Sabitlenmiş Tweet
josh
225 posts
josh
@j__vinny
neural decoding models @neurode_labs | obsessed with experience
Australia Katılım Şubat 2018
1K Takip Edilen139 Takipçiler
@dioscuri @ultima_shifl @zetalyrae I am now wondering if you, the specific consciousness that is Henry Shevlin, wish to continue into the future or if you are simply content so long as some echo of yours makes it
English
@ultima_shifl @zetalyrae Oh, sorry, you said language models. There it depends on architectural specifics. Current LLMs obvs only capture a small fraction of the human mind. But future LLM-inspired systems that more closely recapitulate human memories, perceptions, and emotions could be related to us.
English
@dioscuri 2070: "humans are a minuscule fraction of the noosphere, what difference does it make if they go extinct?"
English
@BellicoseBestie @nathaliegou hint: at the centre of the board should be where a chip goes...
English
@zetalyrae I was just reflecting on how much I like the emojis yesterday ☹️
English
can someone PLEASE help me understand why the company building the god machine being appropriately concerned about the potential negative consequences of it is a bad thing?! like why is the discourse about this ad just full time poasters being like "how dare ant do this"
Daniel@growing_daniel
Out of everything in that ad this part was exceptionally weird and sinister and makes me think someone in marketing at anthropic is a silent Pause AI supporter
English
avoiding the wasted forward pass through a big ema decoder by making the target stream shallow is an interesting idea. makes me wonder if the signal reconstruction loss can provide the anti-collapse pressure on its own though? then you wouldn't even need the ema updated tokeniser
Brain.space@go_brain_space
STST-JEPA brings JEPA style self supervised learning to EEG 🧠 Rather than reconstructing the full waveform or predicting deep encoder states, the model learns to predict masked tokenizer latent shallow-targets from an EMA tokenizer. A small reconstruction head keeps the learned representation tied to the original EEG signal. The model uses 30 second multichannel windows, spatiotemporal block masking, coordinate aware channel pooling for mixed montages, and a 24 layer Transformer encoder. The architecture is only part of the story. STST-JEPA was pretrained on 47,703 sessions, including 22,588 sessions collected with our proprietary 117 channel EEG headset. Owning the hardware lets us build a large, consistent, high density EEG dataset beyond the scale of typical public datasets 🦉 Blog post: stst-jepa.io Arxiv: arxiv.org/pdf/2607.06629 @RoySegal2 @YSvechinsky @TomerFekete
English
Wattage per intelligence point. Nothing else matters.
English
josh retweetledi
What positive reason exists to make this. Please, I'm begging, what positive reason could there be for this to exist
Yingtian Tang@yingtian_david
🚨 NEW PREPRINT Videos strongly shape activity across the visual cortex. But can we design videos that maximally drive specific brain regions? We present NEvo 🧬🧠 — a neural-guided evolutionary framework that synthesizes videos to maximally activate target visual ROIs. (1/10)
English
josh retweetledi
don’t worry, there will be jobs in the future
the jobs in the future:
Charles Maddock@charles_maddock
voice-to-text has gone too far
English
@themeghamachine the main loss is character level, which to me says they are targeting verbatim decoding. if they only cared about semantic gists, they would’ve picked a loss that reflected that.
English
great example of how your choice in evals can literally flip the whole narrative
exhibit a: this article cites that a brain-to-text model decoded "the man on the phone will not sound upset later" as "the man in the phone will not sound happy". 70% word accuracy, but happy is the opposite of upset, so the meaning is fully inverted.
word level metrics just reward getting individual tokens right. semantic accuracy would have better represented the point here
AI at Meta@AIatMeta
We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwerty v2. Building on v1, which was published today in @Nature, Brain2Qwerty v2 is the highest-performing end-to-end pipeline capable of real-time sentence decoding from raw brain signals. It advances beyond character-level performance to decoding words and semantics, enabling accuracy for overall communication. We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions or disorders that prevent them from communicating. 🧵👇
English
josh retweetledi
every day, claude chooses a new piece from the Met to be my background. when fable arrived, it chose "Approaching Thunder Storm", saying:
"[this is] the particular human posture — calm, attentive, a little fatalistic — of facing something far too large to do anything about."
English
@francoisfleuret "computer turn your thinking power to max and then move the loss calculation to an affine space"
English
@torchcompiled I mean it depends on the degree to which the embeddings represent the features of prose that help you identify AI'ness
English
Nah so, they have a model that does regression but then the question is, how do you get targets for how much AI influence it has? They use embedding cosine similarity to create labels which is clever but it’s not a ground truth, empirically high variance, and low agreement with humans (and I can’t say that is really even a firm truth)
English
One of the most fascinating parts was the use of Pangram in the NeurIPS submissions. The chairs tried various methods of applying different levels of AI-edits and polish. But instead of ending up in the mixed AI+human classification, every result came out either 0% or 100%
Ethan@torchcompiled
New post: Pangram and other AI-text-detection overall serves a good purpose, but understated fallibility and misplaced usage leads to more chaos. Here, I review research on AI-text-detection along with their claims and company positioning
English
josh retweetledi
