Let me close by noting that one of the best job benefits is heavily subsidized faculty housing for new professors, located in the aforementioned valley of Mānoa (pictured below).

It is walking distance from the econ department.

5/5

Re: the PF position, the ideal applicant would be interested contributing to our budding research partnership with the HI Department of Taxation, helping them to use their admin data for conducting policy-relevant research.

The ideal applicant also enjoys plumeria trees.

4/5

We can hire at either the assistant or associate level for both positions, and welcome applications from everyone: new PhDs, folks with existing faculty positions, and/or those with policy experience outside academia.

Did I mention how many rainbows you'll see working here?

3/5

Both are academic research positions, with reduced teaching loads, and an expectation of contributing to UHERO's policy outreach work in HI.

There are very real opportunities for impacting public policy through a role at UHERO.

For example: bsky.app/profile/dyla...

2/5

With highly technical material, quality varies a lot, but having a reasoning model do this improves things a lot, especially if the paper is well written.

Customizing the prompt can help a lot too. Mine includes my technical background, preferences on how to discuss technical details, etc.

DIY is best. Pick your fav reasoning model, upload PDF, tell it you want a summary that you can give to a text to speech model. Copy & paste into the ElevenLabs phone app to get amazing text to speech for free.

I use Gemini flash thinking, with a detailed prompt. Free here:
aistudio.google.com

How good are these things? Very much good enough to be useful for my work. If they add LaTeX rendering functionality to ChatGPT's Canvas feature, now that it has o1, it would become my new default workspace for any new task like this.

Ex: in one such prompt, a demand function was needed, o1 and Gemini assumed a linear demand function unless explicitly told not to. By contrast, Deepseek formally examined the implications of a nonlinear demand function without being asked. I'd prefer that approach personally.

So is there anything Deepseek does better? If I ask it to complete a slightly less well defined task, such as "write up a model that has XYZ properties, and derive the solution" it seems to propose slightly more general modeling frameworks than the other models discussed above.

I have encountered a few other cases like this. And given the other advantages of Gemini's "Thinking" model, it would be my go to choice for someone looking to use reasoning models for this kind of task without paying for it.

Note, as with deepseek, you can see the "reasoning" output of the model.

The same error as Claude

But Deepseek-r1 makes an error

Here's o1-mini

Here's Gemini's newest "thinking" model, with a weird formatting issue, but still correct

Here's o1, with the right answer. o1-pro gets it also, naturally.

Here's an example of the kind of task I mean. This is for an actual project, and is certainly simple enough to do, but a tad tedious. It sure would be nice to outsource it to a virtual RA, but only if I can trust the results!

That last point is important because - like Deepseek - this is a model anyone can use for free: aistudio.google.com

And it has a much larger context window than Deepseek. No difficult reading an entire paper, or set of papers, and working through them with you.

For lots of simple math tasks, I'm still finding it better than Claude, but this was an interesting exception. It isn't that it got the setup of the model wrong, it got the execution wrong, and in a relatively basic way.

What's worse, it o1 knew what the final formula needed to look like, so it corrected itself towards the end with some handwaving to make the needed terms magically reappear. When I tried to correct it multiple times, it dug in, insisting it had things right.