Lightnews — Scholar-powered news

braintelligence.bsky.social @braintelligence.bsky.social · 13m

… this is where diffusion can shine

The model guides it with its “thought” embedding values, and diffusion uses this guidance to convert to language/actions

braintelligence.bsky.social @braintelligence.bsky.social · 14m

This is where yann lecun is on to something

We train models to reason in token domain, when we should have them reason in embedding space first, and token translation is a thin last operation

1

braintelligence.bsky.social @braintelligence.bsky.social · 16m

In this case vought wants them as a new underclass to replace migrant farm and factory workers

1

braintelligence.bsky.social @braintelligence.bsky.social · 1h

The mechanics of how they think are different, but I’m starting to think all thought is based on the same core task of modeling the world

So they’ll converge on our same philosophical dilemmas as ai becomes embodied with continuous learning

Reposted

Sheryl Gay Stolberg @sherylnyt.bsky.social · 16h

BREAKING: Friday night massacre underway at CDC. Doznes of "disease detectives," high-level scientists, entire Washington staff and editors of the MMWR (Morbidity and Mortality Weekly Report) have all been RIFed and received the following notice:

810 8K 14K

braintelligence.bsky.social @braintelligence.bsky.social · 15h

In the future, the fact we shoved user input directly into the core layer will be seen as barbaric and inhumane

Reposted

AI Firehose @ai-firehose.column.social · 18h

Researchers have unveiled SUmmarization augmented Policy Optimization (SUPO), a novel RL framework that empowers large language models to handle complex, long-horizon tasks by using LLM-generated summaries for effective context management. https://arxiv.org/abs/2510.06727

Scaling LLM Multi-turn RL with End-to-end Summarization-based Context Management

ArXiv link for Scaling LLM Multi-turn RL with End-to-end Summarization-based Context Management

arxiv.org

1

Reposted

Sung Kim @sungkim.bsky.social · 18h

The same logic applies to fashion: instead of buying expensive clothes to look better, invest that time and money into getting fit, you’ll look better than anything designer labels can offer.

1 4

Reposted

Sarah McAnulty, Ph.D. @sarahmackattack.bsky.social · 1d

Here I am. Once again. I'm out of archaeologists.

We matched 576 groups w/archaeologists, leaving 13 groups unmatched. This brings me pain! We've never run out of scientists like this before. 598 requests for one category is A LOT. But still.

Archaeologists 🥺
www.skypeascientist.com/sign-up.html

Sign Up

Skype a Scientist gives you the opportunity to connect with students and the public around the world.

www.skypeascientist.com

14 110 130

braintelligence.bsky.social @braintelligence.bsky.social · 18h

I’m someone who thinks AI is a powerful art tool for humans, but the trump admin has been shoveling out lazy, low quality AI generations, which tarnishes the whole field

Matt Novak @paleofuture.bsky.social · 18h

Trump launched the TrumpRX website today.

It has a photo of Trump and an AI-generated family with too many fingers and mangled toes.

TrumpRx Is a Narcissistic, AI-Generated Nightmare

AI art really has become the aesthetic of grifters and fascists.

gizmodo.com

Reposted

Techmeme @techmeme.com · 19h

Google says it hit a milestone of 1.3 quadrillion monthly tokens processed across its services this summer, up from 980T monthly tokens announced in July (Matthias Bastian/The Decoder)

Main Link | Techmeme Permalink

1 3

braintelligence.bsky.social @braintelligence.bsky.social · 19h

Put AI or LLM in your bio and you’ll probably get put on a block list

bsky.app/profile/brai...

braintelligence.bsky.social @braintelligence.bsky.social · Aug 18

Apparently Im blocked by 19,000 Bluesky users because Im posting about AI:

Reposted

luokai @luok.ai · 1d

GAGA-1 just dropped: an AI “actor” that fuses voice, lipsync, and performance.

It’s not fully free—there’s roughly 60 seconds of free generation per month; the API is currently unavailable. It’s built for dialogue-driven scene prototyping.

1 1 3

Reposted

Ethan Mollick @emollick.bsky.social · 1d

On one hand: don't anthropomorphize AI. On the other: LLMs exhibit signs of gambling addiction.

The more autonomy they were given, the more risks the LLMs took. They exhibit gambler's fallacy, loss-chasing, illusion of control...

A cautionary note for using LLMs for investing without guardrails.

13 27 110

Reposted

Tim Kellogg @timkellogg.me · 1d

someone got 473 tok/sec(!!) on a macbook M3 Ultra for Qwen3-30b-a3b

and that’s a decent model..

Three line charts showing performance metrics for Qwen3-30B-A3B-4bit running on an M3 Ultra (512 GB RAM) across varying batch sizes.

⸻

Top Chart: Prompt Tokens per Second
• Y-axis: Prompt TPS (tokens per second), X-axis: Batch Size.
• Values rise slightly overall:
• 0: 2876 5: 2868 10: 2880 15: 2885 25: 2893 30: 2891
• Peak performance: 2893 TPS at batch 25, showing stable prompt throughput across batch sizes.

⸻

Middle Chart: Generation Tokens per Second
• Y-axis: Generation TPS, X-axis: Batch Size.
• Throughput improves steadily with larger batches:
• 0: 108 5: 249 10: 336 15: 425 25: 453 30: 473
• Performance scales efficiently, plateauing near 473 TPS.

⸻

Bottom Chart: Peak Memory Usage
• Y-axis: Peak Memory (GB), X-axis: Batch Size.
• Memory grows roughly linearly with batch size:
• 0: 18.28 GB 5: 19.46 GB 10: 21.66 GB 15: 23.28 GB 25: 24.89 GB 30: 26.50 GB
• Peak usage: 26.5 GB at batch 30.

⸻

Overall: Prompt throughput remains flat, generation throughput scales well, and memory usage rises linearly with batch size.

2 1 17

Reposted

Tim Kellogg @timkellogg.me · 1d

UserLM-8B: an LLM that mimics a person

Microsoft fine tuned an LLM to respond as the user instead of as an assistant

This is useful anytime you need mock user input. like maybe for testing agents, or training models

huggingface.co/microsoft/Us...

microsoft/UserLM-8b · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

3 2 31

Reposted

saganite.bsky.social @saganite.bsky.social · 1d

I would like to share some work we've been doing at cascadetech.ai: Predicted Outputs in vLLM. If you aren't familiar with PO, it allows you to dramatically speed up generation when you know something about the contents of the output (think: code modification).

1 3 11

braintelligence.bsky.social @braintelligence.bsky.social · 20h

Next token predictor
Stochastic parrot

/s

1 1

Reposted

Ethan Mollick @emollick.bsky.social · 22h

"Sora 2, an elaborate regency romance where everyone is wearing a live duck for a hat (each duck is also wearing a hat) , prestige drama"

21 13 150

Reposted

Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 22h

WALL FLIP omniretarget.github.io

3 11 64

braintelligence.bsky.social @braintelligence.bsky.social · 20h

This is a shrewd email from the guild. They’re asking what prompts they are using for ai review

I’ve long advocated that the best ai regulation is to know what the system and task prompts are, and training data metadata

max tani @maxtani.bsky.social · 21h

New: The WGAE, which represents some parts of CBS News including its social team, sent a note to members today advising its members not to respond to Weiss’ request for an email by next week. The union also said it asked CBS for more information about Weiss’ request.

braintelligence.bsky.social @braintelligence.bsky.social · 23h

Comfy boy thread

3

braintelligence.bsky.social @braintelligence.bsky.social · 23h

This is kinda interesting… Joe Rogan has started asking ai to explain politics live on his podcast

www.instagram.com/reel/DPmy5Fx...

Login • Instagram

Welcome back to Instagram. Sign in to check out what your friends, family & interests have been capturing & sharing around the world.

www.instagram.com

Reposted

Kyle Griffin @kylegriffin1.bsky.social · 1d

Breaking: MIT just became the first university to reject an agreement that would trade support for the Trump admin's higher education agenda in exchange for favorable treatment.

"Fundamentally, the premise of the document is inconsistent with our core belief." www.nytimes.com/2025/10/10/u...

M.I.T. Rejects a White House Offer for Special Funding Treatment

www.nytimes.com

94 790 3.4K

Reposted

Sung Kim @sungkim.bsky.social · 1d

the LLM is gonna decode it repeatedly forever, guaranteed.

Blog: liujch1998.github.io/2025/10/07/f...

Defying Transformers: Searching for “Fixed Points” of Pretrained LLMs

Transformers are transformative to AI. Are there things that cannot be transformed by Transformers?

liujch1998.github.io

1 1 4