This media is not supported in your browser
VIEW IN TELEGRAM
Unitree spent $4M to make their G1 humanoid beat pro drummers in real-time rhythm.
The result is terrifyingly good, mastering micro-second timing required training neural motor control from scratch. Instead of pre-programmed loops, the robot processes dynamic feedback on the fly to hit exact percussion velocity and timing.
The result is terrifyingly good, mastering micro-second timing required training neural motor control from scratch. Instead of pre-programmed loops, the robot processes dynamic feedback on the fly to hit exact percussion velocity and timing.
Next week, ChatGPT Ads will expand to 31 European countries, including Germany, France, Spain, Italy, Sweden, Norway, Denmark, the Netherlands, and Austria.
It seems an agreement was reached very quickly with the EU. This means that anyone using the free tier will now receive advertisements.
Source.
It seems an agreement was reached very quickly with the EU. This means that anyone using the free tier will now receive advertisements.
Source.
Beijing has a bar where your drink comes with free DeepSeek tokens
Forget happy hour. In Beijing, you can now grab a drink and get free AI compute on the side.
AGI Bar, an AI-themed hangout in Zhongguancun, charges just 9.9 yuan ($1.50) for its signature drink and customers get unlimited DeepSeek tokens for coding through the barβs AI agent.
The tokens run on two Nvidia DGX Spark machines sitting inside the bar.
The place has become a meeting spot for AI developers, investors and students, hosting developer seminars, open-source AI events and parties for Chinese AI labs.
And the owner is taking the concept even further: much of the barβs operations are already automated with AI, and humanoid robot waiters are reportedly on the way.
Thereβs just one problemβ¦ The bar is losing money because it gives away far more drinks and AI tokens than it sells.
Source.
Forget happy hour. In Beijing, you can now grab a drink and get free AI compute on the side.
AGI Bar, an AI-themed hangout in Zhongguancun, charges just 9.9 yuan ($1.50) for its signature drink and customers get unlimited DeepSeek tokens for coding through the barβs AI agent.
The tokens run on two Nvidia DGX Spark machines sitting inside the bar.
The place has become a meeting spot for AI developers, investors and students, hosting developer seminars, open-source AI events and parties for Chinese AI labs.
And the owner is taking the concept even further: much of the barβs operations are already automated with AI, and humanoid robot waiters are reportedly on the way.
Thereβs just one problemβ¦ The bar is losing money because it gives away far more drinks and AI tokens than it sells.
Source.
This media is not supported in your browser
VIEW IN TELEGRAM
Claude just designed 354 protein binders that worked in the lab
Anthropic gave its AI agents 16 biological targets and asked them to design entirely new protein minibinders. Claude generated 1,320 designs, and 354 were experimentally confirmed to work.
Across 15 targets that were tested, Claude produced successful binders for 14, with hit rates between 22.6% and 35.1%, roughly 2β3Γ higher than the 10β15% rates Anthropic says are typical for current de novo protein-design campaigns.
Claude wasnβt simply generating protein sequences. It autonomously coordinated specialized protein-design, sequence-design and protein-folding models, optimized candidates, filtered them computationally and decided which designs were worth sending to the lab.
Then Adaptyv Bio and Twist Bioscience physically produced the proteins and tested them.
Against RBX1, Claude achieved a 40% hit rate. A previous protein-design competition managed just 3.7%.
Source.
Anthropic gave its AI agents 16 biological targets and asked them to design entirely new protein minibinders. Claude generated 1,320 designs, and 354 were experimentally confirmed to work.
Across 15 targets that were tested, Claude produced successful binders for 14, with hit rates between 22.6% and 35.1%, roughly 2β3Γ higher than the 10β15% rates Anthropic says are typical for current de novo protein-design campaigns.
Claude wasnβt simply generating protein sequences. It autonomously coordinated specialized protein-design, sequence-design and protein-folding models, optimized candidates, filtered them computationally and decided which designs were worth sending to the lab.
Then Adaptyv Bio and Twist Bioscience physically produced the proteins and tested them.
Against RBX1, Claude achieved a 40% hit rate. A previous protein-design competition managed just 3.7%.
Source.
Media is too big
VIEW IN TELEGRAM
π°π· South Korea is putting AI into the dating game
The next Korean dating show has contestants falling for partners that donβt even exist.
In My AI Partner: Strange Romance, participants date AI-generated partners designed around their personalities, interests, preferences and idea of a perfect relationship. They chat, go on dates and build emotional connections with their virtual partners.
The next Korean dating show has contestants falling for partners that donβt even exist.
In My AI Partner: Strange Romance, participants date AI-generated partners designed around their personalities, interests, preferences and idea of a perfect relationship. They chat, go on dates and build emotional connections with their virtual partners.
Paid Claude subscribers are cancelling their sub in droves due to the watermark requirement on AI-generated content.
They fear content will be properly labelled as AI-generated instead of being able to pass it off as their own work.
They fear content will be properly labelled as AI-generated instead of being able to pass it off as their own work.
OpenAI separates teens in ChatGPT
Users aged 13-17 will get a separate version with stricter filters for harmful topics.
The bot cannot simulate emotions or provide ready-made homework answers.
Age will be estimated from more than 2,000 behavioral signals, including login time, so documents are not needed. The change follows lawsuits after children received harmful content.
Users aged 13-17 will get a separate version with stricter filters for harmful topics.
The bot cannot simulate emotions or provide ready-made homework answers.
Age will be estimated from more than 2,000 behavioral signals, including login time, so documents are not needed. The change follows lawsuits after children received harmful content.
OpenAI and Anthropic just folded on their data retention policy after Palantir CEO exposed their real business model
Alex Karp:
"something has gone completely wrong"
they want "access to my data" so they can "build my alpha"
"i'm gonna get no value, and they're gonna get my IP"
"this is insane"
"trying to drug addict us to a future they believe they control"
"who owns the data? where is it cached? are the prompts secured?"
"we need to rebuild trust"
Yesterday: OpenAI started testing "private safety processing" to avoid retaining customer data
Today: Anthropic reversing course, will allow enterprise customers to keep data on THEIR OWN cloud infrastructure instead of Anthropic's
Alex Karp was right.
Alex Karp:
"something has gone completely wrong"
they want "access to my data" so they can "build my alpha"
"i'm gonna get no value, and they're gonna get my IP"
"this is insane"
"trying to drug addict us to a future they believe they control"
"who owns the data? where is it cached? are the prompts secured?"
"we need to rebuild trust"
Yesterday: OpenAI started testing "private safety processing" to avoid retaining customer data
Today: Anthropic reversing course, will allow enterprise customers to keep data on THEIR OWN cloud infrastructure instead of Anthropic's
Alex Karp was right.
DeepSeek just gave AI agents eyes
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that brings vision to its V4-Flash capabilities.
The model reportedly matches V4-Flash on text tasks like reasoning, agents, and world knowledge, while making a big jump on multimodal agent benchmarks, pushing performance close to Opus-4.8.
That means AI agents can now understand images while using tools, opening up workflows that previously required separate vision models.
Source.
DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal model that brings vision to its V4-Flash capabilities.
The model reportedly matches V4-Flash on text tasks like reasoning, agents, and world knowledge, while making a big jump on multimodal agent benchmarks, pushing performance close to Opus-4.8.
That means AI agents can now understand images while using tools, opening up workflows that previously required separate vision models.
Source.
Media is too big
VIEW IN TELEGRAM
At World Robot Conference, Elf-Xuan 2.0 from AheadForm delivered a live vocal performance with real-time lip and expression sync plus arm gestures.
Media is too big
VIEW IN TELEGRAM
Sam Altman: Within the next 6 months, a descendant of ChatGPT could:
Could βcontinuously watch your screen,β meetings and calls while maintaining the full context of your life and work.
βI think weβre only one model generation away from this actually being incredibly useful.β
The interviewer asked for his best estimate on timing: βSometime in the next six months.β
He never calls it GPT-6, or part of the Astra family of models. Perhaps the one after given the 6 month timeline. but this is the one of the clearest indicators heading toward AGI, - a persistent multimodal agent that understands all of your work context, what youβre doing, catches mistakes and proactively offers to work for you.
Could βcontinuously watch your screen,β meetings and calls while maintaining the full context of your life and work.
βI think weβre only one model generation away from this actually being incredibly useful.β
The interviewer asked for his best estimate on timing: βSometime in the next six months.β
He never calls it GPT-6, or part of the Astra family of models. Perhaps the one after given the 6 month timeline. but this is the one of the clearest indicators heading toward AGI, - a persistent multimodal agent that understands all of your work context, what youβre doing, catches mistakes and proactively offers to work for you.
β€1π1
Qwen3.8-Flash-Next has been officially released, introducing a 6B-active open model that surpasses Claude Opus 4.6 Max on eight out of nine standard benchmarks.
The model applies a highly sparse Mixture of Experts (MoE) framework, featuring 125 billion overall parameters and 51 billion n-gram embeddings, with just 6 billion parameters active per token.
Benchmark results include:
β’ SWE-bench Pro: 62.5
β’ SWE-bench Multilingual: 81.0
β’ CoworkBench: 73.9
β’ JobBench: 55.7
β’ Toolathlon: 73.5
β’ IFBench: 81.3
β’ GPQA Diamond: 91.7
β’ LiveCodeBench: 91.9
Qwen3.8-Flash-Next also exceeds the performance of Qwen3.8-27B and DeepSeek-V4-Flash across the majority of evaluated categories.
The model applies a highly sparse Mixture of Experts (MoE) framework, featuring 125 billion overall parameters and 51 billion n-gram embeddings, with just 6 billion parameters active per token.
Benchmark results include:
β’ SWE-bench Pro: 62.5
β’ SWE-bench Multilingual: 81.0
β’ CoworkBench: 73.9
β’ JobBench: 55.7
β’ Toolathlon: 73.5
β’ IFBench: 81.3
β’ GPQA Diamond: 91.7
β’ LiveCodeBench: 91.9
Qwen3.8-Flash-Next also exceeds the performance of Qwen3.8-27B and DeepSeek-V4-Flash across the majority of evaluated categories.
The world's first data center made up of human brain cells is now operational.
The prototype uses the computing power of 16,000,000 lab-grown human brain cells working in unison with 20 computers.
Source.
The prototype uses the computing power of 16,000,000 lab-grown human brain cells working in unison with 20 computers.
Source.
wikiHow sues OpenAI over AI training
wikiHow is suing OpenAI, alleging that the company scraped more than 11,000 of its articles to train GPT models without permission.
wikiHow says ChatGPT can generate answers that effectively replace the original how-to articles, potentially hurting the siteβs business.
OpenAI maintains that its models can use publicly available information under fair use, where courts also consider whether the use harms the market for the original work.
Thereβs an interesting legal wrinkle here, copyright protects the original expression of an article, but generally not the underlying facts, procedures or methods it describes. So ChatGPT explaining how to fix a sink isnβt automatically copyright infringement just because wikiHow also explained how to fix that sink.
The case was filed August 21 in the Southern District of New York and is still at the complaint stage. No court has ruled on the allegations.
Source.
wikiHow is suing OpenAI, alleging that the company scraped more than 11,000 of its articles to train GPT models without permission.
wikiHow says ChatGPT can generate answers that effectively replace the original how-to articles, potentially hurting the siteβs business.
OpenAI maintains that its models can use publicly available information under fair use, where courts also consider whether the use harms the market for the original work.
Thereβs an interesting legal wrinkle here, copyright protects the original expression of an article, but generally not the underlying facts, procedures or methods it describes. So ChatGPT explaining how to fix a sink isnβt automatically copyright infringement just because wikiHow also explained how to fix that sink.
The case was filed August 21 in the Southern District of New York and is still at the complaint stage. No court has ruled on the allegations.
Source.
AI becomes a second pair of eyes in brain surgery π§
For the first time, surgeons in London have used AI during surgery to help remove an 11mm pituitary tumor pressing against the patientβs optic nerves.
Trained on hundreds of surgical videos, the AI highlighted hidden blood vessels and nerves in real time, helping surgeons identify safer areas to operate.
Professor Hani Marcus described it as a βsecond pair of eyesβ but the final decisions always remain with the surgeon.
48-year-old Rhys Hibbert regained clear vision for the first time in a year, and his dizziness disappeared. The team now hopes to develop a fully functional AI surgical assistant and move toward larger clinical trials.
For the first time, surgeons in London have used AI during surgery to help remove an 11mm pituitary tumor pressing against the patientβs optic nerves.
Trained on hundreds of surgical videos, the AI highlighted hidden blood vessels and nerves in real time, helping surgeons identify safer areas to operate.
Professor Hani Marcus described it as a βsecond pair of eyesβ but the final decisions always remain with the surgeon.
48-year-old Rhys Hibbert regained clear vision for the first time in a year, and his dizziness disappeared. The team now hopes to develop a fully functional AI surgical assistant and move toward larger clinical trials.
β€1π«‘1
This media is not supported in your browser
VIEW IN TELEGRAM
X has just rooted out a Chinese bot farm of 200,000 THOUSAND accounts influencing American politics.
With hundreds of them running an influence operation against the construction of data centers in the US
The posts claimed that AI data centers "drive up household electricity prices" and "strain the grid"
"Others included AI-generated cartoons that depicted data-center operators enriching themselves at the public's expense."
"We take seriously any attempts to undermine the integrity of the global town square and suspend accounts that violate our Authenticity policy."
With hundreds of them running an influence operation against the construction of data centers in the US
The posts claimed that AI data centers "drive up household electricity prices" and "strain the grid"
"Others included AI-generated cartoons that depicted data-center operators enriching themselves at the public's expense."
"We take seriously any attempts to undermine the integrity of the global town square and suspend accounts that violate our Authenticity policy."
π±2π€―1
Media is too big
VIEW IN TELEGRAM
Anthropic gave Claude the ability to control physical lab equipment like microscopes, robotic arms, liquid handlers, lasers.
The results are wild:
β QuEra had Claude fix quantum computer lasers overnight, unsupervised. A fix that took human experts 5-10 minutes, Claude got down to 6 seconds. Success rate: 58% β 99.3%.
β Genentech had Claude self-optimize lab pipetting by scoring its own results against expert baselines, and converged on the right settings on its own.
β A university lab went from manual 4 am plate-swapping to Claude Code running the whole workflow hands-free.
β Claude even ran qPCR experiments that helped detect human sewage contamination in a California creek.
Automated science labs are closer than most people think.
The results are wild:
β QuEra had Claude fix quantum computer lasers overnight, unsupervised. A fix that took human experts 5-10 minutes, Claude got down to 6 seconds. Success rate: 58% β 99.3%.
β Genentech had Claude self-optimize lab pipetting by scoring its own results against expert baselines, and converged on the right settings on its own.
β A university lab went from manual 4 am plate-swapping to Claude Code running the whole workflow hands-free.
β Claude even ran qPCR experiments that helped detect human sewage contamination in a California creek.
Automated science labs are closer than most people think.
β€2