Christian Laforte
@chrlaf.bsky.social
ML Engineer at NVIDIA. Previously: Stealth GPU startup; Stability AI; AMD; Autodesk; CEO of 2 startups (3D + AI). Toronto, Canada
They are preparing to rename to the Department of Forced Labor
January 13, 2026 at 5:37 AM
They are preparing to rename to the Department of Forced Labor
Reposted by Christian Laforte
Got our RL research codebase at @ai2.bsky.social running on my Nvidia DGX Spark. Fun times ahead and lots of learnings to share :)
January 13, 2026 at 1:47 AM
Got our RL research codebase at @ai2.bsky.social running on my Nvidia DGX Spark. Fun times ahead and lots of learnings to share :)
Reposted by Christian Laforte
Oh wow, deepseek is starting to make serious progress on LLMs that offload memory to external storage: github.com/deepseek-ai/...
github.com
January 12, 2026 at 6:44 PM
Oh wow, deepseek is starting to make serious progress on LLMs that offload memory to external storage: github.com/deepseek-ai/...
Reposted by Christian Laforte
One of my favorite findings: Positional embeddings are just training wheels. They help convergence but hurt long-context generalization.
We found that if you simply delete them after pretraining and recalibrate for <1% of the original budget, you unlock massive context windows. Smarter, not harder.
We found that if you simply delete them after pretraining and recalibrate for <1% of the original budget, you unlock massive context windows. Smarter, not harder.
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
January 12, 2026 at 4:12 AM
One of my favorite findings: Positional embeddings are just training wheels. They help convergence but hurt long-context generalization.
We found that if you simply delete them after pretraining and recalibrate for <1% of the original budget, you unlock massive context windows. Smarter, not harder.
We found that if you simply delete them after pretraining and recalibrate for <1% of the original budget, you unlock massive context windows. Smarter, not harder.
Reposted by Christian Laforte
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
January 12, 2026 at 4:07 AM
Introducing DroPE: Extending Context by Dropping Positional Embeddings
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
We found embeddings like RoPE aid training but bottleneck long-sequence generalization. Our solutionβs simple: treat them as a temporary training scaffold, not a permanent necessity.
arxiv.org/abs/2512.12167
pub.sakana.ai/DroPE
I agree. There are super low hanging fruits like this paper that each can give 5X-10X speed ups and improvement in accuracy. arxiv.org/abs/2510.06824
Efficient numeracy in language models through single-token number embeddings
To drive progress in science and engineering, large language models (LLMs) must be able to process large amounts of numerical data and solve long calculations efficiently. This is currently only possi...
arxiv.org
January 11, 2026 at 7:25 PM
I agree. There are super low hanging fruits like this paper that each can give 5X-10X speed ups and improvement in accuracy. arxiv.org/abs/2510.06824
So at best, that paper is inconclusive. Hopefully the author will share his code soon so we can reproduce his findings and improve upon them.
January 8, 2026 at 4:16 PM
So at best, that paper is inconclusive. Hopefully the author will share his code soon so we can reproduce his findings and improve upon them.
... they measured worse performance:
"""The original paper claims Grassmann flow layers achieve perplexity "within 10-15% of size-matched Transformers" on Wikitext-2. ... using the exact architecture specified in the paper, reveals a 22.6% performance gap - significantly larger than claimed."""
"""The original paper claims Grassmann flow layers achieve perplexity "within 10-15% of size-matched Transformers" on Wikitext-2. ... using the exact architecture specified in the paper, reveals a 22.6% performance gap - significantly larger than claimed."""
January 8, 2026 at 4:15 PM
... they measured worse performance:
"""The original paper claims Grassmann flow layers achieve perplexity "within 10-15% of size-matched Transformers" on Wikitext-2. ... using the exact architecture specified in the paper, reveals a 22.6% performance gap - significantly larger than claimed."""
"""The original paper claims Grassmann flow layers achieve perplexity "within 10-15% of size-matched Transformers" on Wikitext-2. ... using the exact architecture specified in the paper, reveals a 22.6% performance gap - significantly larger than claimed."""
The paper seems promising... I think this could work much better for continuous signals like image and video diffusion models.
Unfortunately the author didn't share any code and so far I only found one attempt at reproducing his findings.
elliotarledge.com used Claude Code and...
Unfortunately the author didn't share any code and so far I only found one attempt at reproducing his findings.
elliotarledge.com used Claude Code and...
Elliot Arledge
Systems engineer and educator. Building and teaching GPU programming, CUDA, and low-level ML systems.
elliotarledge.com
January 8, 2026 at 4:15 PM
The paper seems promising... I think this could work much better for continuous signals like image and video diffusion models.
Unfortunately the author didn't share any code and so far I only found one attempt at reproducing his findings.
elliotarledge.com used Claude Code and...
Unfortunately the author didn't share any code and so far I only found one attempt at reproducing his findings.
elliotarledge.com used Claude Code and...
Reposted by Christian Laforte
When will we learn our lesson?? Our continued dependency on fossil fuel will not only kill people in the long term, it creates geopolitical instability and kills people in the short term. All to satisfy the greed of the broligarchy. We have the capability to move to carbon free sources.
January 6, 2026 at 12:31 PM
When will we learn our lesson?? Our continued dependency on fossil fuel will not only kill people in the long term, it creates geopolitical instability and kills people in the short term. All to satisfy the greed of the broligarchy. We have the capability to move to carbon free sources.
Reposted by Christian Laforte
βcorrupt and kleptocratic governmentβ fits Trumpβs government eminently well, especially if you add βrogue aggressor with zero regard for international lawβ.
January 6, 2026 at 1:21 PM
βcorrupt and kleptocratic governmentβ fits Trumpβs government eminently well, especially if you add βrogue aggressor with zero regard for international lawβ.
Reposted by Christian Laforte
We will either end fossil fuel dependency and the predatory extractivist basis of global production, or they will end us through wars and ecological collapse - all at the same time.
January 3, 2026 at 3:34 PM
We will either end fossil fuel dependency and the predatory extractivist basis of global production, or they will end us through wars and ecological collapse - all at the same time.
Reposted by Christian Laforte
If your president bombs Nigeria on Christmas just to distract from his crimes, you live in a shithole country.
December 26, 2025 at 3:56 AM
If your president bombs Nigeria on Christmas just to distract from his crimes, you live in a shithole country.
Reposted by Christian Laforte
This whale lives for centuries: its secret could help extend human lifespan www.nature.com/articles/d41...
This whale lives for centuries: its secret could help extend human lifespan
A cold-activated protein that mends damaged DNA could play a part in keeping the bowhead whale in tip-top shape.
www.nature.com
October 30, 2025 at 8:15 AM
This whale lives for centuries: its secret could help extend human lifespan www.nature.com/articles/d41...
Reposted by Christian Laforte
The world might be pulling back on climate truth: but I'm stepping it up.
How? I'm adding new videos, deep dives & behind-the-scenes stories to Talking Climate.
Why? Because climate honesty and hope matter more than ever right now.
Join me on Patreon: bit.ly/47pCLaf
Or Substack: bit.ly/4ntfvhJ
How? I'm adding new videos, deep dives & behind-the-scenes stories to Talking Climate.
Why? Because climate honesty and hope matter more than ever right now.
Join me on Patreon: bit.ly/47pCLaf
Or Substack: bit.ly/4ntfvhJ
October 28, 2025 at 3:41 PM
The world might be pulling back on climate truth: but I'm stepping it up.
How? I'm adding new videos, deep dives & behind-the-scenes stories to Talking Climate.
Why? Because climate honesty and hope matter more than ever right now.
Join me on Patreon: bit.ly/47pCLaf
Or Substack: bit.ly/4ntfvhJ
How? I'm adding new videos, deep dives & behind-the-scenes stories to Talking Climate.
Why? Because climate honesty and hope matter more than ever right now.
Join me on Patreon: bit.ly/47pCLaf
Or Substack: bit.ly/4ntfvhJ
Reposted by Christian Laforte
Climate action is not just about future generations. The choices we make today will help determine the climate impacts that people alive today, people in our lives, will experience over the rest of the century (and beyond).
June 25, 2025 at 4:24 PM
Climate action is not just about future generations. The choices we make today will help determine the climate impacts that people alive today, people in our lives, will experience over the rest of the century (and beyond).
Reposted by Christian Laforte
π₯ Real-time *infinite* video from just 1 frame?
Meet Seaweed-APT2: an 8B-param autoregressive GAN that streams 24fps video at 736Γ416βall in *1 network forward eval*. Thatβs 1440 frames/min, live, on a single H100 π
Project: seaweed-apt.com/2
Meet Seaweed-APT2: an 8B-param autoregressive GAN that streams 24fps video at 736Γ416βall in *1 network forward eval*. Thatβs 1440 frames/min, live, on a single H100 π
Project: seaweed-apt.com/2
Seaweed APT2
Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation
seaweed-apt.com
June 23, 2025 at 3:46 PM
π₯ Real-time *infinite* video from just 1 frame?
Meet Seaweed-APT2: an 8B-param autoregressive GAN that streams 24fps video at 736Γ416βall in *1 network forward eval*. Thatβs 1440 frames/min, live, on a single H100 π
Project: seaweed-apt.com/2
Meet Seaweed-APT2: an 8B-param autoregressive GAN that streams 24fps video at 736Γ416βall in *1 network forward eval*. Thatβs 1440 frames/min, live, on a single H100 π
Project: seaweed-apt.com/2
Reposted by Christian Laforte
got scammed by this practice only by Adobe: no other SaaS did this with me before or after. I swear I signed up to a monthly plan: but in the fine print apparently it was βannual, paid monthly, cancel for a fee (that we wonβt tell you how much it is).
Meanwhile the FTC: www.ftc.gov/news-events/...
Meanwhile the FTC: www.ftc.gov/news-events/...
FTC Takes Action Against Adobe and Executives for Hiding Fees, Preventing Consumers from Easily Cancelling Software Subscriptions
The Federal Trade Commission is taking action against software maker Adobe and two of its executives, Maninder Sawhney and David Wadhwani, forΒ deceiving consumers by hiding the early terminati
www.ftc.gov
June 20, 2025 at 8:16 AM
got scammed by this practice only by Adobe: no other SaaS did this with me before or after. I swear I signed up to a monthly plan: but in the fine print apparently it was βannual, paid monthly, cancel for a fee (that we wonβt tell you how much it is).
Meanwhile the FTC: www.ftc.gov/news-events/...
Meanwhile the FTC: www.ftc.gov/news-events/...
I only skimmed the paper but⦠it sounds like the subjects were asked to write essays about uninteresting topics, that no one would read. Can you really blame the kids for just copy pasting the output of ChatGPT in that circumstance?
June 19, 2025 at 12:30 PM
I only skimmed the paper but⦠it sounds like the subjects were asked to write essays about uninteresting topics, that no one would read. Can you really blame the kids for just copy pasting the output of ChatGPT in that circumstance?
I guess the reasoning is that Apple and Google (ads) are consumer driven businesses. Average consumers will likely suffer under current administration. In contrast, NVIDIA and Microsoftβs corporate customers will continue to grow rapidly as they reap AI-driven savings and new revenue streams.
June 18, 2025 at 3:07 PM
I guess the reasoning is that Apple and Google (ads) are consumer driven businesses. Average consumers will likely suffer under current administration. In contrast, NVIDIA and Microsoftβs corporate customers will continue to grow rapidly as they reap AI-driven savings and new revenue streams.
Reposted by Christian Laforte
My French side is always super proud when we are able to successfully export our national hobby to other countries! π Now let me introduce you to the concept of "unlimited general strike" π
June 15, 2025 at 7:03 AM
My French side is always super proud when we are able to successfully export our national hobby to other countries! π Now let me introduce you to the concept of "unlimited general strike" π
Reposted by Christian Laforte
Dear Americans, on behalf of all Canadians and citizens of the free world, thank you to everyone who turned out, for standing up against tyranny! π
June 14, 2025 at 8:55 PM
Dear Americans, on behalf of all Canadians and citizens of the free world, thank you to everyone who turned out, for standing up against tyranny! π
I suspect the real attendance was much lower. I paid to attend, but I cancelled my trip. I found it unconscionable to attend CVPR in a Republican state. Like most Canadians, I plan to NEVER, EVER buy American unless we absolutely have no other choice.
June 15, 2025 at 4:35 AM
I suspect the real attendance was much lower. I paid to attend, but I cancelled my trip. I found it unconscionable to attend CVPR in a Republican state. Like most Canadians, I plan to NEVER, EVER buy American unless we absolutely have no other choice.