I make companies AI native. Codex, claude, Hermes & local LLMs are my jam.
Austin, TX
Joined October 2021
- Tweets1.1K
- Following183
- Followers1.1K
- Likes1K
When I use nordvpn + tailscale together, tailscale stops working. TIL you can setup tailscale exit nodes so traffic routes through my desktop machine at home. Same security benefits and no need for nordvpn anymore.
When you use grok bot a lot, it’s hard to go back to slack or telegram with your agents. The ui is smooth and fast. Example: bot drafts an email and you get 2 buttons (send/discard). In slack that’s another reply you have to type. I get shit done faster now with bot.
GLM 5.3 flash is good at frontend design. So I wondered if it can be as good as Opus 5.5 at video editing & animations. With enough tweaks, it can! Here's my current workflow with @HyperFrames_ (made this with glm + @ElevenLabs).
Thanh Pham retweeted
Vibe coded this last week and now use it all day — global keyboard shortcut to summarize what's on your screen in 100ms, using a Cerebras model. Helpful in a world awash with the tokens of overly verbose agents flocrivello.com/summarize/
Currently training my glm to have same taste as opus so it can make badass videos and animations. With opus I used 700 million tokens last few days on coding and video stuff. Hopefully custom playbooks/skills will be good enough before I make lora adapters.
Right now grok bots and chat's dot is my vibe. UI is so much better than slack/telegram with hermes (and I love hermes).
Set up my grok bot with a phone number with cosvoice and it called my hotel asking for a late checkout and it worked. Holy hell this is going to unlock so many use cases.
Seeing others and myself where bot and muse became less useful because major sites are now blocking them. Lots of use cases go away. Ugly interim outlook as most don’t want bot traffic. Opportunity for those who do embrace agents to capture early market share.
Current models:
Opus 5.5 = cracked 🤯
Astra 6 = daily non-tech driver
Fable 5.1 = orchestrator
Grok 4.7 = meh
Gemini who?
Opus editing videos with hyperframes & thumbnails. Codex + computer use for Youtube management. This would normally take FOREVER. This workflow saved me easily weeks of procrastination and days of work. Crazy times we are in.
Thanh Pham retweeted
Here's a 45-second TL;DR on Jev.
I find the core idea beautifully simple, but the video made it really hard to understand.
Hope you find it helpful.
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Everyone who runs local ai can relate: the models have flashes of brilliance and sometimes beat the frontier models on similar tasks, but then sometimes are also subpar. This is where size/parameter counts do matter. Frontier can handle more edge cases with their size.
I despise filing tx franchise controller stuff every year. Mindless filing. Told codex to do it for me. With computer use it was done in < 5 min. Even organized my local folders right after. The chad is too good to me.
Ran my personal bug hunting & fixing benchmark on deepseek v4.1 flash and it beat v4 (and qwen 3.8 flash and glm 5.3 flash). It seem like a good upgrade so far, running on @MiaAI_lab recipe. Running on thinking=high is potent. More tests coming.
With the recent herdr update I can now launch claude/codex/grok sessions on my Mac mini and studio and see it in one view. Distribute the ram usage and I remote login to manage agents. Don’t see need to have powerful mbp in future.