Notices by Jed Brown (jedbrown@hachyderm.io)

@inthehands It's important to note that search indexing is considered "transformative" and thus fair use *because* it does not supplant the market for the original content. That goes out the window when the product functions to capture traffic that would otherwise go to the cites. They are acting with impunity, but existing copyright law addresses this if courts find it to be not transformative.

@thomasfuchs And the new image is a semver violation.
https://hachyderm.io/@jedbrown/116086730296716484

@thomasfuchs I can relate to this sentiment, but the paper refutes that "perceived ethicality" (or "perceived capability") explains the discrepancy.

https://doi.org/10.1177/00222429251314491

@inthehands @moritz_negwer
They see it as a matter of pride to not learn or know anything.
https://cluely.com/manifesto

@inthehands If one believes the bubble has to pop, what does this say about effective communication strategy amid the boosters in tech and the media? I find myself often torn about how charitably to view their claims and recent phase of lashing out at critics. One of my fears is that the bubble does not pop, not because the tech becomes fit for purpose or "good", but by sheer force of financial power and its partnership with fascism.

@inthehands Thanks for your thread. I think the article nails the incentive structure and conceptual rift. I'd like to examine this proposal about better-designed tools:

> Some people may hope that better models will eventually meet expectations and narrow the gap on their own. My stance is that rather than anchoring coding agent design into ideals of science fiction (magical, perfect workers granting your wishes), they should be grounded in actual science. The gap would be narrowed much more effectively then. AI tool designers should study how to integrate solutions to existing dynamics, and plan to align with known strength and limitations of automation.

I conjecture that when approached from this angle, it would be more apparent that the tools are not fit for purpose. When the unreliable foundations (LLMs) are scaffolded into more reliable and less anthropomorphized interfaces, I think they would under-stimulate the dopamine receptors. The people who are all-in on "agentic" vibe-coding are getting a thrill out of the perceived lack of restraint and their perceived ability to achieve desired outcomes. Even if the new grounded-in-science, well-scaffolded form was more useful than today's (in a holistic engineering sense), I think it would serve to deflate the bubble, challenging the investor-driven narrative and early-adopter thrill. If such a tool is built, even with stellar design and execution, I expect power to marginalize it rather than embrace it. The "magic" is a necessary feature to justify the financials.

@dalias @astraluma @hipsterelectron CI is not a substitute for writing portable code with reliable local incremental builds. But I need to know that SVE vectorization isn't spoiled and that the code remains correct when targeting GPUs from 3 vendors, that the build isn't broken on Windows or MacOS, etc. In a project with more than a couple contributors, having maintainers switch branches and manually build and test on a suite of different hardware and OS environments is untenable.

@inthehands @mcc @lambdageek Hah, yes, the ethical uses are of that variety, but "fun art project" doesn't support trillion dollar bubbles. The promise of exploitation and extraction is what hooked the investors, not mere technical marvel or fantasy.

@ntnsndr This blurry line is inevitable since "AI" is not well-defined. I think it's necessary to sharpen our shared understanding of the purpose of writing (communicating substance to people) and what good writing entails (substance, not elegant sentence structure that dupes the reader into associating it with highly-regarding writings of the past). Passing generated content as human breaks a social contract similar to passing counterfeit money.

@ntnsndr Writing is sometimes an artifact meant to provide evidence of understanding. If you take a document and transform it with a thesaurus and grammatical isomorphisms to obfuscate the source, that is plagiarism. The victim is not just the person who wrote the original content, but also the reader who is being deceived into believing you understand and communicated your understanding. Generative models are automating that obfuscation and everything that comes out is tainted.

@ntnsndr Some say "you are accountable for everything you submit", but we've always made room for honest mistakes (and it would be very anxiety-provoking if human-authored content that made an arithmetic error when summarizing statistics resulted in an honor code violation, for example). But the generated content is never right for the right reasons; it can only be incidentally correct.

@inthehands
Odds there are internal documents that would refute this claim?
> “The goal is to make sure people see what they will find most meaningful – not to keep people glued to their smartphone for hours on end.”

How did it come to be that stable evaluation of log(1 + x), called log1p in C99 and logp1 in IEEE-754, are apparently still not available in #Fortran as of #Fortran2023?

@inthehands @celesteh @adamshostack @ct_bergstrom Absolutely, though the reason we can positively identify it is because they were so sloppy with basic screening. They'll cosmetically "fix" that and then we'll be even more awash with incoherent papers making false statements alongside people hurling false accusations, not meant to stick, but to cast doubt. I don't believe Elsevier is going to lead the charge to reform structural issues that make this outcome possible.

🚨 Are you a PhD student interested in software for #CFD and/or fluids data analytics? The NSF Fluid Dynamics Software Infrastructure (FDSI) team at CU Boulder has 10-week paid internships (with housing) for PhD students to conduct V&V for complex flows like the Common Research Model (CRM), wind turbine blades, and fundamentals like the "speed bump" geometry. This work will involve model/workflow intercomparison and validation using experiments.

https://www.colorado.edu/center/aerospacemechanics/openings

PSA: Outlook apparently automatically generates image descriptions in outgoing email. A colleague sent out an email announcement about an upcoming talk on AI ethics, equitable NLP, and bias, along with a photo of the (woman) speaker. The description:

[A person with a beard Description automatically generated with medium confidence]

It appears it's possible to opt out of this "feature".
https://support.microsoft.com/en-us/office/add-alternative-text-to-a-shape-picture-chart-smartart-graphic-or-other-object-44989b2a-903c-4d9a-b742-6a75b451c669

@ekuber @cliffle It's so important I use this tweet in academic talks that aren't about Rust at all.

All the hype about souped up developer productivity using LLMs for coding reminds me of the original title of this 2014 paper, before it was milquetoasted in 2015 acceptance.

LLMs can help you rapidly acquire semi-plagiarized fragments of well-traveled code instead of using a quality library with vision of the problem domain. Might be great for KPIs, but this debt will come back to bite you, unless you're already gone. Will be painful for orgs to adapt.
https://research.google/pubs/pub43146/