Been reviewing my work at the end of day with voice AI and it's pretty good at finding blindspots if i push for it but bad at holding my attention with long explanations
Voice bot can now search autonomously across the web, extracting details and maintain a summary in any format controlled by voice. Here it is looking up reddit reviews to come up with pros and cons of a product
Built an AI voice bot that can guide through PDFs. Here, it is guiding through an insurance policy document by highlighting specific details and updating a running summary with citations
Experienced the midwit curve where had to research all kinds of techniques to control long seq activations from blowing up memory. Didn't have to apply any of it once i got access to a bigger gpu couple days later. That said, ring attention is really cool
@mynamebedan Would be quite useful for debugging neural nets. A lot of metrics are already visualized to understand training behaviour but they're all proxies for figuring out what's blocking gradient flow or loss
Tuned Mistral 7B to guide one step at a time for coding tasks with SFT on a synthetic dataset. It seems to get stuck in reasoning loops for long multi step tasks tho
Open sourcing code for sharing screen with gpt-4. Also, added options for open vision models like Llava and Bakllava.
Aim's to develop open agents that can specialize and assist with coding and using complex tools. You can try the demo here - https://t.co/S5uJVmQDBl
Iteration 2 of screensharing with gpt-4. Reduced latency, converted the cube into a sphere.
Also, it can summarize and take notes for later use.
https://t.co/gPQFTuSnpo