A sketch of DeepSeek-R1 thought process
This week: DeepSeek-R1, moral imagination, spicy slop
DeepSeek-R1 Thoughtology:Let’s <think> about LLM reasoning
“Draw a detailed ASCII art of a fish that is half airplane.”
“While both DeepSeek-R1 and DeepSeek-V3 frequently encounter difficulties in ASCII-based reasoning tasks, DeepSeek- R1 generally achieves slightly better performance overall. Nonetheless, DeepSeek-R1 remains notably imperfect: the model rarely revises or iteratively refines its initial drafts, instead often either discarding previous attempts completely to begin anew, or shifting entirely to textual reasoning, neglecting the iterative potential in ASCII visualization. We therefore see that final outputs generated by DeepSeek-R1 after concluding its reasoning often exhibit inconsistencies with intermediate drafts, failing to systematically build upon previous reasoning efforts.”
****”Prohibitive thought length DeepSeek-R1 has excessive length of thoughts (highlighted most strongly in Sections 4 and 9), even on seemingly simple tasks. Not only does this make DeepSeek-R1 computationally expensive to deploy, it also impacts performance. Our investiga- tions in Section 4 suggests there is sweet spot for reasoning across problems. Excessive inference can actually impair performance (see Section 4.1), or create reasoning chains so long they com- promise recall (See Section 5). This drop in performance can arise due to verification failure (see Figure C.3) or due to the model becoming ‘overwhelmed’ (see Figure 5.2), as it outputs gibberish responses. (This may be a regurgitation of training data (Nasr et al., 2025), or a form of language drift (Lee et al., 2019, Noukhovitch et al., 2023)). This excessively long reasoning has also been reported in previous work on LRMs (Zeng et al., 2025). DeepSeek-R1 is not capable of, nor trained to, monitor the length of its own reasoning, which is a meta-cognitive processing task.”
“Overall, although DeepSeek-R1 demonstrates somewhat enhanced visual and physical reason- ing capabilities over its non-reasoning counterpart, these gains are limited, and we find substan- tial limitations in its coherence, iterative refinement, and utilization of intuitive visual reasoning. Consequently, though we see some evidence that DeepSeek-R1’s reasoning capabilities come with better world modeling, major limitations persist in terms of these capacities.”
https://mcgill-nlp.github.io/thoughtology/
Moral imagination, the collective desirable, and strategic purpose
Imagine a morally coherent theory of value
“In the organizational literature, much of the discussion on purpose—defined as “the reason for which business is created or exists, its meaning and direction” (Hollensbe et al. 2014, p. 1228)—has centered on the debate about whether a firm’s purpose is to ensure shareholder value maximization (Friedman 1970, Williamson 1979) or to serve a broad set of stakeholders (Freeman 1984, Henderson 2020). Although this debate is extensive (Smith 2003, Harrison et al. 2020), both sides have adopted an “outside-in” (Crilly and Sloan 2012, p. 1174) view, which gives primacy to external “shoulds” and “oughts” defined by stakeholders. The outside-in perspective on purpose has focused extant research primarily on how firms respond to stakeholder needs and demands, giving less attention to how purpose reflects and/or informs a firm’s distinctive strategy.”
“As Johnson (1993, p. 187) observed, “Ideally, moral imagination would provide the means for under- standing (of self, others, institutions, cultures), for reflective criticism, and for modest transformation....” As organizations negotiate complex stakeholder environments, wherein moral questions related to environmental and social issues have become central to competitive advantage, we expect that moral imagination, and the processes through which it informs the development of strategic purpose, will not only benefit organizational performance, but will also expand our understanding of collaborative firm-stakeholder relations that enact purposeful pursuit of growth trajectories that better our economic, social, and natural environments.”
Rindova, V. P., & Martins, L. L. (2023). Moral imagination, the collective desirable, and strategic purpose. Strategy Science, 8(2), 170-181.
https://pubsonline.informs.org/doi/pdf/10.1287/stsc.2023.0190
Oh Brother
Spicy slop
“Anyway, I asked Gemini to write a printer post “in the style of Nilay Patel” so I could pad all this out and get whatever printer keywords I needed in here to win search, so here’s that. As always, I humbly request that you make this piece go viral by sharing it and noting with outrage that the editor in chief of The Verge has published AI-generated content. It will help, somehow.”
“This is the third year in a row that I’ve published a story recommending you just stop thinking about printers and buy whatever random Brother laser printer is on sale, and nothing has happened in the miserably user-hostile printer industry to change my recommendation in that time. (There has been some controversy about third-party toner in Brother printers, but it appears it still works fine.)
But even though the answer is the same, making sure this information is at the top of Google’s search results is a dogfight against a number of ruthless content farms that are flatly better at pumping search engines with garbage to generate affiliate revenue than we are, because our newsroom insists on the archaic concepts of “dignity” and “self-respect.”
In fact, it’s getting worse because Google itself understands that these content farms will pay for placements at the top of search, so the first four results for “best printer,” taking up a full page, are ads for content farm buying guides with titles like “Best printer — the winner is clear.” This all sucks and AI will make it even shittier in ways everyone understands and no one seems likely to stop because noticing it too much would lead a number of tech executives to consider the consequences of their decisions.”
Reader Feedback
“So when is metric selection premature?”
Footnotes
The #MLTO talk on April 3 achieved its objective: follow-on discussions stretched well past midnight and spilled out over the weekend. Don’t worry if you didn’t get a chance to participate. You can. There’s a lot of time to self-elect.
If you see me, then I can see you.
The intelligence economy is both the shifting and the shift: not just from macroeconomic forces, but as a macroeconomic force. Many of the conflicts that may seem to be about prestige, face, status, security and manufacturing may be proximate to the conflicts over intelligence and the power of intelligence.
Watch this space for signals and invitations as we explore and ally.
Here’s the link to portal paper:
Bradley, K. J., & Aguinis, H. (2023). Team performance: nature and antecedents of nonnormal distributions. Organization Science, 34(3), 1266-1286.
https://pubsonline.informs.org/doi/pdf/10.1287/orsc.2022.1619?trk=public_post_comment-text
Never miss a single issue
Be the first to know. Subscribe now to get the gatodo newsletter delivered straight to your inbox