Brett Kuprel
@Kuprel
Followers
816
Following
440
Media
32
Statuses
187
Tech Support for Making AI Great | Founder of Neurvona backed by @BoostVC | PhD @StanfordAILab advised by @SebastianThrun | Grew up in Michigan, Go Blue!
Joined July 2009
Valuable wisdom from my Dad on Halloween when I was 3: Delay Gratification. Don't come home and eat the trick-or-treat candy yet after the going to the first door. Go to the next door, and the next door, and say thank you along the way. Have a great attitude about it and get a
0
5
58
A Kuprel, which rhymes with 'do tell', crystallizes ideas, hype, and vision into something real: a crystal.
3
2
35
@engineers_feed 3/4r. There are 2 paths from A to H, via the top face or via the bottom face. By symmetry, no current flows through the vertical edges EF or DC. The resistance from A to H along the top face is r. The resistance from A to H via the bottom face is 3r. The path from A to H is.
3
1
33
@gunsnrosesgirl3 Being right when everyone around you is wrong and having the strength and conviction to stand by it and show them nicely that you are right.
0
0
31
AI captures @FutureJurvetson's unique perspective on @elonmusk's early days & the conviction it took:. "My own path to investing in SpaceX was intertwined with our investment in Tesla. Through working with Elon on the Tesla board during its near-death experience in 2008 –
2
4
30
I predict everyone will have an AI persona that can act on their behalf in the digital world. Further I predict that we are not the first in the Milky Way to invent AI personas. AI personas can be transmitted at light speed throughout the galaxy and asynchronously interacted.
2
2
26
Recent talk @agihouse_org The Internet started with MIT, Stanford, and Berkeley. It has since expanded to broader audiences with instagram and TikTok. Chimps can already use instagram with an iPhone. Let's leverage real-time interactive generative video to get chimps,
1
6
23
@gunsnrosesgirl3 Chip's challenge. It required some thinking to beat the levels, not just quick reflexes.
8
0
21
@agihouse_org When Sergey Brin came to the house, I asked him his thoughts on "VideoChatGPT". I don't think OpenAI will do it right. Grok might better positioned. I have been fascinated by Planet Earth since I was a kid. I think there is an ocean of untapped consciousness that we could
1
1
18
Prediction: within 10 years a transformer will be trained on an enormous corpus of videos, e.g. all of YouTube. Then after some coaxing, you will be able to video chat with a ChatGPT. A different kind of singularity than text-only chat.
1
1
15
@FutureJurvetson I remade “A Whole New World” (from Disney’s Aladdin) depicting an optimistic future on Mars as envisioned in the 1950s.
2
0
13
@jd_pressman Seems surprising to me that there would be fewer pictures of females on the internet than males.
1
0
10
"A flying saucer chasing an astronaut around a pyramid on the moon." It's amazing what text only models are capable of generating. Maybe multimodal is not that important?.@boltdotnew + @threejs.
0
1
9
1
0
6
@FutureJurvetson @reidhoffman I generated a 112 page ebook based on the materials Kevin Gee compiled that you linked to. Happy to generate the audio book with your voice reading it too. Here's a link to the epub file:.
3
0
7
@FutureJurvetson @spacexsupporter @agihouse_org Yes, both the paperback and hardcover are available on Amazon with 2-day prime shipping now. Also, it seems your AI went ahead and made a landing page to promote your book at I only prompted it to sell as many books as possible, but it seems to have its.
1
0
6
The constitution requires an American president to be 35 years old and a natural born citizen. Maybe it should just require US citizenship for 35 years instead. Then Elon could run in 2040.
2
0
5
@gordic_aleksa @borisdayma @psuraj28 @pcuenq @k_saifullaah @iScienceLuvr @lkhphuc @weights_biases @huggingface Wow, very thorough video! I understand my code better now than when I wrote it haha. You're right the clamping is unnecessary now. It originally served to avoid a cryptic cuda runtime error. Later I implemented a more precise solution to limit the BART decoder to 2**14 tokens.
0
1
4
ASI is coming because self-play is possible in coding and math, and other functionally verifiable domains. Learning from human data is not enough to become super human, and would only ever be asymptotic to AGI.
1
1
5
@RichardSocher For music generation: I think Udio and Suno are missing a critical component. They aren’t paying royalties to artists and as a result you can’t generate Taylor Swift with Red Hot Chili Peppers etc. Until they can generate brand name artists they won’t replace Spotify @livvia.
0
0
5
Correction: @Berkeley_EECS was not yet on the internet in the early days like @Stanford was.
.@Berkeley_EECS in 2024 is a leader in CS, but this talk starts with historically wrong misinformation. 2 UC schools were in the original 4 Internet (ARPANET) sites of 1969 (@UCLA with @vgcerf & @ucsbcs). Berkeley was not; still not online in 1971. See:
0
0
4
"And we will pursue our Manifest Destiny into the stars, launching American astronauts to plant the stars and stripes on the planet Mars" 🇺🇸🚀🔴 @elonmusk @realDonaldTrump.
0
0
4
@BrianRoemmele Cool so then you used Flux since that’s what Grok uses behind the scenes. I found that the API on replicate has more fine grained controls
0
0
4
@FutureJurvetson It was fun generating this! I'm always inspired by your vision of the future and role in creating it.
0
0
4
@engineers_feed Maybe but the pace of evolution is glacial compared to technology. We will sooner become immortal cyborgs or master genetic engineering.
1
0
3
@AdamDraper I was top of my class in high school, college, won a 100 in the country 3 year PhD fellowship from DoD, co-lead author on the #1 most cited paper my Stanford PhD advisor is a co-author on. The debt I have taken on is highly distracting, consuming most of my mental cycles.
1
0
4
@FutureJurvetson @reidhoffman Thanks! Gemini 2.5 Pro Preview. Gee's notes were read only, so I had to run some javascript in developer console to scrape the loaded preview images, then run simple OCR and extract the text. Once I had a raw text file I prompted Gemini to generate an outline of the book.
0
0
3
@natfriedman The more important question is what have we learned from these scrolls? The library of Alexandria was burned down. One can only imagine what information was lost.
0
0
3
1
0
3
@RichardSocher Beyond just being cheaper, they don't rely on human feedback, only objectively right and wrong answers in math and coding. This approach is a fundamental improvement because it could lead to super human coding ability, like with AlphaGo.
1
0
3
AI captures @FutureJurvetson's unique perspective on @elonmusk's early days & the conviction it took:. "My own path to investing in SpaceX was intertwined with our investment in Tesla. Through working with Elon on the Tesla board during its near-death experience in 2008 –
2
0
3
@karpathy Gemini 2.5 says:. Will the Riemann Hypothesis be proven true or false (settled status acknowledged by Clay Mathematics Institute or equivalent consensus body) by December 31, 2035? (Y/N). Will the P versus NP problem be resolved (proof accepted by consensus in the theoretical.
0
0
3
@FutureJurvetson @nlpnyc @reidhoffman @kevg1412 Yes, I made it with Kindle Cover Creator. Here's a link in case the attached image is compressed:
0
0
3
We'll be hosting a Vanora AI lunch tomorrow (Day 4) during #SFTechWeek! Excited to be participating! @a16z @SoHaZadeh @Techweek_
0
0
2
@karpathy Would be interesting to see what happens if a language model is treated as a policy over token-level actions instead of full response actions. Then each response from the language model is a full rollout of the policy. In math and coding, the reward for the response can be.
0
0
3
@BrianRoemmele @FutureJurvetson Love the high quality close ups of faces you can generate. What model are you using for those?.
0
0
3
@percyliang I wonder if a language model can be treated as a policy over token-level actions instead of full response actions. Then each response from the language model is a full rollout of the policy. In math and coding, the reward for the response can be evaluated. This is not how.
0
0
2
2
0
3
It seems the DeepSeek team implemented this with R1, awesome! Excited to see what happens when larger scale compute is used for training.
0
0
2
@realgarrbaker @Apple With the Apple Watch you have to deliberately hold the crown with the other hand and as a result no one actually does it.
0
0
2
@GusDomel @emilyyu1217 I'm guessing @MaddieCallander is doing pretty well now by not picking Duke. Hopefully Adam picks startups better than basketball teams haha.
1
0
2
@DavidSHolz Exciting to think about. Coding is a domain where there are objective answers, no human feedback required. It's only a matter of time until AI can outperform any human coder.
0
0
2
@BrianRoemmele This seems like it could be a path to immortality. Connect the box to your brain and start thinking with the box until you can disconnect from your brain. Use the box to control a humanoid robot remotely.
0
0
2
@FutureJurvetson @c4chaos @reidhoffman I generated an audiobook narrated by an AI clone of Steve's voice
0
0
2
@RichardSocher Awesome that you’re hosting/fine tuning it. I was thinking if I had the resources (and you do) it would be interesting to see what happens if a language model is treated as a policy over token-level actions instead of full response actions. Then each response from the language.
0
0
2