Here are things on the web that I found particularly interesting or insightful in February.
The metaphors we use constrain the experiences that we build, and model-as-person is keeping us from exploring the full potential of large models.
I hope that this piece is actully something along the same lines of a piece that just popped into my feeds from Allison Gopnik. In her piece she talks about these systems as being cultural artefacts more than being just agents. At the moment we interact with LLMs as if they are operating lien we do, we picture them as agents that are a little bit like humans. It’s because they chat to us. But this is a critical and deeply limiting constraint. That constraint is cemented by the current generatation of tooling that sits between us and the models. That will change. The way to think about this is to think about how to change the way we can drive our interactions with information. Or rather that we should realise that this can be rethought, even if how is not yet clear to us.
found on 2025–02–02
One lens to approach that question is to put decision-making tradeoffs on axes, and debate which direction we should head from here
A very nice read on how some sensible ideas about how to make research funding operate better. Nothing controversial, well evidenced, pragmatic. Linked when the Trump administration was just getting started with their defenstration of rationality. The point of view in this article is probably already antiquated.
found on 2025–02–02
Wikenigma is a unique wiki-based resource specifically dedicated to documenting fundamental gaps in human knowledge.
Curiosity drives innovation. This is a super nice resource of open curiosities that science has not yet answered.
found on 2025–02–02
A hint to the future arrived quietly over the weekend
This was one of the first write ups of using deep research from GPT. Since then a slew of other models have come out, and I’ve had a chance to try it out. It’s verbose but clearly capable. I need to spend more time thinking about how I might use this type of tool.
found on 2025–02–03
by Gaoije Lin et al
OmniHuman significantly outperforms existing methods, generating extremely realistic human videos based on weak signal inputs, especially audio.
This thing is wild, I can a lot of potentially very useful applications for teaching and communication, but it is also going to blow the doors off of what deep fakes can do.
o1 will just take lazy questions at face value and doesn’t try to pull the context from you. Instead, you need to push as much context as you can into o1.
A very good overview of the difference between chat models and report use. This goes some way into answering my above question about how we might use deep research models, but one thing I’m still not sure about it how to easily feed large context when the tool is a web based text box.
found on 2025–02–13
Kreuzberg is a Python library for text extraction from documents. It provides a unified async interface for extracting text from PDFs, images, office documents, and more.
could be useful for creating text analysis pipelines.
found on 2025–02–15
Despite the hype, the marketing, the tens of thousands of media articles, the trillions of dollars in market capitalization, none of this feels real, or at least real enough to sustain this miserable, specious bubble. People like Marc Benioff claiming that “today’s CEOs are the last to manage all-human workforces” are doing so to pump up their stocks rather than build anything approaching a real product.
This is a refreshingly negative take on deep research and GenAI in general. My own use of these tools remains positive, so to what extent they will up end our economies remains an open question to me.
found on 2025–02–18
Towards frictionless, portable, and sustainable reproducibility with Binder | 2i2c by Chris Holdgraf
After listening to folks across the open science and publishing ecosystem, I noticed a common challenge:
Publishers care about reproducibility of computational narratives and the interactivity that computation can provide. But they lack the capacity to manage computational infrastructure in a way that is flexible enough for all of their authors. This post is a reflection on how ecosystems like Jupyter and managed community hubs could solve some of these challenges.
good thoughts on an important topic.
found on 2025–02–19
I am delighed to see a startup in this area. I probably won’t have a chnace to use this, but this looks like a good use of generating value from LLMs
found on 2025–02–19
Human bottlenecks become more important, the more productive is AI. Let’s say AI increases the rate of good pharma ideas by 10x. Well, until the FDA gets its act together, the relevant constraint is the rate of drug approval, not the rate of drug discovery.
This is a very good take, in my view, on some of the structural limits that change from AI is working in. One consequence is that organisations should work to get everybody familiar and working with these tools.
found on 2025–02–23
But there’s a second dimension to this attack. Because these update buckets are abandoned, the developers who are using them also no longer have the power to patch them automatically to protect them
Our software systems are complex, and the landscape is at a scale beyond what most people can reason about leading to interesting attack vectors like this one.
found on 2025–02–25
Just looks plain useful for designing SQL schema.
found on 2025–02–27
This is the central aspect of this site. This is not a Finnegans Wake museum, a static collection of notes; this is a work in progress, a dynamically-growing repository of elucidations.
I have a small desire to engage with Finnegns Wake at some point, and if I ever get to there, this site may be helpful.
found on 2025–02–28
My tags:
- linkpost::unposted