this is a very important paper, although not easy to read
it takes hyperconnections (HC), which is basically just a smarter way to do back propagation, and stabilizes it so that it’s actually usable at scale
Hear me out: IPA in @duolingoverde.bsky.social
In the era of GenAI, we should realize and teach young people that the main product of intellectual work is not only the artifact, like lines of computer code or an essay, but also the bonds the author made with the artifact through labor. Without the latter, the artifact is dead.
Is there a better way to show how LLM chatbots hallucinate than giving them a YouTube video link and asking for the name of the background music?
Do you ever think about whether robots in fictions are conscious?
Typing non-code text in VSCode with Copilot enabled is an "interesting" experience. The predicted texts usually are multiple sentences long and look plausible and contextualized based on the preceding text. But it is almost always hallucination that ends up being more distracting than helpful.
The vaccine against Shingles helps protect against dementia, results of a natural experiment, adding to prior evidence
"implications are profound"
New @nature.com
www.nature.com/articles/d41...
nature.com/articles/s41...
Done.
InterSpeech paper reviews: 2 down, 1 to go.
As a reviewer for journals and conferences, have you come across article submissions that are grammatically flawless but still make you wonder whether you have dyslexia when you read them? I am looking at one of those articles and I suspect it is an AI-written fake.
CDC source:
“We just had word that all our fellows and post doc staff are laid off effective immediately. The famous Epidemic Intelligence Service, aka the Disease Detectives, is no more. That’s 1260 staff.
They are calling this ‘Phase 1’.”
The most impactful open-source project of today (dixit Vercel VP of AI)
=> huggingface.co/blog/open-r1
Locally: using ollama
DeepSeek-R1's thinking steps contain a lot cautionary self-checks. It often looks over-cautious to the extent of being mechanical, but it's still really interesting and impressive.
GRPO can be understood as trying solving a problem many ways and learning from feedback (reward) signsls which ones work and which ones don't.
Quick summary of the DeepSeek-R1 paper that everyone is talking about: Do away with supervised fine-tuning on instruction-following or CoT data after pre-training. Instead go straight to RL, specifically by using the Group Relative Policy Optimization on reasoning-intensive tasks.
Can feel the tension building between human-in-the-loop (HITL) vs. machine-in-the-loop (MITL) camps building up surrounding human-AI interactions in recent days.
@fchollet's latest blog post on the latest progress by OpenAI o3 on the ARC-AGI benchmark:
The examples of simple tasks that are unsolved by o3 all involve spatial reasoning and generalization, a task that humans find natural but AI still struggles at.
arcprize.org/blog/oai-o3-...
Until we truly solve the hard problem of consciousness, we will never be able to resolve the debate of whether an AI is self-aware. Discussing AI self awareness, for now at least, is a waste of time. We should instead focus on the much more measurable goal of grounding to the external world.
In his NeurIPS speech, Ilya Sutskever muses that the next generations of large neural networks might be "self aware". Instead of making them aware of themselves, I think it will be more helpful to first make them aware there is an external world.
https://youtu.be/1yvBqasHLZs
"... we tend to be most conscious of the things our intelligent brain does least well, the recent things in evolutionary history such as logic, mathematics, philosophy, and general problem-solving and planning, and we tend to be quite unconscious of its true powers ..." -- Tomaso Poggio (1990)
Interesting post on the current status of open-source LLMs developed by Chinese companies
xianbao-qian.medium.com/dec-2024-chi...
🧪 E11 Bio is excited to share a major step towards brain mapping at 100x lower cost, making whole-brain connectomics at human & mouse scale feasible (🧠→🔬→💻). Critical for curing brain disorders, building human-like AI systems, and even simulating human brains.
Read more: e11.bio/news/roadmap
#ReadingLog
a.co/d/1xPudVC
Designers of artificial languages should take this use case into account.
Collaborating with Team Gleason on applications of generative AI to accessibility is one of the most deeply transformative experiences of my career so far. While we have made progress in shedding light on paths forward, much work and many opportunities await us.
teamgleason.org/speakfaster/
