iiinova — where your digital self keeps evolving

Nicky (They/Them) @sickynicky.bsky.social · 18h

Not much, what is allosexual with you?

Reposted by Nicky (They/Them)

Seattle Bike Blog @seattlebikeblog.com · 23h

These streets have never worked for area residents. In 1954, a group of Pinehurst dads went out to 5th Ave NE with garden tools and created a path between 125th and 130th so their kids could walk to school safe from traffic.

Newsclip image of a group of men digging in ferns and other plants next to a road. The caption calls them "safety engineers."

0132

Nicky (They/Them) @sickynicky.bsky.social · 18h

Ok but you also repost a ton of pictures of delicious hot deli sandwiches I only get around to seeing when i’m already in bed so…

010

Nicky (They/Them) @sickynicky.bsky.social · 18h

You tricked me into reading this for chuckles and now I want one too. This should be illegal. It’s like eldritch knowledge!

010

Reposted by Nicky (They/Them)

Crow's Nest Comics @crowsnest.quest · 05/10/2026

We've been working with Books to Prisoners Seattle to help get graphic novels into prisons for years and a generous donor has committed to matching all book donations this week in honor of banned book week!!! Go to shop.crowsnest.quest/a/listify/re... to get them some comics and double your impact.

red background, white text "books to prisoners donation match", angry black crow flying over the text as if descending to attack

16548

Nicky (They/Them) @sickynicky.bsky.social · 18h

Cannot recommend enough. Brain fog during my commute hours robs me of the chance to listen to most podcasts but when I am feeling well enough on the train, I can’t think of an episode I ever was anything but greatly enthusiastic about.

110

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Ah, got it now. You're very intentionally underselling the reality of how bad the enshittification is to make your point seem stronger. Good luck to you. I'm no longer interested in a discussion with you.

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Gonna need a evidenced support on the claim we can't return to functional technology we previously had through legislative market control.

000

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Unsure the relevance here. I'm not contending it's a language model. I could contend it's not the pivot point for LLM developments but that's not relevant to discussion either.

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

The other is worse than the product offered at google when it only offered one. We don't need either current products, despite the currently more functional one having better desired metrics. We just need to not have industry collusion to force market capture.

000

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Google's specific enshittification began rollout in mid to late 2018, BERT rollout to production was late 2019. Regardless of rollout timing (and which products were started when for repealing don't be evil) there are definitively two products even within google being offered. One is more useable.

200

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

> > Web searches* are not simply web crawlers and websites. I didn't assert this. Web search enshittification and web crawlers empowered by LLMs tiering information are separatable issues. I recognize they are complex systems and one is a component of another.

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Ah! I appreciate that clarification.

000

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

And for clarity, I am no where as knowledgable as either of these actual experts, nor immune to falling into the same disinformation traps even in my opinion about them. Writing is fucking hard and I recognize both are where they are because they are way above my skills in many areas.

000

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Professor Shepard has elsewhere shared he thinks he could have been more clear or effective in his messaging but I think Professor Bender's claim of it being weird is more about the goal of the point. Fundamentally, this conversation needs case by case context to engage in productively.

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

We can both admit it's really hard to intentionally avoid these products in purist sense and that it's important to try without reducing to a futilist sense. Aside from how it landed in the medium, we strip nuace from these conversations 300 characters at a time. I doubt many claim either extreme.

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

Word! I also read it as kind of a bizzaro counter to the attached comic, which I do think has merit in the conversation. I do agree with Professor Bender as I think this vague posting can apply "hit dogs holler" to both the response and the claim. Maybe not helpful/relevant, even aside the reality

4 panel comic meme of a derogatory critic of critics
of systems that are users or participants of said system.

Ends with a peasant stating "We should improve society somewhat" and the strawman character emergening for a well with gleeful look on their face replying "yet you participate in society. Curious! I am very intelligent."

100

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

I would argue we as academics or followers of academia can give grace here. Feedback for sure but not suspicion. The AI booster language is intentionally manipulative to make discourse hard. Also, microblogging is just weird and people publish things that maybe need another draft often.

120

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

There's also a slippery slope of feeding into the intentional misnomer of the "AI" products as being indivisible. A web crawler and image generator stuck to the end of bad chatbot on a stick is definitely dividable. We have had menu inputs for the first two for a long time. We can go back.

010

Nicky (They/Them) @sickynicky.bsky.social · 05/10/2026

The relevance, I'd argue. Machine learning is decades old but the compounding issue of neural networks and LLMs with web crawlers and information tiering wasn't as obvious to the average power user before enshittification of web searches. The break through happened after things got worse.

220

Reposted by Nicky (They/Them)

Matunos @matunos.bsky.social · 04/10/2026

Four-panel comic by The Perry Bible Fellowship (pbfcomics.com).

Panel 1: Two bald, featureless figures face each other in matching yellow T-shirts. One shirt reads "PRO SKUB," the other "ANTI SKUB." They eye each other warily.

Panel 2: Pro Skub scowls and clenches his fist as Anti Skub lunges into him. The two collide and start grappling.

Panel 3: The fight has erupted into a brawl. Pro Skub and Anti Skub trade blows while other people in colored shirts join in, punching and shoving.

Panel 4: The object of the fight: a small pink jar labeled "SKUB," its lid off, holding a pale yellow substance. It's an ordinary, unexplained product.

151

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

For summary: We can communicate in low language intersections. We know if LLMs train on bad data, output goals collapse. Those goal standards have bias to collapse. If your language, dialect, or vernacular is different than training data, what constitutes good or bad data for your standards?

000

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Linguistics standards on transcription are glacial slow but they're important on data set generation. We just can't scale to what's needed. Rather than passing validation of proxy, it's up to personal writing styles and the biases that govern that skill outside of performed language social norms.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

We're using a proxy of a proxy in personally transcribed written language. Microblogging is a great example of how nobody talks or signs like this. It enflames and rewards argumentative and performative outcomes 300 characters at a time (or whatever). Most social media fits here by extension.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

The hope amongst computational linguists is we can increase the quality and quantity of data available to better reflect and query these tools for their intended purpose: insight on statistical trends within language production. My personal research is tied to this broad goal.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

This supports the claim against these programmatic tools somehow having a proxy for reasoning. Humans are also capable of being static and context insensitive but it is inherent to these tools design. Adding webcrawlers and powerful search function access doesn't undo this.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

There's some fair pushback in that. To engage here we need to measure degrees of difference between the trend of human decisions that punish dialect and vernacular bias vs what LLMs are amplifying. But it is detectable and there is significant support in this area. Also, automating that bias is bad

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Extending edge cases of writing with little to no representation in training data, these issues compound. Machine learning is pretty old and we have decades of showing how these statistical tools repeatedly amplify marginalization but we don't need to look beyond LLMs to see impacts.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Using metrics of synthetic use cases where often human participants drop out, accuracy shows things are possible for LLM use in medical contexts. In live contexts using measure to check for false positives and negatives, it's not even close. Appropriate where diagnosis could be false or missed.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Where training includes contextual language that fits well within proxy of the world (Basically anything with large amounts of standardized written digitalized data), it can synthesis said trends likely comparable within accuracy/possibility of comparable human output.

pmc.ncbi.nlm.nih.gov

Reliability of LLMs as medical assistants for the general public: a randomized preregistered study

Global healthcare providers are exploring the use of large language models (LLMs) to provide medical advice to the public. LLMs now achieve nearly perfect scores on medical licensing exams, but this d...

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Inherent to success matching data fit of the trend, application to edge cases in training data has poorer outcomes. This makes LLMs inherently context insensitive on individual levels, hence the "check all answers" included in all major products.

200

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Got it! The stochastic part of the analogy is pretty vital here. These are statistically well tested tools executed using exhaustive programming techniques across incredible scales (relative to our understanding). The outputs are probabilistically arrived via trends of linguistic data.

200

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

For clarity, when you say original question are you talking about stochastic parrots, language's ability to serve as proxy for the world, or how LLMs perform worse at contextual exercises with minority and marginalized speakers?

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Here's another more recentish piece but using simple accuracy across queries measured across subject domains for artificially generated AAVE paired with their source in SAE. arxiv.org/html/2503.04...

arxiv.org

Disparities in LLM Reasoning Accuracy and Explanations: A Case Study on African American English

000

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Here's a reasonable compromise in being modern but testing for false positives and negatives. Unsure if the synthetic generation is as strong a case but the AAVE is transcription. Discrimination in housing applicants. arxiv.org/html/2609.18...

arxiv.org

From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale

200

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

These data structures contain linguistic data, even when applied over cardinality or geographic information. The structures shapes may be interpreted as parallels to other information structures that also conveniently follow the rules of syntax and base use of semantics. Hard to prove when muddy.

000

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Sure, which is where the probability of studies speak louder to their respective moments in the moving target. Fundamentally, we have a breakthrough in lingusitics with these tools that are letting us dispell a lot of old theories. And we can query/probe these data structures as complex as they are.

100

Nicky (They/Them) @sickynicky.bsky.social · 04/10/2026

Word! I can get you some before the end of my day. Though to clarify, most of what I'd share would be: - Historical/Not as recent as these papers. - Measured in possible and/or probable manners including accuracy spikes vs precision, recall, etc.. - Respective to authors' nuanced questions.

100

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

Outside of LLM contexts, pragmatics has amazing tools to demonstrate the limits of seeing the world through the lens of language. Likewise, psycholingustics and neuro-psychology, particulary in context of childhood development, disability, and multilingual contexts.

000

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

Contextual language is something LLMs dramatically under perform at, especially notably in marginalized dialects and vernaculars. This makes sense as their mechanism is probablistic trends in language through exhaustive programming, not individual context senstive dialogue.

200

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

The above studies could imply semantic quality may not be a factor at all in their training at all. Semantics sits a close barrier in linguistic subfields between quantifiable data and qualifiable. Notably what sits closer to qualifable is pragmatics/contextual intersection which transcription fails

200

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

Language IS the world in the context LLMs is an intentionally narrow arguement as LLMs don't actually use the whole of performed language in training it uses written transcription. Furthermore, it uses these transcriptions in quantifiable ways. There's some promising stuff with ontology but a gap.

100

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

A common criticism of the simplicity of stochastic parrots analogy is the parroting/linguistic mimicking. There's two general approaches against it: LLMs have moved beyond language and now model the world; language IS the world. Unrelated: Attacking stochastic is another weirder idea (IMO).

111

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

Fair enough! I'm trying to be careful here because quantitatively, how we track and measure language has a lot of overlap. Information ontology has a lot of developments around this. I don't want to say "has no semantic content" in this case but it's semantically detached from how language works.

000

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

In short, they use context-free grammar (CFG) algorithmic generation for semantically empty content. This does find high probability matching against linguistic models for generation. However, the claim this reinforces a world model only ignores the linguistic mechanism it inherently uses: syntax.

000

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

Corrupting and partial randomizing syntax and interrupting language as a proxy for the world yielded similarity in shorter generations but increasingly fell apart beyond. Their topic is more contrary to generative grammar in scope because of Chomsky's complaint which doesn't hold true.

200

Nicky (They/Them) @sickynicky.bsky.social · 03/10/2026

I appreciate the response and clarification! Are you at all interested in conclusions from studies for LLMs dependency on linguistic features? I have two which are pretty applicable.

100