
Full-Stack Software Engineer | Dev at @FCPEuro | Founder of @nditcommunity | Volunteer at @codethedream | Alum of @turingschool | 👩💻
Denver, CO
Joined March 2010
- Tweets804
- Following768
- Followers172
- Likes1.8K
Hayley in the Stack retweeted
Okaaaayyyyyy let me juuuuuuuust shaaaaaaaaaare my screeeeeeen....
okaycanyouseeit?
Hayley in the Stack retweeted
BREAKING: Widespread outages are disrupting Microsoft Outlook, Teams and 365, with users reporting service issues across platforms.
Hayley in the Stack retweeted
We are tracking a breach of a 3rd-party JS host impacting hundreds of high value websites.
This is staggering.
Hayley in the Stack retweeted
Got off a call with the an MS team lead who was on the bridge for this incident on what exactly happened in the outage:
1. This was a "latent bug" - a commit going in ~4 months ago to a Windows build, that only activated much later. Super hard to spot
2. Zero AI, this code was human written + committed
3. Machines got saturated and unreachable at a result
4. So many other things went wrong during this outage: a series of unlucky events
5. It was a dev who found the root cause, poking around and seeing a dump that reminded them of an error seen here. AI diagnostic tools were used and helped progress with the outage
A detailed, public postmortem is on its way. MS wants to make sure they have everything patched globally before the postmortem. I'll add my two cents when it's out, and those details will be VERY educational
Hayley in the Stack retweeted
Impressed. I’ve only seen one other system beat AI detection before and it was closed source.
Source here github.com/EdisonScientific/…
My workplace is hiring! @FCPEuro
lnkd.in/p/gnahurYM
1
22
Hayley in the Stack retweeted
Blame still flows downhill. We just added a layer that can't be fired and doesn't know it did anything wrong.
Hayley in the Stack retweeted
"AI is a bubble."
Maybe.
The dot-com bubble was a bubble too.
It burst in 2000. Trillions wiped out.
Everyone who had called it a fad took a victory lap.
Then the internet went on to eat every industry on the planet anyway.
A bubble is a statement about stock prices.
It says nothing about the technology.
The prices can crash tomorrow and the model on your laptop will still write code, still read your contract, still teach your kid math.
If you're waiting for the crash so things can go back to the way they were, you're going to wait for the rest of your life.
Hayley in the Stack retweeted
Replying to @ben_sage
Marketing is the new coding.
In a time when anyone can make an app that looks great and functions well, the way you succeed is based on ability to market your brand.
New coders will never know the pain of pulling an all-nighter in school, solving a bug, only to find out it was just from a missing comma. 😭
Hayley in the Stack retweeted
So, what exactly is LLM writing? How is it different from how humans write?
If you look at it closely, it's not just the groups of three, or "it's not X, it's Y"; it's the whole set of characteristics that usually go together.
To study this, I used the same algorithm that Pangram uses to train its classifier. It's called "LLM twins."
It works like this: I downloaded 200 samples of verifiably human writing across all genres, then asked a language model to create a short description of each, and then to fire up separate agents, which had to re-create the pieces from the descriptions alone and make them the same length as the originals.
As a result, I obtained 200 pairs of pieces of writing, each pair sharing a genre, a length, and a rhetorical goal.
For example, a human-written press release about a Boca Raton condominium hosting the filming of "Selling New York," and an LLM-written press release about the same event.
Pangram generates millions of those. I only had 200, but that allowed me to measure the main differences between humans and LLMs.
Here are the main differences:
1. LLMs rarely let two adjacent clauses or sentences share a noun. A human would write, "I stepped into a corridor. The corridor was dark and narrow." An LLM would say, "I stepped into a corridor. Dark. Narrow."
2. LLMs show much more diversity in vocabulary. A human would say "red" or "dark-red." An LLM would write "plum-coloured."
3. LLMs write fewer words per clause. A human would write, "We checked the reported numbers, our investigation found no discrepancies." An LLM would write, "We checked the numbers, no discrepancies found."
4. LLMs write shorter sentences. This is self-explanatory, I believe.
5. LLMs seem to avoid proper nouns (people, places) if they can. A human would write,
"A distraught displaced woman, Hajiya Mairo Abdullahi, was at the weekend seen holding the picture of her seven-year-old child."
An LLM would, among other things, omit the woman's name if it isn't needed for the narrative and doesn't recur later in the text. The LLM version would go something like:
"A displaced woman, struggling with grief, was clutching a photograph of her seven-year-old child."
Even if the person is named, an LLM is more likely to avoid using the name when attributing speech to them. A human would write, "Ms Abdullahi said," while an LLM would write, "The woman said."
6. LLMs also avoid repeated 3-word sequences, whereas humans are okay with them. A human would write "the Kenya cases"; an LLM would rephrase it as "cases from Kenya" or "the cases," or structure the sentence differently to avoid repetition.
7. LLMs operate in a narrower band of word frequencies: more often than humans, they avoid very common words (they are more likely to say "plum-coloured" instead of "red"), but they also avoid rare words (they are unlikely to say "incarnadine" instead of "red").
8. LLMs use fewer numbers. This could be specific to news articles, but LLMs prefer to say "many," "several," or "thousands" when the exact number isn't essential beyond conveying the ballpark. A human would put an actual number.
And finally, an aspect that captures the specific feeling of hollowness that encountering LLM writing brings.
9. Each new sentence in an LLM twin is more likely to introduce a new common noun ("displacement," "war," etc.). But it's less likely to introduce a new, specific, verifiable fact.
A human would write: "Parents want to know if their children are alive and when they will come home. An inquiry has been launched by the Nigerian government..."
An LLM would write: "For the parents, the question is simple. They want to know if their children are alive and when they are coming home."
This is, of course, from studying 200 samples. Pangram has millions, and its neural networks, as unexplainable as this technology is, can probably be contextualised as checking tens of thousands of such "tics."
Importantly, none of these is consequential on its own, but in LLM writing all these features tend to go together. That's why people who work a lot with LLMs can often (but not always) recognise their style.
Hayley in the Stack retweeted
I'm glad we're debating the proper uses of Pangram. Here are my views about it:
1. Awesome tool with a very low false positive rate.
2. Teachers should use it freely and so should journals.
3. It needs subtler categories in many cases, and more of a gradient between them.
4. We need a better public understanding of what a partial score means.
5. Some people use Pangram for apt social shaming, some use it for inapt social shaming, and we need ways to distinguish between its apt and inapt uses.
Hayley in the Stack retweeted
Unpopular opinion:
In 5 years, "made with Al" will mean as little as "made with Photoshop."
Hayley in the Stack retweeted
I regret to inform you all that we have found our first long false positive from 2006
it's a poem where the same 6 words are repeated four thousand times
from 2006, earliest archive is 2013
somehow i don't think you're going to count this one though
web.archive.org/web/20130214…























