are those who believe it is a violation to email people off-hours, or on weekends or holidays, or whatever. I don’t agree: it’s an asynchronous communication mechanism, so that’s sort of the whole point. I can have personal rules about when I read email and these depend in no way on my rules (or lack thereof) about when you send them. Expecting people to know and abide by your Email Reading Rules FAQ is just as silly.

I have an executive function deficit, diagnosed as a child (you know the one), and if you’re lucky, they teach you strategies to cope. I think non-impaired people should just model one of the best: email can’t be allowed to linger. If it’s unimportant, you need to archive it. If it’s important you need to respond to it. You should not have a mass of unopened, unarchived emails at any point in your life. It’s really that easy.

When LLMing goes wrong

[The following is a guest post from Daniel Yakubov.]

You’ve probably noticed that industries have been jumping to adopt some vague notion of “AI” or peacocking about their AI-powered something-or-other. Unsurprisingly, the scrambled nature of this adoption leads to a slew of issues. This post outlines a fact obvious to technical crowds, but not business folks; even though LLMs are a shiny new toy, LLM-centric systems still require careful consideration.

Hallucination is possibly the most common issue in LLM systems. It is the tendency for an LLM to prioritize responding rather than responding accurately, aka. making stuff up. Considering some of the common approaches to fixing this, we can understand what problems these techniques introduce. 

A quick approach that many prompt engineers I know think is the end-all be-all of Generative AI is Chain-of-Thought (CoT; Wei et al 2023). This simple approach just tells the LLM to break down its reasoning “step-by-step” before outputting a response. While a bandage, CoT does not actually inject new knowledge into an LLM, this is where the Retrieval Augmented Generation (RAG) craze began. RAG represents a family of approaches that add relevant context to a prompt via search (Patrick et al. 2020). RAG pipelines come with their own errors that need to be understood,  including noise in the source documents, misconfigurations in the context window of the search encoder, and specificity of the LLM reply (Barnett et al. 2024). Specificity is particularly frustrating. Imagine you ask a chatbot “Where is Paris?” and it replies “According to my research, Paris is on Earth.” At this stage, RAG and CoT combined still cannot deal with complicated user queries accurately (or well, math). To address that, the ReAct agent framework (Yao et al 2023) is commonly used. ReAct, in a nutshell, gives the LLM access to a series of tools and the ability to “requery” itself depending on the answer it gave to the user query. A  central part of ReAct is the LLM being able to choose which tool to use. This is a classification task, and LLMs are observed to suffer from an inherent label bias (Reif and Schwarz, 2024), another issue to control for.

This can go for much longer, but I feel the point should be clear. Hopefully this gives a more academic crowd some insight into when LLMing goes wrong.

References

Barnett, S., Kurniawan, S., Thudumu, S. Brannelly, Z., and Abdelrazek, M. 2024. Seven failure points when engineering a retrieval augmented generation system.
Lewis, P., Perez, E., Pitkus, A., Petroni, F., Karpukhin, V., Goyal, N., …, Kiela, D. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks.
Reif, Y., and Schwartz, R. 2024. Beyond performance: quantifying and mitigating label bias in LLMs.
Wei, J. Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., …, Zhou, D. 2023. Chain-of-thought prompting elicits reasoning in large language models.
Yao, S., Zhao, J., Yu, D., Shafran, I., Narasimhan, K., and Cao, Y. 2023. ReAct: synergizing reasoning and acting in language models.

 

Snacks at talks

The following is how to put out a classy spread for your next talk; ignoring beverages and extras, everything listed should ring up at around $50.

  • The most important snack is cheese. Yes, some people are vegan or lactose-intolerant, but cheese is one of the most universally-beloved snacks world-wide. Most cheeses keep for a while with refrigeration, and some even keep at room temperature. Cheese is, as a dear friend says, one of the few products whose quality scales more or less linearly with its price, and I would recommend at least two mid-grade cheeses. I usually buy one soft one (Camembert, Brie, and Stilton are good choices) and one semi-hard one (Emmental or an aged Cheddar for example). The cheese should be laid out on a cutting board with some kind of metal knife for each. The cheese should not be pre-cut (that’s a little tacky). Cheeses should be paired with a box of Carr’s Water Crackers or similar. Estimated price: $15-20.
  • Fresh finger vegetables are also universally liked. The easiest options are finger carrots and pre-cut celery sticks. If you can find pre-cut multi-color bell peppers or broccoli, those are good options too. You can pair this with some kind of creamy dip (it’s easy to make ranch or onion dip using a pint of sour cream and a dip packet, but you need a spoon or spatula to stir it up) but you certainly don’t have to. Estimated price: $10-20.
  • Fruit is a great option. The simplest thing to do is to just buy berries, but this is not foolproof: blueberries are a little small for eating by hand; raspberries lack structural integrity, and where I live, strawberries are only in season in the mid-summer, and are expensive and low-quality otherwise. In Mid-Atlantic cities, there are often street vendors who sell containers of freshly-cut fruit (this usually includes slices of pineapples and mangos and bananas, and perhaps some berries) and if this is available this is a good idea too. Estimated price: $10-15.

This, plus some water, is basically all you need to put out. Here are some ways to potentially extend it.

  • Chips are a good option. I think ordinary salty potato chips are probably the best choice simply because they’re usually eaten by themselves. In contrast, if you put out tortilla chips, you need to pair them with some kind of salsa or dip, and you need to buy a brand with sufficient “structural integrity” to actually pick up the dip.
  • Nuts are good too, obviously; maybe pick out a medley.
  • Soda water is really popular and cheap. I recommend 12oz cans. It should always be served chilled.
  • A few bottles or cans of beer may go over well. With rare exceptions, should be served chilled.
  • A bottle of wine may be appropriate. Chill it if it’s a varietal that needs to be chilled.

If the talk is before noon, coffee (and possibly hot water and tea bags) is more or less expected. There is something of a taboo in the States of consuming or serving alcohol before 4pm or so, and you may or may not want your event to have a happy hour atmosphere even if it’s in the evening.

And here are a few things I cannot recommend:

  • In my milieu it is uncommon for people to drink actual soda.
  • I wouldn’t recommend cured meats or charcuterie for a talk. The majority of people won’t touch the stuff these days, and it’s pretty expensive.
  • I love hummus, but mass-produced hummus is almost universally terrible. Make it at home (it’s easy if you have a food processor) or forget about it.
  • Store-bought guacamole tastes even worse, and it has a very short shelf life.