In the cycle of enchanted, disenchanted, and re-enchanted with the AI world, I’m briefly stuck in the disenchanted space. Mainly in applications to writing. I’m relying on these tools for STEM-related tasks, analysis, etc, but becoming very cynical about its use in (non-technical) writing.

It's interesting that this has gone under the radar: OpenAI began adopting a constitutional approach to alignment in late 2024, updated last week. Their "deliberative alignment" specs tell the model to treat a list of rules deontologically and deliberate how they apply to particular examples.

My recent article explores how colleges can scale AI readiness in First-Year courses, drawing insights from our FYE program at CWI. It also includes links to the training we developed (CC BY). Some of these resources may work in First-Year Writing courses as well. www.linkedin.com/pulse/how-co...

I used the DeepSeek R1 reasoning model to prepare for a new course proposal. These screenshots show with and without the "DeepThink" option turned on--strikingly different. R1 does a lot more synthesis and offers clearer suggestions. It also accepts pdf files, o1 doesn't. Crazy this is open source.

What worked so well in this Nigerian experiment with using AI to boost literacy is how carefully each step is overseen by actual teachers. Perhaps the "deskilling" we see in other studies (students losing skills because of too much assistance) is bad strategy. blogs.worldbank.org/en/education...

Prompt engineering with o1: Interesting to compare this o1 strategy with the CLEAR or RFTC framework (role format task constraints) I currently find myself relying less on “role” and more on “context dump”

This is one of the most elegant definitions of LLMs I’ve seen. (from this post explaining the new o3 model and the ARC benchmark: arcprize.org/blog/oai-o3-...)

I'm shopping for broccoli sprout seeds on amazon and the "most helpful review" is 100% AI-generated. It's hard for me to read because half the words are completely pointless--but apparently it's helpful to others! 🤷‍♂️

My program is collecting data on this (through surveys) and it somewhat tracks. A small percentage of students are definitely “anti-AI”. Most students, OTOH, say they’re uncomfortable with faculty using AI to evaluate their work, but they’re comfortable with AI in ed otherwise.

I love seeing health gurus compete like this

The anecdote in the Hard Fork podcast may have been intended as—and should definitely be interpreted as—an apt and colorful analogy one researcher perceived between Arrival and the simultaneity of processing in Transformers. It won't bear much weight as a literal claim about historical causality.

Tilman Bayer@tilmanbayer.bsky.social · 2y ago

OK, so maybe what Kevin Roose bungled here was that one of the paper's authors (Polosukhin) had made a comparison to "Arrival" ("believed self-attention was a bit like ...") - but without claiming it had been the actual inspiration www.ft.com/content/37bb...

it's weird that the most effective prompt hack STILL is just "approach this in the style of [person]" these models internalize not just the style but the whole vibe - the writer's voice, epistemic stance, worldview, etc. "write like x" tightens everything up so nicely

I'm seeing more AI scanning being done by employers and institutions--checking for GenAI text in resumes, grant applications, etc. It's odd that "preparing for the workplace" now means both NOT using GenAI for some things but being really savvy at other tasks. Anyone publishing on this?

Not sure why it took me 6 months to discover Max Read's famous article where he coined the term "Zynternet". article here: maxread.substack.com/p/hawk-tuah-... perplexity's summary here: www.perplexity.ai/search/zynte...

Hawk Tuah and the Zynternet

Plus, for some reason, some thoughts about the debate

maxread.substack.com

I think a lot of debates over ethical use of AI in the classroom would be more productive and all parties first agreed there are many things happening simultaneously. Context is that which is scarce.

Ethan Mollick@emollick.bsky.social · 2y ago

AI can help learning... when it isn't a crutch. There are now multiple controlled experiments showing that students who use AI to get answers to problems hurts learning (even though they think they are learning), but that students who use well-promoted LLMs as a tutor perform better on tests.

If you're depressed that non-expert readers prefer AI-written poetry to the classics, perhaps try Matt's quiz. You may discover that the real finding here is the huge gulf between your own taste and that of non-expert readers ... a gulf that has likely existed at least since, oh, IA Richards?

Matt Clancy@mattsclancy.bsky.social · 2y ago

I made a Google quiz using the poems listed on the paper’s OSF site if you want to try your luck at guessing which are human and which are LLM. forms.gle/nGCGawDb9c6f...

I like this take from Ben on the future of AI in Hollywood because 1) he’s done his homework on transformer models lol and 2) he clearly understands the possibilities without 3) getting lost in hype.

Rob Watts@online-audit.co.uk · 2y ago

Interesting take from Ben Affleck #ai #lvm youtu.be/ypURoMU3P3U?...

Wow so all of academic twitter is now here, interesting. Nice to see you.