Jeffrey Wang
@wangzjeff
Followers
9K
Following
3K
Media
58
Statuses
1K
cofounder @ExaAILabs prev SPC, Plaid, Harvard
San Francisco
Joined January 2020
My senior year, I was seeing a girl from MIT and showed her this 6 GPU ML/mining rig I had proudly built. She immediately told me that her ex had a bigger one with 8 GPUs. Can’t even make this stuff up 🫠🫠🫠.
my Harvard dorm room ML/crypto mining setup 😂. 4 1080tis, 2 1070s. Many BERTs trained, many ETHs mined
74
257
3K
There's something deeply wrong w/ the Huberman/Blueprint/longevity movement. Last night, friend bailed on a good friend's housewarming because they "needed" to go on an early morning run. Life is too short to be a hypochondriac in your 20s. Let's all remember to live.
I don’t know who needs to hear this, but you don’t need a daily routine. You can eat when you’re hungry, you can sleep when you’re tired, you can drop all commitments if a special opportunity comes up. Wellness podcasters are just promoting a lifestyle for people who can’t make.
164
70
2K
Sophomore year at Harvard, my team was first place in a competition that our machine learning class was having, and I was feeling pretty good. Then at 1am, Scott Wu made a few submissions and wrecked us. Best of luck to this amazing team!.
Today we're excited to introduce Devin, the first AI software engineer. Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork. Devin is
14
34
997
I thought I was good at prompt engineering before putting together this group. crazy knowledge being dropped. Again, the goal is to foster the best conversations about absolutely maxing LLM performance, with production in mind. Lemme know if you wanna join!
Just made "LLM Performance Enthusiasts" - a private Discord for people who are focused on absolutely maximizing LLM performance. There is a lot of noise on Twitter/internet about finetuning X model or using Y technique. What's the common theme? The vast majority of posts are.
84
11
341
@che4os 9-5ing your way to ~10M net worth with your main recreations being omakase dinners, board game nights, and simping for the 2 girls you know is my idea of hell.
16
14
259
As someone who went there, I know as fact that many Harvard grads with >3.8 GPA suck.
I know heritage American Harvard seniors with >3.8 GPA who don’t have a full time job offer after >30 applications. This is degrading. The social contract has been violated. The system is not working for us, the youth. The boomers betrayed us in their love for cheap foreign.
10
2
247
Feedback on Grok @elonmusk @TheGregYang :. I used to use Grok every day because I find it to be the least biased LLM. However, my usage has gone to ~0 because it now tries to inform every response with web RAG, which ruins the ability to get pure answers.
18
3
201
If your heart skipped a beat reading the topics below, you're probably a good fit for our LLM community🫡. LLM Performance Enthusiasts is for real talk from the best, nerdiest hackers. If you want an invite code, reply here with an LLM performance tip!
Just made "LLM Performance Enthusiasts" - a private Discord for people who are focused on absolutely maximizing LLM performance. There is a lot of noise on Twitter/internet about finetuning X model or using Y technique. What's the common theme? The vast majority of posts are.
133
5
199
🥳🥳🥳. Mega kudos to all our investors and also the Exa team. Back to building! (and hopefully some sleep).
Announcing our Series A, led by @lightspeedvp, with participation from @nvidia and @ycombinator! 🚀. Exa is an AI lab redesigning search. Funding will help scale our API product, the first search engine built to power LLMs. Today, we’re also launching big product updates: 🧵
22
2
187
Claude 3 Opus (biggest model) got 2 wrong. That's a perfect 800 🤯.
OpenAI claims GPT4-turbo is “better” than GPT4, but I ran my own tests and don’t think that's true. I benchmarked on SAT reading, which is a nice human reference for reasoning ability. Took 3 sections (67 questions) from an official 2008-2009 test (2400 scale) and got the
2
20
168
jokes on her now - at @metaphorsystems we have a $1M cluster of A100s :). (if this excites you, DM me 😛).
11
1
148
Our office nap pods have arrived. I’m hosting an open house this Sunday in the Exa SF office so that people can check them out and make orders!. Event link below: ⬇️
I'm buying Japanese capsule beds as nap pods for our new SF office. If you want to bulk order with me, they should be something like $1000/bed. Please DM me if interested!. Comparable products sold in the US cost like $15K
14
3
145
@Jamesthesnake maybe. last stint I did in SF was when I was a new grad. but is mid twenties really that different from early twenties?.
11
0
104
@xhfloz 🙌. If you want to help us crawl, train novel search models, organize the web, check out :).
3
0
107
@brianryhuang I think it's more that a lot of these people are embarrassed that they've chosen a profession that wastes their talent and is by definition a zero sum game?.
5
0
100
Exa is super powerful - vector search over the web is just the first step on a quest to build the best search tech out there. But before today, it was mainly in the hands of mega nerds + devs. Websets changes that because it's a super useful LLM product built on top of our.
8
3
90
@natanielruizg I completely agree! not everyone has to drink. but if you don't, you better not be boring :p.
3
1
88
@lailachima 😂. I think it's actually healthier, was gonna go to a bar and now I'm drinkin' sparkling water :).
4
0
79
@TrevorCampbell_ yeah - fun fact, if you don't feed GPUs protein powder, `nvidia-smi` will show lower and lower max wattages.
1
0
76
@caitlinner Before COVID, I never went to parties where over half the people were not even holding a drink. Nowadays, this is typical for me. People don't even pretend to drink by sipping a white claw over the course of a few hours.
1
0
65
Couldn't be more excited about what we're building at M̶e̶t̶a̶p̶h̶o̶r̶ Exa! New contents features should be super useful for any AI app that needs access to high quality information. Also if you're considering a rebrand, heads up that it will probably take 5 years off your life.
Announcing big changes to Metaphor:. 1. We’re releasing a suite of content retrieval features using new models trained for RAG. 2. We’re renaming to Exa to reflect our mission. 3. We’ve revamped our UI 🎉. Altogether, we're taking big steps toward "search over knowledge": 🧵
8
2
67
Airlines could become dating apps. - Opt in to United Dating, swipe.on people on your flight.- Meet before boarding for a quick date. Maybe one person even takes the other to a lounge.- Sit together on plane (maybe offer ability to opt out if airport portion doesn’t go well).-.
8
0
55
🫡. Calling all hackers, if you're building something with Exa, we're happy to support you with any free credits! Just shoot me a message, or email jeff@exa.ai.
it's finals week at stanford but i want to build something with @ExaAILabs 🙄. maybe during spring break.
1
3
53
Y'all keep asking me how much $ I made from running 6 GPUs in my dorm room. - Full setup was ~$4600.- I was VERY cheap when buying parts and bought GPUs before market got hot (2017).- Electricity was free, so pure profit.- I made something between $4600 and Harvard tuition :)
my Harvard dorm room ML/crypto mining setup 😂. 4 1080tis, 2 1070s. Many BERTs trained, many ETHs mined
6
1
52
@daniel_petri1 Europe is waking up haha. Unfortunately this is impossible/extremely inconvenient in many if not most American cities.
1
0
46
Did not expect to get demand for 100+ units. Here's the plan everyone - I ordered some with express shipping arriving in 2 weeks, will test them out and then figure out logistics for a big order after I'm assured of the quality!.
I'm buying Japanese capsule beds as nap pods for our new SF office. If you want to bulk order with me, they should be something like $1000/bed. Please DM me if interested!. Comparable products sold in the US cost like $15K
6
1
49