Arbaaz Khan (@graphonfounder) header
Arbaaz Khan (@graphonfounder) profile picture

AI Scientist. @GraphonAI Founder. ML Hacker. Robotics PhD

San Francisco
Joined October 2013
Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)Photo by Arbaaz Khan (@graphonfounder)
After years working on robot memory at @NASAJPL, @GoogleDeepMind , @Meta and @nvidia, @clarkzhango and I kept asking: why does AI still forget? Today, with @deepakmishraVC, we’re announcing @GraphonAI and our $35M Series A just 6 months after our seed. • #1 on ATM-Bench, ATM-Bench Hard • 94% on LoCoMo • 84% on DolphinBench, 10x cheaper • A new cost-accuracy frontier across 7 benchmarks businessinsider.com/graphon-…
148
158
141
1,885
2,215,232
Context windows are getting larger. So… shouldn’t we just give the model everything? Not necessarily. On the LoCoMo benchmark, full context answered 90.4% of questions correctly at ~19K tokens each. Graphon got 94% right at ~7.9K. At Claude Sonnet 5.5 rates, that’s ~$22 less in input costs per 1,000 questions. I explain how Graphon achieves this here ↓
5
9
332
Replying to @graphonfounder
Vector search cuts your data into chunks and returns the nearest few.The reasons, the sequence and the relations between them are gone. Graphon builds one self-supervised graph across everything you have, so AI can answer questions that span all of it. Entire E2E workflow from data to answer wrapped in one powerful s3 like api.
1
41
Graphon enables new developer paradigms; for example video as a database. By re-thinking of video as something that can be consumed into a searchable graph we are able to deliver ingest and search over terabytes of raw video files. Read about it here graphonlabs.ai/blog/video-in…
After years working on robot memory at @NASAJPL, @GoogleDeepMind , @Meta and @nvidia, @clarkzhango and I kept asking: why does AI still forget? Today, with @deepakmishraVC, we’re announcing @GraphonAI and our $35M Series A just 6 months after our seed. • #1 on ATM-Bench, ATM-Bench Hard • 94% on LoCoMo • 84% on DolphinBench, 10x cheaper • A new cost-accuracy frontier across 7 benchmarks businessinsider.com/graphon-…
1
1
3
255
Graphon is able to achieve higher accuracy while consuming significantly lesser tokens and compute. For both LoCoMo and DolphinBench, graphon resets the cost-accuracy pareto frontier and drives accuracy even higher than feeding frontier models full context. LoCoMo: 94% at 7.9K tokens vs Full context: 90.4% at 19K DolphinBench: Graphon 84% of tasks for $11.84. mem0ai/dolphinbench#12
1
149
1,381
We named Graphon Labs after graphon theory, a way to analyze the limits of large graphs invented by our advisors @jenniferchayes and @cborgs Here’s the full story, from vector RAG to graph theory for persistent intelligence: graphonlabs.ai/blog/understa…
2
150
1,280
Intelligence must draw on memory. The mind must not simply compile data — as an algorithm now does more quickly than we can — but recall lived experiences, which contain depths of meaning that only the human soul can recognize, and connect more deeply with the goodness of these experiences. Everything that is good begins “with the Spirit” and thus possesses a breath, a dynamism and a sense of communion that, once experienced, are engraved upon the heart. vatican.va/content/leo-xiv/e…
3
466
Got my lifelong wish. My startup is finally a wrapper. Atleast you can get a free ride during SF Tech Week: luma.com/xll5ryfk
3
2
13
138,100
Why is all agent memory markdown files?
5
158