Dave @php100

I'm running older versions of Qwen like 2.5 Coder 7B MLX on a M2 Pro 16GB machine and for the majority of my use cases I'm covered with this model. If the terminal command fails with a local model it runs an auxiliary model with Gemma 4 on Cerebras (Until they replace it with Qwen3.8 27B on September 3rd) or Groq has Qwen3.8 27B at 500 tokens per second. Works well. I've learned by optimizing a short output length I can use very little in resources because each call is being used for a micro unit-level sub-task like reading directories, searching for specific functions, searching for specific lines in files, etc. Most of this is already mapped out in my tool directory so it's blazing fast with a 500+ tps model with structured output For local model, I could use a larger model my hardware allows it, but I prefer to optimize for smallest possible model to complete the task and live in an environment that will understand when it fails a task. This allows for me to run secondary embedding services on one machine or parakeet for voice to text and kokoro for text to speech.

0 0 0 172 0

View Details

Dave @php100

3 days ago

I cracked it! I put 5B tokens through voice-to-voice via Codex over the 25-26th. I was able to recover, analyze, and properly store all traces and artifacts for further benchmarking and data analysis. After the post-mortem, all I can say is this: the abilities I’ve unlocked through pure voice interaction are making me realize the power of 100% voice orchestration with my operating systems. It feels like being a wizard who can move mountains with the flick of my tongue. We’re unlocking a new paradigm of possibility. It’s only up from here. Faster, leaner, more efficient systems. Thank you, Codex, for allowing me to use third-party harnesses, constantly resetting my usage limits, and releasing features that truly improve my day-to-day workflow across personal and business tasks. ...Now all you need to do is stop encrypting your model reasoning tokens.

Tibo @thsottiaux

3 days ago

Your Codex and ChatGPT Work reset will land at 6pm PST. I challenge you to burn as much usage as you can and discover and the latest awesome features. Time to go /fast

3K 577 14K 2.3M 1K

0 0 2 816 1

View Details

Tencent Hy @TencentHunyuan

6 days ago

🚀 Hy4 preview is here. 770B, 49B active, 1M context. Built for productivity. Open source frontier. Consistent affordable price. Use it. Tell us what breaks. More on Hy blog：hy.tencent.ai/research/hy4-p… HuggingFace：huggingface.co/tencent/Hy4-pr… Github：github.com/Tencent-Hunyua…

336 561 5K 2.4M 1K

View Details

Thariq is on vacation @trq212

6 days ago

life is stranger than fiction

93 135 3K 178K 453

View Details

BullBear.News @bullbear_info

a week ago

@TheAhmadOsman my terminal is my entire personality so locking me into a web ui is a federal offense. need that tui yesterday

0 1 7 651 0

View Details

Dave @php100

a week ago

@0xSero subscribed tysm for your work - lurker

0 0 2 100 0

View Details

Mia @MiaAI_lab

2 weeks ago

Best local AI models for 8-256gb of VRAM 8gb VRAM: Qwen3.8-27B IQ1_S huggingface.co/unsloth/Qwen3.… 16gb VRAM: Qwen3.8-27B Q2_K_XL huggingface.co/unsloth/Qwen3.… RTX 3090: Qwen3.8-27B Q4_K_M huggingface.co/unsloth/Qwen3.… RTX 5090: Qwen3.8-27B NVFP4 github.com/MiaAI-Lab/Qwen… RTX PRO 6000: Qwen3.8-27B NVFP4 github.com/MiaAI-Lab/Qwen… DGX Spark: DeepSeek v4 Flash / Qwen3.8-27B github.com/MiaAI-Lab/Deep… github.com/MiaAI-Lab/Qwen… 2x DGX Sparks: DeepSeek v4 Flash github.com/MiaAI-Lab/Deep… Qwen3.8-27B and DeepSeek v4 Flash are running the show, but this might change soon 👀

165 262 3K 181K 4K

View Details

Dave @php100

2 weeks ago

@colbyboles @ekremcetinkaya_ @0xSero In Obsidian, folders (directories) are vaults, and vaults are folders. They are exactly the same thing on your hard drive.

0 0 0 22 0

View Details

Dave @php100

2 weeks ago

@thsottiaux Hey Tibo I just wanted to say thank you and the Codex team for such a fantastic product and service. My favorite experience is 5.6 Sol as orchestration via Codex Voice then delegating multi-project level tasks while having discussion in a single thread. It's a great experience.

0 0 4 134 0

View Details

Dave @php100

2 weeks ago

Exa's free tier is great, $10 mo gets you: ~1,428 basic search queries ~10,000 full-text page contents ~833 deep search queries I avoid implementing their service into production flows for a few logical reasons (scale cost, privacy, etc) I use Exa MCP for in-the-moment ideas that I want to quickly probe into. I also use Exa as a research tool for deterministic tasks running on daily cron jobs.

2 0 4 643 1

View Details

Dave @php100

2 weeks ago

I implemented the core idea of the Autoprompt skill but not the skill itself. The value for me is because it formalized this loop: scope → implement → independent review → repair → verification I use Hermes which already gives me kanban state, bounded delegation, strict TDD, review gates, and verification-before-completion. Running this skill directly on top of Hermes would add duplicate orchestration for my system. github.com/Spielewoy/auto…

0 0 2 155 4

View Details

Dave @php100

4 weeks ago

@_sunnymind @goodhunt omp mentioned

0 0 3 138 0

View Details

Dave @php100

a month ago

it's codex szn

1 0 8 1K 0

View Details

Dave @php100

3 months ago

@ExaAILabs Awesome marketing 20k searches per month on free tier is very generous but hard to justify your pricing at scale.. Does the improvement in result quality justify paying 4-5x more than say.. BrightData for most production use cases?

1 0 2 358 0

View Details

Dave @php100

3 months ago

@davis7 Why do you store private/sensitive personal and business information on Discord? Curious because Discord lacks e2e encryption and they have automated systems constantly scanning your server for policy violations.