Brad Hilton
@bradthilton
Followers
790
Following
79K
Media
295
Statuses
3K
Reinforcement Learning Research Engineer • Sometimes Political Commentator • Husband and Father • Believer in Jesus Christ
Orem, UT
Joined February 2013
just occurred to me this morning that this has very promising potential ramifications for distributed and memory-efficient training. needs to be studied more.
lr=1.0, clip=1e-10 → 99.13% of parameters unchanged compared to the base model. (orange run).lr=0.01, clip=1e-8 → 95.21% of parameters unchanged (purple run). both converge to very similar points, except one is significantly more sparse in what it chooses to update
0
0
0
the democratization of cinema is nearly here.
Untold - The Immortal Blades Saga. So beyond excited to share this concept trailer - Untold is a story I have worked on for over 8 years and next year, we will be making it into a reality with season 1 of the series. This project is probably the one I hold most close to my heart
0
0
3
it’s a matter of national security and pride. give this man the gpus.
i humbly request 10k gpus and am publishing my post "The American DeepSeek Project" on the 4th.
0
0
5
RT @NickJFreitas: Thomas Sowell is now 95. and still doesn't have a Presidential Medal of Freedom!. Respectfully requesting @realDonaldTru….
0
3K
0
sure, maybe in your experience, but i’m not sure how this would hold out-of-distribution.
You can fake it pretty far in this industry just by saying, “Hrmm, that’s cool but I’m worried it won’t generalize,” whenever you’re presented with literally any information.
1
0
2
RT @polynoamial: AI researchers will literally negotiate $100 million comp packages by themselves but they won’t play poker for more than $….
0
51
0
my mom's really smart. graduated in cs and probably one of the few people on earth who has programmed in both cobol and vue.js. meta would be incredibly lucky to have her.
BREAKING: meta poached your mom.
2
0
10
magnificent vibes. hard to tell if he's bruce wayne or a bond-villain, but i'm loving this arc.
0
0
0
RT @corbtt: It's becoming increasingly obvious that mass model customization is the future. Even big labs, which have traditionally pushed….
0
15
0
i think i find myself spending more on ai as the models get better though. jevons paradox is crazy.
@kimmonismus This is wild because looking at the costs of LLM models has shown us that the cost goes down significantly with most model upgrades. The cost will near 0 in the near future.
1
0
2
this is how you win the vertical:. train a model that dominates the *entire* pareto frontier.
We're taking a big step towards medical superintelligence. AI models have aced multiple choice medical exams – but real patients don’t come with ABC answer options. Now MAI-DxO can solve some of the world’s toughest open-ended cases with higher accuracy and lower costs.
0
0
4
yep, i've been frustrated about this too. i don't want gemini to compose emails for me, i primarily want it to be able to help me search and archive emails, as well as identify all my current newsletter subscriptions and allow me to painlessly unsubscribe. just help me get to.
Gemini **inside Gmail** just told me it doesn't have the ability to search my emails. How can Google be simultaneously so good and so terrible?
0
0
0
the future of general versus specific AI hinges on many of the unanswered questions in this thread.
Recently had a good chat with @tamaybes. He thinks we aren’t yet in the GPT-3 era of RL and as it scales, cross-task OOD generalization will emerge. It’s difficult to empirically study this at current scale, but let’s take it as true—what does this mean for custom RL plays? 🧵
0
0
1
perplexity would be sky high on that summary.
0
1
2
many such cases
RL == easy gains with a small amount of compute, but if you really want to push the needle you need a shit ton of compute
0
0
2
amazing the prescience you can have when you take the world economy hostage.
0
0
1
the best agent framework is your own repo.
my favorite agent framework is python.
0
0
2