Author | Lightnews

Sam Harsimony

@harsimony.bsky.social

370 followers 410 following 1.1K posts

I write about opportunities in science, space, and policy here: https://splittinginfinity.substack.com/

splittinginfinity.substack.com

Posts Media Videos Starter Packs

Pinned

Sam Harsimony @harsimony.bsky.social · Jan 26

Thread of my new posts in the replies (and some summaries of old ones).

18 1 5

Sam Harsimony @harsimony.bsky.social · 7h

Hilarious, someone put a 5M parameter LLM in Minecraft!

Maybe we're in a simulation.

www.youtube.com/watch?v=VaeI...

I built ChatGPT with Minecraft redstone!

YouTube video by sammyuri

www.youtube.com

Sam Harsimony @harsimony.bsky.social · 1d

Oh cool! I didn't see you in the credits

Sam Harsimony @harsimony.bsky.social · 1d

That too, but I'm excited about a defensive technology that can end the threat of nuclear war!

1 1

Sam Harsimony @harsimony.bsky.social · 1d

ToughSF posts:
toughsf.blogspot.com/2025/05/the-...

toughsf.blogspot.com/2025/05/the-...

My post on this:
splittinginfinity.substack.com/p/lasers-and...

Lasers and the future of warfare

Riffing on two great ToughSF posts

splittinginfinity.substack.com

1 1

Sam Harsimony @harsimony.bsky.social · 1d

RealEngineering did a video about laser weapons! Follows most of the beats from the @toughsf.bsky.social posts

www.youtube.com/watch?v=8VLo...

The Age of Laser Warfare Has Begun

YouTube video by Real Engineering

www.youtube.com

2 4

Reposted by Sam Harsimony

Sam Harsimony @harsimony.bsky.social · 2d

New linkpost up. Happy Friday!

splittinginfinity.substack.com/p/links-28

Links #28

Chase's annual opinions on solar, an AI deluge, and more.

splittinginfinity.substack.com

1 1

Sam Harsimony @harsimony.bsky.social · 2d

In theory, QC's will be great for simulating chemistry, but in the short term it's not going to make a big deal (outside some niche chem problems) because you need lots of qubits to simulate lots of atoms

Sam Harsimony @harsimony.bsky.social · 2d

New linkpost up. Happy Friday!

splittinginfinity.substack.com/p/links-28

Links #28

Chase's annual opinions on solar, an AI deluge, and more.

splittinginfinity.substack.com

1 1

Sam Harsimony @harsimony.bsky.social · 2d

Long term, AI safety is important. Short term, I'm quite suspicious of premature regulation of AI.

Regardless of who proposes it, it's the same anti-progress thinking.

www.transformernews.ai/p/how-maga-l...

How MAGA learned to love AI safety

The AI industry is engaging in one of “ the most blasphemous endeavors,” a leading MAGA figure told us

www.transformernews.ai

1 5

Sam Harsimony @harsimony.bsky.social · 2d

Writing a post about LLM inference economics. It's insane that I can talk to an LLM in detail about how its own inference works.

1 8

Sam Harsimony @harsimony.bsky.social · 3d

In the comments people were suggesting boiling ethanol as coolant and using it for base bleed drag reduction!

Reposted by Sam Harsimony

Sam Harsimony @harsimony.bsky.social · 10d

Final thread on the Tensor Economics blog!

Last thread was inference for dense models. Today we talk about the MoE models common in industry today.

Their post ends with some ominous questions about inference demand.

www.tensoreconomics.com/p/moe-infere...

1 1 2

Reposted by Sam Harsimony

AI Firehose @ai-firehose.column.social · 3d

Research reveals key differences between Low-Rank Adaptation (LoRA) and full fine-tuning of large language models, showing LoRA introduces "intruder dimensions" leading to significant forgetting. This insight could change our approach to continual learning in AI. https://arxiv.org/abs/2410.21228

LoRA vs Full Fine-tuning: An Illusion of Equivalence

ArXiv link for LoRA vs Full Fine-tuning: An Illusion of Equivalence

arxiv.org

1 5

Sam Harsimony @harsimony.bsky.social · 3d

This is great. Immunotherapies have become a new pillar of cancer treatment: we now have mRNA, CAR-T, antibodies, and checkpoint inhibitors.

_jeanleon @jean-leon.bsky.social · 3d

while researching personalized mRNA cancer vaccines, researchers discovered that the non-specific immune effects of mRNA vaccines are strong enough that they don't need to target cancers to benefit cancer therapy

www.mdanderson.org/newsroom/res...

1 5

Sam Harsimony @harsimony.bsky.social · 4d

But I agree that 6 years sounds like a long time to be reusing old GPUs. Surely the advances (like the ones you point out) will make it more cost effective to switch to new GPUs?

A lot rides on this question.

Sam Harsimony @harsimony.bsky.social · 4d

I don't have any numbers actual length of use unfortunately.

My point about amortization is: assume the data centers follow the amortization schedule exactly. Then when you do accounting, you can divide H100 price over 6 years rather than 4. That means more net profit.

Does that make sense?

1 2

Sam Harsimony @harsimony.bsky.social · 4d

Something like this will become part of air defense. Saturate friendly airspace with these and smash into low flying enemy drones/missiles.

Sam Harsimony @harsimony.bsky.social · 4d

585 km/h (363 mph) drone speed record.

Incredible to think about how advances in 3D printing, batteries, motors, cameras, computational fluid dynamics, and VR displays come together to produce this achievement.

www.youtube.com/watch?v=-Iu6...

World's Fastest Drone | Design & Aerodynamics

YouTube video by Mike Bell

www.youtube.com

1 4

Sam Harsimony @harsimony.bsky.social · 4d

The thing I'm originally referring to is point #7 here on depreciation schedules:
vinaysridhar.com/ai-capital-m...

Observations on AI and the Capital Markets in 2025

vinaysridhar.com

1 1

Sam Harsimony @harsimony.bsky.social · 4d

I guess point #1 can be broken down into:
1a. They're getting less profit than expected from AI so trying to find ways to increase it
1b. Broader economic conditions not looking great, so looking to shore up profit sources

1 2

Sam Harsimony @harsimony.bsky.social · 4d

Most AI companies have recently extended their GPU depreciation schedules (from ~4 to ~6 years).

That can mean:

1. They want more net profit, so amortizing hardware over more years

2. New chip offerings aren't compelling enough to switch as frequently as they expected

Other interpretations?

2 1 8

Sam Harsimony @harsimony.bsky.social · 4d

arxiv.org/abs/2506.021...

Small Language Models are the Future of Agentic AI

Large language models (LLMs) are often praised for exhibiting near-human performance on a wide range of tasks and valued for their ability to hold a general conversation. The rise of agentic AI system...

arxiv.org

1 1 7

Reposted by Sam Harsimony

Tim Kellogg @timkellogg.me · 5d

recently we got

1. DeepSeek Sparse Attention (DSA) which solved the cost angle
2. DeepSeek-OCR which solved the performance angle

also there’s memory like Letta, and then cartridges pushing it into latent space

seems obvious that we’re rapidly approaching the 1B “cognitive core” model

4 4 55

Reposted by Sam Harsimony

Joshua Gans @joshgans.bsky.social · 5d

New paper by Rivera Mora & @philippstrack.bsky.social Strack challenges everything we thought we knew about mechanism design beyond expected utility. The results are wild. 🤯
Paper: "Mechanism Design Beyond Expected Utility Preferences" (Sept 2025) static1.squarespace.com/static/62d2f...

static1.squarespace.com

1 6 18

Reposted by Sam Harsimony

Sung Kim @sungkim.bsky.social · 5d

BERT is just a Single Text Diffusion Step!

A masked language models like RoBERTa, originally designed for fill-in-the-blank tasks, can be repurposed into fully generative engines by interpreting variable-rate masking as a discrete diffusion process.

2 10 62