I still build cloud stuff and brew beer.
Strongsville, OH
Joined March 2009
- Tweets4.1K
- Following539
- Followers1.2K
- Likes942
I tried the exact same prompt from Deepseek API and Opencode Go and got 15min 40s @ 131.7 tok/s with 137K tokens used vs 3 hr 20 min @ 21.3 tok/s and 284K tokens used. Very similar quality result however. Deepseek API is MUCH faster.
🧵 DeepSeek-v4.1-Flash-1017 - Same prompt, same harness, same reasoning effort, but two different providers.
DeepSeek Platform (45m 38s):
• Cache hit: 99.7%
• Tokens: 136,590
• Cached input: 44,212,736 tok
• Output: 226,572 tok
OpenCode Go (3h 25m 3s)
• Cache hit: 96%
• Tokens: 1,409,656
• Cached input: 33,611,264 tok
• Output: 224,378 tok
Qwen3.8 27B quality was so-so. It took many prompts to finally get it right, was only getting about 16 tok/sec so painful.
Finally got my side task bars back in Windows. I stopped using Windows almost altogether with Win11 because I loathed having the taskbar on the bottom. Better late than never, but still probably not going to use much Win11 as the change made me get much more comfy with Mac.
Just tried this on older RTX 3080 12GB machine with 64GB RAM. Getting about 40 tok/s in quick testing with Qwen3.8-Flash-Next. Impressive - almost as good as my M5 Max 128
Just doubled the prompt processing speeds from 539 token/s to 1290 tokens/s.
Qwen3.8-Flash-Next on consumer GPU is now blazing fast.
Next on the list is further optimizations on kernels and swapping which should boost results by 3X-4X.
github.com/Niko1221/Strata
"be_inaudible". Job well done. Couldn't understand a word of what she said.
Here's a demo of a conversation between two people at Deveillance @be_inaudible
Watch the live transcription fall apart.
I really like @FluidVoiceApp. I would love it if they added a deepgram compatible API endpoint running locally too. That way I can integrate Anarlog's transcription option to run through FluidVoice.
I can't provide enough power to my M5 mac during local AI usage when GPU is pegged at 100%. It will slowly drain the battery on Automatic mode and very quickly on High Powered mode. How are people handling this so it can run 24x7? Low-power mode is too slow.
I have 3 non-trivial rust applications. Together, their target directories were over 600MB. It's kinda insane.
Holy shit. Fable refused to work on a prompt that was literally: "Ok, now tell me what should build next?". What a fucking joke @AnthropicAI
Fable for new features is not bad. Fable for reviewing what you have already done: dog shit. Absolute dog shit. It will refuse to review your own code for fear of security bollocks. Fix this @AnthropicAI
Fable has been a disaster. It keeps failing in OpenCode and pi and won't continue. Each time it fails and I ask it to restart it costs me about $1 to watch it fail again after 8 secs. @AnthropicAI I want my money back. I have paid over $100 for virtually nothing.
I just drained $90 in OpenCode credits for Fable in a code review and Fable failed to return the review. I got literally nothing for $90. It keeps saying "session reports returned empty" and dying.