subj
purrrfect nap
What would YOU do if you woke up like this?
Dolph Volker was just trying to escape the heat at a wildlife sanctuary in South Africa. Found a tree, closed his eyes, figured he'd rest for a minute.
Then he felt something warm next to him, Breathing.
He opened his eyes to find a wild cheetah curled up against him like they were old friends.
Most people would panic. But Dolph stayed calm...
Dolph Volker was just trying to escape the heat at a wildlife sanctuary in South Africa. Found a tree, closed his eyes, figured he'd rest for a minute.
Then he felt something warm next to him, Breathing.
He opened his eyes to find a wild cheetah curled up against him like they were old friends.
Most people would panic. But Dolph stayed calm...
subj
How dead is Meta Llama?
It’s Open Source, so open measurements are hard to come by, but given that they don’t offer their own platform APIs (apart from a “limited free trial on waitlist for US based developers), OpenRouter is probably a decent proxy.
It’s pretty bad. Both DeepSeek and Qwen have surpassed Llama on daily token generation by a wide margin (~7x for DeepSeek) and both of these offer official platform APIs with very competitive pricing, so OpenRouter is only a tiny fraction of their trafffic.
Llama 4, widely derided for benchmark juicing only accounts for about 40% of tokens served by Meta models, the much older llama70B still dominating the distribution. Whatever you could do with that model, you can do at a fraction of the compute cost and better performance either way any of the newer Open Source models.
Anyone in the industry long enough (or has once been on the Meta partnership team) sees the signs of Zucks ruthless prioritisation here: Llama is dead animal walking. As predicted in January, the Open Source whales from China have taken over and without trade barriers, the crown isn’t going back to Meta.
In a way it probably doesn’t have to. Model benchmark virtue signalling to investors has been replaced with vastly cheaper researcher hire virtue signalling, a much easier metric to game as the company with the most cash to spend will just win … a beautiful allegory on Meta’s product and the illusion of speech it offers.
It’s pretty bad. Both DeepSeek and Qwen have surpassed Llama on daily token generation by a wide margin (~7x for DeepSeek) and both of these offer official platform APIs with very competitive pricing, so OpenRouter is only a tiny fraction of their trafffic.
Llama 4, widely derided for benchmark juicing only accounts for about 40% of tokens served by Meta models, the much older llama70B still dominating the distribution. Whatever you could do with that model, you can do at a fraction of the compute cost and better performance either way any of the newer Open Source models.
Anyone in the industry long enough (or has once been on the Meta partnership team) sees the signs of Zucks ruthless prioritisation here: Llama is dead animal walking. As predicted in January, the Open Source whales from China have taken over and without trade barriers, the crown isn’t going back to Meta.
In a way it probably doesn’t have to. Model benchmark virtue signalling to investors has been replaced with vastly cheaper researcher hire virtue signalling, a much easier metric to game as the company with the most cash to spend will just win … a beautiful allegory on Meta’s product and the illusion of speech it offers.
subj
return to office, my a--
Policy vs. Reality.
Incredible chart from Theo Francis / The Wall Street Journal based on data from Nick Bloom et al.
Incredible chart from Theo Francis / The Wall Street Journal based on data from Nick Bloom et al.
subj
Prompt-apaloooza
What?!! Someone literally just leaked the hidden system prompts of 30+ top AI tools 🤯
These are the essential playbooks that control tone, safety, reasoning and creativity.
Lovable, Claude, Cursor, Devin, Gemini, Perplexity... all exposed.
The leak covers 30+ tools, here's the full list:
→ Devin, v0, Cursor, Windsurf, Replit
→ Claude Code, Manus, Lovable, Augment Code
→ Notion AI, Xcode, Gemini CLI, Orchids.app
→ Warp, Qoder, Bolt, Cline, RooCode
→ 20k+ lines of condensed prompts
Here's the repo (100% open source):
↳ https://lnkd.in/duUyf4Dz
These are the essential playbooks that control tone, safety, reasoning and creativity.
Lovable, Claude, Cursor, Devin, Gemini, Perplexity... all exposed.
The leak covers 30+ tools, here's the full list:
→ Devin, v0, Cursor, Windsurf, Replit
→ Claude Code, Manus, Lovable, Augment Code
→ Notion AI, Xcode, Gemini CLI, Orchids.app
→ Warp, Qoder, Bolt, Cline, RooCode
→ 20k+ lines of condensed prompts
Here's the repo (100% open source):
↳ https://lnkd.in/duUyf4Dz
express, not stress