Recent searches

No recent searches

Search options

Not available on photog.social.
photog.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
A place for your photos and banter. Photog first is our motto Please refer to the site rules before posting.

Administered by:

Server stats:

246
active users

photog.social: AboutProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.4.3

#llm

86 posts81 participants3 posts today
Reilly Spitzfaden (they/them)

“Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels.”

One of those studies where the result seems self-evident, but it's nice to have research demonstrating it for when people try to argue that AI Is Great Actually.

arxiv.org/abs/2506.08872

Definitely going in my regularly-updated list of AI-critical writing and research: reillyspitzfaden.com/wiki/read

arXiv logo
arXiv.orgYour Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing TaskThis study explores the neural and behavioral consequences of LLM-assisted essay writing. Participants were divided into three groups: LLM, Search Engine, and Brain-only (no tools). Each completed three sessions under the same condition. In a fourth session, LLM users were reassigned to Brain-only group (LLM-to-Brain), and Brain-only users were reassigned to LLM condition (Brain-to-LLM). A total of 54 participants took part in Sessions 1-3, with 18 completing session 4. We used electroencephalography (EEG) to assess cognitive load during essay writing, and analyzed essays using NLP, as well as scoring essays with the help from human teachers and an AI judge. Across groups, NERs, n-gram patterns, and topic ontology showed within-group homogeneity. EEG revealed significant differences in brain connectivity: Brain-only participants exhibited the strongest, most distributed networks; Search Engine users showed moderate engagement; and LLM users displayed the weakest connectivity. Cognitive activity scaled down in relation to external tool use. In session 4, LLM-to-Brain participants showed reduced alpha and beta connectivity, indicating under-engagement. Brain-to-LLM users exhibited higher memory recall and activation of occipito-parietal and prefrontal areas, similar to Search Engine users. Self-reported ownership of essays was the lowest in the LLM group and the highest in the Brain-only group. LLM users also struggled to accurately quote their own work. While LLMs offer immediate convenience, our findings highlight potential cognitive costs. Over four months, LLM users consistently underperformed at neural, linguistic, and behavioral levels. These results raise concerns about the long-term educational implications of LLM reliance and underscore the need for deeper inquiry into AI's role in learning.
#AI#GenAI#LLM
Maj - 🇨🇦

I like that spellcheck now doesn't give a flying rats ass through a doughnut about whether a word is in the dictionary or not. It is just a possible combination of letters for an English word (side note: it is a word in Tagalog!).

I moved away from computation linguistics in the mid 90s because I didn't like how hard it leaned into statistical analysis vs semantic and I am still / more mad now. Furious as a colorless green idea.

#llm#generativeAI#spelling
Veronica Olsen 🏳️‍🌈🇳🇴🌻

"A lifetime of hearing humans explain their actions and thought processes has led us to believe that these kinds of written explanations must have some level of self-knowledge behind them. That's just not true with LLMs that are merely mimicking those kinds of text patterns to guess at their own capabilities and flaws."

#AI #LLM

arstechnica.com/ai/2025/08/why

The Thinker by Auguste Rodin - stock photo
Ars Technica · Why it’s a mistake to ask chatbots about their mistakesBy Benj Edwards
Yogthos

Open source 4b model Jan v1 delivers 91% SimpleQA accuracy, slightly outperforming Perplexity Pro while running fully locally.

Love to see open source approach continue to trounce proprietary models.

arxiv.org/abs/2508.00360

arXiv logo
arXiv.orgLucy: edgerunning agentic web search on mobile with machine generated task vectorsSmall language models (SLMs) are inherently limited in knowledge-intensive tasks due to their constrained capacity. While test-time computation offers a path to enhanced performance, most approaches treat reasoning as a fixed or heuristic process. In this work, we propose a new paradigm: viewing the model's internal reasoning, delimited by <think> and </think> tags, as a dynamic task vector machine. Rather than treating the content inside these tags as a mere trace of thought, we interpret the generation process itself as a mechanism through which the model \textbf{constructs and refines its own task vectors} on the fly. We developed a method to optimize this dynamic task vector machine through RLVR and successfully trained an agentic web-search model. We present Lucy, a 1.7B-parameter SLM that leverages this dynamic reasoning mechanism with MCP integration to achieve 78.3% accuracy on the SimpleQA benchmark, performing on par with much larger models such as DeepSeek-V3. This demonstrates that small models can rival large ones when equipped with structured, self-constructed task reasoning.
#llm#technology
Replied in thread
Cody Boone Ferguson

@madeindex the fact a blocker of ‘#AI ‘ exists shows just how rampant the scam is and it is astonishing how bad. Companies now think if you add AI to something it is somehow better even though at best it’s a mediocre #LLM (and often much worse and even deadly like Google’s AI Overview). The truth is if you have to embellish your product it’s not as good as you think. #AppleIntelligence is incredibly bad but #Apple has built products on it. Probably its beta won’t even leave.

Continued thread
*
Amy Dentata

LLMs will never be what they're currently advertised as. They are not answer machines. They are not search engines. They do not think. They do not provide data.

All they do is generate text via statistical weights. They do not generate thoughts, or information. Just characters in sequence. They cannot be stopped from "hallucinating." Because they do not "hallucinate." They just generate text. Whether that text reflects reality or not is only ever due to coincidence.

These companies won't tell you this, because they want your money. So they lie, which primes you to accept more BS from the automated BS machine.

#OpenAI#LLM#ChatGPT
*
Amy Dentata

Every story where someone is driven mad by ChatGPT begins with lies about the technology, what it is, and what it does—lies that originate from OpenAI itself.

The very framing of "having a conversation with a computer" is artificial, deceitful, and constructed by the company in order to turn statistical text generation into a consumer-facing product. Deceit lies at the core of what OpenAI does with this technology. The people being "fooled" by ChatGPT are first fooled by OpenAI. People are being actively coached—by real, live human beings—into being easily manipulated by this software.

#OpenAI#LLM#ChatGPT
Misinformation-Superhighwayman

GitHub will be folded into Microsoft proper as CEO steps down
arstechnica.com/gadgets/2025/0

Today's news about GitHub should not come as a big surprise. The writing was on the wall in 2018 when Microsoft bought GitHub for $7.5 Billion. Microsoft is asserting full control over a company they've consumed, and they're proceeding to run it into the ground. And, as with practically every Tech company these days, the motivation is to build hype about integration with AI/LLM technology.

What I find surprising is the timing of GitHub CEO Thomas Dohmke's departure — it was not part of an immediate post-takeover clearout, nor after a transitional period that's traditionally one or two years, nor did he settle-in comfortably until the end of his career.

Concerns that GitHub is progressing along the path to enshittification are not unfounded. Microsoft has a habit of this — look at what they did to Hotmail, Visio, Nokia, Skype, and more (LinkedIn is, IMHO, an exception to this pattern; it came pre-enshittified).

I'm tipping that most companies currently using GitHub will continue to use it. The news isn't serious enough for them to put themselves through the discomfort of migrating to a different repository platform, or establishing and committing to maintain their own in-house system. Small businesses, community projects, students, hobby coders, and some Open Source Software projects might be inspired to migrate? Maybe? Startups might think a bit harder about whether to go with GitHub or one of the alternatives?

I think the real lesson here is that any platform you get into needs to have a smooth path to migrate in your content, and to migrate out as a safe and easy exit strategy. In the meantime, the importance of interoperability is as strong as ever. The same goes for not getting locked into an ecosystem that looks convenient but turns out to be a "Walled Garden". That, and the warning that if a product or service is presented to you as "Free" then it's likely that you and your work are the product.

The GitHub logo.
Ars Technica · GitHub will be folded into Microsoft proper as CEO steps downBy Andrew Cunningham
#GitHub#Microsoft#GitHubCopilot
Mix Mistress Alice💄

“I do wonder what that does if you have this sycophantic, compliant [bot] who never disagrees with you, [is] never bored, never tired, always happy to endlessly listen to your problems, always subservient, [and] cannot refuse consent. What does that do to the way we interact with other humans, especially for a new generation of people who are going to be socialised with this technology?”—Dr Raphaël Millière

AI chatbots are becoming popular alternatives to therapy. But they may worsen mental health crises, experts warn >

theguardian.com/australia-news

The Guardian · AI chatbots are becoming popular alternatives to therapy. But they may worsen mental health crises, experts warnBy Josh Taylor
#chatbots#therapy#mental_wellbeing
Michael :donor:

Apparently GitHub will become part of Microsoft's "AI Team" [1]. Even reading the CEO's post about their departure reads like a fever dream. [2]

«I am more convinced than ever that the world will soon see one billion developers enabled by billions of AI agents»

I am not sure if it is even possible to have more mentions of "AI" in a single post. Those sillicon valley people are really out of their minds.

Thankfully, there's @forgejo and @Codeberg

[1]: arstechnica.com/gadgets/2025/0
[2]: github.blog/news-insights/comp

The GitHub logo.
Ars Technica · GitHub will be folded into Microsoft proper as CEO steps downBy Andrew Cunningham
#github#microsoft#enshittification
🌱🏴‍🅰️🏳️‍⚧️🐧📎 Ambiyelp

#Mastodon is beginning to see #llm chatbots that write plausible sounding sign up applications and mix advertising with generated content to launder authenticity. The default mode of operation here is deception, the more they can convince you thats a real person the better their #Advertising works presumably. We need a #Privacy preserving solution to human verification. Right now images involving long swathes of text continue to be hard to generate, maybe thats a good place to start #AI

Fiodar Sazanavets

Semantic Kernel can be used to effortlessly integrate LLMs with .NET apps. I just wrote a post showing you how. And it's only the first post of the series. Next time, I will show you how to integrate it with .NET Aspire to build a fully functioning clone of ChatGPT hosted in the browser.

#llm #dotnet

fiodar.substack.com/p/building

Fiodar’s Tech Insights · Building a ChatGPT clone in minutes with Semantic Kernel and OllamaBy Fiodar Sazanavets
*
Steve Dustcircle 🌹

"I tested #GPT5's #coding skills, and it was so bad that I'm sticking with #GPT4o (for now)"

According to my #LLM testing, #ChatGPT has been the gold standard of #programming tools, but GPT-5 stumbled badly, delivering broken #plugins, flawed #scripts, and errors that could derail projects without careful human oversight.

zdnet.com/article/i-tested-gpt

ZDNET · I tested GPT-5's coding skills, and it was so bad that I'm sticking with GPT-4o (for now)By David Gewirtz
Replied in thread
Uwe Sinha

@NicolasSustr Sehr viel Liebe für diese Analogie:

"Das System arbeitet mit einem Large-Language-Modell. Letztlich bedeutet das, dass es auf Basis eines riesigen Bergs an Beispielsätzen eine Wahrscheinlichkeit berechnet, was die menschliche Person eigentlich wissen will. Letztlich hat das etwas von Dart spielen, wo auch ein Profi häufiger mal das Ziel verfehlt."
#LLM

Live feeds
About