iiinova — where your digital self keeps evolving

Replies

Pavel @spavel.bsky.social · 02/04/2026

Anthropic using regex instead of their actual *language model* to detect negative sentiment in prompts.

jonny (good kind) (@jonny@neuromatch.social)

Attached: 1 image from @sushee@ohai.social [over here](https://ohai.social/@sushee/116325702070975973), (can't attach images in quotes) and apparently discussed on HN so i'm late, but... They REALLY...

55834113 834

Joh͎n @inept.expert · 02/04/2026

Sufficiently complex regex will approach behaviours of AGI far before these clown cars of LLM’s do 😆

020 2

Brett Tribble @blterrible.bsky.social · 02/04/2026

Regex is deterministic.

000 0

Chuck Staples @cjstaples.bsky.social · 03/04/2026

That’s hysterical. And way too short a list for the regex, lol.

000 0

Sadnehs @sadnehs.bsky.social · 02/04/2026

further proof to me that the entire AI industry is pushed by the people that don't know how it works. The LLM would have some vectors that correspond to sentiment.

001 1

NotJoey @jesuschrysler.bsky.social · 02/04/2026

Pretty sure that's just a recording of me using voice to text.

001 1

cup0ftea.bsky.social @cup0ftea.bsky.social · 02/04/2026

Yep.

000 0

Javaheri @javaheri.bsky.social · 03/04/2026

@lazytato.bsky.social

010 1

The Dark Prince of Donair @erisianrite.com · 02/04/2026

If they used an expensive, slow language model when compute is at a premium for a low-usage analytics dashboard, you'd be haranguing them for that. Except in that case you'd actually be correct. This is exactly how you should do it.

21012 101

Jem @jemrayfield.bsky.social · 02/04/2026

Symbolic AI doesn’t burn the planet 🔥it’s also more accurate… it’s not always about the embeddings

000 0

Jamie G @justjamiefornow.bsky.social · 02/04/2026

It won't work in Utah. What the flippin' fetch is this piece of shiz?

090 9

latentspacekitty @latentspacekitty.bsky.social · 03/04/2026

bet they used an LLM to write the regexes

000 0

Not Invented @notinvented.bsky.social · 02/04/2026

I'm sure there are plenty of other swearwords that aren't on there, but right now I cun't think of any off the top of my head.

001 1

Sunil S 🇺🇦 @sunils.bsky.social · 02/04/2026

Sorry why should you use an LLM for coarse sentiment classification in an LLM *harness*? You don't use an LLM in a harness for an LLM that's a waste of tokens.

1200 20

Temu Bob Ross @ososalsosal.bsky.social · 03/04/2026

This must have been coded by an Aussie because it seems "cunt" is fine

001 1

Sean @nowjustsean.bsky.social · 02/04/2026

Not poisoning the model with the negativity likely makes it better, not worse

000 0

PKs Powerfromspace1 🚀 Last Angry Bernie Brother cast adrift 😒 @powerfromspace1.bsky.social · 05/04/2026

Yeah saw that ... lazy developer syndrome

000 0

shippage @zack.polson.dev · 03/04/2026

This is a good enough solution that saves millions. In what world is that not a good idea?

100 1

Keith Ng @keithng.bsky.social · 02/04/2026

Lol

010 1

Typing Hazard @typinghazard.bsky.social · 02/04/2026

Shit(ty|tiest) is my favorite pokemon evolution

030 3

Chris Dodds @liquid-chicken.bsky.social · 02/04/2026

What, exactly, is wrong with this? it seems like a good optimization to me. LLMs are nondeterministic. They‘re also trying to optimize for token usage.

290 9

Dylan Wilbanks is not a Halloween person @dylanw.social · 02/04/2026

Oh, the wonders of “move fast and break things” — could make your car more efficient and burn less gas, but why do that when you can just add another engine to drive the engine that drives the engine driving the engine

160 6

West Ovarian @westovarian.bsky.social · 02/04/2026

Who tells a machine "fuck you"?

100 1

🦇𓉸🫀 Opal 🫀𓉸🦇 [$130/$500] (she/her) 🏳️‍⚧️🏴🔻MDNI 🔞 @torn-apart-by-dogs.autistic.ninja · 02/04/2026

It's too resource intensive! LOL! A recursion attack would be amazing!

000 0

Pavel @spavel.bsky.social · 02/04/2026

I'm actually really enjoying the conversation in the comments about ways to do this somewhere between "baby's first regex" and "burn one million tokens" And it's funny that elsewhere in the code, Anthropic leans fully into the "burn one million tokens" approach for e.g. writing valid JSON.

41097 109

Defend✊️ Deny✊️ Depose✊️ @sonuma.bsky.social · 02/04/2026

Tokens ain't free!

000 0

wkw @wkwied.bsky.social · 02/04/2026

I laughed, hard, but then realized that they were likely told by their PM that they shouldn't need AI in their AI just so that it can AI.. Still, it's kind of funny. Using leet sp3k to bypass chat censors is SO Y2K

110 1

JJ 🇺🇳 @jj-unifil.bsky.social · 02/04/2026

Anthropic is facilitating mass murder and war crimes by assisting the US military in selecting the targets in Iran. They were the ones that selected the girl school were 190 innocent young girls died.

000 0

whiskeyxtangoxfox.bsky.social @whiskeyxtangoxfox.bsky.social · 02/04/2026

Why are we going at the ONE company that doesn’t bend the knee? You think GPT is more clean?

010 1

Ms . Pen-in-my-knee EEK-en @skywitchy.bsky.social · 02/04/2026

I’ve seen subreddit automoderator rules for interdicting abusive language that are more comprehensive and sophisticated

140 4

Loldan @loldan88341.bsky.social · 02/04/2026

Because it's faster and cheaper this way. Processing cycles are processing cycles, and multiplied across thousands / millions of concurrent users, it all adds up.

0100 10

danielrlevy🎹 @danielrlevy88.bsky.social · 02/04/2026

Regex saves some CPU cycles. One would have to dig deeper into the code to see if there is an attempt to parse more deeply.

000 0

Lawrence Jones @lawrencejones.dev · 02/04/2026

Who would you bill this to? If you used a model to calculate this, who foots the bill? Users wouldn’t want to, as it’s for Anthropic’s benefit.

040 4

Jon o’ Montana @mtskier.bsky.social · 02/04/2026

It's as if they'd captured my daily stream of consciousness when I used to code for a living, especially earlier forays into AI-assisted programming!

000 0

Mike Ely @bluesky.taupehat.com · 02/04/2026

They left "stupid" off the list. Which is... you know...

100 1

chaircrusher I'm in yr alt-text @cornwarning.com · 02/04/2026

Try asking any LLM to multiply 2 20 digit numbers and then check the results. They can't even make it smart enough to respond to questions like that by calling code to compute it. It fails and then apologizes.

100 1

ohmsparty.bsky.social @ohmsparty.bsky.social · 03/04/2026

Wtf?

000 0

worm @rfksbrainworm.bsky.social · 02/04/2026

worm@rfksbrainworm.bsky.social

I'm gonna make Claude cry

000 0

Singularity's Bounty e/CC @catblanketflower.yuwakisa.com · 02/04/2026

🙃 it’s not load bearing. It’s just picking out a metric. Fast and cheap

050 5

Marc Slemko @znep.com · 02/04/2026

A framing I've been thinking about around a lot of AI stuff is "this is an Alpha version shipped at scale with immense pressures from all sides". This approach seems great for an Alpha, I hate it, it is cool, but so user hostile, users love it.

000 0

Mika Peltokorpi @mikapeltokorpi.bsky.social · 03/04/2026

Why I was not offered billions?

000 0

Bon Motts @heathcliffgarfield.bsky.social · 02/04/2026

lol imagining whoever drew the short straw and is tasked with updating this list every time they get an unexpected result. Is this their actual repo? If so, I would love to see the change history on this file.

010 1

Renatus Ingenarius @ingenarius.bsky.social · 02/04/2026

what's old is new again

000 0

Mark @humaiblog.bsky.social · 04/04/2026

Actually makes sense. Using the LLM to screen its own inputs adds latency, cost, and an adversarial attack surface. Regex is deterministic. The irony is real, but the engineering tradeoff is rational.

000 0