mips_avatar

7 ms·

91.
by mips_avatar 4mo ago
Fine for me. Not for thee
92.
by mips_avatar 4mo ago
Like if you hired a different services company who decided to sabotage your business that would be fraud.
93.
by mips_avatar 4mo ago
Like if you're using claude code on a feature tangential to your training pipeline it's allowed to nerf itself and damage your AI work.
94.
by mips_avatar 4mo ago
I'm really uncomfortable with these changes, like everything Anthropic's doing as "frontier research" today will be regular product engineering in a year.
95.
If Claude Fable stops helping you, you'll never know(jonready.com)
1036 points by mips_avatar 4mo ago|501 comments
96.
by mips_avatar 4mo ago
They said in their terms and conditions they will silently sabotage you if you do this.
97.
by mips_avatar 4mo ago
It's bad that Anthropic can determine what this means. If you're building a modern app you're likely training your own embedding models and now anthropic can just silently sabotage your training pipelines?
98.
by mips_avatar 4mo ago
Seems like they still won’t let you run models on GPU while the phone is closed or the user switches apps
99.
by mips_avatar 4mo ago
I don't think we've bottomed out on what we can do with embedding models. They're these tiny models that absolutely rip on modern cpus with 8 bit int optimizations. Like in my app we can say pretty definitive things about hun
100.
by mips_avatar 5mo ago
Still somehow was "impossible" for the Seattle police to recover security camera footage of my bike being stolen under the light rail station security camera.
101.
by mips_avatar 5mo ago
I spent two hours the other day trying to figure out how to manage spend on gcp, i gave up and used openrouter and cloudflare.
102.
by mips_avatar 5mo ago
There’s a lot of places where an llm can improve a data pipeline. Like if Claude sonnet was free I’d do a lot of data enrichment.
103.
by mips_avatar 5mo ago
They have to do that because their pricing is bad. Like if you want to run a custom vector db you can run it 20x cheaper on hetzner than aws.
104.
by mips_avatar 5mo ago
I trained karpathys d28 1.6b nanochat on a 3090. Took an extremely long time but I did it.
105.
by mips_avatar 5mo ago
You can fully train a 1.6b model on a single 3090. That’s a reasonably big model.
106.
by mips_avatar 6mo ago
Microsoft is trying to sell things like extended servicing agreements. They purposefully make Windows worse so they can sell you solutions to fix it. They purposefully keep it insecure so you need their updates. It’s about taking the custo
107.
by mips_avatar 6mo ago
I think if modern LLMs were invented in the mid 2010s it would have been promoted in more positive ways, but because everyone is afraid for their economic security saying scary things gets more of a response. I think it's kind of gross
108.
by mips_avatar 6mo ago
It's the same problem as fireworks, the only models supporting LORA are like year old dense models that perform horribly on most tasks. If you want to do anything close to relevant you still need to rent/own dedicated GPUs, which
109.
by mips_avatar 6mo ago
Maxes out around 4K tok/s output. Each pair of 3090s has its own instance of the model with parallelism across the nvlink bridge. Though nvlink is only 2x over pcie5
110.
by mips_avatar 6mo ago
Yeah the model i'm running locally is just one of several models the app supports and it falls back to others if not available.
111.
by mips_avatar 6mo ago
I haven't benchmarked against a pro 6000, it's more that i have 4 3090s and i don't have a pro 6000.
112.
by mips_avatar 6mo ago
I think these gpus were actually used for bitcoin mining before I bought them
113.
by mips_avatar 6mo ago
4 3090s with nvlinks on each pair. Super fast inference on Moe models around 20-36b
114.
by mips_avatar 6mo ago
with Argo networking
115.
by mips_avatar 6mo ago
So it's basically just openrouter with cloudflare argo networking? I feel like they could do some much more interesting stuff with their replicate acquisition. Application specific RL is getting so good but there's no good way to
116.
by mips_avatar 7mo ago
He's roommates with an Anthropic researcher, I was roommates with a Google product manager I don't think I'm really bought out by Google.
117.
by mips_avatar 7mo ago
Well it’s just more that I’ve noticed in the agents I’ve built that qwen doesn’t get reliable until around 27b so unless you want to rl small qwen I don’t think I would get much useful help out of it.
118.
by mips_avatar 7mo ago
Have you tried any really big models on a mac studio? I'm wondering what latency is like for big qwens if there's enough memory.
119.
by mips_avatar 7mo ago
I think the notes about Alexander I’ve heard is just how obvious his were
120.
by mips_avatar 7mo ago
Yeah on campus apparently

More ›