Lightnews — Scholar-powered news

Todd Johnson @johnsontoddr.bsky.social · 51m

I believe that Stigltiz book published in the 90s discussed how the IMF simply copied and pasted reports from one country to another. Sloppy people produce sloppy work regardless of tools.

Todd Johnson @johnsontoddr.bsky.social · 58m

That’s fantastic. I use LLMs in similar ways to those you report. I have to actively remind myself to ask LLMs instead of doing a web search because with an LLM I can ask what I’m looking for and get a stab at an answer rather than links.

1

Reposted by Todd Johnson

Scott Horton @robertscotthorton.bsky.social · 15h

Trump said he deserved a cut for brokering the TikTok deal, and at once there was agreement that his son Baron would have a top executive slot... though he hasn't asked for it, and has no apparent qualifications to hold it. This is how Trump 2.0 works. www.telegraph.co.uk/us/news/2025...

Barron Trump tipped for top TikTok job

US president’s 19-year-old son could be appointed to the app’s board

www.telegraph.co.uk

230 550 1.2K

Reposted by Todd Johnson

Peter Gleick @petergleick.bsky.social · 1h

Here's how extreme the gaslighting has gotten. Trump sycophants are calling Trump the best president the world has ever seen, when objective reality is that he's absolutely the worst president the world has ever seen.

7 46

Todd Johnson @johnsontoddr.bsky.social · 1h

Out of the box, current LLMs do not augment cognition, though there may be a way to use them to do so. At this point, at least for me, they are just speeding up tasks (and sometimes slowing them down).

Todd Johnson @johnsontoddr.bsky.social · 1h

I bring this up because augmentation seems to be fundamentally different from your three categories. Augmentation changes how we think about a task and about what we CAN think. As Victor argues, tools that augment cognition allow us to think unthinkable thoughts.

1

Todd Johnson @johnsontoddr.bsky.social · 1h

And here is a pre-LLM paper on augmenting cognition using AI: distill.pub/2017/aia/

Using Artificial Intelligence to Augment Human Intelligence

By creating user interfaces which let us work with the representations inside machine learning models, we can give people new tools for reasoning.

distill.pub

1

Todd Johnson @johnsontoddr.bsky.social · 1h

Bret Victor's Media for Thinking the Unthinkable expands on this. In human history, we have perhaps made the most advances in cognition through tools and frameworks that augment cognition, such as algebra and calculus. worrydream.com/MediaForThin...

Media for Thinking the Unthinkable

worrydream.com

1

Todd Johnson @johnsontoddr.bsky.social · 1h

Bret Victor's Media for Thinking the Unthinkable expands on this. In human history, tools and frameworks we have perhaps made the most advances in cognition through tools and frameworks that augment cognition, such as algebra and calculus. worrydream.com/MediaForThin...

Media for Thinking the Unthinkable

worrydream.com

1

Todd Johnson @johnsontoddr.bsky.social · 3h

I also think that my use of LLMs helps prevent issues. I use LLMs mostly in the context of putting together simulations to help me and hopefully others understand some area of interest. The measure of success is not the output of the LLM, but whether it helps me produce a good explainer.

Todd Johnson @johnsontoddr.bsky.social · 3h

Where I question LLM use is when it is writing large chunks of code in an area that I am unfamiliar with. Claude can easily overdo a simple task. But this means that those use cases need MORE critical thinking, not less.

1

Todd Johnson @johnsontoddr.bsky.social · 3h

There is no vibe science here, at least in what I'm using the LLMs for. Coders regularly use and modify existing code from others, rather than start from scratch. I use LLMs for small snippets that I then modify as needed. LLMs are just tools. They can be used or abused, just like stats.

1

Todd Johnson @johnsontoddr.bsky.social · 4h

In my area, I am very concerned about the use of LLMs in healthcare. This is already starting and the companies pushing them have every incentive to hype them, whereas the people buying them have very little understanding of what they are and their limitations and harms.

1 2

Todd Johnson @johnsontoddr.bsky.social · 4h

Both are of interest. I spent a decade or more working in cognitive science and with cognitive psychologists. My current work in informatics is still, in part, informed by cognitive science. I also still teach some relevant cog sci/psych principles and theories in class.

1 1

Todd Johnson @johnsontoddr.bsky.social · 4h

I have to critically evaluate all output and the LLM starts going bad after a few prompts. I also use LLMs like Google search on steroids, then again evaluate everything it gives me. Still saves a lot of time. I am still learning to use LLMs this way and there are times when they waste time.

Todd Johnson @johnsontoddr.bsky.social · 4h

Will do; however, my experience is very different from your claims, perhaps due to the field I'm in now (I used to do cognitive science, but now am in informatics). For instance, it is much faster for me to tell Claude to give me some code for a simulation model for causal inference.

3

Todd Johnson @johnsontoddr.bsky.social · 4h

Some of your other points here are concerning and well-warranted. Even with architectures, such as ACT-R or SOAR, that were designed to mimic human subjects, using LLMs to replace subjects seems inappropriate.

Todd Johnson @johnsontoddr.bsky.social · 4h

I completely disagree with the last sentence of the abstract. That really depends on how you use AI. There is plenty of misconduct already without AI, such as major medical journals banning causal language except for RCTs, p-hacking, associational papers that are really seeking causal effects, etc.

3 1

Todd Johnson @johnsontoddr.bsky.social · 4h

When number of grants and papers becomes the main incentives that distorts what science gets done and what gets written up. When leading researchers ask, "How can we turn this into a paper?" it is a sign that we have put the cart before the horse.

Todd Johnson @johnsontoddr.bsky.social · 4h

This decline is not just with students or education in higher ed in the US, but with science as well. The incentive system has put grants and papers over science itself, such that too many researchers work backward from grants and papers to science, instead of putting science first.

1

Todd Johnson @johnsontoddr.bsky.social · 4h

This show was first aired in June 2005. No AI needed to dumb down education. Ultimately, the effects of AI will rest solely on how we use it in education and the standards we hold students to. We have not been doing very well at all, even before AI.

1

Todd Johnson @johnsontoddr.bsky.social · 4h

www.youtube.com/watch?v=BcxD...

Declining by Degrees: Higher Education at Risk

YouTube video by Bamboo Invasion

www.youtube.com

1

Todd Johnson @johnsontoddr.bsky.social · 4h

Link here. Worth watching. I've been teaching in grad school since 1992. The pressure to admit as many students as possible and graduate as soon as possible is real. I know of faculty who have had to find new jobs because "they flunked" too many of their students.

1

Todd Johnson @johnsontoddr.bsky.social · 5h

It probably matters that I grew up without the internet and finished my PhD just when the internet was getting started. I also use AI for topics and tasks that I am genuinely interested in, whereas many students these days just seem to want a degree.

Todd Johnson @johnsontoddr.bsky.social · 5h

Perhaps its because I am in a technical field, but I see this very differently. I agree that uncritical adoption is harmful, but by being very careful I am finding that LLMs (starting about a year ago) are now quite helpful at increasing critical analysis and also automating chunks of coding.

1