dean (@deankepler) header
dean (@deankepler) profile picture
los angeles, ca
Joined February 2022
Still feeling Oct 11. Pre Oct 11, outside a real black swan, $MMT almost never held a sixty percent positive premium on Binance perps. Arb desks would hit it in seconds and close the spread. Now outsized positive premiums show up daily. After Oct 11 many passive market makers were wiped. Books are thin. It is easier to push price and there is little resistance. Because the books are thin, arbers get more cautious when a huge premium prints. It used to be a fire alarm. A few player would step in together, put balance sheets on the line, and kill it. Now there are fewer player. If you rush in alone and no one follows, you are the one who gets burned...
1
1
1
4,510
cheapest cloud GPUs to run Llama 4 maverick by @vllm_project
1
10
4,872
Identical post about Deepseek R1 70B on 8*RTX 3080: Twitter: 3k views, 0 comments Reddit: 18k views, dozens of technical discussions Guess where the real tech crowd hangs out 😅
1
1
13
3,486
Replying to @MistralAI
you made it!
1
1
12
847
tbh we’ve been underestimating how consumer-grade gpu/cpu clusters can power ai. memory bandwidth might be the only real bottleneck, but for day-to-day tasks—like 70b model inference or quick fine-tuning—it’s really not that big of a deal. there are actually way more of these clusters out there than we realize. if we figure out how to coordinate them— stuff like data distribution, scheduling, and robust networking—there’s a real shot at decentralized ai. it won't replace high-end data centers, but it could totally fill the gap for mid-scale or specialized workloads, making ai more accessible and flexible. (this basically lines up with what nvidia’s project digits is aiming for, too).
1
10
1,577
been testing Deepseek-R1 671B on CPU-only servers (@AMD EPYC 7543, 1TB DDR4-3200). memory bandwidth dominates performance over compute - our FP8 tests achieve 1.91 tokens/s, with theoretical 5.01 tokens/s possible on DDR5. the real bottleneck isn't compute power, but memory access patterns and bandwidth utilization. This suggests three optimization paths: - NUMA-aware scheduling - Access pattern optimization - Memory controller tuning consumer deployment of 671B models might be closer than we think - just needs smarter memory management, not necessarily faster hardware.
1
2
1
17
5,238
DeepSeek's breakthrough has sparked concerns about GPU demand. Many believe their efficient training approach threatens GPU markets - I disagree. First misconception: people think DeepSeek found a way to train with fewer GPUs. actually: • Their method is highly scalable • More GPUs = Direct performance improvements • Solved the diminishing returns problem from PRM (Process Reward Model) • Each additional GPU now yields proportional gains DeepSeek fundamentally reimagined the training pipeline. Before, large models using PRM hit an efficiency wall - they needed extra GPUs just for reasoning supervision. DeepSeek restored linear scaling, making each additional GPU deliver meaningful performance gains. This makes GPUs more valuable, not less. Another misconception: some believe this is about inference optimization. In reality: • It's pure training innovation • Post-training approach that needs similar GPU resources as pre-training • Made traditional inference-time scaling less relevant • Found PRMs useful for inference but not cost-effective the real concern to GPU markets and scaling laws isn't architectural - it's data scarcity. Current large models have consumed almost all available internet data. Getting quality training data will become exponentially harder. We might soon face a situation where we have abundant GPU capacity but insufficient data to utilize it.
2
8
1
36
14,005
It’s ironic that a non-US team is the one truly keeping OpenAI’s original vision alive—real openness, frontier research, and empowering everyone. It’s not what you’d expect, but maybe that’s what makes it so fitting. Sometimes, the unexpected turns out to be the most impactful. DeepSeek-R1 is more than just open-sourcing a stack of models; they’re sharing all the training secrets too. It’s probably the first OSS project showing real, sustained RL flywheel growth. Impact doesn’t need flashy names like “ASI” or “Project Strawberry.” Sometimes, it’s as simple as dropping raw algorithms and matplotlib learning curves for everyone to see. This is what open AI should be.
🚀 DeepSeek-R1 is here! ⚡ Performance on par with OpenAI-o1 📖 Fully open-source model & technical report 🏆 MIT licensed: Distill & commercialize freely! 🌐 Website & API are live now! Try DeepThink at chat.deepseek.com today! 🐋 1/n
1
2
13
7,487
Just completed internal testing of a text-to-image model running locally on an NVIDIA RTX 2080. The throughput and quality delivered results beyond expectations — easily reaching a level suitable for daily use.
5
9
4,476