Paul Frazee
@pfrazee.com
310K followers 540 following 35K posts
CTO at Bluesky. I'm on Germ DM 🔑 https://ger.mx/A6lLhakn-kJcja1Rlx6gOuwFvCEyrvK4y9lDSo6anFmU#did:plc:ragtjsm2j2vknwkz3zp4oxrd
Posts Media Videos Starter Packs
pfrazee.com
Tangled has hit the bigtime
pfrazee.com
Can I implement lexui and create an entire bluesky client using server-driven ui that's exchanged via json? Let's find out
pfrazee.com
My #1 usecase for claude code is to just wild out on an idea while I'm doing something else
pfrazee.com
that would be smart
pfrazee.com
Defending it by saying it's a bad experience for anybody and lying that it didn't just happen to me
pfrazee.com
Throwing down wild mandates to the product team that I should never scroll down and see a better post on a topic than the one I just did
pfrazee.com
dot boston is a TLD? that's cool
pfrazee.com
my alt has a real girlfriend. or something. is that a good post
pfrazee.com
Don't tread on my robogirlfriend. is that a good post.
pfrazee.com
I don't recall how I showed up in that debate the first time around, but when I saw this again recently I wasn't really stoked. I get the reasoning but it feels complex to me, and makes lexicon publishing more difficult
pfrazee.com
The purpose is to enable separation of repositories for different security needs. You can reuse the same repo, but app risks (posting to bluesky) are less intense than lexicon publishing risks.

We use two separate repos, and the bsky-lexicons one is the applicaiton-level handle for our lexicon repo
pfrazee.com
This was somewhat hotly debated internally, and I personally have mixed feelings about it, but I get why it was done this way.
pfrazee.com
See here: atproto.com/specs/lexico...

Lexicon domain names are mapped using a different dns record than handle resolution
Lexicon - AT Protocol
A schema definition language.
atproto.com
pfrazee.com
typically around 4k, it can vary per model though
pfrazee.com
lol no, not to replace me postes
pfrazee.com
It was also, typically, seven numbers. Long distance calls cost money! And with landlines you'd change numbers when you moved
pfrazee.com
I just ask it some questions. In the power user or dev modes, it tells you the throughput
pfrazee.com
I'll have to look into it
pfrazee.com
If anybody gets better let me know!
pfrazee.com
deepseek-r1-052b-qwen-8b
35 tok/s

qwen-coder-30b
59 tok/s

gemma-3n-e4b
42 tok/s

gpt-oss-20b
57 tok/s

gpt-oss-120b
27 tok/s
pfrazee.com
Got the framework desktop working with 96GB allocated to GPU. The working software is LM Studio using Vulcan. ROCm crashes the model and ollama crashes gnome

Benchmarks following:
pfrazee.com
I’m not sure how to react