Like a lot of people, I have been trying out the new gpt-oss:120b, I am getting around 68 tokens/sec. The model is highly censored (guard rails) though.

Back to reality after an enjoyable 4 weeks in British Columbia, Canada.

Join millions who have switched to Grok.

Anyone know what style of painting this might be? It originates from China circa 1900.

After much messing around and a failed attempt at using AI to write the Verilog code for me I can now frequency lock a 40 MHz VCTXO to GPS using an FPGA, I know a bit of an overkill.

Best day so far this year for Solar (52.8 kWh generated).

If you specifically tell it to use the key names or their icons instead of [] it produces a usable output.

New AI trick, using Google notebookLM, upload your radio's manual to it then ask it how to do something (in my case tune a TH-D74 radio to 70 MHz USB). It came up with the right procedural steps, but it referred to every key as the [] key. So it almost got it. 😕.

I have just been using Github Co-pilot to implement code for a 40 MHz GPS conditioned oscillator in Verilog, the code looks good. Now I just need to save up to buy a robot to do the soldering for me. The AI actually wrote the code using Verilog I hadn't thought of using before.

and a nuclear reactor to power them.

It has taken me a week to get ollama / webUI / Proxmox / 2 GPUs / pfSense (firewall / router) and WiFi all to talk together so I can simply query a private LLM from my mobile. I am quite pleased with the Epyc Server motherboard I bought from China. Think I need more GPUs though.

I have just seen an NVIDIA DGX Spark for sale on eBay from a UK seller. I didn't think they were available yet, even from NVIDIA.

Blimey! JORN was an active project when I worked at Marconi

This April was a very sunny month. My PV system generated 1.132 MWh (of which I exported 702 kWh to the grid), the best month last year was June when it generated 1.174 MWh (695 kWh export).

Just bolted Dia-1.6B TTS (Text To Speech) model into my local AI assistant server code. Spent most of the time trying to figure out how to get my Honor Android mobile phone into debugging mode. The manufacturer seems to have added yet another level of security to the process.

I think I have met my own 'bridge too far'. I have been trying to get Oculink to work with zero success so far.

Blimey, these multimodal LLMs are getting pretty good.

It outputs at 23 tokens per sec.

Just as a follow-up here is a simple 50 MHz clock divider example. I also got it to write an I2C controller (for an upcoming project). The model is Mixtral:8x22b (80 GB in size). It has taken most of the day to download the model and get the Web interface to work.

After 2 days of Linux tomfoolery I managed to remove Proxmox from my multiple GPU thread ripper system and get Ubuntu 24.04.2 and LM Studio running on it. Turns out AI is pretty good at writing Verilog modules, which will save me time when writing boring boiler plate FPGA code.

Out on my daily walk, lovely again today.