Conrad Lippert-Zajaczkowski
@conradlz.bsky.social
Leader | 30K+ Hours coding | Computer Scientist | Innovator @ Enterprise | Founder-Mode
9 of the top 10 github trending repos currently on Github are from deepseek-ai
Wow
Wow
January 29, 2025 at 4:37 PM
9 of the top 10 github trending repos currently on Github are from deepseek-ai
Wow
Wow
Feel really good to enjoy bluesky with all the other platforms changing with political climate.
January 25, 2025 at 4:33 AM
Feel really good to enjoy bluesky with all the other platforms changing with political climate.
CUDA and HIP both aren't good programming languages. Triton exists because the smartest kernal devs were wasting days manually optimizing code that ended up buggy anyways.
My money is on one of the ai asic startup companies making a programming language that sticks in the near future.
My money is on one of the ai asic startup companies making a programming language that sticks in the near future.
December 24, 2024 at 1:13 AM
CUDA and HIP both aren't good programming languages. Triton exists because the smartest kernal devs were wasting days manually optimizing code that ended up buggy anyways.
My money is on one of the ai asic startup companies making a programming language that sticks in the near future.
My money is on one of the ai asic startup companies making a programming language that sticks in the near future.
Happy Holidays!
Thought provoking idea: the most studied pure functions are ones which operate on numeric input and produce a numeric output.
This is an incredibly small portion of all possible functions and operations that could be programmed.
Thought provoking idea: the most studied pure functions are ones which operate on numeric input and produce a numeric output.
This is an incredibly small portion of all possible functions and operations that could be programmed.
December 23, 2024 at 3:26 PM
Happy Holidays!
Thought provoking idea: the most studied pure functions are ones which operate on numeric input and produce a numeric output.
This is an incredibly small portion of all possible functions and operations that could be programmed.
Thought provoking idea: the most studied pure functions are ones which operate on numeric input and produce a numeric output.
This is an incredibly small portion of all possible functions and operations that could be programmed.
Let's treat this one with enthusiastic participation.
People will certainly be affected, but only those made aware will voice their opinions.
People will certainly be affected, but only those made aware will voice their opinions.
It's time that journalists start taking the discussion regarding the potential benefits and risks of AGI seriously, to not repeat the missed opportunity with climate change decades ago:
www.niemanlab.org/2024/12/the-...
www.niemanlab.org/2024/12/the-...
The media reckons with AGI
"Rather than treat AGI as a fringe concern, we must be proactive and ambitious: taking the possibility seriously, considering the implications, and starting a public, democratic conversation."
www.niemanlab.org
December 20, 2024 at 11:23 PM
Let's treat this one with enthusiastic participation.
People will certainly be affected, but only those made aware will voice their opinions.
People will certainly be affected, but only those made aware will voice their opinions.
The ARC-Prize competition concluded with mindsai scoring 55.5% on ARC-AGI o3 would have won by a huge margin.
arcprize.org
arcprize.org
December 20, 2024 at 6:20 PM
The ARC-Prize competition concluded with mindsai scoring 55.5% on ARC-AGI o3 would have won by a huge margin.
arcprize.org
arcprize.org
Reposted by Conrad Lippert-Zajaczkowski
OpenAI skips o2, previews o3 scores, and they're truly crazy. Huge progress on the few benchmarks we think are truly hard today. Including ARC AGI.
Rip to people who say any of "progress is done," "scale is done," or "llms cant reason"
2024 was awesome. I love my job.
Rip to people who say any of "progress is done," "scale is done," or "llms cant reason"
2024 was awesome. I love my job.
December 20, 2024 at 6:08 PM
OpenAI skips o2, previews o3 scores, and they're truly crazy. Huge progress on the few benchmarks we think are truly hard today. Including ARC AGI.
Rip to people who say any of "progress is done," "scale is done," or "llms cant reason"
2024 was awesome. I love my job.
Rip to people who say any of "progress is done," "scale is done," or "llms cant reason"
2024 was awesome. I love my job.
Hyper Connections
Can nueral networks learn the optimal strength of connections in order to improve performance?
Excellent question by ByteDance and even better answer in the paper:
arxiv.org/abs/2409.19606
Can nueral networks learn the optimal strength of connections in order to improve performance?
Excellent question by ByteDance and even better answer in the paper:
arxiv.org/abs/2409.19606
Hyper-Connections
We present hyper-connections, a simple yet effective method that can serve as an alternative to residual connections. This approach specifically addresses common drawbacks observed in residual connect...
arxiv.org
December 20, 2024 at 5:43 PM
Hyper Connections
Can nueral networks learn the optimal strength of connections in order to improve performance?
Excellent question by ByteDance and even better answer in the paper:
arxiv.org/abs/2409.19606
Can nueral networks learn the optimal strength of connections in order to improve performance?
Excellent question by ByteDance and even better answer in the paper:
arxiv.org/abs/2409.19606
What does this say about our economy and society?
December 19, 2024 at 5:35 PM
What does this say about our economy and society?
HunyuanVideo from Tencent
A new large video foundation model with open source weights, training, and model code
github.com/Tencent/Huny...
A new large video foundation model with open source weights, training, and model code
github.com/Tencent/Huny...
GitHub - Tencent/HunyuanVideo: HunyuanVideo: A Systematic Framework For Large Video Generation Model
HunyuanVideo: A Systematic Framework For Large Video Generation Model - Tencent/HunyuanVideo
github.com
December 18, 2024 at 6:25 PM
HunyuanVideo from Tencent
A new large video foundation model with open source weights, training, and model code
github.com/Tencent/Huny...
A new large video foundation model with open source weights, training, and model code
github.com/Tencent/Huny...
Interesting idea for simultaneous training in multiple environments with a shared replay buffer.
github.com/Albiemc1303/...
github.com/Albiemc1303/...
GitHub - Albiemc1303/TLR_Framework-: TLR (Triple Layer Training) is a multi-environment reinforcement learning framework that allows AI agents to train simultaneously across multiple environments.
TLR (Triple Layer Training) is a multi-environment reinforcement learning framework that allows AI agents to train simultaneously across multiple environments. - Albiemc1303/TLR_Framework-
github.com
December 17, 2024 at 4:24 PM
Interesting idea for simultaneous training in multiple environments with a shared replay buffer.
github.com/Albiemc1303/...
github.com/Albiemc1303/...
According to Ilya, the era of pre-training is beginning to sunset. What era do you want next?
December 14, 2024 at 9:49 PM
According to Ilya, the era of pre-training is beginning to sunset. What era do you want next?
Reposted by Conrad Lippert-Zajaczkowski
Pre-training as we know it will end - Dr. Ilya Sutskever at NeurIPS 2024
December 13, 2024 at 11:17 PM
Pre-training as we know it will end - Dr. Ilya Sutskever at NeurIPS 2024
Reposted by Conrad Lippert-Zajaczkowski
Don't worry everyone, I have figured out a way to delay any potential AI takeover indefinitely. If you know you know. (Turn sound on)
December 13, 2024 at 4:26 AM
Don't worry everyone, I have figured out a way to delay any potential AI takeover indefinitely. If you know you know. (Turn sound on)
Impressive and elegant approach to causal discovery in this competition-winning work by github user "thetourney".
thetourney.github.io/adia-report/
thetourney.github.io/adia-report/
Adia Lab Causal Discovery
thetourney.github.io
December 12, 2024 at 10:54 PM
Impressive and elegant approach to causal discovery in this competition-winning work by github user "thetourney".
thetourney.github.io/adia-report/
thetourney.github.io/adia-report/
I want to train AI to build Gundam, and have them compete in a tournament to see which is the ultimate fighter.
December 10, 2024 at 11:06 PM
I want to train AI to build Gundam, and have them compete in a tournament to see which is the ultimate fighter.
Shoutouts to Meta for open sourcing SPDL(Scalable and Performant Data Loading).
I can easily imagine a future where software 2.0 (model-based software) has its own frameworks, languages, and tooling where these pieces will be foundational.
github.com/facebookrese...
I can easily imagine a future where software 2.0 (model-based software) has its own frameworks, languages, and tooling where these pieces will be foundational.
github.com/facebookrese...
GitHub - facebookresearch/spdl: Scalable and Performant Data Loading
Scalable and Performant Data Loading. Contribute to facebookresearch/spdl development by creating an account on GitHub.
github.com
December 10, 2024 at 9:54 AM
Shoutouts to Meta for open sourcing SPDL(Scalable and Performant Data Loading).
I can easily imagine a future where software 2.0 (model-based software) has its own frameworks, languages, and tooling where these pieces will be foundational.
github.com/facebookrese...
I can easily imagine a future where software 2.0 (model-based software) has its own frameworks, languages, and tooling where these pieces will be foundational.
github.com/facebookrese...
The open-source reasoning model QwQ (Qwen with Questions) is actually making the cost of reasoning cheap.
December 8, 2024 at 4:31 PM
The open-source reasoning model QwQ (Qwen with Questions) is actually making the cost of reasoning cheap.
Reposted by Conrad Lippert-Zajaczkowski
Announcing 🥂 FineWeb2: A sparkling update with 1000s of 🗣️languages.
We applied the same data-driven approach that led to SOTA English performance in🍷 FineWeb to thousands of languages.
🥂 FineWeb2 has 8TB of compressed text data and outperforms other datasets.
We applied the same data-driven approach that led to SOTA English performance in🍷 FineWeb to thousands of languages.
🥂 FineWeb2 has 8TB of compressed text data and outperforms other datasets.
December 8, 2024 at 9:19 AM
Announcing 🥂 FineWeb2: A sparkling update with 1000s of 🗣️languages.
We applied the same data-driven approach that led to SOTA English performance in🍷 FineWeb to thousands of languages.
🥂 FineWeb2 has 8TB of compressed text data and outperforms other datasets.
We applied the same data-driven approach that led to SOTA English performance in🍷 FineWeb to thousands of languages.
🥂 FineWeb2 has 8TB of compressed text data and outperforms other datasets.
Will research engineering become more delegated to agentic systems?
In this paper we have a benchmark to hill-climb with agentic systems.
arxiv.org/abs/2411.15114
In this paper we have a benchmark to hill-climb with agentic systems.
arxiv.org/abs/2411.15114
RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
Frontier AI safety policies highlight automation of AI research and development (R&D) by AI agents as an important capability to anticipate. However, there exist few evaluations for AI R&D capabilitie...
arxiv.org
December 7, 2024 at 8:24 PM
Will research engineering become more delegated to agentic systems?
In this paper we have a benchmark to hill-climb with agentic systems.
arxiv.org/abs/2411.15114
In this paper we have a benchmark to hill-climb with agentic systems.
arxiv.org/abs/2411.15114