Defalt

Industrial automation
AI Researcher since 2008
Python Dev
Build and refine Local Models

[Rohan Paul @rohanpaul_ai]

This Yale + University of Chicago paper shows that real gap between LLM generated research ideas vs humans is not idea quality, but idea range: LLMs think narrower than human researchers.

The researchers built a controlled test from 11,683 real papers, using each paper’s nearby prior work as the shared starting point.

They asked models to propose a new motivation and method from those same prior papers, then compared those ideas with the real human paper ideas.

Instead of asking whether 1 idea looked novel, they labeled each idea by what gap it noticed and what kind of contribution it made.

Human ideas spread across many patterns, such as explaining mechanisms, testing failures, measuring evidence, building systems, and improving efficiency.

Only 12.1% of human ideas were mainly about connecting separate work, but 47.1% to 64.2% of LLM ideas did that, meaning models used this move about 4 to 5 times more often.

Even extra reasoning made this pattern stronger, suggesting models often polish a familiar recipe instead of finding more varied research moves.

---

– arxiv. org/abs/2607.01233

Title: "Measuring the Gap Between Human and LLM Research Ideas"

[Ice Universe @UniverseIce]

If Samsung doesn’t know how to make the most of the S Pen in the AI era, it should take a look at this video.

[Defalt @def4It]

@thomasunise @kimmonismus As if Claude himself hadn't been built with millions of books and software stolen from all over the world, how do you think he learned Mandarin? With Duolingo?

[Defalt @def4It]

@izzynobre Eu iniciei com Sherlock Holmes, uma das barreiras que já percebi é preço da IA, gerar imagens em 360º 4k que não ficam pixeladas em VR é muito caro caro ainda infelizmente, mas o app vai ser código aberto

[Defalt @def4It]

@izzynobre

[Alexandr Wang @alexandr_wang]

First, Mark was clearly talking about the industry’s progress on agentic capabilities on the whole.

But, while we’re on the topic: Our next Muse Spark update is coming soon. Big improvements in coding and agentic capabilities to be more competitive with other leading models.

Excited to get these into your hands—will be rolling out to Meta AI and our new API!

[Matthias Niessner @MattNiessner]

📢 FaceAnything (ECCV 2026) Code Release 📢

Turn any image sequence into high-fidelity 4D face reconstructions, without controlled capture rigs.

Try it on Hugging Face & reconstruct your face in 4D!

🔗https://t.co/1t9msPtfHw
🤗https://t.co/zzoRyurIlM
🌐https://t.co/qRqh3rZaOB

[Alex Finn @AlexFinn]

Everyone on planet Earth is talking about local AI right now

And for good reason

Governments are banning models. Hardware prices are 10xing

You NEED to be getting into local AI. The number 1 questions everyone has though is which computer to buy?

Here's your answer:

You basically have 3 options:

1. MAC STUDIO (high memory, low bandwidth)-

Mac Studios are excellent devices for local AI. They can run MASSIVE models. I'm running GLM 5.2 right now on a single Mac Studio. The model is Opus 4.8 level

The issue is, Mac Studios have very low memory bandwidth. Meaning, the models run very slow

Mac Studios are a good choice for you if you want frontier level intelligence, but are fine running the intelligence passively

Meaning you get top intelligence, but it runs more in the background rather than on demand

As an example, I have GLM 5.2 running security checks on my codebase every hour. It creates a report. I review this later in the day

2. POWERHOUSE NVIDIA CHIPS (RTX 5090, 6000 Pro)

Nvidia is the most valuable company in the world, and for good reason

They make the world's best GPUs.

They have decent VRAM (32gb on the 5090, 96gb on the 6000 Pro) and INSANE bandwidth. Meaning the local models run at unbelievable speeds

I'm running Qwen 3.6 locally on a 5090 and it's just as fast as cloud models

I'd go this route if you want to run an AI agent like Hermes off a local model, still get decent intelligence, but have it able to work lightning fast

3. AI WORKSTATIONS (DGX Spark type computers)

The DGX Spark is an excellent AI computer

It has high memory (128gb unified memory) and has decent speeds because of the Nvidia CUDA architecture

It is basically the sweet spot between a cutting edge Nvidia chip and a Mac Studio

You can run medium sized models, and get usable speeds out of them

You're not going to get the same performance as cloud models, but it will allow you to offload small secondary tasks to your local models for them to handle

They are also the absolute easiest to get up and running

You plug it in, then tell your agent on your main computer to go onto it and set it up. You don't even need it connected to a monitor

CONCLUSION

Here's what it comes down to: how high intelligence do you need, what speeds do you need, and how plug and play do you want?

Want the highest speeds, like you are used to with cloud compute? Build a computer around an RTX 5090

Want to run frontier level intelligence, and don't mind slow speeds, go with a Mac Studio

Either way, it's never been more important to get into local AI

[SkalskiP @skalskip92]

INSID3 generalizes across natural, medical, underwater, and aerial images

the same pipeline works for object-level, part-level, and personalized segmentation

[Defalt @def4It]

@AnthropicAI Their models were created based on content stolen from millions of books and code repositories without authorization. Their technology does not belong to the American government, but to the world.

[Sean C @Garron217]

@AnthropicAI

[snowboat @snowboat84]

你有没有发现，AI里模型的诞生其实相当随意？拿语言模型举例子：先是RNN，再到LSTM，某天突然说Transformer效果好就全换上，后来又拆成Encoder和Decoder，一会儿说BERT一桶浆糊，一会儿又说GPT可以有涌现能力，Scaling Law。整个过程几乎没有理论指导。现在试图在大语言模型之外异军突起的"世界模型"也一样，没什么理论根基，先把模型搭起来。

说白了，现在造模型基本就三招:（1）在别人的模型上改一点点，（2）自己以前做过类似的模型，就惯性套到别的场景试试，（3）突发奇想攒出一个架构，也不知道行不行，先试试看。模型能不能打，全靠benchmark + 算力 + 运气，不是因为它更接近某种真理。虽然今天Transformer看起来一统江湖，但是谁知道它能撑到什么时候，Transformer的边界又在那里。说不定过了几年，大语言模型到达了边界地带，我们又要模出新模型。

我们的想法是：应该给备选模型建立一个资料库，这样我们就可以把随意选模型的整个过程给工程化。

至于模型本身，可以从物理学里获得灵感。其实不少AI模型本来就来自于物理的多体系统模型。比如从Ising获得启发制造出来的玻尔兹曼机和限制玻尔兹曼机，从统计力学启发得到的能量模型，以及生成图片和视频的扩散模型，背后全是物理模型的灵感。

物理学是一个巨大的资源库，里面还躺着大量没人试过的模型。Transformer看着一统江湖，但即便今天，它在很多任务上做得其实并不好，比如把训练时没见过的更长输入丢给它，它就崩。这些Transformer啃不动的地方，恰恰可能是某个物理模型的主场。

与其继续靠运气摸彩票，不如把这件事工程化：把物理学里那些成熟的模型，一个个翻出来、做成可训练的网络，拿去跟主流硬碰，记下哪一个擅长哪一类任务。这样下一次需要新模型时，我们不是凭感觉攒一个，而是去库里查、去挑。

我先把所有可能的备选物理模型整理成了一张表，开了个repo，准备一格一格做下去。👇

[Defalt @def4It]

@eusouomatt Todo esse esforço do governo americano para tentar regular e limitar os modelos terá consequências negativas para os laboratórios, enquanto a China vai se aproveitar da situação e diminuir a distância nos resultados, fora a agressividade dos valores.

[💾 Izzy Nobre🐁 @izzynobre]

O jogo de tiro online que eu vivo postando aqui se chama Forefront

É um Battlefield em realidade virtual. Eu praticamente só jogo isso hj em dia. É brutalmente dopamínico e eu adoro que há uma comunidade imensa de brasileiros jogando (tem 3 servidores da América do Sul, chilenos e brasileiros dominam lá.)

Faço novas amizades lá diariamente, conheci ontem um argentino q eu ia tinha visto nos lobbies várias vezes. Ontem pela primeira vez conversamos. Super gente boa o cara. É super legal pra praticar idiomas tbm.

O único jogo que vai curar meu hyperfoco em Forefront será GTA6

[Defalt @def4It]

@izzynobre Caramba que exepriência foda! Que jogo é esse Izzy?

[Defalt @def4It]

@idare @ihtesham2005 Man, he’s Yann LeCun. A huge part of modern AI was built on his research; you can be sure he could secure funding for any project he wanted, maybe even more than Musk himself.

[Junior @Bldinsilence]

@MTSlive Dario probably couldn’t go 5 minutes without saying AI was going to take everyones jobs lol

[Defalt @def4It]

@RealNeo2026 @Rainmaker1973 It really looks like grapes, we even make wine with it.

[Defalt @def4It]

@eusouomatt Matt na sua opinião, os grandes laboratórios são uma bolha? Considerando que os LLMs são operações matemáticas bem estruturadas, serão commodities como os próprios softwares estão se tornando, assim que o mundo todo aprender de fato as equações, os big labs terão diferenciais?