At the edge of human reasoning

Decoding-based Regression

Interesting research, commercially…I’d want to see the costs come down

“Empirically, we showed that it can be competitive as, or even outperform traditional pointwise heads for regression tasks. Furthermore, it is also capable of density estimation over a variety of conditional distributions 𝑝(𝑦|𝑥), and can further outperform common baselines such as Gaussian mixtures and Riemann distributions.”

”Numerous ways to extend decoding-based regression include improved tokenization schemes or other basis distributions besides piecewise constants. Further applications exist in other domains such as computer vision, where the encoder may be a convolutional network, or for multi-target regression, where the regressor needs to predict multiple different 𝑦-value targets. More broadly however, we hope this will also be a valuable reference for the language modeling community, and provides a principled explanation for the use of supervised fine-tuning over numeric targets.”

Song, X., & Bahri, D. (2025). Decoding-based Regression. arXiv preprint arXiv:2501.19383.

https://arxiv.org/abs/2501.19383

https://github.com/google-research/optformer/tree/main/optformer%2Fdecoding_regression

Data Visualization and the State of the Industry

Unstable diffusion

Artificial Intelligence (AI) technology is shifting much of how we work, including within the field of data visualization. This is the second year the survey asked data visualizers if they’ve used AI in their data visualization work in the past year (Figure 16). This year:

• 60% (492 people) did not use AI in their data visualization work. The percentage decreased by 14 percentage points compared with 2023.
• 37% (305 people) used AI in their data visualization work. The percentage increased by 13 percentage points compared with 2023.

3% (28 people) were not sure if they used AI in their data visualization work. The percentage increased by 1 percentage point compared with 2023.

https://www.datavisualizationsociety.org/soti-report-2024

Pre-entry experience, knowledge inheritance, and entrepreneurial resource mobilization

Clusters matter

“The study shows that new ventures can benefit from showcasing not only their founders’ prior experience in the industry but also their knowledge linkages to their former employers in building the new venture’s technology. Many technology entrepreneurs face a Darwinian struggle to secure resources. In this regard, our findings reveal that leveraging knowledge linkages to the erstwhile employers of the founders affords new ventures one more advantage in the competition for resources. By leveraging knowledge from the founders’ former employers, new ventures signal to external resource providers their ability to learn and build upon inherited knowledge, thereby enabling external resource providers to assess the venture’s future potential better and more positively.”

Chila, V., Devarakonda, S. V., & Martin, X. (2025). Pre-entry experience, knowledge inheritance, and entrepreneurial resource mobilization. Small Business Economics, 1-24.

https://link.springer.com/content/pdf/10.1007/s11187-025-01003-8.pdf

Chance, probability, and uncertainty at the edge of human reasoning: What is Knightian uncertainty?

Dark knight of the soul

“Knight's prophetic insights about the nature of true uncertainty remind us even today that the agentic choices of human actors, as well as the evolving indeterminism and dynamism of the environment, remain potent sources of real indeterminism and human unknowingness, but also the inspiration and impetus for entrepreneurial action to actualize a desirable future.”

Townsend, D. M., Hunt, R. A., & Rady, J. (2024). Chance, probability, and uncertainty at the edge of human reasoning: What is Knightian uncertainty?. Strategic Entrepreneurship Journal, 18(3), 451-474.

https://onlinelibrary.wiley.com/doi/10.1002/sej.1516

Reader Feedback

“We see conflicting interests in tech transfer in open source projects all the time.”

Footnotes

Suppose an agent only knew the world up 1350. It had the sum of written knowledge up to that point.

How would it act?

Never miss a single issue

Subscribe now to get the gatodo newsletter delivered straight to your inbox