Post
1451
CoT: 💬💬💬💬
Astra: 🧠🧠🧠💬
Astra: 🧠🧠🧠💬
AI & ML interests
Benchmark, Code Generation, LLM
Post
980
My take on device-side inference: it's all about high bandwidth memory (thinking about it, this holds for the cloud too).
MacBooks enjoy incidental capacity of apple silicon, but per-device RAM is too low (16 to 24GB), only sufficient for a decent SLM. 512GB is the highest you can go (Kimi K2*). Counting MLX downloads of Kimi K2* on Huggingface, I estimate the user base to be <25K.
On the other hand, the newly debuted DGX station (Nvidia) has 748GB, which can fit in the latest Kimi, DS, and Qwen. Also the quantization options of CUDA is way better than MLX.
For high-end inferencing, I place my bet on workstations over Macs.
MacBooks enjoy incidental capacity of apple silicon, but per-device RAM is too low (16 to 24GB), only sufficient for a decent SLM. 512GB is the highest you can go (Kimi K2*). Counting MLX downloads of Kimi K2* on Huggingface, I estimate the user base to be <25K.
On the other hand, the newly debuted DGX station (Nvidia) has 748GB, which can fit in the latest Kimi, DS, and Qwen. Also the quantization options of CUDA is way better than MLX.
For high-end inferencing, I place my bet on workstations over Macs.
- 7 replies
Post
158
My biggest takeaway from Ben Thompson interview is the legacy of this round of "bubble".
For railroad, it was the track, and dot com, fiber. Compared to these two, GPUs depreciate too fast.
The answer is power plants.
For railroad, it was the track, and dot com, fiber. Compared to these two, GPUs depreciate too fast.
The answer is power plants.
Post
2008
Back of the envelope calculation: Ox Alpha has been out for a week with 130K users and 7T tokens burned.
Method 1: Assuming Ox Alpha is the GLM 5.3 class, it has roughly the same active param count as DeepSeek V4 Pro (~40B vs 49B), derive the cost by the floor price DS has ever published.
Method 2: Assuming the users are concentrated within an 8-hour working window each day, derive number of H800 nodes needed (~700 nodes at $2/GPU-hour).
In both methods, I assume 90/10 IO split and 60% cache hit. Both methods come to $2M.
I think the ROI is awesome: (1) publicity and (2) data harvesting.
Method 1: Assuming Ox Alpha is the GLM 5.3 class, it has roughly the same active param count as DeepSeek V4 Pro (~40B vs 49B), derive the cost by the floor price DS has ever published.
Method 2: Assuming the users are concentrated within an 8-hour working window each day, derive number of H800 nodes needed (~700 nodes at $2/GPU-hour).
In both methods, I assume 90/10 IO split and 60% cache hit. Both methods come to $2M.
I think the ROI is awesome: (1) publicity and (2) data harvesting.
- 1 reply
Post
107
I for one believe DeepMind will ship one more frontier model.
In any case GCP is irrelevant here. GDM should still have more compute than DeepSeek, Kimi, or GLM, maybe all of them combined.
In any case GCP is irrelevant here. GDM should still have more compute than DeepSeek, Kimi, or GLM, maybe all of them combined.
Post
3230
Lots of attentions are now on GPU residual value. I find car analogy to be useful.
* Both new and used cars can do the same job (you can run the latest model on 6-year-old A100)
* New cars are more efficient (higher performance per power draw)
* It's not the year, but mileage and maintenance
The last point is undeveloped. We need Carfax and KBB for GPUs.
And there will be lemon GPUs.
* Both new and used cars can do the same job (you can run the latest model on 6-year-old A100)
* New cars are more efficient (higher performance per power draw)
* It's not the year, but mileage and maintenance
The last point is undeveloped. We need Carfax and KBB for GPUs.
And there will be lemon GPUs.
- 1 reply
Post
2767
This is an easy-to-remember pattern.
GLM and Kimi are in Beijing, DeepSeek and Qwen are in Hangzhou.
OpenAI and Anthropic are in SF, xAI and Meta are in the peninsula.
In both China and the Bay Area, token price of one area is ~1/3 of the other area.
GLM and Kimi are in Beijing, DeepSeek and Qwen are in Hangzhou.
OpenAI and Anthropic are in SF, xAI and Meta are in the peninsula.
In both China and the Bay Area, token price of one area is ~1/3 of the other area.
- 1 reply
Post
1756
I think contributor pricing is a good idea. But things will get sophisticated very quickly.
* data sharing tiers from exclusive to syndication
* provenance and lineage
* scoping such as duration and geofencing
and their levels of enforcement (best practices, standards, laws)
* data sharing tiers from exclusive to syndication
* provenance and lineage
* scoping such as duration and geofencing
and their levels of enforcement (best practices, standards, laws)
Post
2672
Looking forward to the Qwen 3.8 model drop, and congratulations on joining the trillion parameters club.
But my eyes are on the promised 27B model. Small (<50B) models decline on OpenRouter, because they are being run on local devices. If people see the family trees here on HF they will understand.
But my eyes are on the promised 27B model. Small (<50B) models decline on OpenRouter, because they are being run on local devices. If people see the family trees here on HF they will understand.
- 2 replies
Post
2435
DeepSeek plans to raise token prices. I don't think this is because they are bleeding, but they are overwhelmed. If your price is 1/10 of your affiliate vendors, you can't leverage their resources. Markup is the only way to diverge traffic away.
Sadly I haven't found discussions on differentiators enabling DS to balance cost at such low prices. All software solutions (that we know of) are accessible by other vendors. If you attribute it to electricity or hardware, you can't explain why GLM and Kimi charge so much for their APIs.
This is where our attention should be (but distracted by things above).
Sadly I haven't found discussions on differentiators enabling DS to balance cost at such low prices. All software solutions (that we know of) are accessible by other vendors. If you attribute it to electricity or hardware, you can't explain why GLM and Kimi charge so much for their APIs.
This is where our attention should be (but distracted by things above).
- 5 replies
Post
145
Jeff Dean departure is 100% unexpected. It's impossible to process this.
And yes, it matters more than other exec shakeups today.
And yes, it matters more than other exec shakeups today.
Post
2355
Many developers discovered that the native DeepSeek API has higher cache-hit rate than neocloud APIs hosting the same DS models.
My speculation is that DS aggressively kills its old models. It has released 18 models thus far, and only 2 are being served now (v4 pro and flash).
This is tough to customers who don't want to upgrade (migrate or leave), but effectively boost the serving capacity to the same model, i.e. more woods behind fewer arrows.
My speculation is that DS aggressively kills its old models. It has released 18 models thus far, and only 2 are being served now (v4 pro and flash).
This is tough to customers who don't want to upgrade (migrate or leave), but effectively boost the serving capacity to the same model, i.e. more woods behind fewer arrows.
Post
2229
There has been a leaked memo (now struck down) from the founder of DeepSeek. I'm not here to circulate it, but comment on the minimum-effort evolutionary path he proposed.
LLM->CoT->Agent->Self-improvement->Singularity->Physical
This makes sense to me: even at the agent stage I learn world models much faster than when I learned LLM at the LLM stage.
But this means humans are still needed beyond the digital singularity, until robots can close their own loop: eval, manufacturing, self improvement, i.e. physical singularity.
LLM->CoT->Agent->Self-improvement->Singularity->Physical
This makes sense to me: even at the agent stage I learn world models much faster than when I learned LLM at the LLM stage.
But this means humans are still needed beyond the digital singularity, until robots can close their own loop: eval, manufacturing, self improvement, i.e. physical singularity.
Post
447
The Nvidia paper came down to this: remove synchronization barriers. DeepSeek has already done that with DeepEP (which this paper cited) one layer above.
Every Microsecond Matters: Achieving Near Speed-of-Light Latency in GPU Collectives (2607.16100)
Nevertheless this is great. More users will benefit from this:
Nvidia NCCL > SGLang/vLLM DeepEP
Every Microsecond Matters: Achieving Near Speed-of-Light Latency in GPU Collectives (2607.16100)
Nevertheless this is great. More users will benefit from this:
Nvidia NCCL > SGLang/vLLM DeepEP
Post
114
This is definitely a minority opinion. World model MIGHT be the answer to enterprise AI.
https://arxiv.org/abs/2602.23643
* Tacit knowledge won't be captured in words, much less accessible dataset
* RL is not efficient at all, and labeling gets more and more expensive
* AGI won't cover your specialized tasks
PS: paper folks you missed one from Yann LeCun!!
https://arxiv.org/abs/2602.23643
* Tacit knowledge won't be captured in words, much less accessible dataset
* RL is not efficient at all, and labeling gets more and more expensive
* AGI won't cover your specialized tasks
PS: paper folks you missed one from Yann LeCun!!
Post
115
Hi, it's been a while 🤗
K3 stunned the world (again). But if you followed my post on K2, the architectural lineage is super clear
https://onekq1.substack.com/p/beyond-the-bonsai-how-kimi-k3s-quantization
I will next discuss the K3 optimizer (Muon again)
K3 stunned the world (again). But if you followed my post on K2, the architectural lineage is super clear
https://onekq1.substack.com/p/beyond-the-bonsai-how-kimi-k3s-quantization
I will next discuss the K3 optimizer (Muon again)
Post
424
GPT 5.1 codex didn't make SOTA either. This should conclude 2025. No model has ever reached above 0.8.
onekq-ai/WebApp1K-models-leaderboard
Can this leaderboard be saturated in 2026?
onekq-ai/WebApp1K-models-leaderboard
Can this leaderboard be saturated in 2026?