Lightnews — Scholar-powered news

LightNews

Shawn Reed

@shekari.bsky.social

140 followers 68 following 63 posts

Self proclaimed good guy with some bad habits including Porsche, Ducati & sugar. Also very lucky to have @sparklyseaster.bsky.social in my life.

Posts Media Videos Starter Packs

Shawn Reed @shekari.bsky.social · 12h

I can’t make this new tool work, it must be the tools fault.

Shawn Reed @shekari.bsky.social · 1d

Open source that can be audited and consensus between a few different models would be a start.

Shawn Reed @shekari.bsky.social · 8d

microwho?

Shawn Reed @shekari.bsky.social · 10d

Dedicated button, no. Specific chip to support? Yes and proof. bsky.app/profile/shek...

Shawn Reed @shekari.bsky.social · 11d

After a day of tuning and testing models, another variable has entered the fray. #AMD #RYZENAIMAX #GMKtec

Shawn Reed @shekari.bsky.social · 11d

After a day of tuning and testing models, another variable has entered the fray. #AMD #RYZENAIMAX #GMKtec

Shawn Reed @shekari.bsky.social · 11d

Some incredible results after tuning gpt-oss a bit. an FP16 GGUF Thanks Unsloth! with a 320k context all in VRAM and >100 tok/sec! #GPT-OSS #llama.cpp #AMD #ROCM

Shawn Reed @shekari.bsky.social · 11d

I also found that ROCm 6.4 was way faster on prompt processing but Vulkan as a bit faster on the streaming speed so Gemini Flash helped me determine the break even point.

Shawn Reed @shekari.bsky.social · 11d

I think I will add in a gpt-oss and produce my own IQ4_NL quant of the Dolphin model as I did for the Gemma 3.

Shawn Reed @shekari.bsky.social · 11d

Will try matching the kv cache quants to the model quant and some more common like q8_0 next.

Shawn Reed @shekari.bsky.social · 11d

Shawn Reed @shekari.bsky.social · 11d

Some fun building llama.cpp with ROCm 6.4, ROCm 7.0 RC2 and Vulkan in addition to testing some kv cache quants. Note, don't bother with iq4_nl for k & v , it pushes prompt processing to CPU. Without quantizing kv, ROCm is beating vulkan in my testing. Dolphin-Mistral-24B-Venice-Edition-Q5_K_M.gguf

Reposted by Shawn Reed

James C. Burchill @jamesburchill.com · 13d

Spring 2026 marks 50 years since I wrote my first lines of code. #AI - What was once sci-fi is now a daily reality. It fuels my work, my creativity, and my return to #solopreneurship. The journey’s come full circle ~ and it still feels like the beginning.

vault.jamesburchill.com/mindset/50-y...

50 Years In, and Still Learning ~ Why I Rebuilt Everything From Scratch (and What It Taught Me) - The Vault

After nearly five decades of tinkering with computers--from the days of beige boxes and clunky keyboards to today’s cloud-driven AI—I’ve come full circle. What started as a childhood fascination with ...

vault.jamesburchill.com

Shawn Reed @shekari.bsky.social · 12d

59 here and it was a Timex Sinclair 1000 but the same experience adding AI into my workflows. I have so many ideas to try out that I can’t even sleep some nights.

Shawn Reed @shekari.bsky.social · 18d

I think I'll get the letters in the words to slowly fade and get refreshed by the matrix rain. Then clean it up and push it up to #GitHub for others to play with.

Shawn Reed @shekari.bsky.social · 18d

I modified the world clock example of the Pimoroni Presto to give it some much needed Matrix. @pimoroni.com #micropython #RaspberryPiPico #Pimoroni

Shawn Reed @shekari.bsky.social · 18d

One example of the errors I was seeing is
```shell
error: Failed to read metadata for: packaging==25.0
Caused by: Failed to parse METADATA file: .venv/lib/python3.12/site-packages/packaging-25.0.dist-info/METADATA
Caused by: Metadata field Name not found
```

Shawn Reed @shekari.bsky.social · 18d

I had something wrong with several dependencies that were in use so I could have / will try the 'uv cache clean' next time but somehow I picture that being identical functionality to the rm . If/when it happens again I will also first try the prune and see if it resolves the issue.

Shawn Reed @shekari.bsky.social · 19d

Never underestimate the healing powers of an occasional rm -rf ~/.cache/uv #python #softwaredevelopment

Shawn Reed @shekari.bsky.social · Aug 25

I spent some time with Gemini CLI today. Made great progress on a local project until I ran out of free Gemini Pro 2.5 requests and it switch me to Gemini Flash. Then it was more hassle fixing the errors it cause... Even the Premium plan is only 1500 a day. #Gemini #Google #SoftwareDevelopment

Shawn Reed @shekari.bsky.social · Aug 23

Cline is WAY better than GitHub Copilot. For my current workflow at least #python #Neo4j #Gemma3-27b-q5

Shawn Reed @shekari.bsky.social · Aug 17

Just a call out that gpt-oss:20b got an update and it now can be shown the error of it's ways. And for some specific tasks it really is a great tool. 30k context and fully loaded into VRAM in 18GB, 87.93 tok/sec, 1.20s to first token in LM Studio - ROCm llama.cpp (Linux) v1.46.0 #ROCm #AMD

Shawn Reed @shekari.bsky.social · Aug 16

#ai #aiartwork #aiartist #Aiartcommunity #aiart

Young girl with headphones not in use down around her neck.

Shawn Reed @shekari.bsky.social · Aug 14

Up most of the night because your brain won’t stop coming up with new software solutions? Yea me too, you’re not alone.

Shawn Reed @shekari.bsky.social · Aug 10

We need people to use critical thinking rather than force safety guardrails to be baked into a model. You don’t dull a knife because it might cut you.

Shawn Reed @shekari.bsky.social · Aug 10

And guardrails can make models perform so poorly that they are worthless for intended purpose. Try to point out to gpt-oss:20b that it is wrong about something. In its thinking response you can see its policy preventing it from listening to you. And it is frequently wrong.