founder at @beam_cloud (YC W22). I like hackathons, container runtimes, and steak medium rare
NYC
Joined June 2017
- Tweets1.5K
- Following1K
- Followers4.7K
- Likes11.9K
everyone dumping all their data into a company brain is unintentionally building the biggest honeypot of all time
things thailand is good for:
- eating great food
- "finding yourself"
- medical tourism
things thailand is bad for:
- shipping products
- being around ambitious people
- doing literally anything productive
theres literally nothing stopping you from moving to Thailand, splitting a full years worth of rent for a penthouse ($8,000) with six other people, and completely locking in on ai/ml.
Eli Mernit retweeted
We’re betting our $5.5M seed round that the next leap in AI coding won’t come from OpenAI or Anthropic.
It’ll come from a new feedback loop that learns from you.
Introducing Beacon: open-source, self-improving memory for coding agents.
Your sessions capture how you debug and solve problems. Today, that context gets lost and you have to re-explain things every time you switch tools.
Beacon continuously distills your sessions into reusable skills across Claude Code, Cursor, Codex, and 20+ harnesses, so your next agent can learn from the last.
Beacon is free and OSS 🌟 github.com/Asymptote-Labs/ag…
Why should the frontier labs be the only ones winning with your sessions?
You already did the work. Now let your coding agents learn from you.
people hate data centers because they feel like nothing is happening
if you built an automotive factory in ohio, it'd be the pride of the town - workers would eat lunch and socialize together and take selfies with the cars rolling off the assembly line
but data centers have almost no humans in them
no one eats lunch together and no one sees the product they produce, so people fill their imaginations with existential fears
if you had people riding power-generating stationary bikes in every data center, you'd probably wipe out all community opposition to them within a month
We trained our Muse foundational models, including Muse Wonder, at Prometheus, our gigawatt-scale supercluster in New Albany, Ohio.
Across a supercluster with hundreds of thousands of GPUs, even one underperforming host can slow an entire training run. We engineered Prometheus for resilience at scale, achieving 85%+ goodput on pretraining, 90%+ on reinforcement learning and ~47K GPU hours mean time between interruptions during Muse Wonder training.
Take a look at the tech powering Prometheus: go.meta.me/4yXXApZ
stop coding on your PC
just setup a T3 code instance on a beam.cloud VM with 4 vCPU and 8gb RAM
multiple agents running with all my files mounted to a 50gb NVMe drive
expose a public or private URL and run all your coding agents remotely
all this for just $0.00004 per second
seriously changed how I program forever
6am flight out of SFO
- Leave 1h before flight leaves
- 15 min Uber to airport, views of the Bay
- TSA agent smiles asks if you like the Grateful Dead
- Ads for AI agents that cure cancer
- Multiple food options spanning global cuisines
- 2 min walk to gate
- Get upgraded to first class
- Depart 5 min early
6am flight out of JFK
- Leave house 3h before flight leaves
- 75 minute Uber to airport, bumper to bumper traffic
- TSA agent hates you
- Ads for zyns
- No food options besides Dunkin Donuts and hardboiled eggs
- 17 min walk to gate
- Get downgraded to seat next to bathroom
- Depart 2 hours late
reflections on 36 hours in philly:
- 2 hours from NYC
- pretty good pizza
- people really love their football team
- feels like a small town in a big city
- no one talks about how data brokering is the new drop shipping
PR review environments are such a cheat code
we review dozens of PRs a day and spin up a remote environment for each one
each gets an entire backend including Postgres + Redis, all in one beam.cloud VM
we snapshot the repo after clone + deps install so it's not starting from scratch every time
we’re no longer spinning up 97 services on our computers just to review something
and it literally costs a fraction of a cent each time we do this
there’s a misconception that building great products is mainly about knowing how to build thoughtful software
which is why many people on X like saying that software is dead
it was never about that
it was always about deeply understanding your users
many mistake the ability to build for the ability to build something people want
sandboxes are a compute primitive
just like a cloud function or a database or a cache, a sandbox is another component of a broader product suite
the same way Postgres, Redis, Kubernetes are primitives, not platforms
there’s no winner take all Postgres company beyond the framework itself, which is open-source anyway
the winning infra company builds a platform around those primitives and builds an ecosystem around it
you don’t “win” the market for Postgres, or containers, or VMs
you win by building the platform that runs all of it
I’ve been thinking about this a lot.
and I would put the odds at 50/50 of if the sandbox market is a winner takes most market.
I see a lot of momentum from folks like Ivan. From the outside looking in it seems like Ivan has built their business on great friendships and fat margins, which are two very powerful combinations. I look at this and completely understand his take away of “winner takes most”. If a coffee shop or cafe has these two things going for it in a local town it becomes THE coffee shop or cafe.
Then I look at some of the work by folks like @deok_filho or @PrimeIntellect and they come up and bum rush the market. CoreWeave having one set of strengths and Prime having another set. always overlapping in use cases, but also kind of differentiating.
This leads me to think that perhaps the sandbox market is much bigger than we expected even when we were saying it was big.
Perhaps it’s the last compute primitive we need and most workloads will move to it or baremetal. Perhaps lambdas and containers aren’t needed anymore.
And this is how i get to 50/50.
POV: your agent has a desktop in the cloud
4 vCPU, 8Gi RAM, $0.00004/second
the future is here
most dev tools are priced like gym memberships
charge $20 a month for usage that most people won't actually use, and price the usage component absurdly cheaply
some small group scales and give you your margin
and the people who never actually use the service subsidize everyone else
POV: your agent wants a desktop in the cloud
Introducing Beam Virtual Machines
4 vCPU, 8Gi RAM, 100Gi NVME, $0.144/h
- Scale-to-zero
- Memory + Disk snapshots
- Connect Postgres, Redis, S3
- GPU support
the future is here
the world isn't ready for vera rubin
the transition from B300 to Vera Rubin is a generational hardware upgrade where the architecture, cooling, and data center facilities are fundamentally different from the prior generation
consider a few things:
1. vera rubin deliveries are being prioritized to only the closest Nvidia partners and won't be widely available until 2027
2. less than 10% of data centers in the US will be equipped to run vera rubin *12 months* from now
3. of those that have the cooling and facilities to support vera rubin, only a small fraction of that group even has power contracted
combine this with the current demand for inference, and it's a perfect storm for GPU prices to continue climbing
It's interesting that every depreciation model assumed the A100 would be nothing but scrap metal by now.
Its utilization is climbing and none are flooding into the secondary market yet which means contracts are being renewed.
If this continues to be the case for Hoppers, there is a serious bull case to consider for companies that have financed these chips on standard 6 years.
Else, there's going to be a huge change in this market with disaggregated compute supply.
sandbox providers loved to quote ultrafast startup speed
but it doesn't really matter how fast you can launch a container, what matters is how quickly you can mount an image to said container
we spent years doing serverless GPU at beam.cloud and forced to optimize for <10s container boot time with 10-50Gi images
after getting a 30Gi image to cold start on a GPU in 5s, launching a 1Gi image in a sandbox is like kindergarten
starting to think sandboxes aren't exactly a commodity.
there's actually alot you can do on top of it (seen some incredible demos recently)
> spinning them up with an image in milliseconds
> checkpointing every state difference
> treating sandboxes as durable actors
> making sandboxes more agent friendly.
@puffleai uses daytona which takes ~19 seconds to load a stock image onto the VM, and we're actively trying to solve for it.
they gave me lots of credits so i don't want to churn.