Linkstream

https://eprint.iacr.org/2025/118 interesting preprint re ZK vulnerability

for all (surveyed) large enough models, embeddings converge to projections legible/interpretable without access to the model itself, not even as a blackbox 🤯
https://arxiv.org/abs/2505.12540

Terence Tao shows how to use LLMs for math.

Interesting because making formal spec drafts from informal descriptions to feed into discrete solvers is the way to get results that are exact, not approximate.
https://www.youtube.com/watch?v=zZr54G7ec7A

turns out pre-training on 4chan is an important part of aligning the LLMs
https://arxiv.org/abs/2505.04741

self-contained software engineering company as a benchmark
https://arxiv.org/abs/2412.14161

programmable matter/software defined labs is the future of R&D
https://arxiv.org/abs/2408.09171

https://github.com/jessevig/bertviz is an interactive llm visualizer; be sure to play with it while i'm trying to find the time to >_<

Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

nice paper, what's interesting is that graph databases make almost no difference in their tests, compared to just a vector store of facts distilled from history
https://mem0.ai/research

i'd say if you take the model far enough OOD from the sequence of chat sentences it will do whatever you want

in other words, the space of jailbreaks is infinite.

https://hiddenlayer.com/innovation-hub/novel-universal-bypass-for-all-major-llms/

this looks like a nice middle ground between rigid electrode arrays and [very finicky] external EM field/electricity BCI interfaces.

still needs survery but i guess we might eventually get to the implants that just grow into you [just a speculation though]

https://actu.epfl.ch/news/soft-brainstem-implant-delivers-high-resolution--2/

llm "hallucinations", human anxiety and a lot of other things have one thing in common: trying to do something when the optimal decision would be to do nothing this time
https://www.evanmiller.org/attention-is-off-by-one.html

Karpathy delivers: while LLMs ship slop 90% of the time they empower individuals with access to skill diversity previously available to corporations only.

the catch? your job now is to filter thru all that slop.
No, you can't delegate that.

https://x.com/karpathy/status/1909308143156240538?s=46

backpropagation-free training, this time much more efficient. small datasets tho.
https://arxiv.org/abs/2503.24322

glimpse of ASI (narrow domain, still interesting as a workflow)

https://www.pnas.org/doi/10.1073/pnas.2406675122

build the code that builds physical objects!
also a thread with a lot of solid experience

https://x.com/seveibar/status/1905443905979715725?s=46

instantly wishlisted. portable Raman spectrometer is like a sixth sense, except it's real

Also, "Minimal ignition risk" is an absolutely lovely detail 😂

https://fixupx.com/jwt0625/status/1904562738833367531?s=46

that box is something.
https://x.com/austinbv/status/1903276706699546958

if you add x-middleware-subrequest header to your request to the next.js website it would be treated like it has already passed through the auth wall middleware. cozy php vibes

https://security.snyk.io/vuln/SNYK-JS-NEXT-9508709

https://thingofthings.wordpress.com/2017/01/10/meditation-for-people-who-hate-meditating/
interesting.

https://ezyang.github.io/ai-blindspots/

AI Blindspots – Blindspots in LLMs I've noticed while AI coding

regardless if you use AI, almost all of these are simply good coding practices to adhere to anyway :) and there's quite a list, so take a look