James Cunningham

Once nearly an economist, I now work in that liminal space between IT and applied stats. Will add caffeinated syrup to energy drinks without regard to health.

[James Cunningham @jamescunningham]

@edzitron Did you really just permanently ban me from your subreddit for telling you not to get bitter about getting insufficient credit? After you complained bitterly about not getting credit? Seriously?

[James Cunningham @jamescunningham]

@DeryaTR_ @aravartanian What is that supposed to mean, exactly? "Not for long"? Are you imaging pogroms or some such?

[James Cunningham @jamescunningham]

@bigolhumbo @jesawyer In addition to what others have said, yes, there are two Pathfinder games that offer both options: Kingmaker and Wrath of the Righteous.

TB is nice to have but battles were designed around RTwP, thus many of them have so many mobs that they take *forever*. Some > 1h.

[James Cunningham @jamescunningham]

@jesawyer Wrath of the Righteous is marginally better. But only because either no micromanaging is needed, or else I turn on TB mode. Unfortunately TB is a complete slog for most fights in the PF games … from which I conclude that encounter designs should be tailored to either choice.

[James Cunningham @jamescunningham]

@jesawyer For what it’s worth, RTwP does sort of turn me off the original PoE, which I’m playing now. My brain just doesn’t cope — I forget to use abilities on cooldowns, I can’t target AoEs properly, positioning mucks me up.

A “me” problem, I suppose.

[James Cunningham @jamescunningham]

I asked various AIs what the coolest psychiatric medication was. Overwhelmingly their answer was (...drumroll...) ketamine. Good ol' vitamin k.

[James Cunningham @jamescunningham]

@RaptorBricks @simoncarless I don't think it's a terrible question necessarily. If AI continues to improve it could become increasingly difficult to tell whether a game is using it, especially without many false positives. I suspect Valve could pour a lot of money into this and not get it perfectly right.

[James Cunningham @jamescunningham]

@FrayedBond @sleepy_oli @soumithchintala I don't see how this helps open source obliterate any moats for the time being, as the resources required to train a 1.7T parameter model are far, far out of reach for anything but wealthy companies. Even 65B parameter models are too high an expense in general.

[James Cunningham @jamescunningham]

@reactive_dude @rowancheung API key cheaper? If you queried GPT-4 as much as possible via ChatGPT, you could do so thousands of times throughout a month. But a single query via the API could easily chew through 20 cents or more, which could mean substantially more money than $20 over the span of a month.

[James Cunningham @jamescunningham]

@rowancheung Poe gives you, what, 2 queries against GPT-4 per day? It's hardly enough to do do anything, unfortunately. Of course, you can pay $20 to get 600 queries per month (which feels a bit pricey to me but YMMV).

[James Cunningham @jamescunningham]

For all its faults, Bard is the only language model so far to pass one of my go-to coding tests (without handholding): write a function to implement a salted join in PySpark. It's a simple problem but GPT-4, Claude, etc all get it wrong. #LLMs #Bard #GPT4

[James Cunningham @jamescunningham]

I've found something ChatGPT is horrible at: providing code to generate a mock dataset based on specifications that are even mildly complex. It just doesn't *get it*, no matter how I phrase the question. #ChatGPT

[James Cunningham @jamescunningham]

Its source for this is a Plotly-related StackOverflow question. If I had to take a WAG, maybe it picked up on some physics-related stack-exchange link that was on the page when it was scraped. But who knows. Maybe it just has a weird sense of humor.

[James Cunningham @jamescunningham]

After asking Perplexity (which AIUI uses GPT-3.5 under the hood) for troubleshooting suggestions regarding Plotly plots not displaying correctly in Jupyter notebooks, it gave reasonable suggestions. It followed this up with:

[James Cunningham @jamescunningham]

I guess this might explain why my plots aren't showing up ...

[James Cunningham @jamescunningham]

But does it give any insight into all as to how to make such things more intelligent?

[James Cunningham @jamescunningham]

Can someone explain to me why so many people fear ultra-intelligent AI, when all recent advances in AI derive from models *designed from the ground up* to generate facsimiles of texts written by broadly average-IQ humans?

[James Cunningham @jamescunningham]

We still don't understand *how* to create intelligent machines. The availability of huge amounts of training data (i.e., text written by mostly average IQ humans) permits a brute force hack to create a human-esque text generator. Great work ...

[James Cunningham @jamescunningham]

I don't think it works like that, though.

[James Cunningham @jamescunningham]

I understand the propensity toward extrapolation. We went from zero chatbots that could plausibly write like a human to several in what seems like an instant, so maybe an instant later we'll go from human-like chatbots to advanced super-intelligent chatbots!

[James Cunningham @jamescunningham]

I don't voraciously consume all research into LLMs, but most of the research I'm aware of goes into developing better or faster facsimiles of (again) mostly average-IQ text. Where do we suppose a super-intelligence is going to come from?