Hacker News: WASDx

New comment by WASDx in "DeepSeek V4 Pro 0813"

WASDx — Thu, 13 Aug 2026 17:02:29 +0000

On DeepSWE it's now 53% vs 63% which is one of the coding benchmarks I trust the most. DS own measurements also show a more significant increase so I suspect AA might update when they release an article.

Surprisingly DeepSWE currently shows a lower total cost for pro so that might also update I guess. As usual, don't trust the benchmarks and try for yourself.

New comment by WASDx in "DeepSeek Harness developer preview"

WASDx — Thu, 13 Aug 2026 16:54:23 +0000

I needs to be harness+model combination, https://artificialanalysis.ai/agents/coding-agents

New comment by WASDx in "ChatGPT Desktop (Codex Desktop) for Linux"

WASDx — Thu, 13 Aug 2026 09:52:22 +0000

Right. I'm also curious about those use cases. I don't want an AI clicking through my mailbox.

New comment by WASDx in "ChatGPT Desktop (Codex Desktop) for Linux"

WASDx — Thu, 13 Aug 2026 08:52:06 +0000

I upgraded my workflow a few months ago from "copy-paste things in and out of ChatGPT" to "use an agent that edits my project files and runs tests on its own" and the ergonomics are just so much better and enables automating bigger tasks. I still monitor everything it does and do manual adjustments so I feel ownership of the code.

New comment by WASDx in "DeepSeek V4 Pro 0813"

WASDx — Thu, 13 Aug 2026 06:56:13 +0000

This was a disappointed to me. Why would I use pro over flash now? Is there some area where the difference is significant?

New comment by WASDx in "Nvidia Nemotron 3.5 Lightning and NeMo Switchyard"

WASDx — Tue, 11 Aug 2026 21:06:57 +0000

If you take 10 turns with a model A, it has to read the cache 10 times and write a lot of tokens (the expensive part). Switching to model B is just prefilling the diff + your new message, which is still just one turn. So total number of turns does not increase for a long session even with many switches.

I didn't understand this before reading the sibling comments so I'm not sure I got it fully right but I think the total cost becomes like this:

* Input tokens: Pay for both models * Output tokens: Pay for the model that generates * Cached tokens: Pay per turn, so in total a weighted average over both models?

Since output tokens are the most expensive, I can see how this is an overall win for many use cases as benchmarks also show. The hard part is routing correctly.

New comment by WASDx in "Muse Code and Muse Spark 1.2"

WASDx — Thu, 06 Aug 2026 06:59:38 +0000

And they are all TUI's installed via curl | bash.

New comment by WASDx in "DeepSeek V4 Flash on a Single AMD MI300X"

WASDx — Tue, 04 Aug 2026 11:17:12 +0000

At 830tok/s * 1 hour that's almost 3M tokens which is just $0.54 worth of tokens at Deepseeks current output price.

New comment by WASDx in "The new rules of context engineering for Claude 5 generation models"

WASDx — Sun, 26 Jul 2026 11:04:35 +0000

Do you know why they don't just cache the system prompt for everyone? It seems so wasteful not to.

New comment by WASDx in "Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber"

WASDx — Tue, 21 Jul 2026 16:34:11 +0000

DeepSWE and FrontierCode are more realistic if you read up on what they actually measure. But the most realistic is to try it yourself. Benchmarks can only vaguely represent typical usage, and how you judge the result. Giving the same real task you have to a few models will make you understand them better than chasing benchmarks.

New comment by WASDx in "Kimi K3: Open Frontier Intelligence"

WASDx — Sun, 19 Jul 2026 20:08:14 +0000

Great explanation, thanks!

New comment by WASDx in "Kimi K3: Open Frontier Intelligence"

WASDx — Fri, 17 Jul 2026 21:51:57 +0000

> On top of that, doing research in the open amortizes the cost.

Can you elaborate on this? I appreciate the open models but don't see the economics behind just giving them away like now.

New comment by WASDx in "OpenAI unveils its first custom chip, built by Broadcom"

WASDx — Wed, 24 Jun 2026 20:31:25 +0000

I see only these two possibilities:

1. If LLMs keep improving, burning models onto silicon becomes obsolete too fast and is not worth doing. Outcome: We keep getting better LLMs. 2. If LLM improvements slow down, they will be burned onto silicon. Outcome: We get faster, cheaper and energy-efficient LLMs.

Either way sounds great to me. It will certainly be a mix so we can even get both.

New comment by WASDx in "GLM-5.2 is the new leading open weights model on Artificial Analysis"

WASDx — Wed, 17 Jun 2026 16:19:07 +0000

Are you suggesting it should summarize the image in text or generate it in HTML or something else?

New comment by WASDx in "Running local models is good now"

WASDx — Tue, 16 Jun 2026 17:30:01 +0000

Looking at some benchmarks, the latest ~30B Gemma/Qwen score similar as Claude or GPT versions that were released just one year earlier. That's crazy progress. I can't imagine how it will be in a few years.

New comment by WASDx in "Open source AI must win"

WASDx — Sat, 13 Jun 2026 19:45:42 +0000

I think this is inevitable. Sooner or later, model-specific ASIC's will make economical sense. We're already seeing it happening with Taalas/Cerebras so I think it's sooner than 5 years. And inference is order of magnitude faster which is amazing.

New comment by WASDx in "Open source AI must win"

WASDx — Sat, 13 Jun 2026 19:25:22 +0000

> distributed LLM inference

This seems extremely inefficient considering data transfer between model layers if the model is distributed. I found this project called Petals that claim up to 4 tok/s for a 180B model although its repository hasn't been updated in two years.

https://petals.dev/

New comment by WASDx in "Claude Fable 5"

WASDx — Tue, 09 Jun 2026 20:15:44 +0000

I like this one, although its data seem to overlap with ECI.

https://artificialanalysis.ai/trends

New comment by WASDx in "Real-time LLM Inference on Standard GPUs: 3k tokens/s per request"

WASDx — Fri, 29 May 2026 16:44:18 +0000

https://chatjimmy.ai/ from Taalas also feels like that.

New comment by WASDx in "Claude Opus 4.8"

WASDx — Thu, 28 May 2026 22:02:16 +0000

I think their "code" ranking is biased towards visual aesthetics more than raw coding as the voters are just asked which generated website they prefer.