IvmeLabs/Ivme-Conversate-22M-Base
Text Generation • Updated • 1
Eren Ekşi PRO
ereniko
AI & ML interests
None yet
Recent Activity
updated a model 5 days ago
IvmeLabs/Ivme-Conversate-22M-Base updated a Space 5 days ago
IvmeLabs/README liked a Space 16 days ago
wop/fontendOrganizations
updated a
model 5 days ago
updated a
Space 5 days ago
liked a
Space 16 days ago
updated a
Space 17 days ago
published a
Space 17 days ago
New model ivme
#13 opened 18 days ago
by ereniko
published a
model 18 days ago
replied to AxionLab-official's post 18 days ago
I mean, like, we are talking about a 50M model here. That thing, already in full precision, can barely do basic math. On 2 bits, even a single question being right is impressive.
published a
Space 19 days ago
liked a
Space 20 days ago
Can the space be restarted?
#5 opened 24 days ago
by ereniko
updated a
Space 27 days ago
updated a
Space about 1 month ago
published a
Space about 1 month ago
posted an update about 1 month ago
Post
180
Something concerning is happening with Nvidia's GPU pricing and benchmark transparency.
Look at how other hardware generations work: a new iPhone costs roughly what the previous one did. Same with Samsung Galaxy phones, same with AMD Ryzen CPUs. Price increases are modest and the specs are straightforward to compare. Nvidia has broken from this pattern. Each new generation costs meaningfully more than the last, while the benchmarks get harder to interpret, not easier.
Look at what's happening with the numbers. BF16 and FP16 performance figures, the ones most relevant for real workloads, are increasingly buried or absent. Sparsity is baked into headline numbers without always being clearly flagged. The B200 is being benchmarked as an NVLink pair rather than a single card in some comparisons. Each of these is defensible in isolation, but together they make it genuinely hard to do apples-to-apples comparisons across generations.
The inference vs training tension is worth flagging too. Lower precision formats are good for inference but introduce real tradeoffs for training. As Nvidia pushes further into inference optimization, NVFP4 being the latest example, it raises a fair question about where training workloads go. NVFP4 training exists, but documentation and real-world results are thin enough that most teams aren't seriously evaluating it yet.
The "ROCm is immature" narrative is also getting a bit stale. I've personally run workloads on the MI300X and honestly it felt smoother than I expected, even compared to CUDA in some cases. And the price to performance ratio isn't even close. AMD isn't a consolation prize anymore.
At some point, labs and enterprises doing the actual buying will have enough alternatives like AMD MI300X, custom silicon, and inference-optimized ASICs to push back. Nvidia is standing on thin ice, and a few more steps like this might be enough to crack it.
Look at how other hardware generations work: a new iPhone costs roughly what the previous one did. Same with Samsung Galaxy phones, same with AMD Ryzen CPUs. Price increases are modest and the specs are straightforward to compare. Nvidia has broken from this pattern. Each new generation costs meaningfully more than the last, while the benchmarks get harder to interpret, not easier.
Look at what's happening with the numbers. BF16 and FP16 performance figures, the ones most relevant for real workloads, are increasingly buried or absent. Sparsity is baked into headline numbers without always being clearly flagged. The B200 is being benchmarked as an NVLink pair rather than a single card in some comparisons. Each of these is defensible in isolation, but together they make it genuinely hard to do apples-to-apples comparisons across generations.
The inference vs training tension is worth flagging too. Lower precision formats are good for inference but introduce real tradeoffs for training. As Nvidia pushes further into inference optimization, NVFP4 being the latest example, it raises a fair question about where training workloads go. NVFP4 training exists, but documentation and real-world results are thin enough that most teams aren't seriously evaluating it yet.
The "ROCm is immature" narrative is also getting a bit stale. I've personally run workloads on the MI300X and honestly it felt smoother than I expected, even compared to CUDA in some cases. And the price to performance ratio isn't even close. AMD isn't a consolation prize anymore.
At some point, labs and enterprises doing the actual buying will have enough alternatives like AMD MI300X, custom silicon, and inference-optimized ASICs to push back. Nvidia is standing on thin ice, and a few more steps like this might be enough to crack it.
posted an update about 2 months ago
Post
318
I don't know why, but lately there's been a growing problem on HuggingFace: the platform is filled to the brim with datasets of reasoning traces from large AI models. I feel like someone should address this, but the dataset tab is just full of reasoning and output traces from models like Claude Opus and the model tab is full of fine-tunes trained on these.
What scares me most are the legal consequences of this, and the possibility that all models will start converging on the same tone because everyone is just fine-tuning on everyone else.
What scares me most are the legal consequences of this, and the possibility that all models will start converging on the same tone because everyone is just fine-tuning on everyone else.
- 1 reply
updated a
model 2 months ago
published a
model 2 months ago
published a
Space 2 months ago
updated a
Space 2 months ago