But in any case, what this tells me, is that either of 2 things happen soon:
1. this whole "AI will do your white-collar job" thing dies out…
*or*
2. MS Office goes extinct, replaced by tools that can actually use LLM Agents.
But in any case, what this tells me, is that either of 2 things happen soon:
1. this whole "AI will do your white-collar job" thing dies out…
*or*
2. MS Office goes extinct, replaced by tools that can actually use LLM Agents.
But ask it to work with .doc and .xls, and watch as it flounders trying to use half-baked Python packages to parse/write these formats… 🤷
")…
Except that to do that remotely well, they need to go through a layer of deterministic scripting: they aren't replacing the intern, they're replacing the one-use 100 line script you might write…
It took massive community efforts (or large corps like Google) to reverse-engineer enough of it, to provide semi-passable conversion tools.
These formats were explicitly designed *NOT* to be easily modified by external scripts: use MS' own exploit-riddled scripting, or get lost.
Long before LLMs, MS Office formats were already notorious for being a jumbled underdocumented (often proprietary) mess, that barely worked with its own designated app, and had zero interoperability whatsoever.
xls…) are the *polar* opposites of what works well with LLMs…
https://mastodon.social/@deivudesu/116205894730402456
Translation: "It's not AI if there was a human somewhere that clicked on the 'vibecode this for me' button". 🤡
https://www.cnbc.com/2026/03/10/amazon-plans-deep-dive-internal-meeting-address-ai-related-outages.html
#ai #vibecoding #amazon #aws
I would have literally gone faster writing the full python script myself. 🤡
My point here is not to say that LLM agents can't do that sort of stuff well (they can).
But clearly the version being sold to corps: absolutely cannot. 😆
… Copilot had proudly produced the most minimal, badly-formatted, unusable garbled mess of a doc.
And this after I manually broke down the steps, made it go through formats it could handle (csv, md) and babysat it through every step of the process.
The task was a poster-child for what a half-baked LLM agent should do in its sleep: take legal paragraphs stupidly written in xls cells, and extract them to a word doc with header structure…
50 mins and a lot of handholding later… 🥁
2. can't really persist any document or anything by itself: you download its output, then re-upload it and ask it to do something with it.
3. will regularly flush *all* its context with no attempt at summarising, to the point where it will lose the file pointer you started the convo with.
... 🧵
(Needless to say: absolutely no other tools are permissible. In typical finance corp fashion, you can barely even copy-paste paragraphs out of your cloud documents)
Knowing I do occasionally make these things work for me, she asked me for a hand today………
2/ 🧵
Speaking as someone who uses LLM agents on a daily basis: if this is what the average corporate drone is being forced to use as "the future of AI-assisted work"… This AI bubble is going to explode in a billion pieces any day now, to everyone's delight.
→ 🧵
<2006> that whole `seize the means of production` marxist class warfare schtick is starting to sound dated and a bit silly tbh…
<2026> Hi there! A dozen corporations own all AI infrastructure in the world, and are hard at work convincing everyone else that the only job left is AI prompt operator.
I'd say that's a new low, and ought to be a wake up call, but I don't think the vast majority of maga voters has even heard of my lai…
RE: https://mstdn.social/@Bellingcat/116195203740479781
Just the US, caught trying to hide the obvious fact that they murdered a 100 schoolgirls as part of an unprovoked attack on a sovereign country…
https://fed.brid.gy/r/https://bsky.app/profile/did:plc:b2p3lxhgjnvlfuweycqv4tjn/post/3mg3qsr253a2j
PS: To me it’s really a funny little anecdote that was very much a side note to my life at the time.
Someone who’d be entitled to be a lot saltier about it (yet isn’t, which is a testament to how cool a bloke he is), is Mike White:
But discovering completely by chance, that WP uses comment `karma` internally (“we reimplemented our own proprietary spam filtering algorithm from scratch!”)…
ngl: it does feel like a pretty strong middle finger from way back. 😂
/🧵
Not to worry, though: it's completely free for most use. You just need to pay if you are a big company with lots of $$$ to spend on comment spam.
And this is the point where Matt started blurring the line between the OSS project and his own company
8/ 🧵
Then after a few months, Matt unveils a grand new Wordpress update: *Akismet*, WP's very own built-in Spam Filter! 🎉
7/ 🧵
I, and many others, gently suggest on the #wordpress IRC that maybe Spam Karma should be moved into WP Core…
6/ 🧵
Things were apparently so dire, that I got an email from Anil, suggesting creating an "Anti-Spam consortium". Enthusiastically replied. Never heard back. 🤡
(to be fair, this coincided with some unrelated drama around MT's commercial status and an ill-fated attempt at monetisation)
5/ 🧵
It's simple, but also very efficient, and much harder to work around by spammers.
Pretty quickly, it becomes the standard anti-spam option for Wordpress: a large chunk of the entire WP ecosystem runs it.
4/ 🧵
So I come up with a pretty basic spam filter plugin:
3/🧵
Within a few weeks, all blogs are under a constant barrage of spam.
2/🧵
#wordpress
Which is a funny reminder of that very precise point in time when Matt hijacked an OSS project to turn it into his private $ billion company, and the small role I played in that…
Gather round, web younglings,] Story Time: 🧵
#foss #wordpress
While doing some housekeeping around my Wordpress blog, I randomly learnt that WP stores a comment "karma" score in its Database model.
https://wordpress.stackexchange.com/questions/16676/sort-comments-by-karma
I know there are plenty of ways around these things (for tech people), but I'd happily spend time making it dirt-easy for everyone to bypass these surveillance state enablers.
Is there any open-source effort out there to hack video "age check" solutions? The kind that wants you to send a copy of your ID, credit card, or "just" smile at the camera in order to satisfy some idiotic legal requirement (and hand-on-their-heart will never keep that data anywhere).