On the Opportunities and Risks of Foundation Models

A certain gov'ment agency —
I'm not inclined to tell you which —
Awarded a large grant to me
That made me comfortably rich.

The lavish contract that I got
Involved decisions that they make
I thought I could improve a lot.
My clever plan was, we would take

Some data — rather personal —
Downloaded from the internet
And feed it to a big ML1
And use the answers we would get.

And even if (I got a laugh)
The output wasn't super-fine
The agency could fire staff
And thus improve its bottom line.

One thing, however, wasn't clear,
Indeed, a little worrying:
I didn't have the least idea2
How I could build the bloody thing.

So I decided on a plan:
I'd pay a nice consulting fee
To scientists at Leland Stan-
ford Junior University.

The people at CRFM3
Agreed to sign an NDA.4
Accordingly I met with them,
And I heard what they had to say.

"It's necessary to pretrain
A neural network on a task
That is complete for the domain
Then it will do just what you ask.

And then you might want to fine-tune,
Choose carefully your prompts, and tweak
Hypér-parameters; and soon
You're in Fat City, so to speak".

I asked, "Would there be any gain
Or would it be at all worthwhile
From studying the task domain?"
They gave a condescending smile.

"There's really not the slightest need.
Don't worry! It will work out fine.
Foundation models will succeed
At any task that you assign."

"Can I be sure the system won't
Display illegal prejudice?"
"Indeed! It's vital that you don't
Appear to be ignoring this.

It's most important that you cite
The warnings of Timnit Gebru.
But once she's referenced, you're all right:
Proceed as you intended to."

"And could there be a data leak
In ways prohibited by law?"
"Be reassured that, as we speak,
A clever postdoc, Mildred Shaw5

Is working hard on that. To date
She's proved some theorems, it appears.
But you'd be ill-advised to wait.
They won't be practical for years."

I got a billion terabytes.
I trained 5000 GPUs
For forty days and forty nights,
And built a model I could use.

The agency deployed it. I'm
As pleased as punch that I can say
It worked quite well, most of the time.
And I received my bonus pay.

Although, the agency admits,
From time to time the thing does fail
Some people lost their benefits,
And others have been sent to jail.

One contract clause of mine was very
Careful, prudent, and astute:
The model is proprietary,
So it's hard to bring a suit.

Now, having built my own foundation
Model, I'm hot property!
And thanks to my huge reputation
Everyone is hiring me!

1 ML = machine learning program.
2 In Rhode Island, where I grew up, "idea" rhymes with "clear".
3 CRFM = Center for Research on Foundation Models
4 NDA = Non-disclosure agreement.
5 The name "Mildred Shaw" is invented to fit the meter and rhyme. No reference is intended to any actual person.


This is part of the collection Verses for the Information Age by Ernest Davis