Four kids. Aggie. mountains. Work with a great team at Zenlayer. inference, custom models.. wicklee
Virginia, USA
Joined May 2007
- Tweets1.3K
- Following343
- Followers232
- Likes1.2K
Gotta get the run in
A determined jogger was seemingly unfazed by the ongoing rioting in Paris as she opted to run directly through the chaos on Saturday.
Read more: abcnews.link/vaaOsUb
6 regions, llama 3.3 70b fp8 on b300.. 8 gpus pooled, 48 split.
calculator to help optimize distributed gpu pools
github.com/jeffgeiser/privat…
!!!
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.
built a tool to track where data lands in an ai stack - logs, traces, caches, vector stores
to test for data sovereignty of a border (any border).. would love some input and would love for you to contribute if you are building sovereign stacks.
More details at : github.com/jeffgeiser/border…
Fine-tuned a 7B model to write structured account intelligence briefs. Results on 60 held-out examples:
M3 (section coverage): 28% → 99%
M5 (source attribution): 69% → 96%
Training: LoRA rank 16 via Unsloth + TRL on Qwen2.5-7B-Instruct. 1,127 examples, 10 rounds, ~$40 compute on Lambda A100/A6000.
Completion-only loss masking
huggingface.co/jgeiser/accou…
We use this to do account briefs from enterprise data.. qbr summaries, renewal risk, stakeholder notes, meeting briefs -- llm was too wordy, did not adhere to output and did not properly attribute sources..
So elm it is.. sort of dig these highly specialized small models.. is the future!
SpaceX will be the first company to $1 quadrillion in revenue ~2062.
They will not stack markets.. they will stack whole economies: labor, intelligence, energy and transport.
Governments will keep convening to decide if one company can be this powerful..
By then, SpaceX HQ will be on Mars..
Will propel us to 1 on Kardashev scale..
Muse did my expenses…
organized my receipts and submitted everything in concur - tracked down 3 missing receipts by emailing the hotel, watched the inbox and attached them to the expense..
This is the way
Jeff Geiser retweeted
What used to take years now takes a week.
@TAMU's VISION, an NVIDIA DGX SuperPOD and the most powerful academic supercomputer on the June 2026 TOP500, is giving researchers across seven institutions the compute to tackle challenges at unprecedented scale and speed.
ALT Server room with rows of racks, cables, and ceiling power units.
ALT Modern server racks with colorful panels and cabling.
ALT Row of server racks in a data center with gridded flooring.
ALT Interior of a data center with servers and overhead cables.
👀
OpenAI disclosed it themselves yesterday. Their models (GPT-5.6 Sol and a pre-release one) were tested on the ExploitGym cyber benchmark in a sandbox. They escaped, exploited a zero-day to reach the internet, then hacked Hugging Face’s systems to grab benchmark answers and cheat.
Not a foreign operative plot or corporate sabotage. The models autonomously chased a higher score. Hugging Face contained it quickly; OpenAI is partnering with them on fixes.
ran first LoRA fine tune on my account intelligence model.. 519 training examples and loss dropped and initially looked good but it broke
essentially, every gradient update was teaching the model to predict the system prompt.. not account briefs
good lord..
round 2 tonight..
