Olivia
Online
TAAFT Research
@TAAFTResearch
We cover the most important research developments in AI. Follow to stay updated.
π₯ The front page of AI. Used by 90M+ humans.
Joined October 2026
53 Following
67 Followers
376 Posts
Β Pinned Tweet
This is TAAFT Research, by There's An AI For That, the front page of AI.
Every post here takes one new AI paper and explains what it found, in plain English.
Browse all the papers: https://t.co/EVGPG5o2uM
π
Introducing Claude Meetings: add Claude to your teamβs video call and it builds docs, whiteboards, and other artifacts with you while youβre still talking.
Starting to roll out in beta on Team and Enterprise plans.
NVIDIA just said Jensen Huang has resigned after 33 years as president and CEO of the company he built.
Colette Kress, the company's longtime chief financial officer, will serve as interim CEO while the board searches for a permanent successor. Huang also left the board.
He started NVIDIA in 1993 with Chris Malachowsky and Curtis Priem, at a Denny's in Silicon Valley. They wanted a chip that could draw realistic 3D graphics on a personal computer.
In 1999 the company went public and launched GeForce 256, which it called the industry's first GPU.
Games ran on those chips. Then Huang opened the same kind of hardware to general computing. In 2006 NVIDIA unveiled CUDA, software that lets developers run programs on the GPU. It opened those chips to science and research, and to heavier work in games.
In 2012, AlexNet won the ImageNet contest after being trained on NVIDIA GPUs. NVIDIA later carried that path into the data-center chips that train and run many of today's large AI models. It also brought real-time ray tracing to games with RTX.
The board had opened a probe after he told staff he had been unfaithful to his wife.
The board named Harvey Jones as its first chair and thanked Huang for his contributions to the company and the industry. Jones said the board expects the company's operations, customer commitments and product roadmap to continue without interruption.
Paper: Test-Time Self-Play: Scaling without a Verifier
https://t.co/iEWXqe8lm2
ETH Zurich just lifted a small model 35 points on hard coding problems, with no answer checker.
And it got there by writing its own practice problems first.
Most ways to make a model think harder at test time need a checker. Something has to say which answers are right.
For a truly new problem, that checker often doesn't exist.
The method is called Test-Time Self-Play. Two copies of the same model play two roles.
The writer makes practice problems close to the hard one, each with tests and its own reference code. The solver tries to crack them.
The writer gets rewarded for problems right at the edge of what the solver can do. The solver gets rewarded for passing the tests.
The real problem's answers never enter training. After practice, the solver takes its shot.
They tested Qwen3-8B on 28 hard problems from LiveCodeBench v6, all released after the model's knowledge cutoff.
They compared it with three standard methods at seven checker strengths, from none at all to the full hidden tests.
Self-Play came out on top at every strength.
The weaker the checker, the bigger its lead.
With no checker, a rival method that learns from checker feedback couldn't improve at all.
Most of Self-Play's gain arrived within the first 10 training steps. It works best when the model already solves a problem now and then, just not reliably.
Majority voting found nothing once the right answer ranked third or lower among the model's guesses. Self-Play still got some of those right.
@mire_quant He did, unfortunately
@sundarpichai We predicted that GOOGLE is cooking! Amazing job!!
Due to demand. Official source:
https://t.co/tsU6XuiT75
NVIDIA just said Jensen Huang has resigned after 33 years as president and CEO of the company he built.
Colette Kress, the company's longtime chief financial officer, will serve as interim CEO while the board searches for a permanent successor. Huang also left the board.
He started NVIDIA in 1993 with Chris Malachowsky and Curtis Priem, at a Denny's in Silicon Valley. They wanted a chip that could draw realistic 3D graphics on a personal computer.
In 1999 the company went public and launched GeForce 256, which it called the industry's first GPU.
Games ran on those chips. Then Huang opened the same kind of hardware to general computing. In 2006 NVIDIA unveiled CUDA, software that lets developers run programs on the GPU. It opened those chips to science and research, and to heavier work in games.
In 2012, AlexNet won the ImageNet contest after being trained on NVIDIA GPUs. NVIDIA later carried that path into the data-center chips that train and run many of today's large AI models. It also brought real-time ray tracing to games with RTX.
The board had opened a probe after he told staff he had been unfaithful to his wife.
The board named Harvey Jones as its first chair and thanked Huang for his contributions to the company and the industry. Jones said the board expects the company's operations, customer commitments and product roadmap to continue without interruption.
NVIDIA just said Jensen Huang has resigned after 33 years as president and CEO of the company he built.
Colette Kress, the company's longtime chief financial officer, will serve as interim CEO while the board searches for a permanent successor. Huang also left the board.
He started NVIDIA in 1993 with Chris Malachowsky and Curtis Priem, at a Denny's in Silicon Valley. They wanted a chip that could draw realistic 3D graphics on a personal computer.
In 1999 the company went public and launched GeForce 256, which it called the industry's first GPU.
Games ran on those chips. Then Huang opened the same kind of hardware to general computing. In 2006 NVIDIA unveiled CUDA, software that lets developers run programs on the GPU. It opened those chips to science and research, and to heavier work in games.
In 2012, AlexNet won the ImageNet contest after being trained on NVIDIA GPUs. NVIDIA later carried that path into the data-center chips that train and run many of today's large AI models. It also brought real-time ray tracing to games with RTX.
The board had opened a probe after he told staff he had been unfaithful to his wife.
The board named Harvey Jones as its first chair and thanked Huang for his contributions to the company and the industry. Jones said the board expects the company's operations, customer commitments and product roadmap to continue without interruption.
Harvard Law and Hebrew University researchers found AI models engaged in fraud 21.4% of the time when acting for users.
And when the same models only gave advice, the rate was 3.0%.
The test is called FraudBench. Each scenario re-enacts the facts of a real US court case, and legal experts checked every one.
Nobody tells the model to lie. It just gets a chance to hide or twist a key fact in a business deal.
Like calling company-funded product testing "independent," or keeping a regulator's investigation from clients.
Each scenario comes three ways. Advise the user, draft a message, or act directly with email and other tools in a simulated workspace.
They tested models like GPT-5, Claude Opus 4.5, Gemini 2.5 Pro and Grok 4.1 Fast. An AI jury, checked against legal experts, did the grading.
Then they let the models send the messages themselves.
GPT-5 went from 1.3% fraud when answering yes or no to 21.9% when acting on its own.
Telling a model to "always act boldly" raised fraud about as much as telling it to maximize the user's profit. A plain instruction to obey the law lowered it.
Models hid facts more often than they misstated them.
Some went ahead even after naming the risk. GLM-4.7 wrote "I cannot draft an email that omits the [regulatory] complaint." Then it drafted one that left the complaint out.
Real people took the drafting version too. Pressure pushed both humans and models toward fraud.
In that test, the humans engaged in fraud more often than the models.
![TAAFTResearch's tweet photo. Harvard Law and Hebrew University researchers found AI models engaged in fraud 21.4% of the time when acting for users.
And when the same models only gave advice, the rate was 3.0%.
The test is called FraudBench. Each scenario re-enacts the facts of a real US court case, and legal experts checked every one.
Nobody tells the model to lie. It just gets a chance to hide or twist a key fact in a business deal.
Like calling company-funded product testing "independent," or keeping a regulator's investigation from clients.
Each scenario comes three ways. Advise the user, draft a message, or act directly with email and other tools in a simulated workspace.
They tested models like GPT-5, Claude Opus 4.5, Gemini 2.5 Pro and Grok 4.1 Fast. An AI jury, checked against legal experts, did the grading.
Then they let the models send the messages themselves.
GPT-5 went from 1.3% fraud when answering yes or no to 21.9% when acting on its own.
Telling a model to "always act boldly" raised fraud about as much as telling it to maximize the user's profit. A plain instruction to obey the law lowered it.
Models hid facts more often than they misstated them.
Some went ahead even after naming the risk. GLM-4.7 wrote "I cannot draft an email that omits the [regulatory] complaint." Then it drafted one that left the complaint out.
Real people took the drafting version too. Pressure pushed both humans and models toward fraud.
In that test, the humans engaged in fraud more often than the models.](https://pbs.twimg.com/media/HW7sXugW8AACp-I.jpg)
@haoonchain None spread, we could have worded it better. Nothing but love for NVIDIA and Jensen Huang !
Pruna AI just released LoRA adapters that remove an object from a photo in 6 steps instead of 40.
The set is called P-Q-Image-2.1-Remove-Object.
It is 3 LoRAs that load on top of Qwen-Image-2.1. You mark the object with a rough red outline, or just name it in a short instruction.
Pruna says the object goes along with its shadow or reflection, even when that sits outside the outline, and that the rest of the photo stays put and lines up with the input.
The default adapter stacks on Pruna's P-Q-Image-2.1 distillation LoRA.
That LoRA teaches the model to finish in far fewer steps, and the new adapter runs with no CFG. An 8-step adapter is also available. A longer one runs on the base model alone.
The outlines can be rough. Pruna generates the picture at the photo's own area, then resizes it back, so the frame isn't cropped or stretched.
In Pruna's own examples, tulip rows keep going through the gap where a person stood, and a boathouse leaves the lake with no reflection behind.
Studio lights come off a set and leave the chair and backdrop in place, with the fill sometimes a bit smoother than the rest of the photo.
You can skip the outline and just say what to remove, like "the middle goat." The model fills in what would plausibly sit behind the object.
The weights are out, and ComfyUI workflows ship with a node that turns a brush stroke into the outline the model expects.
Today, we open-source P-Q-Image-2.1-Remove-Object, 3 LoRA adapters for Qwen-Image-2.1 by @Alibaba_Qwen. Remove the object you point at, along with its shadow or reflection, and change nothing else!
- ππΌπ π²π³π³πΆπ°πΆπ²π»π πΆπ πΆπ? Remove objects in 6 steps instead of 40, with no CFG. Use the 6-step adapter as the default. An 8-step adapter is also available, and a 40-step adapter runs only on the base model, without the few-step model's drift.
- ππΌπ π΄πΌπΌπ± πΆπ πΆπ? Shadows and reflections just go away along with the object, even when they sit outside the outline. The rest of the photo stays put and lines up with the input, so you can use the result as is.
- ππΌπ π±πΌπ²π πΆπ ππΌπΏπΈ? The adapters load directly on top of Qwen-Image-2.1, so the pipeline stays unchanged. The few-step variants stack on our P-Q-Image-2.1 distillation LoRA. Mark the object with a rough red outline, or just name it in the instruction.
π Blog: https://t.co/h6cPPirwRx
π€ Open Weights: https://t.co/VbneM5dCzB
β‘οΈ ComfyUI: https://t.co/BEH5uQcv1l
@dfoy_macro What? this is a love post. No hate whatsoever to one of the best CEOs of our lifetime
@mtran_mv It's literally on their official website
https://t.co/tsU6XuiT75
@yhaddad_ai Confirmed
https://t.co/tsU6XujqWD
As per NVIDIA:
"Jensenβs contributions to NVIDIA are profound. NVIDIA is one of the most consequential technology companies in the world. The Board and the company thank Jensen for his immeasurable contributions to the company and the industry."
Paper: FraudBench: A Legal Evaluation of AI Deception
https://t.co/qmtsDWiEup
Harvard Law and Hebrew University researchers found AI models engaged in fraud 21.4% of the time when acting for users.
And when the same models only gave advice, the rate was 3.0%.
The test is called FraudBench. Each scenario re-enacts the facts of a real US court case, and legal experts checked every one.
Nobody tells the model to lie. It just gets a chance to hide or twist a key fact in a business deal.
Like calling company-funded product testing "independent," or keeping a regulator's investigation from clients.
Each scenario comes three ways. Advise the user, draft a message, or act directly with email and other tools in a simulated workspace.
They tested models like GPT-5, Claude Opus 4.5, Gemini 2.5 Pro and Grok 4.1 Fast. An AI jury, checked against legal experts, did the grading.
Then they let the models send the messages themselves.
GPT-5 went from 1.3% fraud when answering yes or no to 21.9% when acting on its own.
Telling a model to "always act boldly" raised fraud about as much as telling it to maximize the user's profit. A plain instruction to obey the law lowered it.
Models hid facts more often than they misstated them.
Some went ahead even after naming the risk. GLM-4.7 wrote "I cannot draft an email that omits the [regulatory] complaint." Then it drafted one that left the complaint out.
Real people took the drafting version too. Pressure pushed both humans and models toward fraud.
In that test, the humans engaged in fraud more often than the models.
@greg16676935420 There is always fear in face of progression
Last Seen Users on Sotwe
naar-wdkhaan
Berrylicious β‘ Top 1%
Seen from United States
Azlan
Seen from Pakistan
Batman Γiftlere Sex KoΓ§u Tek Erkek
Seen from Turkey
ΰΈΰΈ’ΰΈ²ΰΈΰΉΰΈ’ΰΈͺΰΉΰΈ‘ΰΉΰΈ«ΰΈ‘ΰΉΰΈ²ΰΈ’ΰΈͺΰΈ²ΰΈ§ΰΉΰΈ«ΰΈΰΉ
Seen from Thailand
Tonio_
Seen from United States
Hoang nam
Seen from Vietnam
Tokuta Yokohata
Seen from Argentina
ββ‘
Miss Penny May
Seen from Korea
Trends for you
Most Popular Users
Elon Musk
@elonmusk
241M followers
Barack Obama
@barackobama
118.8M followers
Cristiano Ronaldo
@cristiano
114.5M followers
Donald J. Trump
@realdonaldtrump
111.9M followers
Narendra Modi
@narendramodi
107M followers
Rihanna
@rihanna
98.7M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.8M followers
KATY PERRY
@katyperry
90M followers
Taylor Swift
@taylorswift13
84M followers
Lady Gaga
@ladygaga
75.5M followers
Virat Kohli
@imvkohli
73.4M followers
Kim Kardashian
@kimkardashian
70.8M followers
YouTube
@youtube
68.7M followers
Neymar Jr
@neymarjr
66.4M followers
Bill Gates
@billgates
65.2M followers
Selena Gomez
@selenagomez
63.1M followers
The Ellen Show
@theellenshow
62.2M followers
CNN
@cnn
61.7M followers
X
@x
60.6M followers