@azeem.bsky.social
Inception Labs launched a Diffusion LLM, Mercury, that is 5x faster at coding than traditional LLMs.
Traditional LLMs write sentences word by word. Diffusion LLMs work differently. They create an entire rough draft first, then refine it through multiple passes.
Traditional LLMs write sentences word by word. Diffusion LLMs work differently. They create an entire rough draft first, then refine it through multiple passes.
February 28, 2025 at 5:00 PM
Inception Labs launched a Diffusion LLM, Mercury, that is 5x faster at coding than traditional LLMs.
Traditional LLMs write sentences word by word. Diffusion LLMs work differently. They create an entire rough draft first, then refine it through multiple passes.
Traditional LLMs write sentences word by word. Diffusion LLMs work differently. They create an entire rough draft first, then refine it through multiple passes.
Alibaba released Wan 2.1, now open-source!
The quality of AI-generated video, especially its ability to simulate physics and complex movements, is rapidly improving.
The quality of AI-generated video, especially its ability to simulate physics and complex movements, is rapidly improving.
February 27, 2025 at 5:00 PM
Alibaba released Wan 2.1, now open-source!
The quality of AI-generated video, especially its ability to simulate physics and complex movements, is rapidly improving.
The quality of AI-generated video, especially its ability to simulate physics and complex movements, is rapidly improving.
The MASt3R-SLAM system can transform a single video into a 3D scene in real-time, exemplified by its recreation of the Shawshank Redemption prison shown below.
via bilawalsidhu on X
via bilawalsidhu on X
February 27, 2025 at 12:00 PM
The MASt3R-SLAM system can transform a single video into a 3D scene in real-time, exemplified by its recreation of the Shawshank Redemption prison shown below.
via bilawalsidhu on X
via bilawalsidhu on X
For the first time in years, IEA expects electricity demand to rise in advanced economies.
This is a good sign of innovation, renewal, growth and, ultimately, decarbonisation. The future energy system will be led by electricity.
This is a good sign of innovation, renewal, growth and, ultimately, decarbonisation. The future energy system will be led by electricity.
February 26, 2025 at 12:00 PM
For the first time in years, IEA expects electricity demand to rise in advanced economies.
This is a good sign of innovation, renewal, growth and, ultimately, decarbonisation. The future energy system will be led by electricity.
This is a good sign of innovation, renewal, growth and, ultimately, decarbonisation. The future energy system will be led by electricity.
Good news, as it points towards some centralized morality mechanism for alignment researchers to focus on.
February 26, 2025 at 9:55 AM
Good news, as it points towards some centralized morality mechanism for alignment researchers to focus on.
Dark factory in China. No workers, just robots. So might as well save the electricity and keep the lights off.
via @XH_Lee23 on X
via @XH_Lee23 on X
February 25, 2025 at 5:00 PM
Dark factory in China. No workers, just robots. So might as well save the electricity and keep the lights off.
via @XH_Lee23 on X
via @XH_Lee23 on X
Gotta catch 'em all!
Anthropic's new reasoning model, Claude 3.7 Sonnet, can catch a lot more Pokémon and other game items than previous models.
Anthropic's new reasoning model, Claude 3.7 Sonnet, can catch a lot more Pokémon and other game items than previous models.
February 25, 2025 at 1:00 PM
Gotta catch 'em all!
Anthropic's new reasoning model, Claude 3.7 Sonnet, can catch a lot more Pokémon and other game items than previous models.
Anthropic's new reasoning model, Claude 3.7 Sonnet, can catch a lot more Pokémon and other game items than previous models.
Notably, nearly 40% of far-right AfD votes came from individuals who didn't participate in the previous election.
February 24, 2025 at 3:25 PM
Notably, nearly 40% of far-right AfD votes came from individuals who didn't participate in the previous election.
Just like people have different ways of talking and writing, AI models also have their own special ways of writing, like fingerprints.
Researchers could correctly identify which AI wrote a text 97.1% of the time when choosing between five different AIs.
Researchers could correctly identify which AI wrote a text 97.1% of the time when choosing between five different AIs.
February 21, 2025 at 3:12 PM
Just like people have different ways of talking and writing, AI models also have their own special ways of writing, like fingerprints.
Researchers could correctly identify which AI wrote a text 97.1% of the time when choosing between five different AIs.
Researchers could correctly identify which AI wrote a text 97.1% of the time when choosing between five different AIs.
My copy of Silicon Valley Engineer, August 1989. Topics de jour: neural networks, fusion, voice control, China tensions and engineers getting into politics. Plus ça change.
February 20, 2025 at 10:09 AM
My copy of Silicon Valley Engineer, August 1989. Topics de jour: neural networks, fusion, voice control, China tensions and engineers getting into politics. Plus ça change.
The field of robotics has been busy lately. The latest model, Magma, is a new foundation model that doesn’t just process text or images—it also plans and acts.
February 19, 2025 at 6:05 PM
The field of robotics has been busy lately. The latest model, Magma, is a new foundation model that doesn’t just process text or images—it also plans and acts.
Tonight's launch of Grok 3 will be a major test of AI scaling laws.
Built on the Colossus supercomputer with 100k Nvidia H100s, probably 10 times more compute than GPT4.
If scaling laws hold, s Grok 3 should be substantially smarter--or is the era of predictable scaling coming to an end? #ai
Built on the Colossus supercomputer with 100k Nvidia H100s, probably 10 times more compute than GPT4.
If scaling laws hold, s Grok 3 should be substantially smarter--or is the era of predictable scaling coming to an end? #ai
February 17, 2025 at 8:50 AM
Tonight's launch of Grok 3 will be a major test of AI scaling laws.
Built on the Colossus supercomputer with 100k Nvidia H100s, probably 10 times more compute than GPT4.
If scaling laws hold, s Grok 3 should be substantially smarter--or is the era of predictable scaling coming to an end? #ai
Built on the Colossus supercomputer with 100k Nvidia H100s, probably 10 times more compute than GPT4.
If scaling laws hold, s Grok 3 should be substantially smarter--or is the era of predictable scaling coming to an end? #ai
The race for pharmaceutical weight loss is on. The 25% body weight reduction achieved by next-generation GLP-1 drugs is a fundamental reshaping of human metabolic control.
Paper in Nature: www.nature.com/articles/d41...
Paper in Nature: www.nature.com/articles/d41...
February 13, 2025 at 3:13 PM
The race for pharmaceutical weight loss is on. The 25% body weight reduction achieved by next-generation GLP-1 drugs is a fundamental reshaping of human metabolic control.
Paper in Nature: www.nature.com/articles/d41...
Paper in Nature: www.nature.com/articles/d41...
A new AI benchmark, “Humanities Last Exam,” saw its best score jump by 17.5 percentage points with the release of OpenAI’s Deep research.
February 6, 2025 at 2:50 PM
A new AI benchmark, “Humanities Last Exam,” saw its best score jump by 17.5 percentage points with the release of OpenAI’s Deep research.
Goldman’s CEO says that AI drafts 95% of an IPO prospectus in minutes, not weeks. The remaining 5%—human judgment and nuance—is where modern banking's value resides.
January 20, 2025 at 5:07 PM
Goldman’s CEO says that AI drafts 95% of an IPO prospectus in minutes, not weeks. The remaining 5%—human judgment and nuance—is where modern banking's value resides.
My year end message: www.exponentialview.co/p/the-year-e...
December 31, 2024 at 8:11 PM
My year end message: www.exponentialview.co/p/the-year-e...
The ChatGPT gift.
November 30, 2024 at 9:14 AM
The ChatGPT gift.
An important result. AI agents are now competitive with human engineers on tough AI r&d tasks when given a time budget of less than 4 hours.
That boundary will only shift out.
Sets the terrain for faster acceleration.
t.co/kbUgxUxWdp
That boundary will only shift out.
Sets the terrain for faster acceleration.
t.co/kbUgxUxWdp
November 23, 2024 at 9:03 AM
An important result. AI agents are now competitive with human engineers on tough AI r&d tasks when given a time budget of less than 4 hours.
That boundary will only shift out.
Sets the terrain for faster acceleration.
t.co/kbUgxUxWdp
That boundary will only shift out.
Sets the terrain for faster acceleration.
t.co/kbUgxUxWdp
Is bluesky finally going to get where other x-alternatives have not?
November 20, 2024 at 11:23 AM
Is bluesky finally going to get where other x-alternatives have not?
did not expect my daily AI use to look like this
November 16, 2024 at 11:10 AM
did not expect my daily AI use to look like this