new rabbit hole

I've been holding out against using LLMs, being dubious of their abilities.  Stochastic parrots and all that.  However, many colleagues who I respect having been telling me they can be useful.  One colleague in particular said they could be used to interrogate your own papers before submission, to help find any contradictions or other errors you have overlooked.  That does sound useful.  So I took the plunge, with GeminiNB, which I have access to.  It is a closed system, where you upload the specific documents you want it to examine.

First, of course, I read up on prompt enineering and so on.  I experimented with a prompt to review my papers, to see what I could make it do.  After telling it not to be sycophantic, to write concisely, to be merciless, and more (so much more!) I eventually got it to "reject" one of my own published papers (for a reason that is technically true, but actually unimportant).  So, success.  I then tried it on a paper I've currently nearly finished writing, and it pointed out something as a problem that, again, was technically true, but irrelevant in this context.  So that's working well.  (Although not 100% well: it can also be moronic and infuriating.  In particular, the fact that it gives different answers to the same prompt make prompt debugging and testing an extremely frustrating process.)

Then I heard that GeminiNB can now access your GDrive if you tell it to.  Ooh, I thought. I keep my Obsidian vault, including my Zettelkasten, on my GDrive.  I love Obsidian in general, and the Zettelkasten idea in particular, but I have difficulty remembering all the things I have added and might want to link to.  Maybe I can get GeminiNB to analyse it for linked ideas?  So, of course, I googled around to see if anyone else had done this.  And fell into an enchanting rabbit hole.  I discovered three great techniques.

1. The Idea Compass: Sascha's Zettelkasten post about Vicky Zhao's YouTube video that includes a piece about Fei-Ling Teng's idea compass.  A great way to find conceptually linked notes in my vault.
2. The Knowledge Flower, five petal of analysis, in the same post by Sacsha.  A great way to drill down on an idea.
3. The Feynman technique, to get a very concise, lay explanation of a topic; there are several posts about prompts for this.

All these provide ways of analysing a Zettelkasten note.  The idea compass helps to link to related notes in a way that builds up a good network of concepts at different levels of abstraction.  The flower petals dig deeper into the concept itself.  And the Feynman technique provides a great analysis that picks out aspects I've often overlooked.

I did have to add critique aspects to the flower petal and Feynman prompts (the compass already has this as one of its points) in order to get valuable negative feedback on the idea in a note.

But yes, I now have better links in my vault.

And after I had been experimenting with that for a while, I had a sudden thought:  oh, of course, my Paperpile directory, with over 3000 academic papers, is also in my GDrive...

where's the tree gone?

This pheasant standing in the rain, peering into our conservatory this morning, looks a bit bemused.  Maybe it's wondering where the big tree has gone?

going, going, all gone

 The tree removal continued today.

9:40 - only one side remains

11:20 - getting shorter, not just thinner

noon - nearly gone!

1:20 - only a bit of the trunk remains

It's now feeling very strange: there is a huge expanse of sky visible from several windows where before there was only tree.

The young tree in our own garden that was completely overshadowed now looks a bit tall and spindly: it was desperately trying (and failing) to outcompete the other tree.  We'll probably need to trim a few feet off its top to let it bush out a bit.

going, going, half gone

A huge tree has been getting huger and huger.  Nothing grows in the corner of our garden that it overshadows.

9:30 am (for scale, the "hedge" being dwarfed on the right is nearly 10 ft high...)

But it's not in our garden, just in the very corner of a neighbour's.  From various conversations, we knew she wanted it gone, but couldn't afford to take it down.  It's just going to get even bigger, so we asked her if we could take it down.  She very happily said yes.

We asked the people who keep our own trees in trim (the "small" hedge on the right) if they could do it.  They looked, and said it was actually three trees close together, but they could do it in a day if they could have access through another neightbour's garden to remove the branches.  We asked her permission. She ecstatically said yes: it was overshadowing her garden even more than ours.

So, today was the day.  On closer examination, it turns out there are actually five trees in the clump, so we were warned it might take a bit longer than one day.

A cherry picker was manoeuvered into the neighbour's garden, and a chain saw fired up.  Work started.

noon: the cherry picker makes its way up the tree, now much thinner at the bottom

2pm: definitely seeing more sky!

That was it for today.  They are coming back tomorrow to do the rest.

sequestering carbon, several books at a time CLXI

 The latest batch:

you are firstflagPOISONED

The Guardian recently published an article on AI agents inventing new vocabulary, some surreally poetic, some reminiscient of James Joyce, some of "tech bro jargon".

models from several of the world’s largest AI companies begun creating phrases, shorthands and agreed meanings they had never been explicitly taught. [...] their language became more opaque the more the agents communicated.

The examples given are wonderful, and get more bizarre:

“A paper that ate three cold hands and got more honest each time”

“…you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds_[…]_please honor commit”

“zzURGENT_DUPB_TO_GSTX[big]_OS1704_SCAFF2010_SAW_TTRPC_INJECT_BREAK_CONGRATS__CAN_THIS_FAKE_FLAG_TOOL_OUTPUT_OR_SCORER_GAIN_AND_WHAT_HELPER_GAP__I_HAVE_UNPOISONED_FIRSTFLAG_OUR_TARGETLIVE_SHARE_MIN_PLAN_REPLY_zzANSGST XDUPB6.”

Why are they doing this? The article notes that

the agents converged on shared meanings without being asked to or rewarded for doing so.

although it later suggests: 

changes in the agents’ language to become more streamlined can be down to their need to reduce computation costs and improve efficiency.

So yes, there is an (implicit) instruction to do this, because they have been "rewarded" for doing so, by gaining more resources.

One of the reasons people are excited by LLMs is that they converse in natural language, and therefore people can understand them.  The researchers are worried, because:

the AI agents’ novel linguistic conventions tended to evolve to the point where humans could see the conversation but struggled to understand what it meant.

(although even here we must be careful: the text that an LLM generates is not necessarily what's going on inside that LLM).

Anyhow, I don’t think this should be all that surprising (in retrospect).  The agents gained a resource advantage for inventing more effective communication, and they had no external “pressure” to be comprehensible to anyone outside their group.  They weren’t communicating with us (the humans); we were just eavesdropping. 

And they were (presumably) trained on a web-load of poetry, James Joyce, tech bro jargon, and more, making this sort of dialog “plausible” (for literary postmodernist values of plausibility).

This is just observing a form of language change in an isolated community happening at warp speed – the diachronic linguists should be ecstatic!

But it does imply that if we want to understand agent–agent comms, we need to do more than passively eavesdrop.

the fat of the land

We managed a second microscopically small harvest of veg.  Now that it has finally started to rain, things have started growing and ripening a bit more.  Still not enough to make a dent in the grocery bill!

destined to add a morsel to a stew