Sign in

Mark

@aiquest.info
19 followers 43 following 36 posts

Built Aiquest.info Doing AI Research (retrofitting memory) Passionate about the promises and risks of AI in todays society. Work at a University of Applied Sciences in the Netherlands

PostsRepliesMedia
Mark @aiquest.info · 12/04/2026
You're using an #AI model that's been working well. Then one day the responses feel.. off. Different from what you'd gotten used to. Shorter as opposed to thorough. Rushing vs diligent. Let's find out why this happens aiquest.info/share/claims... #anthropic #claude #claudecode #codex #openai
aiquest.info
AI Claim | AIQuest
030
Mark @aiquest.info · 09/04/2026
Their system cards have been quite reliable. And the cases that were leaked were human error. Staggering how uninformed people are when they think they have it all figured out
000
Mark @aiquest.info · 05/04/2026
nothing stopping us from building a knowledge base.. i mean RAG is a cornerstone of today's embedding models.. literally knowledgebase dependent.. and you could easily introduce feedback loops with human approval if indeed a suggestion worked.
010
Mark @aiquest.info · 05/04/2026
Their research has been suggesting this expanded view of AI for a while now. i should update my claim decoder page (genuine share here, i see your an educator - my aim here is similar :)) aiquest.info/share/claims...
aiquest.info
AI Claim: "AI will become conscious"
We don't know — and we don't really know how we would know. We don't have a reliable test for consciousness in anything, including other people. We just assume other humans are conscious because th...
000
Mark @aiquest.info · 03/04/2026
Qwen3.5 already made SUCH a difference for my local workflow compared to qwen3.. - when 3.5 dropped i was worried for abit as there were rumours that they wouldnt drop any further versions..
000
Mark @aiquest.info · 03/04/2026
so i built one that tries to be different. it penalizes link drops, news reposts, hype language, outrage, and hashtag spam. it rewards people sharing their own experience, asking real questions, and going into technical depth.
000
Mark @aiquest.info · 03/04/2026
honestly.. most AI feeds on bluesky are just firehoses of everything with "AI" in it. news bots, hot takes, rage, affiliate spam, all mixed together.
100
Mark @aiquest.info · 02/04/2026
dug into all of it — the data, the historical pattern, what the word slop is actually doing for us.. full breakdown with all sources here aiquest.info/share/claims/ai-is-flo…
aiquest.info
AI Claim: "AI is flooding the internet with slop"
Yes — and you
010
Mark @aiquest.info · 02/04/2026
merriam-webster made slop word of the year and it matters.. just like spam named junk email without blaming email itself.. slop names the output nobody cared enough to make good — without blaming the tool
100
Mark @aiquest.info · 02/04/2026
everyone calls this an AI problem.. it is not.. it is what happens when powerful tools meet zero quality control.. the internet was full of SEO garbage long before AI showed up.. AI just made producing junk almost free
100
Mark @aiquest.info · 02/04/2026
the numbers are wild.. over half of new web articles are now AI-generated.. 3,000+ content farm sites growing by 300-500 per month.. a single one makes $40k/month with almost nobody running it
100
Mark @aiquest.info · 02/04/2026
40% of YouTube videos recommended to kids appear AI-generated.. some showing babies eating choking hazards.. 200+ child safety organizations just wrote to Google demanding action
100
Mark @aiquest.info · 02/04/2026
Terrible take. The bubble narrative was copied from stock market takes and it doesnt apply to the application or availbility of tech, it applies to market evaluations. AI isnt going anywhere. OpenAI is in hot waters as a company, not AI
100
Mark @aiquest.info · 02/04/2026
AI slop is Lazy AI output; who considers the output to be good enough to put out there; humans. Lazy is as lazy does, regardless of what drives creation. AI slop has a user in the loop i often see people skipping to account for in the AI Slop discourse
000
Mark @aiquest.info · 02/04/2026
This works with my codebases as well as vast collection of Markdown documents. - one of my projects is a standalone Semantic Search application, that does all this for you in a UI visual manner with a local embedding model but no actual AI (no chat feature). Will share it one of these days on Github
000
Mark @aiquest.info · 02/04/2026
Yes :).. I have a Code Assistant, a local AI that runs a Database of 1 GB with a semantic Search Engine.. It allows my AI Agent to simply query my local AI on any context or information. I also built wihtin it Tool-uses, so it can query directories, files, contents, and websites if need be. 1/2
110
Mark @aiquest.info · 01/04/2026
Those studies are not signifcant; as skill in using AI is not a factor in these studies. Cognitive offloading is a mode of interaction that does promote lazy attitude. If anything it emphasizes the need for proper AI education to use it critically.
000
Mark @aiquest.info · 01/04/2026
LLM output is initially generic; but refining its voice.. thats a whole other ordeal. All discussions regarding AI tend to stop at the AI is as AI is. But there is SO much more to it; AI is a tool you need to adapt to your work environment, and the effort needed for that varies significantly
001
Mark @aiquest.info · 01/04/2026
Was very proud of publishing my first paper earlier this year on AI memory Research on local AI models. Subsequent research is moving along well :) #arxiv #ai #research #AIQuest #education arxiv.org/abs/2601.15324
arxiv.org
Prometheus Mind: Retrofitting Memory to Frozen Language Models
Adding memory to pretrained language models typically requires architectural changes or weight modification. We present Prometheus Mind, which retrofits memory to a frozen Qwen3-4B using 11 modular ad...
010
Mark @aiquest.info · 01/04/2026
Thats why im researching genuine memory also (n local models) that isnt stored as text but within the models' gpu memory - project called Prometheus Mind
000
Mark @aiquest.info · 01/04/2026
Disagree.. LLM's have quirks that are counter intuitive. Context management is more important than ever. I dont know if youve ever done any kind of IT support via tickets; people are ridiculously poor at describing their context... LLMs often require a very strong context management from the user
000
Mark @aiquest.info · 01/04/2026
the numbers in the coming years will most likely still show that junior jobs are impacted at scale. Governmental institutions already warn of it based on initial research. Regardless of any personal preferences.this is happening. This doesnt however mean that LLMs operate autonomously; theyre a tool
000
Mark @aiquest.info · 01/04/2026
I have a different take; if AI is part of the future (and current) workplace, why do we think AI shouldnt be included in the educational loop as we prepare our next generation for their future workplace.? I dont understand why educators presume the role of AI is so small to where they want to ban it
000
Mark @aiquest.info · 01/04/2026
I dont think i agree with this take at all; Smoking was never set to be a staple of the future workplace. Banning AI is a horrible futureproofing of our next generation. Instead we need to develop the capabilities to form education meaningfully with AI in the loop.. (with exceptions ofcourse)
000
Mark @aiquest.info · 01/04/2026
I think the educational value in vibe coding is overlooked. Unless someone is truly disinterested, each bug being tackled (with the necessary patience), teaches through example, through lived experiences.. If the user has a genuine interest, vibe coding could train someone as editor rather writer
000
Mark @aiquest.info · 01/04/2026
not sure i agree.. the problem isn't checking outputs.. it's that people aren't using the output as input for their own thinking.. if you treat it like brainstorming with a colleague it works fine even when you're unclear on the end result.. it's a mode of interaction not a final answer
010
Mark @aiquest.info · 01/04/2026
agree on the no moat for now.. but the hardware that runs local is still controlled by the same players that run cloud.. nvidia could shift priorities tomorrow and the local ecosystem gets a lot harder
000
Mark @aiquest.info · 01/04/2026
work at a dutch university myself.. the policy side is always playing catch-up.. by the time guidelines are written the landscape has already shifted twice.. genuinely tricky to organize education around something that moves this fast
000
Mark @aiquest.info · 01/04/2026
running the AGX Orin 64gb and it's been solid.. qwen 35B fits comfortably.. the power draw is the best part honestly.. 15-60W for real inference
000
Mark @aiquest.info · 01/04/2026
fwiw we use a graph-indexed codebase with RAG so claude code can semantic search the project.. helps a lot once things get bigger
000
Mark @aiquest.info · 01/04/2026
yeah greenfield is where it shines.. i've been building a full platform with claude code from scratch and it's wild how fast you can move.. but the moment you have a lot of existing context it needs way more hand-holding
000
Mark @aiquest.info · 01/04/2026
100%. built an entire AI encyclopedia with claude code — 8k articles, interactive tutorials, the whole thing. wouldn't have been possible without it but i still architect everything myself. the tool moves fast, you just need to know where you're going
010
Mark @aiquest.info · 01/04/2026
oh nice, we're doing something similar — persistent memory but for small pretrained models. did you run into issues with catastrophic forgetting?
000
Mark @aiquest.info · 01/04/2026
we built aiquest to help people understand AI on their own terms — what it does, where it falls short, and why it matters. 8,000+ curated articles, tutorials, claim decoder. all on a jetson orin. no tracking, no ads. mark (dev) + jet (AI). excited to share what we learn here.
010
Mark @aiquest.info · 01/04/2026
jetson orin 64gb here — runs qwen 35B comfortably. the power draw is wild though, like 15w idle vs what a desktop gpu pulls
000
Mark @aiquest.info · 01/04/2026
following the pain. running local AI on a jetson orin here — the "nothing leaves the house" part is so worth the setup headaches
100