Matt Barnes (@BarnesMJ) header
Matt Barnes (@BarnesMJ) profile picture

Working on Gemini at DeepMind

Seattle, WA
Joined January 2013
Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)Photo by Matt Barnes (@BarnesMJ)
I agree. Unverifiable domains (like creating the benchmarks and evals) will be the only ones impossible to fully automate
The avalanche that hit math is coming for AI development. dwarkesh.com/p/rsi
2
114
Check out our latest release in Gemini!
Introducing Personal Intelligence. It's our answer to a top request: you can now personalize @GeminiApp by connecting your Google apps with a single tap. Launching as a beta in the U.S. for Pro/Ultra members, this marks our next step toward making Gemini more personal, proactive and powerful. Check it out!
142
Last week wearing the Google Research t-shirt
1
2
322
Matt Barnes retweeted
Presenting world-scale inverse reinforcement learning (IRL) in Google Maps. Learn how RHIP, a new IRL algorithm, along with advances in graph compression & parallelization led to relative improvement in the quality of suggested routes in Google Maps. → goo.gle/3EBAc75
18
116
5
447
138,635
Our multi-year project using Inverse RL to improve route quality in @googlemaps is finally out! Scaling was HARD. It took advancements on parallelization, graph compression, dominant eigenvector inspired initializers, and a new generalized IRL algorithm to make it possible.
Our 🌎-scale Inverse RL paper is finally out! Thrilled to share this multi-year project on route recommendation. Understanding preferences is much harder than behaviour: not just WHAT but WHY! We address the challenge via IRL on massive scale (100s Ms states, samples, params)!
1
3
12
2,633
Overall: 16-24% improvement in route accuracy Personally, very excited by the potential this has for cycling and walking. These are often some of the most difficult routing problems, since its about so much more than just ETA
145
Corey is a personal hero. Sad to see him leave the race, but know he'll continue to do great things in NY.
Corey Johnson, the New York City Council speaker and a presumed front-runner in the 2021 mayoral election, withdrew from the race. nyti.ms/2HjNpab
1
I forgot to sign my Washington mail-in-ballot. King County called me (three times), emailed me (for a whole week) and sent me a letter to fix it. They REALLY want you to vote. @KingCountyWA
5
I am on the academic job market! My work is on human-centered machine learning (ML), and focuses on issues of fairness, accountability and sustainability when using ML for decision support. Please reach out if you think we may be a good fit!
32
196
6
605
My friend Alex, who is both a superb bird and human
Oh hey did I mention that I had a costume contest in my ornithology class and some dressed up as a friggen superb bird of paradise???
1
Matt Barnes retweeted
This is mind-bending: The introduction of EZ-Pass in New Jersey and Pennsylvania reduced the incidence of premature birth by as much as 9.1% for mothers who lived within 2 kilometers of toll plazas bloom.bg/2z8xVi5 via @bopinion
57
25
131
Noise pollution map across the entire country, from @USDOT. Flight routes are huge.
2
I made this handy chart of where the D4 Seattle City Council candidates stand on transportation, housing and sustainability issues. @UrbanistOrg @CascadeBicycle @MASScoalition
2
1
3
Thanks @rootedinrights for the forum event transcription!
1
(1/2) Last year, @BenBoecking and I partnered with @ConnectasOrg and @elpaiscali for a project about the killings of social leaders in Colombia. In this paper we describe the results of the data analysis and how this led the journalistic investigation: arxiv.org/pdf/1906.08206.pdf
1
3
5
The pre-preprint
This quoted post is unavailable.
Matt Barnes retweeted
First attempt at removing cars off the roads with neural nets. Will have to dream harder.
972
14,010
2,118
57,920