MTS @ https://xitter.twitterapi.xyz/t.co/MowZ6tydCO new media @ https://xitter.twitterapi.xyz/t.co/UPVdiCl5mL gpu stuff @ https://xitter.twitterapi.xyz/t.co/VCTLXhaWRA Scout @trygravityai & investor. Most startup problems can be solved w/ 100M views.
San Francisco
Joined August 2020
- Tweets10.3K
- Following7.1K
- Followers18.5K
- Likes23.7K
i’ve been offline for a bit, can someone please explain why Nvidia stock is down so much?
8
27
2,592
this one is a banger by frontrun
anybody selling data rn is noticing a pattern:
the customers are increasingly enterprises that want to fix their agents
evals are going to be a 100B+ market, because an eval is what makes ai deployable to production yes, but also what lets you hillclimb outcomes, and is also an rl task at the same time
rn the major buyers are frontier labs. but not forever!!
ex-@GoogleDeepMind gemini post-training. ex-@mercor.
now they're fixing ai agents in production, and the company has 2 followers: @Convergent_Dev
@OmkarSavant_ and @debnilsur turn production agent failures into evals, then fix them.
flagged oct 15th on @frontrunvc
thinking of a new startup idea:
obviously servers are good for agents. claude/codex need computers to do things with
what if there was a VPS that could spin up and down very quickly?
you only pay for it when it’s active. i guess I can do this myself with k8s, but that’s a lot of work.
there needs to be a platform that manages all of this for me
not sure why nobody is building this, feels like a pretty obvious play
Extremely sad to hear Vals is shutting down
Steve is a founder I look up to, and it always felt like Vals was one decision away from PMF.
Not quite infrastructure, not quite a vibe coding platform. Val felt like early Replit yet somehow also a love letter to the Javascript ecosystem.
I can’t wait to learn what’s next for the team.
6
40
4,504
NO MORE SLOP
but use more ai. MORE AI. you’re not using enough AI
but no more slop. NO MORE SLOP.
see that engineering team? they do 6k PRs a day. they spend more on CI than our company makes in a year.
WHY ARE YOU SPENDING SO MUCH ON CODEX.
cut costs. no more slop. use less ai. BUT USE MORE AI.
no more slop though
18
89
3,308
“dad, why are we so poor”
“well son, I was in YC in 2026 and everyone was pivoting to sandboxes in the middle of the batch. I decided that there were too many sandbox companies already
so I kept building my b2b saas”
“dad did you not see the MRR charts? were you not on X or something?”
“i was son, but I thought every MRR chart was going to be the last one. but they kept coming
for months, box, boxes, boat, box (this one by upstash). everybody launch post went viral.
modal quietly got to 1B arr. loveable started making sandboxes for some reason.”
“i hate you dad”
“not as much as i hate myself”
there's lots of ways to get inference to be cheaper:
1. get your own GPUs and colo space
2. include ads in the inference
great idea by Gravity + Inference!
We’ve been waiting to release this for a little while, so I’m very excited to announce the world’s first generally available LLMs that are ads-subsidized.
You can now use some of the best Open Source models for up to 25% less through the @inference_net gateway!
Ads only append to a response ~4.2% of the time, and when they do, they’re relevant to you, or the conversation you’re having. They do not bias or modify the core response in anyway.
300k in ARR added in three hours from a couple tweets
interesting.
Introducing Morph Ultra
Unlimited Kimi and DeepSeek for $200/month
1,000 subscriptions available
Details below
morphllm.com/products/ultra
labs are buying absurd amounts of CAD data rn, in million dollar++ batches
we’re gonna see massive improvements in CAD/engineering ability in the next three months
context.dev people search + jev is a game changer for gtm
(paid partnership)
(ai generated)
X is definitely getting saturated
GTM is about finding alpha, and eventually that alpha disappears
Launch videos still work, but it’s now 3x harder and more expensive to go viral.
High value posts still work, but are not necessarily repeatable. For example, I can’t just manufacture 1M views for Context a day.
Field marketing still works, but now you need to be much more creative to fill rooms.
I’m seeing founders find success with UGC, Meta ads, Rednote, guerrilla marketing, podcasts, all kinds of stuff.
But every GTM operator needs to be on their toes: if poker nights/dinners don’t work as well as they used to, what’s next?
bro was at 500k ARR like last week
exponential growth is scary
Paul Graham’s advice is wrong. Growing 30% MoM is boring
Be Anicet and grow 100% WoW
we just reached 1M ARR with boat.dev
it was our best case scenario for the entire YC batch, yet it took only 2/3rd of it to happen
slowly then all at once, we grew from 0 to 1M ARR in 22 months over 6 pivots:
0 -> $10k ARR in 16m, 5 pivots
$10k -> $100k ARR in 4m, 1 pivot
$100k -> $200k ARR in 1m
$200k -> $1M ARR in 1m
we already got way more in rev run rate btw, subscriptions are just a fraction of our total revenue
now road to $100M!
there’s a massive difference between the vector database rush and the sandbox rush
sandboxes make money
Everyone is arguing about AI sandbox pricing, but they're not even comparing the same thing.
@boatdotdev sells a whole VM at a flat hourly rate, $0.018 an hour, so you can predict the bill.
@Agent37HQ sells minutes and sleeps the box down to just disk, so an agent that mostly sits idle costs less.
@vercel Sandbox only bills CPU while your code is actually running, which makes sense because an agent spends most of its life waiting on an LLM response.
@e2b sells the trust layer, a real microVM per session, for running code you do not trust.
@daytonaio sells pause and fork, so you can freeze a running machine and clone it into twenty copies.
And there are about 30 of these on one public benchmark alone, which probably makes it the most competitive space in infra right now. Feels a lot like the vector DB rush in 2023, when everyone shipped roughly the same thing and then spent a year posting benchmarks at each other 😂