mastodon.online is one of the many independent Mastodon servers you can use to participate in the fediverse.
A newer server operated by the Mastodon gGmbH non-profit

Server stats:

11K
active users

#llm

393 posts298 participants19 posts today

I finally understand the appeal of #LLM "#Agents" that's being hyped everywhere after trying out Dagger's example here. https://docs.dagger.io/ai-agents/quickstart

I blindly ran it to see what it does and it was impressive demo.
- It writes code
- Builds it
- Reports error back to the LLM and corrects the code
- build's it again

Does this until the build passes and the final binary is ready.

Just saying "write a curl clone" and letting it do the rest is just altered my perception of what is possible.

While I am aware of the real-world complexities of software, the perception shift that happened in this demo for me is big.

docs.dagger.ioBuild a Coding AI Agent | DaggerEstimated time: 5 minutes

Ars Technica: Why do LLMs make stuff up? New research peers under the hood.. “From a human perspective, it can be hard to understand why these models don’t simply say ‘I don’t know’ instead of making up some plausible-sounding nonsense. Now, new research from Anthropic is exposing at least some of the inner neural network “circuitry” that helps an LLM decide when to take a stab at a (perhaps […]

https://rbfirehose.com/2025/03/29/ars-technica-why-do-llms-make-stuff-up-new-research-peers-under-the-hood/

Mashable: A new AI test is outwitting OpenAI, Google models, among others. “The Arc Prize Foundation, a nonprofit that measures AGI progress, has a new benchmark that is stumping the leading AI models. The test, called ARC-AGI-2 is the second edition ARC-AGI benchmark that tests models on general intelligence by challenging them to solve visual puzzles using pattern recognition, context […]

https://rbfirehose.com/2025/03/29/mashable-a-new-ai-test-is-outwitting-openai-google-models-among-others/

ResearchBuzz: Firehose | Individual posts from ResearchBuzz · Mashable: A new AI test is outwitting OpenAI, Google models, among others | ResearchBuzz: Firehose
More from ResearchBuzz: Firehose

Si ça c'est pas du teasing...

"Une fois n’est pas coutume, vous ne lirez pas sur notre site le même article que celui que nous publions dans la version papier. Dispositif spécial imposé par le sujet: les experts d’une grande variété de domaines scientifiques ont sélectionné 3000 questions pour mettre à l’épreuve les meilleures IA. Mais il ne faut pas que les IA les connaissent: interdiction de les écrire sur Internet"

epsiloon.com/tous-les-numeros/

www.epsiloon.comIA : le quiz ultimeIls appellent ça “Le dernier examen de l’humanité” : les meilleurs experts de leur domaine sont en train de chercher des questions tellement compliquées qu’elles font sécher les meilleurs IA. Nous ...
#IA#Quiz#Science
Continued thread

🧵 …dass die KI Datenklau ist, hatte ich in dieser Tootreihe schon vor über einem Jahr hingewiesen. Nun scheint sogar die Boulevard Medien darauf aufmerksam zu werden.

»Datenklau für die KI – Meta bediente sich heimlich bei Schweizer Literatur:
Stehlen statt kaufen - Um sein KI-Modell Llama 3 zu füttern, bediente sich Meta offenbar illegal bei Millionen von Büchern - darunter Werke von Dürrenmatt und Suter.«

📰 blick.ch/digital/datenklau-fue

Meta klaute Millionen Bücher für KI – auch in der Schweiz
Blick · Zuckerbergs KI-Dilemma: Millionen Bücher geklautStehlen statt kaufen: Um sein KI-Modell Llama 3 zu füttern, bediente sich Meta offenbar illegal bei Millionen von Büchern – darunter Werke von Dürrenmatt und Suter.
#ki#datenklau#daten
Replied in thread

@Catawu @briankrebs I’m not really interested in their frame of reference or what they think about the people impacted. That’s not because I don’t care, but because I think it's irrelevant to the deeper underlying issues.

I’m actually more interested to what extent this situation may violate #HIPAA and other #patientprivacy laws. Part of the functional challenge in what is currently going on at the federal level is that many privacy and #healthcare safeguards such as HIPAA are a complex mixture of laws passed by Congress and regulations defined by the executive branch to implement those laws.

I am not a lawyer, but I do deal with #privacyregulations and #regulatorycompliance issues professionally. To the extent that the administration is arguing that they have constitutional authority to make changes to the implementations developed and overseen by the executive branch itself, the extent of what is being done seems unprecedented but may not be illegal per se. I am not qualified to make that determination, but I think it's the foundational question that needs to be asked.

On the other hand, the parts of HIPAA and other federally-enacted laws regarding #healthcare and privacy are in fact laws established within our country’s constitutional framework. The executive branch can’t simply wish clearly-established laws into the cornfield. Unfortunately, many laws leave a great deal of the implementation details—whether unintentionally or through deliberate delegation—to the executive branch, the states, or various regulatory agencies. In turn, many of those regulators also operate to one extent or another under the executive branch, and that further complicates the picture.

Many federal laws leave a great deal of wiggle room for interpretation to the executive and judicial branches whether not by design, but congressionally-enacted laws and protections provided by the Constitution itself cannot simply be ignored. While there's definitely a difference, separating a "law" from the "regulations" that implement that law isn't necessarily a simple exercise.

The real challenge is that our republic was designed as a Venn diagram of overlapping roles, responsibilities, and authority that were meant to operate in a state of carefully-balanced tension. The republic's framework has never been tested this broadly within my lifetime, if ever. Even though how our three branches of government should work is material covered in any decent highschool civics class, the complexity of statutory vs. regulatory authority requires legal and Constitutional scholarship that is more than the average citizen can bring to bear on the matter. I'd like to think I understand these issues better than most—and I certainly have my own personal and professional instincts about what's right and wrong—but I wouldn't dream of claiming to understand all the nuances involved.

Professionally, I am taking a deliberately apolitical approach to what is a very legitimate set of questions about constitutional authority. Likewise, my apolitical but professional experience tells me that there is entirely too much gray area around the constitutional and legal topics to determine with certainty what is legal as opposed to what is moral or ethical. In my professional experience, what is right and what is lawful aren't always the same.

Unless society as a whole is willing to revisit some of the underlying assumptions collectively made over the past several hundred years about the differences between legislative laws and the administrative regulations that implement them, this problem is unlikely to go away anytime soon. In fact, it is likely to spread to other areas with similar gray areas. As an argument by analogy, the current legal mess around #copyright and #LLM training may be similar in terms of being pure sophistry where the term "fair use" is clearly being used in an intellectually dishonest way, but apparently it's far enough into the gray to pass legal muster right now. Decades or centuries of legislative layering has led to a legal framework that never envisioned modern realities. Revisiting and revising centuries of legal accretion would require a strong moral compass, a great deal of political courage, and in-depth analysis by legal and constitutional scholars (among others) in order to address the very real institutional unraveling we're observing.

Sadly, in a society that frequently classifies expertise as “elitism" such a brutally honest conversation is unlikely to happen soon. A broad reconsideration of how our republic was designed to function and a hard look at how it actually functions would require high levels of both personal and political courage. It's even less likely to be rapidly prioritized without sufficiently clear political self-interest from a majority of those with the remaining authority to materially affect the outcome.

What I’ve said may strike some as political opinion rather than strictly analytical observation. However, my statements are deliberately based on well-established sociological and psychological norms rather than current politics. I feel confident in asserting that the likelihood of Congress or the Supreme Court—much less the general public—addressing these things effectively in the near term is essentially zero. For any elected or appointed official acting alone, the risk of asserting constitutional prerogatives vastly exceeds both the collective will of their respective institutions and the already-ceded institutional powers required to do so effectively.

AI sovereignty is just one part of the desire for broader “tech sovereignty” that’s also been gaining steam, growing out of more sweeping concerns about the privacy and security of data transferred to the United States

#EU #Europe #tech #technology #AI #software #politics #uspol #trump #data #datascience #democracy #usa #llm #generativeAI

technologyreview.com/2025/03/2

MIT Technology Review · Why the world is looking to ditch US AI modelsBy Eileen Guo
Continued thread

‘the protective “messiness” in some jobs comes from the back-and-forth and unpredictability inherent in interacting with other people. There is a certain irony in the realisation that the mantra of rugged self-reliance and workflow optimisation common in Silicon Valley may have made tech roles more, not less fragile.’ #Writers #software #developers #LLM #AI

Tab auto-fill on #Linux which has been around for ages, has done way more for me than any #AI or #LLM ever has. And tab autofill is purely algorithmic. There's nothing more powerful than hitting the tab key and going "you know the one I mean!"