AGI, Inc. research night
Edge Cases
Edge Cases is a research night for on-device AI, hosted by AGI, Inc. in San Francisco. One researcher presents their work to 15 to 25 people, then we have dinner.
Next edition: date to be announced. AGI HQ, San Francisco.
Speakers have come from
Past speakers
Tom Bishop
Glass Imaging, Co-founder and CTO
PhD in computational photography
Previously Apple, iPhone Portrait Mode
Edge Cases
#003
·
Mochamad Asri
NVIDIA, Lead Architect
UT Austin, PhD in Electrical and Computer Engineering
Edge Cases
#002
·
Adam Awale
AGI, Inc., on-device research team
Arizona State University, PhD candidate, MPS Lab
Edge Cases
#001
·
What it is
Edge Cases is for work that has to run on constrained hardware: on-device inference, small models, efficient training, processing-near-memory, computer use and the action layer, agentic systems on phones and edge devices.
Most of our team came out of research labs and we wanted to keep that part of the work alive. About 25 researchers and engineers from labs and OEMs around the Bay come to each edition. One speaker presents, we ask questions, then we have dinner.
The format
1 talk per night
Slides or whiteboard, both fine.
40 to 60 minutes
Plus Q&A off the record.
15 to 25 seats
In person, every one approved.
0 recordings
So you can show work that is not public yet.
SF
AGI HQ. Evenings, about once a month. The address goes out after approval.
Dinner
After the talk, with open discussion.
For researchers, PhD students and research engineers. Free to attend.
Editions
Three editions so far. Every talk was given in person and none was recorded.
#004
Next edition · Date to be announced
Speaker to be announced
Applications are open. Subscribe on Luma to get the date first.
#003
Neural image restoration on device
Tom Bishop
·
Glass Imaging, Co-founder and CTO
As camera pixels shrink, the optics hit a physical wall and traditional image processors cannot fix the blur. His team trains a neural processor on the actual degradation and recovers 2.5 to 3x the resolution of a traditional pipeline.
#002
System Architectures for Near-Memory Execution
Mochamad Asri
·
NVIDIA, Lead Architect
How he thinks about bandwidth, latency, and end-to-end QoS for serving generative AI and LLMs.
#001
Accelerating on the edge: Optimizing PNM
Adam Awale
·
AGI, Inc., on-device research team
DPC, a DRAM-PNM compiler built with Prof. Shrivastava, up to 96% faster processing-near-memory without touching the hardware. Paper under peer review.
Have work that has to run on a device? Present it here.
Topics we host
Anything that has to run on constrained hardware. Recent and upcoming talks cover:
On-device inference and inference engines
Small models, quantization and efficient training
NPUs, accelerators and processing-near-memory
Computer use, mobile use and the action layer
Agentic systems on phones, PCs, wearables and cars
On-device vision and computational photography
Hybrid inference across device and server
Who should apply
You are working on something that has to run on constrained hardware.
Published, preprint, work in progress, or internal-and-shareable all welcome. You do not need a finished paper.
We do not record the talks and the Q&A is off the record, so you can show work that is not public yet.
One presenter per night, so we schedule a few weeks out.
In person in San Francisco. We do not do remote talks.
Researchers, PhD students and research engineers. This is not a pitch slot for a product.
We select for technical depth and a result the audience can question, not product demos.
How it works
1
Apply
About 5 minutes. Tell us what you are working on and what you would show.
2
We read it
We come back to you with a date.
3
You present
40 to 60 minutes, then Q&A and dinner.
FAQ
Present your research at the next Edge Cases.
Published, preprint or work in progress. The application takes about 5 minutes.
